Policy Research Working Paper 8998 The Entertaining Way to Behavioral Change Fighting HIV with MTV Abhijit Banerjee Eliana La Ferrara Victor H. Orozco-Olvera Development Economics Development Impact Evaluation Group September 2019 Policy Research Working Paper 8998 Abstract This paper tests the effectiveness of an entertainment edu- improvements in knowledge and attitudes toward HIV and cation television series, MTV Shuga, aimed at providing risky sexual behavior. Treated subjects are twice as likely to information and changing attitudes and behaviors related get tested for HIV eight months after the intervention. The to HIV/AIDS. Using a simple model, the paper shows that study also finds reductions in sexually transmitted diseases “edutainment” can work through an individual or a social among women. These effects are stronger for viewers who channel. This study is a randomized controlled trial con- reported being more involved with the narrative, consistent ducted in urban Nigeria, where young viewers were exposed with the psychological underpinnings of edutainment. The to MTV Shuga or a placebo television series. Among those trial’s experimental manipulations of the social norm com- exposed to MTV Shuga, the trial created additional varia- ponent did not produce significantly different results from tion in the social messages they received and the people with the main treatment. The individual effect of edutainment whom they watched the show. The study finds significant thus seems to have prevailed in the context of this study. This paper is a product of the Development Impact Evaluation Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at banerjee@mit.edu, eliana.laferrara@unibocconi.it, and vorozco@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team THE ENTERTAINING WAY TO BEHAVIORAL CHANGE: FIGHTING HIV WITH MTV Abhijit Banerjee Eliana La Ferrara Victor H. Orozco-Olvera This research is part of the entertainment-education program of the World Bank’s Development Impact Evaluation Department (DIME), part of the Development Economics Research Group. We thank Oriana Bandiera, Michael Callen, Ruixue Jia, Gaia Narciso, Ricardo Perez-Truglia, Devesh Rustagi and seminar participants at IIES, LSE, Toulouse, Trinity College Dublin, University of Bonn, UCLA, UC San Diego, University of Manheim, University of Oslo, University of Amsterdam, University of Southern California, Yale, UPF, PSE, Wharton and BREAD 2018 conference for helpful comments. Laura Costica and Edwin Ikuhoria did a superb job as research and field coordinators. Tommaso Coen, Viola Corradini, Dante Donati, Francesco Loiacono, Awa Ambra Seck, Sara Spaziani and Silvia Barbareschi provided excellent research assistance. This study was funded by the Bill and Melinda Gates Foundation and the World Bank i2i Trust Fund. La Ferrara acknowledges financial support from ERC Advanced Grant ASNODEV. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research. Victor H. Orozco-Olvera Abhijit Banerjee World Bank Department of Economics, E52-540 Development Impact Evaluation Department (DIME) MIT 1818 H Street 50 Memorial Drive Washington, DC 20433 Cambridge, MA 02142 vorozco@worldbank.org and NBER banerjee@mit.edu Eliana La Ferrara Universita' Bocconi Dept. of Economics and IGIER via Roentgen 1 20136 Milano Italy eliana.laferrara@unibocconi.it A randomized controlled trials registry entry is available at AEARCTR0000509 1 Introduction International organizations and governments in developing countries have invested massively in behavior change campaigns, to improve outcomes in such diverse areas as health, education, …nancial decision making and governance. The evidence on the e¤ectiveness of such campaigns is mixed at best, especially when we consider campaigns targeting HIV prevention.1 The general sense seems to be that complex psychological factors are at play and the usual means of public communication fail to touch deep seated preferences in this domain. The policy community is thus increasingly embracing alternative approaches that combine information provision with entertaining content, an agenda which has come to be known as edutainment (short for ‘ entertainment education’). Edutainment consists of media programs, usually radio, television or …lm, that aim to change attitudes and behaviors by getting the viewer immersed into an entertaining narrative where the educational messages are presented as an integral part of a bigger story. Despite the increased popularity of this approach among policymakers, there is little rigorous evidence on whether edutainment works and, if so, through which mechanisms. There are two distinct reasons why edutainment might work where ordinary behavior change campaigns fail. Pending more detailed discussion later, one reason is that the appeal of the show makes the individual pay more attention to the message and reduces potential resistance to top-down advice (Bandura, 1976). This is what we will call the ‘ individual’e¤ect. A second reason, which we refer to as the ‘ social’e¤ect, comes in because the show portrays alternative lifestyles that viewers could take as a norm. If people conform to what others do or think, then the message on the screen, potentially coupled with the fact that TV shows get seen by a large number of people, could coordinate a shift in the social norm. Given the growing evidence on the importance of conformity (e.g., Bursztyn and Jensen, 2015; Bursztyn, Fujiwara and Pallais, 2017; Perez-Truglia, 2017), it is plausible that changing behaviors might require shifting the norm that everyone conforms to. While any form of public communication can have a similar coordination e¤ect, this is particularly important for popular movies or TV shows, just by virtue of the sheer numbers of people who watch them. At the same time, there are several reasons why edutainment may not work. One is that the educational content may be too ‘ diluted’ in the narrative, and viewers fail to recognize and retain the relevant information. Another is that the …ctional nature of the story may lead viewers to trust the quality of the information less than a more ‘o¢ cial’source. Whether edutainment works or not is thus an open question. This paper proposes a framework for analyzing the impact of edutainment along the chan- nels described above and reports on a …eld experiment that we carried out to test the impact 1 See, among others, Krishnaratne et al. (2016). Padian et al. (2010) conducted a systematic review of interventions aimed at preventing sexual transmission of HIV found that only one in seven of these were e¤ective. 1 of an edutainment TV series. The series was the third season of MTV Shuga, a popular serial produced by MTV Staying Alive Foundation to provide information on HIV and change atti- tudes and behavior relating to HIV and risky sexual behavior more generally. The experiment covered over 5000 young men and women in 80 urban and peri-urban locations in South West Nigeria: 54 locations were randomly assigned to screen MTV Shuga, while the remaining 26 screened a di¤erent serial that involved a similar demographic but made no connections to HIV. The …rst question we ask is whether MTV Shuga had the desired e¤ects. We …nd striking e¤ects on knowledge about sources of transmission of HIV and its treatment, on attitudes towards HIV+ people and on a range of behavioral outcomes (both self-reported and objec- tively measured) eight months or more after the showing. The likelihood of testing for HIV, objectively measured through redemption of a voucher that we distributed at health camps, increased by 3:1 percentage points in the treatment compared to the control group. This corre- sponds to a 100 percent increase over the control group mean. Analogous e¤ects are estimated for the self-reported measure, where the likelihood of testing increases by 2:5 percentage points. Corresponding to this e¤ect is an improvement in treated individuals’knowledge about HIV, including sources of transmission, awareness of anti-retroviral drugs, and the need to take a second HIV test after at least three months from the …rst (window period). These are topics speci…cally covered in MTV Shuga. The e¤ects we estimate are robust to aggregating outcomes into indexes and to correcting for multiple hypothesis testing. We …nd more nuanced e¤ects on risky sexual behavior. On the one hand, the acceptability and reported incidence of concurrent sexual partnerships signi…cantly decreased. On the other hand, MTV Shuga did not induce greater condom use, neither as reported by respondents nor as revealed in an experimental game that our subjects played in health camps. Despite the lack of e¤ect on condom use, we do …nd signi…cant impacts on a biomarker that proxies for unprotected sex with risky partners. The likelihood of testing positive for Chlamydia, a common STD, decreased by 55 percent in response to treatment for women in our sample (the impact on men is in the same direction but statistically insigni…cant). This is consistent with the reduction in the number of concurrent partners, and possibly with a more general shift away from risky behaviors. MTV Shuga worked. The next question is why. To explore the importance of emotional involvement in the narrative, we make use of measures of viewer responses from the com- munication literature (Green and Brock, 2000; Murphy et al., 2011) that, to the best of our knowledge, have not been used in the economics literature before. In our endline survey we asked a battery of questions on how immersed in the story the respondent was while watching the show, and how much he or she identi…ed with the characters. Using these we construct two indexes, ‘Transportation ’and ‘ , and we show that the treatment e¤ects were Identi…cation ’ stronger for viewers that had higher values of these indexes. While these interaction e¤ects 2 need to be interpreted with caution since both Transportation and Identi…cation are corre- lated with other individual characteristics, the results are robust to including the interaction between treatment and a large number of observable controls. The results support the view that edutainment needs to be absorbing in order to work. The last main question that we seek to address has to do with the role of social e¤ects. The views and behaviors portrayed in MTV Shuga could signal a new norm about how one should interact with HIV positive people, with sexual partners, etc. In addition, our viewers may have expected Shuga to soon be released and have wide viewership, and they may have internalized the shift in norms that would take place as a result. To examine this possibility we take several approaches. First, in half of the locations where MTV Shuga was shown (randomly selected), viewers were shown statistics on the attitudes of others like them after viewing Shuga: we call this treatment arm T2. As our theoretical analysis makes clear, T2 should have di¤erent e¤ects on those viewers who in the baseline already believed that others had the kind of attitudes that MTV Shuga implicitly endorses, compared to viewers who believed that most others had di¤erent views from those in Shuga. The latter group is ‘surprised’in a more positive direction by the information (where positive is de…ned to mean ‘ in the direction endorsed by Shuga’ ) 2 and therefore should respond more strongly. To the best of our knowledge, this strategy of announcing post-viewing beliefs to test for conformity e¤ects is novel.3 However, we …nd no evidence of the predicted heterogeneous treatment e¤ect of T2; nor is the mean e¤ect of T2 any di¤erent from that of the ‘ basic’treatment (T1). A possible reason for the lack of a di¤erential e¤ect of T2 is that exposure to T1 may have already conveyed a precise enough signal about the norm, compared to which T2 adds no new information. If this were the case, we should …nd that the basic treatment T1 did not change individual priors regarding social norms in their community of origin. We instead …nd that it did, albeit not in an entirely robust way. However, we …nd that the observed change in individuals’own attitudes in response to Shuga was not mediated by the perceived change in norms in the way that Bayesian updating would have predicted. This suggests that the prevailing norms in the community are not the main driver of individual choices in our context. To explore social e¤ects further, our experiment included a third randomly assigned treat- ment (T3), cross-cut across T1 and T2. In T3 we o¤ered our viewers extra tickets to allow them to bring up to two friends to the screening. This treatment is meant to address two 2 This argument relies on the decision-maker being Bayesian. We discuss what could happen if this assumption fails in section 2.5. 3 In subsequent work Bursztyn, Gonzalez and Yanagizawa-Drott (2018) estimate the e¤ect of announcing others’ beliefs on married men’ s attitudes towards female labor force participation in Saudi Arabia. As will be clear from the description of our experiment, we announce the beliefs of people who watched Shuga, while Bursztyn et al. announce pre-intervention beliefs. 3 distinct concerns. First, if one were concerned that T2 fails to identify a relevant reference group for viewers (we told them that the announcement video and statistics were from “people like them in nearby communities” ), T3 by construction allows participants to bring members of their own social network. Second, in general it may be hard to change one’ s opinions alone, without knowing what one’ s friends will think. In particular, since MTV Shuga in‡ uences be- haviors relating to risky sexual behavior and HIV, it could be that the response is larger when people who are potentially sexual partners attend together. We …nd no evidence that T3 had any di¤erential e¤ect on attitudes and behavior. On the other hand, there is some evidence that when the invited friend was of the opposite sex, and therefore in Nigerian society more likely to be a sexual partner, the viewer is better informed about HIV. This is consistent with the interpretation that participants who attended with potential partners may have discussed the issues more in depth after the viewing. The di¤erential role of friends of the opposite sex also emerges when we test for spillovers from viewers onto other friends who were not invited to the showing (we collected a list of friends at baseline). We …nd weakly positive spillovers on HIV knowledge on average, stronger when the friend and the treated participant are of the opposite sex. There is no evidence of positive spillovers on attitudes and behavior. The last strategy we adopt to explore social e¤ects is non-experimental. We collected a range of standard measures of conformity, adherence to norms and independent thinking drawn from the literature in psychology (Schwartz, 2012). We test whether the e¤ect of MTV Shuga di¤ers depending on baseline values of these traits and we …nd no robust patterns. Taken together, the above evidence suggests that edutainment works by conveying infor- mation while at the same time entertaining viewers. It also suggests that coordination on social norms was not a big part of the e¤ect of MTV Shuga. This could be because viewers do not care about social norms when it comes to very private decisions, like ones relating to HIV and risky sexual behavior. Such interpretation would be consistent with the evidence from Bursztyn and Jensen (2015) and Bursztyn, Fujiwara and Pallais (2017), which are all about public acts of conformity. Compared to existing studies, we consider outcomes that are more sensitive and less discussed in public. Nonetheless, the fact that edutainment can have large individual e¤ects, even in the absence of coordination e¤ects, is quite striking. The enormous potential reach in terms of number of viewers and the low marginal costs of distribution make edutainment communication tools potentially very valuable for development policy. Our paper contributes to a growing literature on the e¤ects of the media on socioeconomic outcomes (see DellaVigna and La Ferrara, 2015 and La Ferrara, 2016 for a review). Part of this literature exploits non-experimental variation to study the e¤ects of commercially oriented TV programs (e.g., Jensen and Oster, 2009; Chong and La Ferrara, 2009; La Ferrara et al., 2012; Kearney and Levine, 2015, 2019). These evaluations typically use expansion of access to television over time as the main source of variation and do not focus on the mechanism 4 questions that we are able to investigate. Ours is the …rst RCT to look at the impact of edutainment on sexual behavior and HIV-related outcomes. Banerjee, Barnhardt and Du‡ o (2015), Ravallion et al. (2015), Coville et al. (2014) and Berg and Zia (2017) are examples of RCTs of the e¤ect of edutainment. These authors evaluate interventions to promote, respectively, the consumption of iron-forti…ed salt, knowledge about a public works program and the latter two on …nancial literacy. Compared to these studies, our goal is to a¤ect outcomes that are more sensitive and less discussed in public. To the best of our knowledge, we are the …rst to experimentally evaluate an edutainment TV series designed to change behavior determined by deep seated preferences such as those pertaining to HIV and sexual behavior. Our study also di¤ers because of the emphasis on trying to identify the underlying mechanisms.4 There are also a number of recent RCTs that randomize exposure to videos or short doc- umentaries containing information on role models (e.g., Bernard et al., 2014; Bjorvatn et al., 2015). These are not, strictly speaking, examples of edutainment and also focus on changing aspirations, an outcome we do not investigate. A di¤erent strand of literature to which our paper relates is that on social pressure and social image (e.g., DellaVigna et al., 2012; Bursztyn and Jensen, 2015; Bursztyn, Fujiwara and Pallais, 2017). While we share with these authors the interest in how individual beliefs and actions respond to the beliefs and actions of reference groups, we di¤er in one fundamental dimension. Participants in our experiment remain ‘ private’in the sense that their choices are not announced to others: the concern that their behavior may be observed and sanctioned by others may emerge in the long run but it is not a direct consequence of our experimental de- sign. What our experiment randomly makes public is information on the beliefs and behaviors of others. The contributions cited above manipulate instead the observability of individual behavior to others and …nd important e¤ects of ‘ . We believe that the two ap- social image’ proaches are strongly complementary and that much can be learnt in the future by enriching research designs to contemporaneously address all these dimensions. Another fruitful approach may be to endogenize the reference group and the emergence of the social norm as a result of individuals’past choices, as in Benabou and Tirole (2011). The remainder of the paper is organized as follows. Section 2 presents a stylized model that guides our empirical analysis. Section 3 describes the experimental design and section 4 the empirical strategy and data. In sections 5 and 6 we present results on our basic treatment and on social e¤ects, respectively. Section 7 contains robustness analysis and section 8 concludes. 4 Our work is also related to that of Paluck (2009), and Paluck and Green (2009), who study the impact of a radio soap opera on post-war reconciliation in Rwanda. These studies report some e¤ects on prescriptive norms but relatively weak impacts on behavior, which may be related to the limitations of radio relative to a more visual medium or to the persistency of the historical legacy of the Rwandan genocide. 5 2 A simple model of learning and conformity In this section we present a simple model to highlight potential channels of in‡uence of edu- tainment and guide our empirical speci…cations. 2.1 Model idea Our model is a variant of the standard Bayesian model of learning with quadratic preferences. Agents have a preference for being right as well as a preference for conformity. We model the e¤ect of our interventions as sources of additional signals about the right thing to do as well as possibly about the social norm. The idea is that Shuga generates engagement, and therefore signals that would have otherwise been ignored actually get through to the viewer. 2.2 Preferences We assume that the respondents in our study want to maximize a utility that depends on three terms: the distance to some objectively correct choice y , the di¤erence between the choice they make and their preferred point a and possibly also on the distance between the observed choice and the average choice in the peer population, Y: Formally at time t individual i chooses yit to maximize Eit [ (yit y )2 + (yit Yi )2 + (1 )(yit ait )2 ] (1) where Et is the expectation operator taken based on the information at time t: This tells us that yit = Eit [y ] + Ei [Yi ] + (1 )ait : Here captures the importance of information about the ‘ truth’, while picks up the degree of conformity. The fact that Yi is assumed not to change over time is based on the idea that while individual choices vary, it all averages out. Allowing Yi to drift over time would not change anything essential. The goal of the intervention was to raise the average choice of yit : MTV Shuga was intended to promote a liberal and informed discourse on HIV and risky sexual behavior, so we are e¤ectively assuming that a higher value of yi represents a more liberal/informed position. 2.3 Information and decisions In making this decision individual i starts from a prior on y , si0 N (y ; 1=py ) and a prior on Yi ; ri0 N (Yi ; 1=pY ), where p indicates the precision of the signal. Therefore the baseline choice yi0 , in both treatment and control, is given by yi0 = si0 + ri0 + (1 )ai0 : 6 We assume that ait evolves as a AR(1) but may be shifted up by exposure to Shuga: ait = ait 1 + T (i) + t where T (i) = 1 if T (i) = 1 (i is treated) and zero otherwise, 0 and t is distributed as N (0; 1=p ). We assume that in control the prior signals are all the information that individuals get. Therefore C yit = si0 + ri0 + (1 )ait : Obviously C yi 1 = si0 + ri0 + (1 ) ai0 + (1 ) 1 = (1 )( si0 + ri0 ) + yi0 + (1 ) 1: (2) In the di¤erent treatments each individual potentially gets a signal Si about y , Si N (y ; 1=pS ) as well as a signal about Yi , Ri N (Yi ; 1=pR ). The two signals are drawn independently, though this assumption is easily relaxed at the cost of some additional notation. We impose no assumptions about correlations of signals across individuals. The updated choice based on the new information is: T py si0 + pS Si pY ri0 + pR Ri yi 1 = + + (1 ) ai0 + (1 )( + 1 ): py + pS pY + pR We can rewrite this as T pS (Si si0 ) pR (Ri ri0 ) yi 1 = + + (1 )( si0 + ri0 ) + yi0 + (1 )( + 1 ): (3) py + pS pY + pR 2.4 Empirical approach 2.4.1 Treatment versus control Di¤erencing equation (2) from equation (3) and taking expectations conditional on yi0 and ri0 gives us T C pS E [Si ] pR E [Ri ] pS E [yi 1 yi 1 jyi0 ; ri0 ] = + E [ si0 jyi0 ; ri0 ] py + pS pY + pR py + pS pR ri0 + (1 ) : pY + pR Note that we are assuming that ri0 is known to the econometrician, though in fact we only have proxies for it. This does not make a big di¤erence since we will use the proxy when we 7 actually estimate the relationship. Si and Ri are also only in the mind of the decision-maker and will have to be proxied for by treatment dummies. However one can use the fact that yi0 ri0 = si0 + (1 )ai0 s expectation of si0 ; to come up with the econometrician’ ea P ey P ea P E [ si0 ] = (yi0 ri0 ) + e y e a; ey ea + P P ey ea + P P ey ea + P P ey ) to si0 and a e; 1=P under the assumption that the econometrician assigns a distribution N (y distribution N (e ea ) to ai0 and de…ning = (1 a; P 2 ) : Plugging this into the above gives us pS E [Si ] pR E [Ri ] pS Pea T C E [yi yi 1 jyi0 ; ri0 ] = + [ (y ri0 )+ 1 py + pS pY + pR py + pS Pea + Pey i0 ey P ea P pR + e y e a] ri0 + (1 ) ; ey ea + P P ey ea + P P pY + pR which is more conveniently rewritten as pS E [Si ] pR E [Ri ] pS ey P T C E [yi 1 yi 1 jyi0 ; ri0 ] = f + + (1 ) [ e y py + pS pY + pR py + pS P ey ea + P ea P pS Pea pR pS Pea e a]g y [ ]r : ea + P P ey ea + P py + pS P ey i0 pY + pR ea + P py + pS P ey i0 (4) It is plausible that Treatment 1 provides both a signal about the state of the world (Si ) and one about the social norm (Ri ). This suggests an estimating equation: yi1 = + T 1 (i) + yi0 + 'ri0 + T 1 (i)yi0 + T 1 (i)ri0 + "i ; (5) where T 1 (i) is a dummy equal to 1 when the status of i is treatment 1 and zero if it is control, pS Pa T 1 (i) represents the term in curly brackets, T 1 (i)yi0 picks up the term py +pS Pa +Py yi0 ea and T 1 (i)ri0 picks up the term [ pY p R +pR pS py +pS P ea +P ey ]ri0 : We expect to be positive if P Si and/or Ri and/or are positive enough; is clearly negative and can go either way.5 A su¢ cient condition for = 0 is that = 0; but it is not necessary. For many of the outcomes in our survey we have yi0 but not rio : In those cases we can take the expectation with respect yi0 alone to get an estimating equation: 5 The reason why the sign of is ambiguous is that a higher ri0 directly increases yi1 in Treatment but, for any given yi0 , it also tells us that si0 must be lower, and this depresses yi0 in Treatment. 8 yi1 = + T 1 (i) + yi0 + T 1 (i)yi0 + "i : (6) Finally we can also take unconditional expectations to get an estimating equation yi1 = + T 1 (i) + "i : (7) 2.4.2 Treatment 2 versus treatment 1 The di¤erence between treatment 1 and treatment 2 was supposed to be that the signal on Y , Ri ;was more precise –formally, the precision of the signal should be p0 R > pR : Pooling the observations from T 1 and T 2 into a single treatment category therefore does not change the estimating equation and we can use (5), (6) or (7) to estimate the overall impact of MTV Shuga (with and without the extra announcement). As for the di¤erence between T 1 and T 2; from equation (3) it is easy to see that T2 T1 p0 R pR E [yi 1 yi 1 jyi0 ; ri0 ] = ( )(E [Ri ] ri0 ): pY + p0 R pY + pR surprise’element of our treatment, i.e., the di¤erence The term (E [Ri ] ri0 ) represents the ‘ between the signal on Y i provided by Shuga and the individual’ s prior. This suggests an estimating equation yi1 = + T 2 (i) + yi0 + 'ri0 + T 2 (i)ri0 + "i : (8) We would expect to be positive and to be negative, unless = 0 or p0 R ' pR : Note that yio T 2 (i) does not enter this equation. This is because we have assumed that T2 provides no additional information about the correct choice y . In fact what is announced in T2 in our experiment is the mean opinion in a similar population after viewing Shuga. Since people know that others watched the same content as they did, they may assume that everyone else got the same signal on y as they did. Under this assumption, the announcement in T2 only contains new information on how people reacted to Shuga, thus yielding a more precise signal on Y . If we assumed instead that T2 viewers also received a more precise signal on the correct choice y ; the estimating equation for the extra e¤ect of T2 compared to T1 should also include an interaction term between yio and the T2 dummy, that is yi1 = + T 2 (i) + yi0 + 'ri0 + T 2 (i)yi0 + T 2 (i)ri0 + "i : (9) In our empirical analysis, we consider both alternatives and estimate (8) as well as (9). 9 2.5 Discussion The key implication of Bayesian learning that the above analysis makes clear is that what matters to the decision-maker is the surprise, the extent to which the signal that he or she s priors with gets from the world di¤ers from his/her priors. This is why the interaction of i’ treatment enters the estimating equation for the marginal e¤ect of T2 over and above T1 with a negative sign. Something similar may also hold in non-Bayesian models of learning, though for a slightly di¤erent reason. Consider one obvious alternative Bayesian learning– an ‘infection’model where the decision-maker, with some …xed probability, forgets his/her prior and adopts the signal he/she is exposed to. Otherwise he continues to hold onto the prior. This model would also generate a similar estimating equation: suppose treatment provides the decision-maker with an addition signal about the state of the world that he does not get in the control group, and T2 is more likely than T1 to provide him/her with a new signal about the social norm. Then the decision-maker’ s prior interacted with treatment will still enter the estimating equation with a negative sign, because treatment reduces the ex ante probability that he sticks to his prior. The fact that we assume that Shuga provides a signal to everyone in treatment is a conve- nient simpli…cation. Suppose instead we assumed that some viewers are engaged by the show but these viewers are all otherwise identical and being engaged is not correlated with their preferences or their prior information. This is equivalent to assuming that a …xed fraction of the treated population actually responds to the intervention. The rest behave as if they were in control, because they are not attentive to the message and e¤ectively assign a precision of zero to it. Because of the linearity of the decision rule and the fact that we are taking averages across the population, this generalization ultimately does not change our estimating equation. However it may be that those who are more likely to be attentive were also those who were more informed to start with (maybe because they had paid attention to previous messages) or more open to new ideas. In other words, there could be heterogeneity in, say, pS and this could be correlated with si0 and therefore yi0 . In that case the interaction between yi0 and treatment would pick up this extra responsiveness to the treatment among those who have a higher initial level of yi . This positive e¤ect would then counteract the negative interaction e¤ect coming out of our model. A large number of other assumptions have also gone into making the preceding analysis tractable and the notation less cumbersome than it would otherwise be. For one, as already noted, we assume that we have a proxy for ri0 even though in reality it is not observable to the econometrician. Speci…cally, in our empirical analysis we estimate ri0 using a survey question where i reports what is the number of people out of twenty in the community who have a particular view or take a particular action on an issue. This is potentially problematic for two 10 distinct reasons. First, there is no very good reason why our measure and not some correlate of it – 2 – is the right measure of the perceived norm. Second, from the point of view of like ri 0 conformity, the relevant peer group does not have to be the one that i is reporting about: he or she may be reporting about the objective reality of the local community (which is what our respondents were asked about) but when it comes to conforming, i may only care about his/her friends. We will try to deal with this issue in the empirical work by using alternative de…nitions of what i’ s reference group might be. Yet another approach would be to ignore all information about ri0 that we collected and use yi0 to pick it up (as we do for si0 ). This would be exactly like the case where we have no measure of ri0 ; discussed above, where we estimate (6). Finally, the simplicity of our estimating equations also derives from the fact that we assume quadratic preferences. Without this assumption how an individual reacts to new information will depend, for example, on how far he or she is from the views of the peer population. 3 Background and experimental design 3.1 Background The edutainment product we evaluate is a TV series called Shuga and produced by MTV Staying Alive Foundation. MTV Shuga is a TV drama designed to raise awareness and change attitudes and behavior related to HIV/AIDS among young people in Africa. It presents young Africans from various socioeconomic strata balancing bright futures with the negative conse- quences of high-risk behaviors. The third season of Shuga, whose impact we evaluate, was …lmed in Nigeria in 2013 and features prominent Nigerian actors and music, making it very appealing for the local public.6 In order to have exogenous variation in the exposure to the show, we conducted the study before MTV Shuga was widely distributed in Nigeria and we organized our own screenings in community centers, schools, and other locations that we could rent and that could accommo- date about 100 individuals. The series consists of eight 22-minute episodes, which we screened in two blocks of four episodes, for a total duration of about 90 minutes per screening. For the control group we chose another TV series …lmed in Nigeria, Gidi Up, which portrays a similar setting as MTV Shuga – urban and relatively upscale compared to the average population– but has no educational content. Also Gidi Up was screened in two blocks of comparable duration to Shuga. In all cases, screenings took place on Saturday or Sunday, and were one week apart. The study sites were 80 urban and peri-urban locations chosen in 7 towns across three 6 http://www.mtvshuga.com/show/?series=series-3 Note that series 1 and 2 were …lmed in Kenya and season 3 was largely unknown during our experiment: at follow-up only 6 percent of respondents in our control group said they had watched Shuga on TV. 11 states of South-West Nigeria. Locations were de…ned by drawing a 2-mile radius around each screening center where the intervention was implemented, and randomly selecting households within this radius.7 We ensured that there were bu¤er zones between communities to minimize the risk of contamination across study groups. These locations constitute our unit of random- ization. Online Appendix A describes the selection of locations and respondents in detail, and Appendix Figure A1 shows the geographic distribution of treatment and control locations. To identify study participants, we adopted a three-step recruitment strategy. First, enu- merators visited a random selection of 200 225 households per location. Second, the research team randomly selected one person aged 18 to 25 in each household to be invited to an initial …lm screening, which we denote as Screening 0. This …lm was di¤erent from MTV Shuga as the intent was to recruit amongst people available to attend …lm screenings during weekends, so as to reduce attrition. The selection was strati…ed by gender, half males and half females. Out of 17; 224 people invited to Screening 0, 6; 613 attended, with a turnout rate of 38:4 per- cent. Appendix Table A1 compares those who chose to attend Screening 0 and those who did not, on a number of socioeconomic characteristics. The two samples are well balanced; a few variables show signi…cant di¤erences, yet of extremely small magnitude: in all cases the normalized di¤erence is below the threshold of 0:25 suggested by Imbens and Rubin (2015). We thus conclude that those who attended Screening 0 were a representative subset of the households invited.8 The third step was the selection of our baseline sample. In each location enumerators randomly selected 64 individuals among those who attended Screening 0 and visited them at home to administer the baseline survey. From now on, we refer to these 64 individuals as ‘ main study participants’ 9 . At the end of the survey, the main study participants received invita- tions to attend two other screenings organized in the two weekends following the interview: Screenings 1 and 2. Those in the ‘ Friends treatment’(to be described in the next section) also received invitations to bring up to two friends of their choice. Note that invitees were not told in advance what they would watch, neither before Screening 0 nor before Screening 1. Attendance to the screenings was relatively high: 78:4 percent of those invited attended at least one of the two screenings, with the shares being 77 percent in the treatment group and 81 in the control group (signi…cantly di¤erent).10 Note that in all our analysis we report 7 The condition for a household to be in the study was that at least one of the members should be in the target age range of our intervention, i.e., 18-25 years old. 8 We also compared our sample to the 2008 Nigeria DHS, restricted to the South-Western region, and found that the two were quite similar in terms of religion (a third Muslim and two thirds Christian), years of education (around 11 years) and television-ownership rates (around 90 percent). 9 Appendix Table A2 shows balance in observable characteristics between the people who attended Screening 0 and were selected into the baseline sample and those that were not. 10 70:5 percent of our sample attended Screening 1, 57:2 attended Screening 2, and 49:4 percent attended both. The likelihood of attending Screening 2 conditional on attending Screening 1 is 70 percent in both treatment and control. 12 intention-to-treat e¤ects, using the initial assignment as our treatment indicator. In each location, all baseline survey and screening activities were concluded in four weeks. Implementation was rolled out so that activities in a given location were completed before mov- ing to the next, in order to minimize attrition due to subjects forgetting about the screenings, travelling or relocating. Online Appendix B reports the timeline of activities. 3.2 Experimental design Since individual-level randomization would run a signi…cant risk of contamination, we imple- mented a clustered randomized trial where our study locations were randomly assigned to treatment and control groups. The experiment was designed to allow us to study the impact of MTV Shuga screenings alone as well as that of being exposed to Shuga plus information on beliefs and values of peers. We created di¤erent treatment arms and strati…ed the sample so that each town would have an equal number of locations in each arm (where possible). Prior to the actual intervention, we piloted the Shuga screenings in some urban and peri- urban locations outside our sample frame. We used these pilots to shoot short videos with exit surveys’containing selected attitudinal interviews of participants, and to administer short ‘ questions. T1. Treatment T1 consisted only of MTV Shuga screenings and was administered in 27 randomly selected locations. Participants were shown the Shuga TV drama in two screenings of four 22-minutes episodes each. We did not organize any discussion at the end of the screenings, to ensure uniformity of the treatment and to make the experience more comparable – albeit not perfectly–to that of a viewer watching TV at home. T2. The second treatment (T2) involved another 27 randomly chosen locations and was the same as T1, except that after the MTV Shuga episodes we showed video-clips containing information on beliefs and values of peers in other communities who had watched Shuga. These video-clips were assembled using material from the pilot screenings and included interviews of youth condemning negative behaviors and praising positive ones after watching Shuga, as well as ‘smart graphs’with statistics. The intent was to raise awareness about how common certain beliefs and attitudes are among other participants, and how willing they may be to change them. T2 thus embeds a …rst type of ‘ , namely young people from other social referents’ communities who look similar to our respondents and who watched MTV Shuga. The choice of this way of implementing T2 represented a compromise between a number of di¤erent imperatives. Announcing the post-screening average – which is what we opted for–is the right thing to do if the population expects that the rest of their peer group will also see Shuga eventually and they want to be close to that post-Shuga consensus. This is a plausible case since MTV is a well-known brand and our viewers may have correctly expected that Shuga would be shown on television eventually. Consistent with this interpretation, below we show 13 that T1 participants updated their expectations on the prevailing values in their communities among other people who had not yet seen Shuga. An alternative possibility for T2 would have been to announce the pre-Shuga average, which would have the advantage that it would be from the same population/community as the one being treated and not from a similar population from elsewhere. On the other hand, our study participants might have discarded this as dated information if they expected their peers to see MTV Shuga and change their mind. Moreover our partners and funders were very worried that we would undermine Shuga’ s message by reminding participants of their and others’pre-Shuga views. Finally, even if our viewers did not believe that the social norm would shift as a result of Shuga, and therefore wanted to get closer to the pre-Shuga norm, they could back this out of the post-Shuga mean that was announced. They could do so using their prior about the distribution of signals about the state of world provided by Shuga, the signal they themselves got from Shuga and their prior on the distribution of preferred points in the population. This, it is easy to show (available from the authors), will generate the same estimating equation as the one we estimate. T3. To half of the treated individuals, randomly selected, we o¤ered the option of bringing up to two friends to the screenings. The goal of this treatment was to determine whether the e¤ect of Shuga di¤ers when individuals can discuss its content with close peers who also watched the drama. This treatment was randomized at the individual level and cut across the other two treatments. We denote this as the ‘ Friends treatment’(T3). Compared to T2, T3 includes a di¤erent type of ‘ : friends who are chosen by the individual and thus social referents’ surely belong to his/her social network. Take-up of T3 was relatively low. Of the 1775 main study participants who received an invitation, only 830 brought at least one friend to at least one screening (47 percent take-up), and 483 brought two friends. In our analysis, we estimate an intention-to-treat e¤ect and focus on an indicator for whether the respondent was given the option to bring a friend (T3). It is worth discussing why we did not include a ‘pure information treatment’in our experi- ment. A large body of evidence already exists on information campaigns to prevent HIV in low income countries (e.g., Fonner et al. 2014, McCoy et al. 2010), including Sub-Saharan Africa (Harrison et al. 2010, Michielsen et al. 2010). These studies consistently …nd limited or no impacts on reducing risky sexual behaviors.11 In a review of systematic reviews, Krishnaratne et al. (2016) concluded that results for HIV behavior change interventions were overall “dis- appointing” . The e¤ective approaches seem to be medical ones or conditional cash transfers (e.g., Bjorkman Nyqvist et al. 2018), which are not based on simple information provision 11 Du‡ o, Dupas and Kremer (2015) evaluate the impact of an HIV education program in Kenya in which primary school teachers were trained to deliver the national HIV/AIDS curriculum. They …nd that, when implemented alone, this program did not succeed in reducing STIs. It is only when combined with an education subsidy that the program was e¤ective. 14 and may also be quite expensive to scale in a large country like Nigeria. Given budget and logistical constraints, we therefore took stock of the evidence on the lack of e¤ectiveness of pure information campaigns and privileged an experimental design that would allow us to shed light on the channels for the e¤ect of edutainment. Summary of interventions. To sum up, we have three treatments: T1 and T2, random- Friends’ treatment (T3), which is cross-cut across T1 and ized at the cluster level, and the ‘ T2 and is randomized at the individual level. These treatments together cover 54 locations. The remaining 26 locations constitute our control group, where we screened the ‘ placebo’TV series Gidi Up. We complemented our data collection on the main study participants with a sample of contacts from their social networks that we use for detecting potential spillovers. In all loca- tions, before taking the baseline survey, study participants were asked to list two friends aged 18-25 to whom they regularly talked and who lived in the community. In each location, we administered the baseline and the follow-up surveys to a random sample of 15 contacts among those indicated by respondents who were not in the ‘ Friends’ treatment. We refer to these individuals as ‘network members’and we will conduct an impact analysis on them too. They should be distinguished from the friends that individuals brought in T3, on whom we have no information other than their gender.12 4 Empirical strategy and data 4.1 Empirical strategy Benchmark speci…cation. To estimate the average treatment e¤ect, we use two speci…ca- tions. One is the cross-sectional speci…cation corresponding to equation (7) in the model: yilc1 = Tilc0 + X0 ilc0 + c + "ilc1 : (10) where yilc1 is the outcome of interest for individual i who lives in location l within city c in the follow-up survey (time 1); Tilc0 is a dummy equal to 1 if the individual was assigned to treatments T1 or T2; Xilc0 is a vector of controls measured at baseline (time 0) that include age, years of education, and dummies for being enrolled in school, single, Muslim, speaking Yoruba as main language at home, speaking English as main or second language at home, not living with one’ s parents, household size, a wealth index, homeownership, and two dummies for father and mother having achieved more than secondary education; c denotes town …xed e¤ects. Appendix Tables A3 and A4 report variable de…nitions and summary statistics. 12 By construction, contacts of those main respondents who had been given extra tickets for friends (and who could have therefore attended the screenings) are not part of the spillover sample. 15 The second speci…cation includes the lag of the dependent variable and its interaction with treatment, as in model equation (6): yilc1 = yilc0 + Tilc0 + (Tilc0 yilc0 ) + X0 ilc0 + c + "ilc1 (11) where yilc0 is the baseline value of the dependent variable, and the remaining controls are de…ned as above. In both speci…cations (10) and (11) we cluster standard errors at the location level, which is our unit of randomization (we have 80 locations). E¤ect of announcement. Next we test whether the provision of information on how other people reacted to Shuga (T2) di¤erentially a¤ected viewers compared to the simple screening of the series (T1). In this case we restrict the sample to treated individuals and estimate a speci…cation corresponding to equation (9) in the model: yilc1 = yilc0 + T 2ilc0 + (T 2ilc0 yilc0 )+ eilc0 + X0 eilc0 ) + Y + (T 2ilc0 Y ilc0 + c + "ilc1 (12) where Yeilc0 is individual i’ s prior (elicited at baseline) on the average realization of outcome y in the community. We also estimate a shorter version of (12) that corresponds to equation (8) in the model and that does not include the interaction between T2 and yi0 : Heterogeneous e¤ects. To shed light on the importance of the edutainment component and of conformism, we exploit individual-level measures that we elicited through our survey. We estimate an augmented version of equation (11) that includes an interaction term between s involvement with the plot (or i’ Tilc0 and i’ s conformism at baseline), plus of course the standalone variable. Friends treatment. To test whether viewers who watched Shuga with a friend exhibited di¤erent responses, we estimate: yilc1 = yilc0 + F F riendilc0 + X0 ilc0 + c + "ilc1 (13) where F riendilc0 is a dummy that takes value one for individuals in T3. This regression is estimated on the treatment sample only, because only treated participants received the friends invitation. If the possibility of talking about the show with a friend reinforced the message in Shuga, one would expect bF > 0 for outcomes for which the main treatment e¤ect is positive and bF < 0 when the main treatment e¤ect is negative. Of course, this need not necessarily be the case, depending on the friend’ s own preferences. Spillovers. To estimate spillover e¤ects, we use a di¤erent sample of respondents j who were referred to us by our study participants i. We use the notation ji to indicate that j is a member of i’ s network. We estimate: yjilc1 = yjlc0 + S Tilc0 + S (Tilc0 yjlc0 ) + X0 jlc0 + c + "jilc1 (14) 16 where the outcome y and the controls X refer to network member j , but exposure to Shuga s friend i. In the presence of spillovers, the estimated coe¢ cient bS is indirect, only through j ’ should have the same sign as b in (11). In other words, if Shuga positively a¤ected i’ s outcome and i talked about it with his/her friend j , then j ’ s outcome would also respond positively (and vice versa). We also estimate a variant of (14) that includes an interaction between Tilc0 and a dummy for whether i and j have opposite sex. The idea is that discussion around the issues covered in Shuga may be di¤erent among people of same vs. opposite sex, and that the latter may be a coarse proxy for whether the pair is potentially a couple. Some behaviors, e.g., condom use, may be easier to adopt if both partners have been exposed to Shuga. Reporting results. Since we have a large number of outcome variables, we present results in two formats. First, to address the issue of multiple hypothesis testing, we group our original outcomes into indexes. This reduces the number of hypotheses actually tested and increases statistical power by reducing errors due to random variation at the level of the individual variables. We describe the construction of the indexes in section 4.2. Aggregation only partially solves the multiplicity problem, as we still have several hypotheses being tested jointly. To correct for this, we adjust p-values according to the free step-down resampling method (Westfall and Young, 1993) so that they can be used to control the family-wise error rate (FWER), de…ned as the probability of rejecting at least one true null hypothesis. Our second way of reporting results focuses on individual outcomes that can be regarded as important on their own. These are selected within the broader list of variables from which we compute the indexes, and we single them out because they are key to the overall message of Shuga (e.g., “you should wear a condom” ; “having concurrent partners can be risky”; etc.) or they capture speci…c messages that are strongly emphasized in certain episodes (e.g., “a young boy should be allowed to play football” ; “you have to take a second HIV test after 3 months” naive’standard errors corrected ; etc.). For our coe¢ cients of interest, we report both ‘ for clustering at the location level, and FWER-adjusted p-values – that adjust for multiple hypothesis testing–based on 10,000 replications. 4.2 Data and descriptive statistics Sample In total we interviewed 5166 main study participants at baseline and 4986 at follow up.13 Since our conditional speci…cation includes the lag of the dependent variable, our working sample consists of the individuals for whom we have both rounds of data, i.e., 4986 observations. 13 We performed power calculations using the Nigeria DHS 2008 and determined that a sample of 64 individuals per location, or 5120 individuals in total, half male and half female, would enable us to detect a change of between 0:15 0:20 standard deviations in our main outcomes of interest. Updated power calculations using our baseline data showed an improved minimum detectable e¤ect of 0:12 0:17 standard deviations. 17 For some speci…cations we use smaller samples because of missing data for speci…c outcomes. In Appendix Table A5 we regress the decision to participate in the follow up on the treat- ment dummy and on baseline values of our outcome indexes and of socioeconomic controls. As we can see, attrition is not selective on our main variables of interest. Indexes We group our outcome variables into two broad topics: those related to HIV and those related to risky sexual behavior. For each topic we elicit responses on knowledge, attitudes and behavior using several survey questions. The individual questions are then aggregated into indexes following Kling et al. (2007), i.e., we construct equally weighted averages of the z-scores of the variables that enter each index. For robustness, we also use a second method based on principal component analysis. Both methods are described in detail in Online Appendix C, and the list of variables entering each index (with their respective signs and loading factors) are shown in Appendix Tables A6 and A7. Variables are oriented so that the impact of treatment on each component of the index should be positive. We construct …ve indexes: HIV knowledge, HIV attitudes, HIV testing, Attitudes towards risky sex, and Risky sexual behavior. These can be brie‡ y described as follows. HIV knowledge measures how aware an individual is of the methods of transmission, the availability of drugs, and the timing of testing for HIV. Higher values of this index correspond to greater awareness. HIV attitudes captures respondents’inclination to potentially reveal their HIV+ status, allow HIV+ people to interact with the community, and not hold negative judgements towards them. A higher value of this index denotes more progressive attitudes (consistent with the message of MTV Shuga). The index HIV testing measures whether the respondent knows where to get tested, if he/she has been tested and when, if he/she picked up the results and if he/she asked for the test him/herself. Increasing values of the index HIV testing correspond to more active testing. A second family of variables relates to risky sexual behavior. The index Attitudes towards risky sex includes opinions regarding multiple concurrent partners, dating sugar daddies, and attitudes towards women who bring a condom. Again, increasing values of this index corre- spond to attitudes more consistent with the message of Shuga. Finally, Risky sexual behavior captures whether the respondent has multiple concurrent sexual partners (and the number), condom use during the last intercourse, and having a main or additional sexual partners. This index is only available for the subset of individuals who are sexually active. Increasing values of the index correspond to less risky sexual behavior, so we should expect a positive treatment e¤ect. Health camps objective’ outcome measures in addition to survey-based ones, we set In order to collect ‘ up ‘ health camps’in 80 schools to which survey respondents were invited when they took the follow-up survey. At the health camp participants were informed about testing by counsellors 18 and were o¤ered the opportunity to test for Chlamydia through urine sample collection. During the same session they also participated in a game that consisted in choosing between a …xed amount of money and a certain number of condom packs. At the end of the session participants received contact details of HIV counselling and testing centres in their town and were given a voucher that would entitle them to free HIV testing at one of these centers. Through the ID number on the voucher we know who took the HIV test, but we do not know the results.14 A second health camp was set up two weeks after the …rst, where participants were informed of the results of the Chlamydia test and –if they tested positive–they were prescribed treatment. Online Appendix D contains further details on the health camps. In our analysis we will use the following outcomes collected at health camps: (i) whether participants attended the health camp and took the Chlamydia test; (ii) whether they tested positive for Chlamydia; (iii) whether they redeemed the voucher to get tested for HIV; and (iv) whether they chose condoms over money when given the choice. Attendance to health camps was relatively high: 77 percent of the study participants at- tended, and on average this share was the same in treatment and control locations. Appendix Table A8 shows how baseline characteristics and baseline values of our outcomes correlate with the decision to participate in the health camp. People currently attending school and living outside the family were less likely to attend, possibly due to con‡ ict with school schedules. Treatment status is uncorrelated with the decision to attend, as are baseline values of our outcome indexes. Expectations Among the variables we collected through our survey, it is worth detailing how we elicited expectations regarding community attitudes, because these variables play an important role in our test for social e¤ects. For the main attitudinal outcomes we elicited two types of responses. The …rst was the individual’s own position, for example: “If you had HIV and you had a boyfriend/girlfriend, would you reveal your status to him/her?” . This type of variable is used as dependent variable in our analysis. The second category relates to the position of community members, for example: “If you picked 20 people of your age from your community who had a partner, how many would reveal their status to their partner if they had HIV?” . From this type of question we construct the share of community members who would choose a certain action (or support a certain statement) and we employ this variable as the ‘ prior’ about the average choice in the peer population (ri0 in the model, Y eilc0 in regression (12)). Due to constraints on the length of the questionnaire, we elicited these priors for some but not all of our outcomes of interest. For this reason, we will be able to estimate speci…cations (10) and (11) for all outcomes, and speci…cation (12) for a subset of them. Randomization check 14 NACA (2015) estimates that in 2012 the South-West region had an adult HIV prevalence rate of 2:8 percent. 19 Table 1 shows summary statistics for our outcomes of interest (Panel A) and the control variables (Panel B) at baseline. We report the mean in the control and in the treatment group, the p-value for the test that the di¤erence is zero, the normalized di¤erence and the number of observations for each variable.15 Our outcome indexes are well balanced: for all …ve the di¤erence in means is never statis- tically signi…cant. When we look at individual outcomes, out of 22 variables 3 have p-values of :05 or less (although these p-values are not corrected for multiple hypothesis testing: the p-values would be much higher if we accounted for that). Even so, the normalized di¤erence in means is extremely small, well below the cuto¤ of 0:25 suggested by Imbens and Rubin (2015). Turning to control variables, Panel B of Table 1 shows that variables such as gender, age, education, religion, language spoken, having a partner and living outside the family are well balanced. We have some imbalance in household size, wealth and parents’ education: on these variables the control group seems to be better o¤ than the treatment. The normalized di¤erences, however, are well below the 0:25 cuto¤, so in terms of economic signi…cance of the imbalance we do not …nd reasons for concern. Furthermore, we control for these variables in all our speci…cations. In Appendix Table A9 we perform an alternative test, regressing our treatment dummy on covariates (Panel A) as well as on covariates and outcome indexes (Panel B). None of the s education. regressors is signi…cantly di¤erent from zero, except for homeownership and father’ The F test for joint signi…cance always yields p-values greater than 0:10. 5 Results: average treatment e¤ects In this section we report our estimates of the impact of Shuga on a variety of outcomes, starting from the average treatment e¤ects and then exploring the role of involvement with the narrative. 5.1 HIV related outcomes Table 2 reports the average treatment e¤ects for outcomes related to HIV knowledge, attitudes and behavior. Panel A considers aggregate indexes as dependent variables, while Panel B 15 The normalized di¤erence is a scale-free measure of the di¤erence in distributions, recommended by Imbens and Wooldridge (2009): XA XB = p 2 2 SA + S B 2 2 where X A and X B are the means of covariate X in groups A and B, respectively, and SA and SB are the corresponding sample variances of X: 20 individual outcomes.16 Columns 1, 3, 5 estimate the cross sectional model (10), while columns 2, 4, 6 include the lag of the dependent variable and its interaction with treatment, as in (11). Appendix Table A11 shows the estimates from a simple ANCOVA model. Individual level controls and city …xed e¤ects are always included but not shown. Standard errors in parenthesis are adjusted for clustering at the screening center level. In square brackets we show p-values corrected for multiple hypothesis testing using FWER. At the bottom of the table we report the mean and the standard deviation of the dependent variable at follow-up in the control group, and the p-value for the test that the sum of the coe¢ cients on T reated and T reated Yt 1 is equal to zero when evaluated at the mean of Yt 1 . The results in Table 2 clearly show that exposure to MTV Shuga signi…cantly improved all HIV-related outcomes. In the top panel, the impact on respondents’knowledge about HIV is positive and signi…cant at the 1 percent level with either speci…cation and either method of correction for the standard errors. The magnitude of the e¤ect in the conditional model (column 2) when evaluated at the mean of Yt 1 corresponds to :13 of a standard deviation of this index. Shuga also improved attitudes towards people with HIV. The e¤ect on the aggregate index is again positive and signi…cant, with an e¤ect size of :10 of a standard deviation when evaluated at the mean (column 4). We detect positive and signi…cant impacts also on the HIV testing index: based on the estimate in column 6, treatment induced an increase in the aggregate index of :08 of a standard deviation when evaluated at the mean.17 As for the other coe¢ cients in the table, the lagged dependent variable is always signi…cantly correlated with current outcomes, while the sign on the interaction between treatment and the lagged dependent variables is negative (as predicted by the model) in two out of three cases, though typically insigni…cant. This may be an indication that there is actually heterogeneity in the response to the intervention along the lines we suggested above, which creates a countervailing e¤ect. In Panel B of Table 2 we consider three individual outcomes that are included in the in- dexes but are also of interest in and of themselves, because they are explicitly targeted in the messages of MTV Shuga. First of all, we compare the results on HIV testing obtained when using objective behavior from health camps (columns 1-2) and when relying on respondent’ s own reports (columns 3-4). Both variables show positive treatment e¤ects. The self-reported probability of testing increases by 2:5 percentage points, over a mean of 8:6 percent in the control group. The impact is even larger when we consider actual testing measured by the re- demption of vouchers received at health camps. In this case the probability of testing increases 16 As a benchmark, we employ the aggregate indexes constructed following Kling et al. (2007). Appendix Table A10 reports estimates for indexes constructed using principal component analysis: the results are qualitatively unchanged. 17 Appendix Table A12 reports separate estimates by gender. Impacts on knowledge and testing are stronger for women, while impacts on attitudes are more pronounced for men. 21 by 3:1 percentage points, which is a 100 percent increase over the control group mean.18 The remaining columns of Panel B consider outcomes that were explicitly addressed in MTV Shuga. In columns 4-5 we show that treatment improved respondents’knowledge about the need to take a second test and that this test should be after at least three months (window period).19 Columns 6-7 show improved support among treated respondents for the claim that HIV positive boys should be allowed to play football. This is noteworthy, as Shuga prominently features a sub-plot about a boy who was born with HIV and struggles to remain part of a football team. Appendix Table A13 shows analogous results for a larger number of individual outcomes that are also explicitly featured in the TV series, such as knowledge about transmission during pregnancy, contagion through sexual intercourse, awareness of anti- retroviral (ARV) drugs. The increase in HIV testing rates is an important result of our intervention and it is worth investigating deeper how it occurs. A …rst underlying factor is that individual knowledge about what testing is and why one should test improves. Secondly, it is possible that Shuga alerted people to the fact that HIV may be relatively common, thus leading them to update their expected risk of contagion.20 We test this hypothesis in Appendix Table A14. We …nd no e¤ects on the expectation that the respondent or the partner have HIV, while we …nd an increase of about 3 percentage points in the subjective probability that someone of the same age in the community is HIV+, representing an 8 percent increase over the control group mean.21 Testing may thus at least in part be a response to higher perceived risk. An additional dimension we investigate is the di¤erence between people who have a stable partner and people who do not. Appendix Table A15 shows that while the nature of the partner does not typically a¤ect the impact of Shuga on other outcomes, it has a strong predictive power when the dependent variable is the index of HIV testing (column 3). The increase in testing in response to treatment is entirely driven by people who do not have stable partners, possibly because they realize the risk involved (one of the messages featured in Shuga). 18 The sample in this regression is smaller because not all respondents attended health camps. Also, given that the option to test for HIV was given at follow up but not at baseline, for this regression we cannot estimate the speci…cation interacted with the baseline value of the outcome. 19 The discussion on the window period is explicitly featured in a scene where the main female character receives the results of her HIV test and the nurse tells her that she cannot consider herself free from risk until she takes a second test at least three months later. 20 For an analysis of testing in response to ‘ external HIV risk’, see Godlonton and Thornton (2013). 21 To elicit subjective probabilities, we followed Delavande and Kohler (2016) and asked respondents to select a number of beads ranging from 0 to 10, with each bead representing a 10 percent increment in probability. As commonly found in the literature, our respondents on average overestimate HIV prevalence rates. Note, however, that 1/4 of the characters in Shuga are HIV+, and in particular the main male character is. This may increase perceptions of how common HIV-AIDS is. 22 5.2 Risky sexual behavior In Table 3 we estimate the e¤ect of MTV Shuga on attitudes towards various sexual behaviors and on risky behavior itself as reported by the respondent. The sample for the behavioral outcomes (columns 3-4) is smaller because it is restricted to respondents who are sexually active. Panel A shows that the impact on the two aggregate indexes goes in the expected direction, namely improvement in attitudes and reduction in risky behavior (recall that our outcomes are constructed in a way that the expected treatment e¤ect is positive). However, the e¤ect is not signi…cant –although Appendix Table A12 shows a signi…cant improvement in attitudes for men. In Panel B we turn to two important variables that enter the index of risky sexual behavior: concurrent partnerships and condom use.22 Appendix Table A16 reports results for a broader set of individual outcomes. We …nd that respondents assigned to watch Shuga are less likely to have concurrent sexual partners. Based on the estimates in column 2, the total e¤ect of treatment on the probability of not having concurrent partners is +3:3 percentage points when evaluated at the mean of the dependent variable. Appendix Table A16 also shows a negative e¤ect on the number of concurrent partners is negative when evaluated at the mean, though the e¤ect size depends on the number of partners at baseline.23 The same table shows that attitudes towards concurrent partnerships also shift in a consistent direction. In the last two columns of Table 3, Panel B we estimate the e¤ect of treatment on the likelihood of having used a condom the last time the respondent had sex. We …nd no di¤erence between people who watched MTV Shuga and those who did not. The same holds if we consider alternative self-reported measures of condom use and future intentions to do so (Appendix Table A16). The general lack of an e¤ect is surprising because the importance of wearing condoms is repeatedly stressed in the show. We tested whether the results di¤ered if we distinguished between ‘ main’ and ‘ secondary’ partners (respondents may view the latter as less safe), but found insigni…cant results in both cases. Also, the result does not seem to be driven by reporting bias, as the next set of results shows that we obtain similar (nil) results when we use behavioral outcomes from health camps.24 In Table 4 we consider outcomes related to risky sexual behavior that are ‘ objectively’ 25 measured at our health camps. Panel A shows the results of a game where participants were 22 These outcomes refer to actual behavior of respondents who are sexually active, hence the smaller number of observations. 23 We also tested if treatment a¤ected the likelihood of being sexually active and found no e¤ects (results available from the authors). 24 Note that access to condoms is not an issue in our study sites, as 4 in 5 respondents said that they could get a condom in less than 10 minutes if they wanted. 25 The sample includes only respondents that attended health camp. Appendix Table A8 shows that the likelihood of attending health is uncorrelated with treatment status and with our index of risky sexual behavior (measured at baseline). 23 o¤ered a choice between the value of one pack of condoms in cash, and one, two or three packs of condoms. The dependent variable takes value one if the respondent chose the condoms over the monetary amount and zero otherwise. While participants were more likely to choose condoms when the relative price was lower (i.e., when o¤ered a higher number of packs), choice behavior did not di¤er among those who watched Shuga and those who did not. Results are equally insigni…cant for women and men (columns 3 to 6). This zero e¤ect of treatment aligns with the results obtained in Table 3 when looking at self-reported condom use. The absence of an e¤ect on condom use is consistent with di¤erent possible explanations. One is that there is strong cultural resistance to condoms in the Nigerian context and Shuga was simply unable to overcome such resistance. Another is an endogenous response to safer sexual behavior by treated individuals. As the incidence of concurrent partnerships is reduced for treated individuals (see Table 3), and possibly the nature of the partners becomes safer, the reduced risk may have induced our treatment group to rely less on condoms. An additional result consistent with the interpretation that the treatment group adopted a safer behavior comes from the prevalence of sexually transmitted diseases. Panel B of Table 4 shows the e¤ect of treatment on the probability of testing positive for Chlamydia. Results are shown for the full sample, for women and for men. While the estimated e¤ect is negative and comparable in size in all three samples, it is only statistically signi…cant for the female subsample. This is not surprising as Chlamydia is more prevalent among women. Also, the fact that the prevalence is very low in general makes it di¢ cult to detect an e¤ect with high precision. On the other hand, the magnitude of the e¤ect is quite sizeable relative to the baseline prevalence rate: exposure to Shuga leads to a 55 percent decrease in the likelihood that women test positive for Chlamydia. As discussed above, even in the absence of an e¤ect on condom use, this improvement may be generated by more careful behavior on behalf of the respondent, e.g., decreasing the number of sexual partners or choosing ‘ safer’partners.26 5.3 The role of involvement with the narrative According to the proponents of entertainment education, a key advantage of these programs is that viewers get engaged with the narrative and this leads them to pay more attention to content, learn from characters and be less defensive against external inputs (Singhal and Rogers, 1999). To test the role played by these factors we included in our endline survey a series of questions created by communication experts to measure two key dimensions. The …rst dimension is what Green and Brock (2000) call ‘ transportation’, and which could be alternatively described as ‘ . Individuals who are transported into the narrative engagement’ of a movie tend to focus their cognitive attention on its messages and to have heightened 26 We tested whether, conditional on showing symptoms, treated respondents were more likely to seek treat- ment for STDs and found that they were not. The estimated coe¢ cient on Treated is insigni…cant and equal to 0:019; where the mean of the dependent variable is 0:15. 24 emotions, which reduces counter-arguing. We use twelve questions proposed by Green and Brock (2000), which include statements about things that happen during the screening and ask respondents to agree or disagree on a scale of 1 to 5. Example of these statements include: “You were distracted by activities in the room around you” ; “You wanted to learn how the story ended”; “It a¤ected you emotionally” ; “You had a clear picture of the characters in the story”. The second dimension we want to explore is the extent to which viewers identify with the characters. Identi…cation is understood to make viewers more receptive to the modeling of behavior and more likely to rehearse the arguments presented (Murphy et al., 2011). We use ten questions proposed by Cohen (2001), also in the form of statements with 5-point scale responses, which include for example: “While viewing the show you felt as if you were part of the action”; “you wanted the characters to succeed in achieving their goals”; “you felt you had experienced the same thing as the character” . We aggregate the above questions into a Transportation and an Identi…cation index, respectively, using principal component analysis. Appendix Table A17 reports the loading factors for the two indexes. While these indexes cannot be considered as exogenous, in Appendix Table A18 we inves- tigate which observable characteristics correlate with transportation and identi…cation. We …nd that the only robust correlate of both indexes is wealth. For identi…cation, also gender and speaking English as a primary or secondary language at home matter.27 In estimating heterogeneous e¤ects according to Transportation and Identi…cation, we therefore control for the interaction between treatment and observable characteristics Xi0 : In Table 5 we estimate the e¤ect of treatment on the …ve indexes of HIV outcomes and sexual behavior, including an interaction term between Treated and Transportation (Panel A) or Treated and Identi…cation (Panel B), plus the standalone variables.28 Odd-numbered columns do not include the interactions T reated Xi0 , while even-numbered ones do. Transportation and Identi…cation enhance the e¤ect of watching Shuga, as shown by the positive coe¢ cient on the interaction term between these indexes and treatment. In the top panel, this coe¢ cient is positive and signi…cant for three out of …ve indexes; in the bottom panel for two out of …ve (after correcting for multiple hypothesis testing). In terms of magnitude, for example, based on the estimates in column 1 of Panel A, a one standard deviation increase in Transportation is associated with a 0:45 standard deviation increase in HIV knowledge for the treatment group compared to the control one. The corresponding e¤ect for a one standard deviation increase in Identi…cation (Panel B) is a 0:3 standard deviation increase in HIV knowledge.29 27 Women tend to identify less with the characters in Shuga, and people who speak English at home identify more (not surprisingly given that the language of Shuga is English). 28 Online Appendix Table A19 reports the cross sectional results (without including the lagged dependent variable and its interaction with treatment), which yield a very similar coe¢ cient on the interaction between treatment and Transportation (or Identi…cation ). 29 Note that the negative coe¢ cient on the standalone indicator Treated in the even-numbered columns of 25 While the results in Table 5 are strongly suggestive of a role for the entertainment com- ponent in inducing behavior change, as mentioned above one should be careful in inter- preting them causally. However, notice that when we do not control for T reated Xi0 and T reated yi0 (odd-numbered columns), the coe¢ cients on Treated Transportation and Treated Identi…cation are very similar both in terms of magnitude and of signi…cance. This increases our con…dence that unobservables may not be driving the correlations we estimate. 6 Results: social e¤ects An important focus of this paper is whether, in addition to understanding if edutainment interventions are on average successful, we can say anything on the extent to which social e¤ects may reinforce or undermine the impact of edutainment. Since identifying the relevant reference group is not obvious, we employ di¤erent strategies to address this question. First, we rely on experimental variation in the announcement of other people’ s reaction to Shuga (treatment T2). Second, we test if the basic version of MTV Shuga (treatment T1) a¤ected viewers’perceptions about norms in their own community, and if this explains the change in their behavior. Third, we exploit experimental variation in the possibility of bringing friends to the show (treatment T3). Finally, we provide non-experimental evidence on treatment e¤ect heterogeneity according to individuals’self-assessed degree of conformism. 6.1 Announcement treatment Our …rst experimental design for testing the importance of social e¤ects relies on complement- ing the basic treatment with an announcement on how other viewers reacted to Shuga. As explained in section 3.2, in half of the treated locations we complemented MTV Shuga with a short video that included interviews with young people who had watched the show, as well as ‘smart graphs’with statistics on their reactions. Our model predicts that, if social e¤ects are important ( > 0) and if our manipulation increases the precision of the signal about other people’ s choices (p0 R > pR ), then the di¤erence between T2 and T1 should depend on the interaction between the T2 dummy and the individual’ s prior about the social norm (see equation 8). Table 6 reports our estimates of equation (12) for the three outcomes for which we an- nounced statistics in T2. The dependent variables are indicators for whether the respondent states that (i) he/she would reveal his/her status to the partner (column 1); (ii) it is not OK to table 7 is only apparently counter-intuitive: given that those speci…cations include a full set of interactions between Treated and Xi0 , the coe¢ cient on Treated is hard to interpret. As we report in the last row of table 7 (panels A and B), under the most conservative speci…cation the fraction of individuals for whom the overall treatment e¤ect is positive ranges from 66 to 98 percent for HIV-related outcomes and is around 50 percent for outcomes related to risky sexual behavior. 26 date a sugar daddy in order to …nance one’ s education (column 2); and (iii) men should only have one partner (column 3). The variable Prior on community (ri;t 1 ) is the respondent’ s baseline expectation of the share of community members who agree with each statement. The sample includes only treated individuals as we aim at comparing the e¤ect of T1 and T2. The results in Table 6 show that T2 did not have a di¤erential e¤ect compared to T1 and that the coe¢ cient on T2*Prior on community is never signi…cant. This suggests that learning about the reactions of other viewers did not elicit signi…cant conformity e¤ects on top of what the basic showing of Shuga may have already done. In Appendix Table A20 we also test whether on average T2 had a di¤erential impact on our …ve outcome indexes compared to T1. We …nd that it did not.30 6.2 Conformity e¤ects of basic treatment The lack of a signi…cant impact of T2 does not necessarily imply the absence of conformity. A possible reason for failing to …nd an e¤ect of T2 is that the ‘basic’treatment T1 may have already conveyed a signal about the prevailing norm in the reference group, compared to which T2 adds no new information. We next test whether this interpretation is supported in the data. The …rst step is to establish whether T1 produced a shift in individual priors regarding social norms in the respondent’ s community. In Panel A of Table 7 we estimate speci…cation (11) using as outcomes the respondents’expectations of the share of community members who would behave in a certain way or support certain views. The set of dependent variables in this table is the one for which we elicited such expectations, as described in section 4.2. We exclude from the sample people assigned to T2, so the regressor of interest is T1 and we compare the basic screening of MTV Shuga to the control TV series. The results in Panel A suggest some degree of shift in perceived norms generated by T1. While Shuga did not a¤ect perceptions about how many community members would reveal their HIV+ status or the status of a family member, it did improve attitudes towards HIV+ people (e.g., shopkeepers or boys playing football). The e¤ects on blame also qualitatively point to a reduction in the stigma associated with HIV, although they are only signi…cant for one out of three variables. In Panel B of Table 7 we test whether the change in individuals’own attitudes was mediated by their perceived change in the prevailing norms. To this end, we turn to model equation (5) and observe that, if people were Bayesian, the coe¢ cient on T 1 ri;t 1 should be negative in columns 1, 3 and 4 (where the main treatment e¤ect should be positive) and positive in the remaining ones (where the main e¤ect should be negative). We only …nd this pattern in one 30 We also tried to test whether T2 increased the precision of the signal about community norms, using as a proxy for precision respondents’ s own assessments of how sure they were about their priors. Unfortunately, we have little variation in this as 92 percent of respondents said that they were “sure” about their estimates. 27 out of seven outcomes (column 7). We thus conclude that changes in perceived norms do not seem to account for the impact of Shuga on individual attitudes and behavior. 6.3 Friends treatment Our treatment T2 provided information on the reactions of “young people in neighboring communities”who had watched Shuga. These people looked similar to our respondents in age and socioeconomic status, but they did not personally know them. We next consider as an alternative peer group the respondents’own friends. Our treatment arm T3 was designed to test if watching Shuga with a friend led to di¤erent outcomes. As described in section 3.2, we randomly provided half of the treated sample with tickets that they could give to up to two friends. Table 8 reports the estimated impacts on our …ve outcome indexes. The main regressor of interest is the dummy T3, which takes value 1 for respondents who received tickets to bring friends: its coe¢ cient thus captures an intention-to-treat e¤ect. In columns 1, 4, 7, 10, 13 we estimate an ANCOVA model, while in columns 2, 5, 8, 11 and 14 we also include the interaction between T3 and the lagged dependent variable. The coe¢ cient of T3 is never signi…cant in these speci…cations. One possible interpretation is that social e¤ects are absent altogether and individuals do not care about what their friends say. Another possibility is that there are social e¤ects, but half of the sample brought friends who were positively inclined towards the messages of MTV Shuga, while another half brought friends who would ‘ talk them out’of those messages, generating a zero overall e¤ect. 31 In columns 3, 6, 9, 12 and 15 we test whether the e¤ect of T3 di¤ers for people who brought at least one friend of the opposite sex, compared to those who brought only friend(s) of the same sex. In the Nigerian context this may be a proxy for the fact that the individual attended the screening with his/her partner. We …nd a signi…cant e¤ect on HIV knowledge, suggesting that people who watched Shuga with a potential partner learnt more, possibly because the incentives to share knowledge and discuss HIV-related issues were greater. No other outcome, however, shows this e¤ect. 6.4 Spillovers The di¤erential impact of treatment on knowledge depending on the gender of the friend also emerges from our analysis of spillovers. As described in section 3.2, in each location we interviewed a random sample of ‘ network members’ who were not part of treatment T3 by 31 As explained in section 3.1, due to logistical constraints we could not collect information at the screening sites on the friends that people brought along, except for their gender. Therefore we do not know the opinions of friends. 28 design, and thus could not have been directly in‡uenced by Shuga. To test whether people who watched Shuga passed on any of the e¤ects to friends who did not watch it, in Table 9 we estimate model (14). The observations refer to network members, but the treatment status is that of the main study participant who ‘ nominated’the respondent. The variable ‘ Friend of Treated ’is a dummy equal to 1 if the respondent was a friend of a treated individual. The odd-numbered columns present the results of estimating equation (14).32 We detect positive spillovers on HIV knowl- edge and negative ones on testing, though neither e¤ect survives the correction for multiple hypothesis testing. There is no signi…cant e¤ect for the other variables. In columns 2, 4, 6, 8 and 10 we test whether the e¤ects are di¤erent for network members who have the opposite sex of the treated individual. Our conjecture is that friends of the opposite sex may include the respondent’s partner, and messages like those conveyed by Shuga may be particularly e¤ective if shared between members of a couple. Column 2 shows that indeed the positive e¤ect on knowledge is stronger for friends of the opposite sex. No e¤ect is found for other outcomes. Overall, the above results uncover the presence of some knowledge spillovers regarding HIV, albeit not signi…cant at conventional levels when correcting for multiple hypothesis testing. People who watched Shuga seem to have passed on some ‘ factual’information to their friends, but no spillovers are detected on attitudes and behavior. This suggests that, while edutainment programs may have trickle-down e¤ects when it comes to information provision, in order to generate attitudinal and behavioral change direct exposure to the program is needed.33 These results can also partly speak to the issue of social e¤ects: if the untreated individuals in our spillover sample conformed to the changed attitudes and behaviors of their friends who were exposed to Shuga, we should observe some impact on their own attitudes and behavior. The fact that we only observe (weak) impacts on knowledge is consistent with a modest role of conformism in the setting we study. 6.5 Self-assessment of conformism Our …nal exercise to gauge the importance of social e¤ects tests whether the impact of treat- ment di¤ers based on respondents’baseline propensity to conform with other people’s views. In the model, the parameter captures how costly it is to deviate from the choices of a reference group. This strategy has the advantage that it would work even if di¤erent people conformed 32 Appendix Table A21 reports the cross-sectional model. 33 A caveat is that the samples of ‘main respondents’and of ‘ network members ’may be di¤erent because the former were (randomly) selected from a population that revealed interest in TV programs by attending screening 0, while the latter were (randomly) selected from lists of contacts provided by the main participants. We tested for balance between the two groups and found that all outcome indexes except Attitudes towards risky sexual behavior are balanced (see Appendix Table A22). 29 with the views of di¤erent types of peer groups (e.g., some compare themselves to the average Nigerian while others only care about their friends). The variation we are using here is speci…c to the individual and tells us how much they care about the particular group that they have chosen to compare themselves to. We use a series of questions aimed at measuring how strongly individuals identi…ed with three of the values categorized by Schwartz (2012): conformity, tradition and self-direction. Online Appendix C describes the construction of the indexes in detail and Appendix Table A23 reports the list of variables and their loading factors from principal component analysis. We construct the following three indexes: (i) Conformity captures how inclined an individual is to restrain his/her own choices if these were to upset others or violate social norms; (ii) Tradition captures individuals’ acceptance and commitment to the values that their culture or religion promote; (iii) Self-direction : captures how inclined an individual is to think and act independently. ‘ Tradition ’and ‘ Conformity ’are similar in that they capture individuals’ willingness to subordinate themselves to what is expected from them, but they di¤er in the group respondents subordinate themselves to: in the case of conformity it is mainly people (e.g., parents or peers), while in the case of tradition it is religious and cultural customs. In Table 10 we estimate a series of regressions having as outcomes our …ve outcome indexes, and as main regressors of interest Treated and the interaction of Treated with Conformity (Panel A), Tradition (Panel B) and Self-direction (Panel C), plus the lagged dependent variable and its interaction with treatment.34 The goal is to test if exposure to MTV Shuga had di¤erential e¤ects depending on viewers’degree of conformity or independent judgement. Save a couple of exceptions, the interaction of treatment with these indexes is never signi…cant at conventional levels. Clearly we cannot interpret these results in a causal sense, as conformism may be correlated with unobservables, but it is instructive that we fail to uncover signi…cant correlations. 7 Robustness In this section we discuss additional results and robustness checks to deal with some potential concerns. Social desirability bias. Some of our dependent variables (notably biomarkers, HIV testing and condom choice) are objectively observed in health camps. Others are elicited through a survey but are factual (e.g., knowledge about transmission and treatment of HIV). Another set of outcomes, however, are self-reported and possibly subject to social desirability bias (e.g., attitudes, number of partners, etc.). Note that if reporting bias were similar across treated and control subjects, this would not be a problem: the concern only arises if treated individuals are more likely to misreport in a direction consistent with the message of Shuga. 34 Appendix Table A24 reports results for the cross-sectional model. 30 Three pieces of evidence lead us to believe that our results are not driven by experimenter demand e¤ects. First, the point estimates of our treatment e¤ect on objective and subjective indicators for HIV testing are very similar at 0:031 and 0:025, respectively (see Table 2). In the presence of experimenter demand e¤ects, one would expect the latter coe¢ cient to be larger. Second, many relevant self-reported outcomes do not display signi…cant treatment e¤ects. Notably, we …nd no e¤ects on self-reported condom use (Panel B of Table 3) and on …ve other condom-related variables (Appendix Table A16). Third, if treated individuals were di¤erentially a¤ected by reporting bias, we should …nd that when they have a chance of pleasing the research team by choosing condoms in the experimental game played in health camps, they should be more likely to do so. Panel A of Table 4 shows that this was not the case. Heterogeneous e¤ects by education and language. An interesting question is whether our e¤ects di¤er depending on viewers’level of education or ability to understand English (as MTV Shuga was in English). In Appendix Table A25 we interact the treatment dummy with the respondent’s years of education (odd-numbered columns) and with a dummy for whether respondent speaks English as the primary or secondary language at home (even-numbered columns). The average respondent in our sample has 11:6 years of education, and 96 percent of the respondents speak English at home as either primary or secondary language.35 We …nd that Shuga had a stronger e¤ect on HIV knowledge and attitudes of relatively more educated respondents, although the coe¢ cient on the interaction term with education is not signi…cant when adjusting for multiple hypothesis testing. This may seem surprising as one may expect television to be a particularly e¤ective means of communication for audiences with low literacy, but it should be noted that ours is a uniformly highly educated sample, as is the region.36 So it is plausible that the education e¤ect we uncover re‡ects the fact that the 37 message of Shuga was relatively progressive. No robust e¤ects are found for the interaction between treatment and speaking English at home, possibly because the latter variable has very little variation in our sample. 8 Conclusions In this paper we have tested the e¤ectiveness of an entertainment education TV series, MTV Shuga, aimed at providing information and changing attitudes and behaviors related to HIV. 35 Note that English is the language of instruction in schools and is the only o¢ cial language in the country. 36 The high level of education is not an artifact of our sample: the average respondent in the 2008 Nigeria DHS (of similar age and living in the South West) had 11 years of education. 37 Note that despite the negative coe¢ cient on the standalone treatment dummy, the overall e¤ect of treatment is abundantly positive when evaluated at the mean years of education. For example, ceteris paribus the impact on knowledge is positive for any respondent with at least 9:5 years of education, which corresponds to the 4th percentile in the distribution of education. 31 The simple model we set up to motivate the analysis captures the idea that edutainment can work through an ‘ social’channel. We conducted a randomized individual’or through a ‘ controlled trial in urban Nigeria where young viewers were exposed to MTV Shuga or to a non-educational TV series. Among those who watched Shuga, we created additional variation social messages’they received and in the people with whom they watched the show. in the ‘ We found that MTV Shuga led to signi…cant improvements in knowledge about and atti- tudes towards HIV and to less risky sexual behavior. Treated subjects were twice as likely to get tested for HIV 8 to 9 months after the intervention. We also found reductions in STDs among women. Our experimental manipulations of the social norm component, on the other hand, did not produce signi…cantly di¤erent results from the main treatment. Finally, we de- tected (weak) spillovers on friends who did not watch Shuga in terms of HIV knowledge, but not on attitudes and behavior. While it is virtually impossible to embed every possible type of social interaction in a single experiment, our experimental design allowed for several di¤erent types of social e¤ects, and we found no evidence that any of those explains the impact of Shuga. We thus learnt two lessons. First, the ‘ individual’ e¤ect of edutainment seems to have prevailed in the context of our study, and this is remarkable because it suggests that -at least in the context of HIV- AIDS- people react to the messages they see on TV regardless of what others say. Second, if policymakers wanted to leverage social e¤ects in edutainment, they should experiment with di¤erent (and potentially larger) reference groups, e.g., schools, villages, etc. Finally, more research is needed to assess the potential role of conformity when the private versus public nature of the message is varied. This seems especially relevant for the edutainment agenda given the growing importance of social networks in today’ s society. References [1] Bandura, A. (1976), Social Learning Theory. Englewood Cli¤s, NJ: Prentice-Hall. [2] Banerjee, A., S. Barnhardt and E. Du‡ o (2015), “Movies, Margins and Marketing: En- couraging the Adoption of Iron-Forti…ed Salt”, in D. Wise (ed.), Insights in the Economics of Aging, NBER. [3] Benabou, R. and J. Tirole (2011), “Laws and Norms”, NBER Working Paper 17579. [4] Berg, G., and Zia, B. (2017). “Harnessing Emotional Connections to Improve Financial Decisions: Evaluating the Impact of Financial Education in Mainstream Media.” Journal of the European Economic Association, 15(5), 1025-55. 32 [5] Bernard, T., S. Dercon, K. Orkin, and A. S. Ta¤esse (2014), “The Future in Mind: Aspirations and Forward-Looking Behaviour in Rural Ethiopia”, CSAE Working Paper WPS/2014-16. [6] Bjorkman Nyqvist, M., L. Corno, D. de Walque and J. Svensson (2018). “Incentiviz- ing Safer Sexual Behaviour: Evidence from a Lottery Experiment on HIV Prevention”, American Economic Journal: Applied Economics, 10(3): 287-314. [7] Bjorvatn, K., A. Cappelen, L. Helgesson Sekeiz, E. Sørensen, B. Tungodden (2015), “Teaching through television: Experimental evidence on entrepreneurship education in Tanzania” , mimeo, NHH Norwegian School of Economics. [8] Bursztyn, L. and R. Jensen (2015), “How Does Peer Pressure A¤ect Educational Invest- ments?”, Quarterly Journal of Economics, 130(3), 1329-1367. [9] Bursztyn, L., T. Fujiwara and A. Pallais (2017), “Acting Wife: Marriage Market Incentives and Labor Market Investments,” American Economic Review, 107(11), 3288-3319. [10] Bursztyn, L., A. Gonzalez and D. Yanagizawa-Drott (2018), “Misperceived Social Norms: Female Labor Force Participation in Saudi Arabia,” NBER Working Paper 24736. [11] Chong, A. and La Ferrara, E. (2009). “Television and Divorce: Evidence from Brazilian Novelas” Journal of the European Economic Association: Papers & Proceedings 7(2-3), 458-468. [12] Cohen, J. (2001). “De…ning identi…cation: a theoretical look at the identi…cation of audi- ences with media characters.” Mass communication and society, 4(3), 245-64. [13] Coville, A., Di Maro, V., Zottel, S., and Dunsch, FA (2014). “Nigeria’s Nollywood Nudge. An entertaining approach to savings” , in Enhancing …nancial capability and behavior in low- and middle-income countries, ch. 10, The World Bank. [14] Delavande, A. and H.-P. Kohler (2016), “HIV-related Expectations and Risky Sexual Behavior in Malawi,” Review of Economic Studies, 83(1), 118-164. [15] DellaVigna, S., J. List and U. Malmendier (2012), “Testing for Altruism and Social Pres- sure in Charitable Giving,” The Quarterly Journal of Economics, 127 (1), 1-56. [16] DellaVigna, S., and E. La Ferrara (2015). “Economic and Social Impacts of the Media” , in Anderson, S., Strömberg, D., and Waldfogel, J. (eds.), Handbook of Media Economics, vol. 1. North-Holland, Amsterdam. [17] Du‡ o, E. P. Dupas and M. Kremer (2015), “Education, HIV, and Early Fertility: Exper- imental Evidence from Kenya” , American Economic Review, 105(9), 2757-97. 33 Reilly and M. Sweat (2014). “School based [18] Fonner, V., K. Armstrong, C. Kennedy, K. O’ sex education and HIV prevention in low- and middle-income countries: a systematic review and meta-analysis,” PLoS One. [19] Godlonton, S. and R. Thornton (2013), “Learning from Others’HIV Testing: Updating Beliefs and Responding to Risk,” American Economic Review: Papers and Proceedings, 103(3), 439-444. [20] Green, M. C., & Brock, T. C. (2000). “The role of transportation in the persuasiveness of public narratives.” Journal of Personality and Social Psychology, 79, 701–721. [21] Harrison, A., M. Newell, J. Imrie, G. Hoddinott (2010). “HIV prevention for South African youth: which interventions work? A systematic review of current evidence,” BMC Public Health. [22] Imbens, G. and J. Wooldridge (2009), “Recent Developments in the Econometrics of Program Evaluation”, Journal of Economic Literature, 47(1), 5-86. [23] Imbens, G. and D. Rubin (2015), Causal Inference for Statistics, Social, and Biomedical Sciences. Cambridge University Press. s Status [24] Jensen, R., and Oster, E. (2009). “The Power of TV: Cable Television and Women’ in India.”The Quarterly Journal of Economics 124(3), 1057-1094. [25] Kearney, M.S., and Levine, P.B. (2015). “Media In‡ uences on Social Outcomes: The Impact of MTV’ s 16 and Pregnant on Teen Childbearing.” American Economic Review, 105(12) , 3597-3632. [26] Kearney, M.S., and Levine, P.B. (2019). “Early Childhood Education by Television: Lessons from Sesame Street”. American Economic Journal: Applied Economics 11(1), 318-50. [27] Kling, J.R., J.B. Liebman and L.F. Katz (2007), “Experimental Analysis of Neighborhood E¤ects.” Econometrica, 75(1), 83-119. [28] Krishnaratne, S., B. Hensen, J. Cordes, J. Enstone and J. Hargreaves (2016). “Interven- tions to strengthen the HIV prevention cascade: a systematic review of reviews,” Lancet HIV, 3(e307– 17). [29] La Ferrara, E. (2016). “Mass Media and Social Change: Can We Use Television to Fight Poverty?”, Journal of the European Economic Association, 14(4), 791-827. [30] La Ferrara, E., Chong, A., and Duryea, S. (2012). “Soap Operas and Fertility: Evidence from Brazil.”American Economic Journal: Applied Economics 4(4), 1-31. 34 [31] McCoy, S., R. Kangwende and N. Padian (2010). “Behavior change interventions to pre- vent HIV infection among women living in low and middle income countries: a systematic review.” AIDS Behavior. [32] Michielsen, K., M. Chersich, S. Luchters, P. De Koker, R. Van Rossem and M. Temmerman (2010). “E¤ectiveness of HIV prevention for youth in sub-Saharan Africa: systematic review and meta-analysis of randomized and nonrandomized trials.” AIDS. [33] Murphy, S. T., L. B. Frank, M. B. Moran, and P. Patnoe-Woodley (2011), “Involved, Transported, or Emotional? Exploring the Determinants of Change in Knowledge, Atti- tudes, and Behavior in Entertainment-Education”, Journal of Communication, 61, 407-31. [34] NACA - National Agency for the Control of AIDS (2015), Global AIDS Response Country Progress Report, Abuja, Nigeria. [35] Padian, N., S. McCoy, J. Balkus and J. Wasserheit (2010), “Weighing the gold in the gold standard: challenges in HIV prevention research”, AIDS, 24(5), 621-35. [36] Paluck, E.L. (2009), “Reducing intergroup prejudice and con‡ ict using the media: A …eld experiment in Rwanda” , Journal of Personality and Social Psychology, 96, 574-587. [37] Paluck, E.L., and Green, D.P. (2009). “Deference, Dissent, and Dispute Resolution: A Field Experiment on a Mass Media Intervention in Rwanda.” American Political Science Review 103(4), 622-644. [38] Perez-Truglia, R. (2017), “Political Conformity: Event-Study Evidence from the United States.” Review of Economics and Statistics, forthcoming. [39] Ravallion, M., D. van de Walle, P. Dutta and R. Murgai (2015). “Empowering Poor People through Public Information? Lessons from a Movie in Rural India”, Journal of Public Economics, 132, 13–22. [40] Schwartz, Shalom H. (2012). “An Overview of the Schwartz Theory of Basic Values”. Online Readings in Psychology and Culture, 2 (1), 1-20. [41] Singhal, A., and E. M. Rogers (1999). Entertainment education: A communication strategy for social change. Mahwah, NJ: Erlbaum. [42] Westfall, P. and S. Young (1993), Resampling-based Multiple Testing: Examples and Meth- ods for P-value Adjustment, vol. 279, John Wiley & Sons. 35 Table 1: Average individual characteristics, pre-treatment Mean Mean Diff=0 Normalized No. Obs. Control Treated (p-value) Diff.(a) (1) (2) (3) (4) (5) Panel A: Outcomes Indexes HIV knowledge -0.049 0.205 0.177 0.029 5166 HIV attitudes 0.036 -0.016 0.625 -0.010 5166 HIV testing -0.048 -0.035 0.916 0.002 5166 Attitudes towards risky sexual behavior 0.006 -0.041 0.638 -0.010 5166 Risky sexual behavior (conditional on sexually active) -0.018 0.079 0.458 0.020 3246 Individual variables HIV transmitted during pregnancy 0.612 0.611 0.962 -0.001 5166 Mentions ARV drugs spontaneously 0.020 0.024 0.330 0.021 5166 Mentions drugs to live longer with HIV 0.619 0.634 0.310 0.021 5166 Recognizes ARV when mentioned by enumerator 0.187 0.213 0.030 0.046 5166 Knows that second test is necessary 0.277 0.287 0.450 0.016 5166 Knows about 3-months window period 0.074 0.089 0.078 0.038 5166 Can get HIV through intercourse 0.947 0.948 0.897 0.003 5166 Would buy from an HIV+ shopkeeper 0.415 0.427 0.398 0.018 5166 An HIV+ boy should play footbal 0.579 0.571 0.571 -0.012 5166 People HIV+ should not be blamed 0.652 0.632 0.165 -0.029 5166 HIV is not punishment for sleeping around 0.433 0.465 0.031 0.046 5166 Would reveal HIV status 0.707 0.694 0.365 -0.019 5166 Tested last 6 months (self-reported) 0.053 0.055 0.675 0.009 5166 Men should have one partner only 0.842 0.860 0.106 0.034 5163 Women should have one partner only 0.880 0.898 0.056 0.040 5166 Not ok date sugardaddy to finance educ 0.760 0.745 0.265 -0.024 5166 Not ok date sugardaddy for money 0.670 0.675 0.713 0.008 5166 Not ok date sugardaddy to go out 0.886 0.866 0.050 -0.042 5166 If a woman brings a condom does not mean she’s not serious 0.579 0.603 0.111 0.034 5166 Has not had multiple concurrent sexual partners 0.776 0.786 0.521 0.018 2933 Number of current sexual partners if sexually active 1.329 1.276 0.119 -0.048 2349 Used condom the last time he/she had sex 0.520 0.495 0.219 -0.036 2690 Panel B: Controls Female 0.473 0.474 0.943 0.002 5166 Age 20.618 20.614 0.962 -0.001 5166 Currently attending school 0.342 0.350 0.565 0.012 5166 Years of education 11.598 11.596 0.950 -0.001 5166 Speaks English 0.958 0.962 0.484 0.015 5166 Single 0.232 0.224 0.509 -0.014 5166 Does not live with the family 0.229 0.246 0.186 0.028 5166 Household size 4.482 4.257 0.001 -0.067 5166 Wealth index 1.781 1.736 0.010 -0.054 5150 Home owner 0.448 0.355 0.000 -0.135 5165 Father’s education > secondary 0.376 0.314 0.000 -0.092 3928 Mother’s education > secondary 0.252 0.214 0.004 -0.065 4393 Muslim 0.370 0.374 0.799 0.005 5166 Yoruba 0.920 0.921 0.894 0.003 5166 Notes: (a) Normalized difference is the difference in the sample means of treatment and control groups divided by the square root of the sum of the sample variances. 36 Table 2: Impact on HIV outcomes Panel A: HIV indexes Dep. Var. (Yt ): HIV knowledge HIV attitudes HIV testing (1) (2) (3) (4) (5) (6) Treated 0.898*** 0.782*** 0.344** 0.339*** 0.353** 0.335** (0.244) (0.215) (0.135) (0.103) (0.148) (0.128) [0.001] [0.001] [0.022] [0.003] [0.022] [0.012] Treated*Yt−1 -0.065* 0.008 -0.033 (0.035) (0.037) (0.032) [0.201] [0.826] [0.523] Yt−1 0.390*** 0.360*** 0.472*** (0.029) (0.030) (0.027) R-squared 0.085 0.198 0.050 0.179 0.092 0.261 P-value test joint sig(a) 0.000 0.001 0.011 Observations 4986 4986 4986 4986 4986 4986 Mean Dep. Var. (Control) 0.0400 0.0400 0.0439 0.0439 -0.139 -0.139 Std Dev Dep. Var. (Control) 6.106 6.106 3.474 3.474 4.145 4.145 Panel B: HIV individual outcomes HIV testing HIV+ boy should be Dep. Var. (Yt ): Objective(b) Self-reported Knows window period allowed to play football (1) (2) (3) (4) (5) (6) (7) Treated 0.031** 0.025*** 0.026*** 0.049*** 0.045*** 0.051*** 0.080*** (0.013) (0.009) (0.009) (0.012) (0.012) (0.016) (0.028) [0.014] [0.014] [0.004] [0.001] [0.003] [0.012] [0.022] Treated*Yt−1 -0.038 -0.007 -0.046 (0.063) (0.062) (0.036) [0.550] [0.955] [0.581] Yt−1 0.253*** 0.286*** 0.300*** (0.054) (0.048) (0.030) R-squared 0.023 0.031 0.059 0.028 0.072 0.044 0.123 P-value test joint sig(a) 0.005 0.000 0.006 Observations 3828 4982 4971 4986 4986 4986 4986 Mean Dep. Var. (Control) 0.033 0.086 0.086 0.129 0.129 0.662 0.662 Std Dev Dep. Var. (Control) 0.180 0.280 0.280 0.335 0.335 0.473 0.473 Controls(c) Yes Yes Yes Yes Yes Yes Yes Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. FWER adjusted p-values in square brackets, based on 10,000 replications. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. (a) P-value test of joint significance: p-value for the null that the sum of the coefficient on Treated and Treated*Yt−1 (evaluated at the mean) is zero. (b) For this outcome we cannot estimate the model including the lag of the dependent variable because testing through health-camps was only offered at follow-up. (c) Controls in each regression include: female, age, currently attending school, years of education, english spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. 37 Table 3: Impact on Risky Sexual Behavior Panel A: Indexes Attitudes towards risky Risky sexual behavior sexual behavior Dep. Var. (Yt ): (1) (2) (3) (4) Treated 0.149 0.148 0.175 0.148 (0.091) (0.090) (0.151) (0.145) [0.205] [0.196] [0.337] [0.317] Treated*Yt−1 0.001 -0.117** (0.034) (0.056) [0.982] [0.085] Yt−1 0.292*** 0.322*** (0.028) (0.046) R-squared 0.021 0.099 0.100 0.152 P-value test joint sig(a) 0.104 0.318 Observations 4986 4986 3070 3070 Sample Full Sample Sexually Active Mean Dep. Var. (Control) 0.00186 -0.0631 Std Dev Dep. Var. (Control) 3.452 3.625 Panel B: Individual outcomes Has NOT had multiple Used condom last time Dep. Var. (Yt ): concurrent partners he/she had sex Treated 0.027* 0.113*** -0.003 -0.021 (0.015) (0.041) (0.019) (0.032) [0.232] [0.018] [0.844] [0.506] Treated*Yt−1 -0.103** 0.038 (0.041) (0.053) [0.028] [0.574] Yt−1 0.288*** 0.319*** (0.035) (0.047) R-squared 0.119 0.172 0.073 0.194 P-value test joint sig(a) 0.007 0.513 Observations 3339 2623 3084 2308 Mean Dep. Var. (Control) 0.780 0.494 Controls(b) Yes Yes Yes Yes Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. FWER adjusted p-values in square brackets, based on 10,000 replications. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. (a) P-value test of joint significance: p-value for the null that the sum of the coefficient on Treated and Treated*Yt−1 (evaluated at the mean) is zero. (b) Controls in each regression include: female, age, currently attending school, years of education, english spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. 38 Table 4: Risky sexual behavior outcomes measured at health camps Panel A: Demand for condoms Full Sample Females Males Dep. Var. =1 if chose condoms over N50 (1) (2) (3) (4) (5) (6) Treated 0.013 -0.024 0.012 -0.006 0.014 -0.045 (0.021) (0.043) (0.026) (0.054) (0.028) (0.066) Treated * # packs offered 0.018 0.009 0.028 (0.019) (0.028) (0.030) # packs offered 0.059*** 0.046*** 0.049*** 0.043* 0.068*** 0.049* (0.009) (0.016) (0.012) (0.023) (0.014) (0.025) Constant -0.061 -0.033 -0.304** -0.292* -0.085 -0.039 (0.116) (0.118) (0.148) (0.160) (0.187) (0.194) Observations 3,827 3,827 1,844 1,844 1,983 1,983 R-squared 0.137 0.137 0.054 0.054 0.063 0.063 Controls(a) Yes Yes Yes Yes Yes Yes Mean Dep. Var. (Control) 0.383 0.383 0.229 0.229 0.520 0.520 Panel B: STD biomarkers Full Sample Females Males Dep. Var. =1 if tested positive for Chlamydia (1) (2) (3) Treated -0.014 -0.017* -0.013 (0.012) (0.010) (0.015) Observations 3,820 1,839 1,981 R-squared 0.010 0.024 0.014 Controls(a) Yes Yes Yes Mean Dep. Var. (Control) 0.029 0.031 0.013 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. 39 Table 5: Involvement with the narrative Dependent variable: HIV knowledge HIV attitudes HIV testing Attitudes towards Risky sexual behav- risky sexual behav- ior (for sexually ac- ior tive) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Transportation Index Treated*Transportation 0.428*** 0.436*** 0.141** 0.146** 0.120 0.117 0.311*** 0.318*** 0.091 0.092 (0.108) (0.107) (0.067) (0.066) (0.076) (0.076) (0.064) (0.065) (0.066) (0.065) [0.001] [0.000] [0.075] [0.061] [0.123] [0.133] [0.000] [0.000] [0.169] [0.167] Treated 1.035*** -7.329** 0.333*** -1.970 0.281* -2.116 -0.009 -1.813 0.085 -3.607* (0.221) (3.561) (0.125) (1.721) (0.161) (2.021) (0.115) (1.905) (0.149) (2.080) [0.000] [0.122] [0.020] [0.439] [0.086] [0.439] [0.936] [0.345] [0.809] [0.168] Treated*Yt−1 -0.053 -0.075* -0.008 -0.019 -0.055 -0.073 -0.027 -0.038 -0.131** -0.123* (0.038) (0.042) (0.038) (0.039) (0.043) (0.045) (0.038) (0.038) (0.065) (0.066) [0.430] [0.222] [0.829] [0.633] [0.430] [0.222] [0.483] [0.328] [0.101] [0.130] Transportation -0.058 -0.069 0.032 0.026 -0.000 -0.001 -0.007 -0.008 -0.038 -0.037 (0.076) (0.076) (0.039) (0.039) (0.048) (0.048) (0.045) (0.046) (0.047) (0.046) Yt−1 0.385*** 0.401*** 0.376*** 0.384*** 0.500*** 0.512*** 0.309*** 0.316*** 0.359*** 0.355*** (0.029) (0.030) (0.030) (0.030) (0.037) (0.037) (0.030) (0.030) (0.051) (0.052) Observations 3,753 3,753 3,753 3,753 3,753 3,753 3,753 3,753 2,279 2,279 R-squared 0.215 0.220 0.186 0.188 0.272 0.274 0.119 0.125 0.175 0.180 Controls (a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls*Treated No Yes No Yes No Yes No Yes No Yes Mean Dep. Var. (Control) 0.0400 0.0400 0.0439 0.0439 -0.139 -0.139 0.00186 0.00186 -0.0631 -0.0631 Std Dev Dep. Var. (Control) 6.106 6.106 3.474 3.474 4.145 4.145 3.452 3.452 3.625 3.625 Share with treatment effect > 0 0.892 0.791 0.892 0.764 0.864 0.661 0.523 0.520 0.426 0.483 Panel B: Identification Index Treated*Identification 0.251*** 0.273*** 0.062 0.063 0.098 0.092 0.162*** 0.160** 0.016 0.010 (0.092) (0.092) (0.059) (0.061) (0.074) (0.075) (0.061) (0.061) (0.060) (0.059) [0.025] [0.015] [0.353] [0.388] [0.353] [0.388] [0.022] [0.019] [0.785] [0.868] Treated 1.061*** -7.357** 0.358*** -2.026 0.285* -2.172 0.013 -1.908 0.102 -3.601* (0.214) (3.567) (0.125) (1.740) (0.166) (2.010) (0.115) (1.914) (0.151) (2.071) [0.000] [0.126] [0.012] [0.426] [0.092] [0.426] [0.915] [0.320] [0.784] [0.163] Treated*Yt−1 -0.050 -0.072* -0.006 -0.017 -0.055 -0.074 -0.022 -0.032 -0.131** -0.125* (0.038) (0.042) (0.038) (0.040) (0.043) (0.045) (0.038) (0.038) (0.065) (0.066) [0.477] [0.249] [0.866] [0.658] [0.477] [0.249] [0.574] [0.397] [0.102] [0.122] Identification -0.021 -0.037 0.037 0.035 0.006 0.007 0.026 0.030 -0.028 -0.021 (0.066) (0.066) (0.041) (0.043) (0.055) (0.055) (0.041) (0.042) (0.039) (0.038) Yt−1 0.385*** 0.401*** 0.376*** 0.384*** 0.500*** 0.512*** 0.309*** 0.316*** 0.358*** 0.356*** (0.029) (0.030) (0.030) (0.030) (0.037) (0.037) (0.030) (0.030) (0.051) (0.052) Observations 3,753 3,753 3,753 3,753 3,753 3,753 3,753 3,753 2,279 2,279 R-squared 0.211 0.214 0.183 0.185 0.271 0.274 0.110 0.115 0.175 0.180 Controls(a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls*Treated No Yes No Yes No Yes No Yes No Yes Mean Dep. Var. (Control) 0.0400 0.0400 0.0439 0.0439 -0.139 -0.139 0.00186 0.00186 -0.0631 -0.0631 Std Dev Dep. Var. (Control) 6.106 6.106 3.474 3.474 4.145 4.145 3.452 3.452 3.625 3.625 Share with treatment effect > 0 0.957 0.834 0.987 0.825 0.883 0.661 0.549 0.511 0.197 0.487 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. FWER adjusted p-values in square brackets, based on 10,000 replications. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. (a) Controls include: female, age, currently attending school, years of education, English spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. 40 Table 6: Effect of announcing social norm (T2) Dep. Var. (Yt−1 ): Would reveal HIV status Not ok date sugardaddy to fi- Men should have one nance education partner only (1) (2) (3) (4) (5) (6) T2 0.016 0.019 0.011 0.069 -0.037 -0.011 (0.027) (0.038) (0.032) (0.050) (0.033) (0.047) Yt−1 0.231*** 0.235*** 0.215*** 0.246*** 0.110*** 0.129*** (0.018) (0.021) (0.020) (0.033) (0.020) (0.031) Prior on community (rt−1 ) 0.125** 0.123** -0.130*** -0.117** 0.001 -0.005 (0.051) (0.050) (0.048) (0.048) (0.036) (0.038) T2 * Prior on community (rt−1 ) -0.073 -0.071 -0.038 -0.065 0.073 0.085 41 (0.067) (0.067) (0.064) (0.065) (0.054) (0.055) T2*Yt−1 -0.006 -0.060 -0.038 (0.037) (0.042) (0.041) Observations 3,402 3,402 3,402 3,402 3,394 3,394 R-squared 0.091 0.091 0.080 0.081 0.044 0.045 Controls (a) Yes Yes Yes Yes Yes Yes Mean Dep. Var. (Control) 0.738 0.738 0.751 0.751 0.903 0.903 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (54 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. Sample includes treated individuals only (T1+T2). (a) Controls include: female, age, currently attending school, years of education, English spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. Table 7: Shuga and perceived norms Would prefer to Would buy fruit Would reveal Think an HIV+ Think testing Think people Think HIV is a keep HIV+ of & veg from HIV+ HIV+ status boy should be for HIV means with HIV should punishment for family member a shopkeeper to partner allowed to play sleeping around be blamed sleeping around secret football (1) (2) (3) (4) (5) (6) (7) Panel A: Dep. var. (ri,t ) is respondent’s expectation on the share of peers who [column heading] T1 0.018 0.003 0.037*** 0.035** -0.019 -0.017 -0.059*** (0.014) (0.025) (0.012) (0.015) (0.020) (0.019) (0.022) [0.886] [0.993] [0.033] [0.184] [0.952] [0.952] [0.096] T1*ri,t-1 -0.023 -0.008 -0.138*** -0.061* 0.020 0.016 0.065** (0.039) (0.033) (0.038) (0.034) (0.032) (0.028) (0.030) [0.998] [0.998] [0.006] [0.537] [0.998] [0.998] [0.335] ri,t−1 (a) 0.149*** 0.139*** 0.268*** 0.204*** 0.104*** 0.163*** 0.102*** (0.029) (0.024) (0.027) (0.023) (0.022) (0.020) (0.023) R-squared 0.035 0.041 0.107 0.127 0.039 0.051 0.059 Mean Dep. Var. (Control) 0.319 0.589 0.260 0.377 0.538 0.570 0.644 Panel B: Dep. var. (Yi,t ) is respondent’s own opinion on [column heading] T1 0.030 0.089* 0.091*** 0.088** -0.066** -0.009 -0.093** (0.040) (0.047) (0.025) (0.035) (0.032) (0.038) (0.036) 42 [0.701] [0.181] [0.005] [0.074] [0.176] [0.822] [0.071] T1*Prior on community (ri,t−1 ) 0.064 -0.008 -0.022 -0.025 0.084 0.007 0.156** (0.069) (0.060) (0.079) (0.051) (0.051) (0.060) (0.063) [0.888] [0.990] [0.988] [0.981] [0.476] [0.989] [0.112] T1*Yi,t−1 -0.043 -0.075* -0.111*** -0.040 -0.049 -0.032 -0.056 (0.034) (0.042) (0.041) (0.044) (0.034) (0.043) (0.042) [0.570] [0.387] [0.006] [0.594] [0.570] [0.594] [0.570] Prior on community (ri,t−1 ) 0.057 0.040 0.081 0.096*** -0.029 0.117*** 0.028 (0.048) (0.040) (0.055) (0.036) (0.038) (0.039) (0.049) Yi,t−1 (b) 0.279*** 0.223*** 0.359*** 0.284*** 0.152*** 0.193*** 0.260*** (0.028) (0.028) (0.031) (0.030) (0.024) (0.032) (0.028) R-squared 0.108 0.086 0.131 0.124 0.058 0.073 0.100 Mean Dep. Var. (Control) 0.726 0.712 0.518 0.694 0.161 0.311 0.487 Observations 3,262 3,262 3,262 3,262 3,262 3,262 3,262 Controls(c) Yes Yes Yes Yes Yes Yes Yes Notes: Standard errors in parentheses adjusted for clustering at the screening centre level (53 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. FWER adjusted p-values in square brackets based on 10,000 replications. Sample includes only T1 + Control. (a) ri,t-1 is the respondent’s baseline expectation of the share of community members who agree with the statement in the heading of each column. (b) Yi,t-1 is a dummy taking value 1 if at baseline the respondent agreed with the statement in the heading of each column. (c) Controls include: female, age, currently attending school, years of education, english spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. Table 8: Friend invitations (T3) Dep. Var. (Yt−1 ): HIV knowledge HIV attitudes HIV testing Attitudes towards risky Risky sexual behavior (for sexual behavior sexually active) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) T3 -0.105 -0.105 -0.149 -0.001 -0.001 -0.131 0.143 0.144 0.119 -0.068 -0.068 -0.161 -0.206* -0.207* -0.285 (0.192) (0.190) (0.235) (0.107) (0.107) (0.128) (0.126) (0.126) (0.153) (0.109) (0.109) (0.140) (0.122) (0.123) (0.172) [0.897] [0.902] [0.693] [0.991] [0.992] [0.693] [0.697] [0.696] [0.693] [0.897] [0.902] [0.693] [0.386] [0.407] [0.412] Yt−1 0.319*** 0.320*** 0.320*** 0.363*** 0.367*** 0.364*** 0.434*** 0.426*** 0.433*** 0.290*** 0.288*** 0.289*** 0.211*** 0.209*** 0.211*** (0.024) (0.027) (0.024) (0.023) (0.031) (0.023) (0.020) (0.025) (0.020) (0.021) (0.026) (0.021) (0.034) (0.045) (0.034) T3 * Yt−1 -0.001 -0.007 0.015 0.003 0.005 (0.034) (0.029) (0.031) (0.035) (0.054) [0.901] [0.999] [0.994] [0.999] [0.999] 43 Friend Invited is of Opposite Sex 0.898** -0.122 -0.258 0.108 -0.413 (0.412) (0.214) (0.265) (0.217) (0.255) Brought a Friend -0.219 0.326** 0.143 0.165 0.315 (0.274) (0.161) (0.235) (0.155) (0.257) Observations 3,402 3,402 3,402 3,402 3,402 3,402 3,402 3,402 3,402 3,402 3,402 3,402 2,117 2,117 2,117 R-squared 0.188 0.188 0.189 0.184 0.184 0.184 0.246 0.246 0.246 0.107 0.107 0.107 0.135 0.135 0.136 Controls (a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean Dep Var. (Control) 0.839 0.839 0.839 0.415 0.415 0.415 0.200 0.200 0.200 0.124 0.124 0.124 0.126 0.126 0.0827 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (54 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. Sample includes only treated individuals (T1+T2). FWER adjusted p-values in square brackets, based on 10,000 replications. T3 is a dummy taking value 1 if the respondent was given extra tickets to bring up to two friends. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. Table 9: Spillovers Dep. Var. (Yt ): HIV knowledge HIV attitudes HIV testing Attitudes towards Risky sexual be- risky sexual behav- havior (for sexu- ior ally active) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Friend of Treated 0.640* 0.540 -0.082 -0.124 -0.512* -0.545* -0.138 -0.085 -0.337 -0.311 (0.332) (0.328) (0.232) (0.228) (0.274) (0.289) (0.207) (0.218) (0.314) (0.314) [0.106] [0.199] [0.812] [0.587] [0.106] [0.178] [0.625] [0.704] [0.455] [0.536] Friend of Treated*Opposite Sex 2.493** 1.317 0.840 -1.383 -1.089 (1.234) (0.961) (1.143) (1.057) (1.735) [0.228] [0.386] [0.519] [0.455] [0.624] Opposite Sex -1.540* -0.072 -0.456 0.562 0.211 (0.786) (0.592) (1.011) (0.735) (0.926) 44 Yt−1 0.434*** 0.438*** 0.395*** 0.397*** 0.301*** 0.300*** 0.273*** 0.271*** 0.300*** 0.299*** (0.060) (0.060) (0.060) (0.061) (0.071) (0.072) (0.060) (0.061) (0.100) (0.101) Friend of Treated * Yt−1 -0.110 -0.113 -0.148* -0.148* 0.013 0.012 0.032 0.034 -0.107 -0.106 (0.072) (0.072) (0.075) (0.075) (0.082) (0.083) (0.072) (0.073) (0.118) (0.119) [0.236] [0.238] [0.154] [0.156] [0.876] [0.881] [0.789] [0.646] [0.789] [0.624] Observations 924 924 924 924 924 924 924 924 580 580 R-squared 0.249 0.250 0.177 0.179 0.222 0.222 0.139 0.140 0.184 0.185 Controls (a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean Dep. Var. (Control) -0.185 -0.185 -0.316 -0.316 -0.164 -0.164 -0.107 -0.107 -0.213 -0.213 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. ”Opposite sex” is a dummy equal to 1 if respondent has the opposite sex as his/her treated friend. Sample includes network friends that have non-missing Y t−1 . FWER adjusted p-values in square brackets, based on 10,000 replications. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. Table 10: Treatment effects and conformism (1) (2) (3) (4) (5) Dependent variable Yt : HIV knowledge HIV attitudes HIV testing Attitudes towards Risky sexual behav- risky sexual behav- ior (for sexually ac- ior tive) Panel A: Conformity Treated 0.775*** 0.345*** 0.335** 0.148 0.149 (0.214) (0.105) (0.128) (0.089) (0.145) [0.001] [0.004] [0.012] [0.188] [0.312] Treated*Conformity 0.192 -0.188* 0.028 -0.044 0.009 (0.132) (0.099) (0.093) (0.076) (0.097) [0.276] [0.167] [0.768] [0.815] [0.926] Conformity -0.052 0.138* -0.077 0.054 0.030 (0.108) (0.076) (0.065) (0.060) (0.084) Yt−1 *Treated -0.067* 0.013 -0.034 0.004 -0.116** (0.036) (0.036) (0.032) (0.033) (0.058) [0.185] [0.715] [0.506] [0.907] [0.089] Yt−1 0.391*** 0.357*** 0.473*** 0.288*** 0.320*** (0.029) (0.029) (0.027) (0.027) (0.048) R-squared 0.198 0.180 0.262 0.099 0.152 P-value of joint significance 0.000 0.001 0.010 0.105 0.305 Panel B: Tradition Treated 0.784*** 0.342*** 0.336** 0.147 0.143 (0.215) (0.104) (0.128) (0.090) (0.147) [0.001] [0.004] [0.011] [0.203] [0.336] Treated*Tradition -0.016 -0.152** -0.007 -0.017 -0.131 (0.141) (0.065) (0.087) (0.087) (0.115) [0.993] [0.063] [0.992] [0.845] [0.445] Tradition -0.031 0.100* -0.036 0.057 0.174* (0.109) (0.052) (0.064) (0.075) (0.104) Yt−1 * Treated -0.065* 0.011 -0.033 0.001 -0.111* (0.035) (0.037) (0.032) (0.034) (0.057) [0.204] [0.767] [0.524] [0.979] [0.113] Yt−1 0.390*** 0.359*** 0.472*** 0.290*** 0.314*** (0.029) (0.030) (0.027) (0.027) (0.047) R-squared 0.198 0.180 0.261 0.099 0.154 P-value of joint significance 0.000 0.002 0.010 0.106 0.355 Panel C: Self Direction Treated 0.762*** 0.333*** 0.332** 0.146 0.153 (0.213) (0.104) (0.128) (0.091) (0.141) [0.001] [0.005] [0.010] [0.211] [0.281] Treated*Self-direction 0.086 0.035 0.142** 0.060 0.140* (0.112) (0.065) (0.058) (0.074) (0.079) [0.685] [0.685] [0.055] [0.421] [0.148] Self-direction -0.358*** -0.098* -0.066* -0.040 -0.161** (0.088) (0.050) (0.039) (0.063) (0.061) Yt−1 * Treated -0.060* 0.009 -0.032 0.004 -0.120** (0.035) (0.038) (0.032) (0.035) (0.056) [0.257] [0.805] [0.560] [0.922] [0.077] Yt−1 0.380*** 0.357*** 0.471*** 0.290*** 0.323*** (0.028) (0.031) (0.027) (0.028) (0.046) R-squared 0.202 0.180 0.262 0.099 0.153 P-value of joint significance 0.000 0.002 0.011 0.112 0.280 Controls (a) Yes Yes Yes Yes Yes Observations 4,986 4,986 4,986 4,986 3,070 Mean Dep. Var. (Control) 0.040 0.044 -0.139 0.00186 -0.063 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. FWER adjusted p-values in square brackets, based on 10,000 replications. (a) Controls include: female, age, currently attending school, years of education, english spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. 45 ONLINE APPENDIX The Entertaining Way to Behavioral Change: Fighting HIV with MTV Abhijit Banerjee Eliana La Ferrara Victor Orozco MIT Bocconi University Oxford University and IGIER and LEAP The World Bank This version: April 2019 A. Sample selection Locations The study sites were 80 urban and peri-urban locations chosen in 7 towns across three states of South-West Nigeria. The distribution of locations across states and towns is as follows. Oyo state: Ibadan (50 locations), Ogbomosho (6 locations), Oyo (4 locations). Osun state: Ile-Ife (3 locations), Ilesha (4 locations), Osogbo (7 locations). Ondo state: Akure (6 locations). The selection of these towns balanced competing requirements such as: (i) excluding states earmarked by MTV as priority states for marketing Shuga (to avoid contamination of the control group); (ii) excluding areas where the integrity of the evaluation could have been compromised by security risks; (iii) choosing contiguous states to facilitate the logistical implementation. Locations were defined by drawing a 2-mile radius around each screening center where the intervention was implemented, and randomly selecting households within this radius. We ensured that there were buffer zones between communities to minimize the risk of contamination across study groups. These locations constitute our unit of randomization. Appendix Figure A1 shows the geographic distribution of treatment and control locations. Main study participants To identify study participants, we adopted a three-step recruitment strategy. First, enu- merators visited a random selection of 200 − 225 households in each location and collected basic demographic information about all young people aged 18 to 25 residing in those households. i The condition for a household to be in the study was that at least one of the members should be in the target age range of our intervention, i.e., 18-25 years old. Second, the research team randomly selected one person age 18-25 in each household to be invited to the screening of a Nigerian movie, different from MTV Shuga. We denote this initial screening as Screening 0. The selection was stratified by gender, half males and half females. Out of 17, 224 people invited to Screening 0, 6, 613 attended, with a turnout rate of 38.4 percent. Appendix Table A1 compares those who chose to attend Screening 0 and those who did not, on a number of characteristics that we collected when we first visited the households, as well as the normalized pair-wise differences and the p-values for the test that the difference in means is zero.1 The two samples are well balanced, with few variables showing significant differences, but of extremely small magnitude: in all cases the normalized difference is below the threshold of 0.25 suggested by Imbens and Rubin (2015). We thus conclude that those who attended Screening 0 were a representative subset of the households invited. The third step was the selection of our baseline sample. In each location enumerators paid home visits to 64 individuals among those who attended Screening 0, randomly selected.2 All were invited to participate in the study and administered the baseline survey if they agreed. Appendix Table A2 shows balance in observable characteristics between the people who attended Screening 0 and were selected into the baseline sample and those that were not. We also compared our sample to the 2008 Nigeria DHS, restricted to the South-Western region, and found that the two were quite similar in terms of religion (a third Muslim and two thirds Christian), years of education (around 11 years) and television-ownership rates (around 90 percent). At the end of the survey, the main study participants received invitations to attend Screen- ings 1 and 2, organized in the two weekends following the interview. Those in the ‘Friends treatment’ also received invitations to bring up to two friends of their choice. Note that invi- tees were not told what they would watch, neither before Screening 0 nor before Screening 1. It is only when attending Screening 1 that they learnt they were watching Shuga or Gidi Up. Network members We complemented our data collection on the main study participants with a sample of con- 1 The normalized difference is a scale-free measure of the difference in distributions, recommended by Imbens and Wooldridge (2009): XA − XB ∆= 2 2 SA + SB 2 2 where X A and X B are the means of covariate X in groups A and B, respectively, and SA and SB are the corresponding sample variances of X. 2 In some locations we have 65 respondents instead of 64, because our teams consisted of eight enumerators working in parallel. Enumerators were given a target number of respondents to interview per day and they did not know how many their colleagues had interviewed until the end of the day. ii tacts from their social networks that we use for detecting potential spillovers. In all locations, before taking the baseline survey, study participants were asked to list two friends to whom they regularly talked and who lived in the community (‘network members’). In each location, we administered the baseline and the follow-up surveys to a random sample of 15 network members among those indicated by main respondents who were not in the ‘Friends’ treatment. No more than one friend per main respondent was interviewed. By construction, contacts of those main respondents who had been given extra tickets for friends (and who could have therefore attended the screenings) are not part of the spillover sample. B. Timeline of activities In each location, all baseline survey and screening activities were concluded in four weeks. Implementation was rolled out so that activities in a given location were completed before moving to the next, in order to minimize attrition due to subjects forgetting about the screen- ings, travelling or relocating. The overall timeline of the project across all locations was as follows: listing, Screening 0 and the baseline survey occurred between mid-September and the first week of December 2014; Screenings 1 and 2 between October and end of December; the endline survey and health camps between end of May and August 2015. In each location, baseline and endline were 8 weeks apart. The typical sequence of activities in a given location was as follows: MONTH 1 • Week 1: listing • Weekend 1: Screening 0 • Weeks 2-3: baseline survey • Weekend 3: Screening 1 • Weekend 4: Screening 2 MONTH 8 • Week 1: Follow-up survey • Weekend 1: Health camp 1 - Chlamydia testing and distribution of referral letters for HIV testing • Weekend 3: Health camp 2 - test results for Chlamydia test and post-test counselling iii C. Construction of indexes C.1 Outcome indexes We construct our outcome indexes related to HIV and risky sexual behavior using two methods. The first method follows Kling et al. (2007): we take equally weighted averages of the z -scores of the variables that enter each index, where the sign of each variable is oriented so that answers consistent with Shuga’s message translate into higher values of the index.3 Appendix Tables A6 and A7 report list of variables contained in each index, with a sign (+) or (−) to denote whether the variable enters the index with a positive or negative sign. Variables are oriented so that the impact of treatment on each component of the index should be positive. The second method uses principal components analysis. Starting from the same lists of variables as above, we extract the first principal component for each family of outcomes. The individual variables and their loading factors are shown in Appendix Tables A6 and A7. C.2 Indexes of conformism In our survey we included a series of questions aimed at measuring how strongly individuals identified with three of the values categorized by Schwartz (2012): conformity, tradition and self-direction. For each category, respondents were read four questions describing people with certain characteristics and were asked how similar each person was to them, with answers on a 5-point scale ranging from “not like me at all” to “very much like me”. We aggregate the four questions in an index using principal component analysis (see Appendix Table A23 for the list of variables and loading factors) and we construct the following three indexes. Conformity : captures how inclined an individual is to restrain his/her own choices if these were to upset others or violate social norms. People with a high value of this index believe that people should do what they are told, be obedient and polite, and they generally have a taste for smooth social interaction, even at the cost of self-restraint. Tradition : captures individuals’ acceptance and commitment to the values that their culture or religion promote. Respondents who identify with this profile believe that people should be humble and be satisfied with what they have. ‘Tradition’ and ‘Conformity’ are similar in the sense that they capture individuals’ willingness to subordinate to what is expected from them, but they differ in the group to which one subordinates him/herself: in the case of conformity it is mainly people (e.g., parents or peers), while in the case of tradition it is religious and cultural customs. 3 To deal with missing values we follow Kling et al. (2007): if a respondent has a non-missing value for at least one of the variables in an index, we impute any missing values for the other variables using the random assignment group mean. This implies that differences between treatment and control means of an index coincide with the average of treatment and control means of the variables in that index (when divided by their standard deviations). iv Self-direction : captures how inclined an individual is to think and act independently. Re- spondents with a high value of this index like to be curious, creative, free to make their own choices and to rely on themselves. D. Health camps We set up 80 ‘health camps’ (one per location) in schools. Respondents were invited to the health camp when they took the follow-up survey. Each health camp was set up for the weekend immediately following the follow-up survey: the data collected at health camps is therefore about 8 months after the baseline survey and about 7.5 months after the last screening of our intervention. Participants were informed about testing by counsellors and were offered the opportunity to test for Chlamydia through urine sample collection. During the same session they also participated in a game that consisted in choosing between N50 (approximately equivalent to 0.25 USD at the time) and a certain number of condom packs. The number was randomly determined and could vary from 1 to 3, with each pack worth approximately N50 on the market. At the end of the session participants received contact details of HIV counselling and testing centres in their town and were given a voucher that would entitle them to free HIV testing at one of these centers. After the specimens were analyzed and the results for Chlamydia were available, participants were invited for a second visit to the health camp, where they were informed of the outcome and –if they tested positive for Chlamydia– they were prescribed treatment. We did not test anyone for HIV hence do not know who is HIV positive or negative, but only if they took the test. From our sample, 3828 individuals attended the health camp, and all got tested for Chlamy- dia and participated in the condom game; 74 of them tested positive for Chlamydia, and 213 redeemed the voucher to get tested for HIV. In our analysis we will use the following outcomes ‘objectively’ collected at health camps: (i) whether participants attended the health camp and took the Chlamydia test; (ii) whether they tested positive for Chlamydia; (iii) whether they redeemed the voucher to get tested for HIV; and (iv) whether they chose condoms over money when given the choice. Attendance to health camp was relatively high: 77 percent of the study participants at- tended the health camp, and on average this share was the same in treatment and control locations. Appendix Table A4 shows how baseline characteristics and baseline values of our outcomes correlate with the decision to participate in the health camp. People currently at- tending school and living outside the family were less likely to attend, possibly due to conflict with school schedules. While treatment status is uncorrelated with the decision to attend (column 1), we also test whether observable characteristics may have played a different role within the treatment and the control samples. We find that higher values of the HIV testing v index at baseline predict attendance among treated individuals, while lower values of the Risky Sexual Behavior index predict it among control ones. In Appendix Tables A6 and A9 we show that these differences do not explain our estimated impacts on outcomes collected at health camps, as our estimates are robust to including the baseline values of these indexes (and their interaction with treatment) among the regressors. vi Appendix Figures and Tables Figure A1: Location of treatment and control centres (a) Ibadan (b) Ife, Ilesha, Akure (c) Oyo, Ogbomosho, Osogbo vii Table A1: Summary statistics on invitees to Screening 0 (S0), by participation status Mean Mean Normalized Variable Name Did not partici- Participated Diff=0 (p-value) Difference(a) pate in S0 in S0 Female 0.47 0.47 0.65 0.01 Age 20.74 20.60 0.00 -0.04 Highest Educ. Level Attained is Primary 0.01 0.01 0.06 -0.03 Highest Educ. Level Attained is Junior Sec. 0.01 0.01 0.12 -0.02 Highest Educ. Level Attained is Senior Sec. 0.87 0.88 0.12 0.02 Highest Educ. Level Attained > Senior Sec. 0.11 0.11 0.55 -0.01 Speaks mostly English at home 0.07 0.06 0.28 -0.01 First Preferred Language is Yoruba 0.81 0.80 0.11 -0.02 First Preferred Language is English 0.17 0.18 0.03 0.02 Second Most Preferred Language is Yoruba 0.15 0.17 0.00 0.03 Second Most Preferred Language is English 0.78 0.78 0.69 0.00 Number of household members aged 18-25 1.03 1.04 0.00 0.04 Muslim religion 0.39 0.37 0.00 -0.03 No. Obs 10.102 6348 Notes: Sample includes individuals invited to attend Screening 0, i.e. a movie unrelated to Shuga. (a) Normalized difference is the difference in the sample means of treatment and control groups divided by the square root of the sum of sample variances. viii Table A2: Summary statistics on Screening 0 attendees, by inclusion in baseline sample Mean Not selected for baseline (i) Mean for Baseline Sample (ii) Normalized Difference(a) (ii)-(i) P-value (i) = (ii) Variable Name Female 0.48 0.47 0.00 0.84 Age 20.46 20.63 0.05 0.02 Highest Educ. Level Attained is Primary 0.01 0.01 -0.04 0.11 Highest Educ. Level Attained is Junior Sec. 0.01 0.00 -0.05 0.03 Highest Educ. Level Attained is Senior Sec. 0.87 0.88 0.02 0.51 ix Highest Educ. Level Attained > Senior Sec. 0.10 0.11 0.01 0.76 Speaks mostly English at home 0.06 0.06 -0.01 0.81 First Preferred Language is Yoruba 0.84 0.79 -0.09 0.00 First Preferred Language is English 0.15 0.19 0.07 0.00 Second Most Preferred Language is Yoruba 0.15 0.18 0.06 0.02 Second Most Preferred Language is English 0.81 0.77 -0.07 0.00 Number of household members aged 18-25 1.04 1.04 0.02 0.47 Muslim religion 0.36 0.37 0.02 0.29 No. Obs 1.182 5.166 Notes: Sample includes individuals that attended Screening 0, i.e. a movie unrelated to Shuga. (a) Normalized difference is the difference in the sample means of treatment and control groups divided by the square root of the sum of sample variances. Table A3: Variable definitions Variable name Definition Panel A: Outcomes Indexes HIV knowledge Index of knowledge about transmission, testing and treatment of HIV HIV respondent’s attitudes Index of attitudes toward HIV positive people HIV testing Index of testing behavior Attitudes towards risky sexual behavior Index of attitudes towards risky sexual behavior Risky sexual behavior (for sexually active) Index of risky sexual behavior (defined for sexually active individuals) Individual variables HIV transmitted during pregnancy Dummy=1 if knows that HIV can be transmitted during pregnancy Has heard of ARVs Dummy=1 if, when specifically asked, respondent says he/she has heard of ARV drugs Second test necessary Dummy =1 if knows that a second test is necessary Window period 3 months Dummy=1 if knows that a 3 month period is necessary before retest Can get HIV through intercourse Dummy=1 if knows that HIV can be contracted via sexual intercourse Would buy from an HIV+ shopkeeper Dummy=1 if would buy food from an HIV positive shopkeeper An HIV+ boy should play footbal Dummy=1 if agrees that an HIV positive boy should be allowed to play football People HIV+ should not be blamed Dummy=1 if agrees that HIV positive people should not be blamed HIV is not punishment for sleeping around Dummy=1 if says that HIV is not a punishment for sleeping around Would reveal HIV status Dummy=1 if would reveal own HIV status to partner Tested last 6 months (self-reported) Dummy=1 if has been tested less than 6 months ago Tested at health camp (observed) Dummy=1 if has attended the health camp and has been tested for STDs Chose condoms over N50 Dummy=1 if chose condoms over money in experimental game at health camp Tested positive for Chlamydia Dummy=1 if tested positive for Chlamydia Men should have one partner only Dummy=1 if agrees that men should only date one partner at a time Women should have one partner only Dummy=1 if agrees that women should only date one partner at a time Not ok date sugardaddy to finance educ Dummy=1 if does not consider appropriate dating a sugardaddy, even if he offers to pay for education Not ok date sugardaddy for money Dummy=1 if does not consider appropriate dating a sugardaddy in exchange for money Not ok date sugardaddy to go out Dummy=1 if does not consider appropriate dating a sugardaddy even if he brings the girl out If a woman brings a condom does not mean she’s Dummy=1 if disagrees with the statement that if a woman brings a condom, her man will think she’s not serious not serious Has not had multiple concurrent sexual partners Dummy=1 if has not had multiple concurrent sexual partners No. of current sexual partners if sexually active Number of current sexual partners Used condom the last time he/she had sex Dummy=1 if used a condom during last sexual intercourse Panel B: Control variables Indexes Transportation Index of immersion in the narrative while watching the show Identification Index of identification with the characters Conformity Index of propensity to subordinate to norms (e.g., instilled parents or peers) Tradition Index of acceptance and commitment to the values that religion or culture promote Self-direction Index of inclination to think and act independently Socioeconomic controls Female Dummy=1 if female Age Age of respondent Currently attending school Dummy=1 if currently attending school Years of education Years of education Speaks English Dummy=1 if speaks english as primary or secondary language at home Single Dummy=1 if does not have a partner Does not live with the family Dummy=1 if does not live with family Household size Number of components of respondent’s family Wealth index Principal component index from dwelling characteristics and durable goods ownership Owns his/her house Dummy=1 if dwelling where respondent lives is owned Father’s education > secondary Dummy=1 if father obtained a level of education higher than secondary Mother’s education > secondary Dummy=1 if mother obtained a level of education higher than secondary Muslim Dummy=1 if Muslim Yoruba Dummy=1 if native language is Yoruba x Table A4: Summary statistics at follow-up No. Mean Std. Dev. Min Max Obs. (1) (2) (3) (4) (5) Panel A: Outcome variables Indexes HIV knowledge 4986 0.585 6.127 -22.006 25.984 HIV attitudes 4986 0.297 3.476 -10.114 7.089 HIV testing 4986 0.093 4.441 -5.052 11.732 Attitudes towards risky sexual behavior 4986 0.085 3.417 -15.113 4.495 Risky sexual behavior (for sexually active) 3618 0.067 3.525 -31.687 5.1 Individual variables HIV transmitted during pregnancy 4986 0.669 0.471 0 1 Mentions drugs to live longer with HIV 4986 0.739 0.438 0 1 Recognizes ARV when mentioned by enumerator 4986 0.332 0.471 0 1 Has heard of ARVs 4986 0.292 0.447 0 1 Second test necessary 4986 0.372 0.484 0 1 Window period 3 months 4986 0.16 0.367 0 1 Can get HIV through intercourse 4986 0.976 0.153 0 1 Would buy from an HIV+ shopkeeper 4986 0.521 0.5 0 1 An HIV+ boy should play footbal 4986 0.696 0.46 0 1 People HIV+ should not be blamed 4986 0.693 0.461 0 1 HIV is not punishment for sleeping around 4986 0.52 0.5 0 1 Would reveal HIV status 4986 0.727 0.446 0 1 Tested last 6 months (self-reported) 4982 0.103 0.304 0 1 Tested at health camp (observed) 3828 0.056 0.229 0 1 Men should have one partner only 4976 0.894 0.308 0 1 Women should have one partner only 4986 0.926 0.262 0 1 Not ok date sugardaddy to finance educ 4986 0.744 0.437 0 1 Not ok date sugardaddy for money 4986 0.696 0.46 0 1 Not ok date sugardaddy to go out 4986 0.905 0.293 0 1 If a woman brings a condom it does not mean 4986 0.625 0.484 0 1 she’s not serious Has not had multiple concurrent sexual partners 3339 0.803 0.398 0 1 No. of current sexual partners if sexually active 4978 0.673 0.867 0 21 Used condom the last time he/she had sex 3084 0.489 0.5 0 1 Chose condoms over N50 3827 0.393 0.488 0 1 Tested positive for Chlamydia 4986 0.302 0.459 0 1 Panel B: Control variables Indexes Transportation 3753 0 1.945 -9.074 3.422 Identification 3753 0 2.051 -8.651 3.621 Conformity 5166 0.036 1.31 -7.009 1.515 Tradition 5166 0.026 1.276 -7.241 1.911 Self-direction 5166 0.004 1.449 -1.846 6.152 Socioeconomic controls (a) Female 5166 0.473 0.499 0 1 Age 5166 20.615 2.362 18 26 Currently attending school 5166 0.348 0.476 0 1 Years of education 5166 11.597 1.085 0 12 Speaks English 5166 0.961 0.194 0 1 Single 5166 0.227 0.419 0 1 Does not live with the family 5166 0.241 0.428 0 1 Household size 5166 4.328 2.362 1 19 Wealth index 5150 1.75 0.587 0 3 Owns his/her house 5165 0.385 0.487 0 1 Father’s education > secondary 3928 0.334 0.472 0 1 Mother’s education > secondary 4393 0.226 0.418 0 1 Muslim 5166 0.373 0.484 0 1 Yoruba 5166 0.92 0.271 0 1 Notes. (a) Summary statistics for these variables are calculated at baseline, because this is how they are included in all regressions. xi Table A5: Attrition between baseline and follow-up Dep. Var. =1 if interviewed at follow-up (1) (2) (3) (4) (5) Treated -0.004 -0.004 -0.004 -0.004 -0.000 (0.005) (0.005) (0.005) (0.005) (0.007) HIV knowledge -0.000 (0.000) HIV attitudes -0.001 (0.001) HIV testing 0.001 (0.001) Attitudes towards risky sexual behavior -0.000 (0.001) Risky sexual behavior (for sexually active) -0.001 (0.001) Female -0.022*** -0.022*** -0.023*** -0.022*** -0.029*** (0.005) (0.005) (0.005) (0.005) (0.007) Age 0.001 0.001 0.001 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.002) Currently attending school -0.013** -0.013** -0.013** -0.013** -0.016* (0.006) (0.006) (0.006) (0.006) (0.008) Years of education 0.001 0.001 0.000 0.000 0.002 (0.003) (0.003) (0.003) (0.003) (0.003) English Spoken 0.002 0.002 0.002 0.002 0.009 (0.009) (0.009) (0.009) (0.009) (0.010) Single -0.003 -0.003 -0.003 -0.003 0.006 (0.005) (0.005) (0.005) (0.005) (0.009) Does not live with the family -0.014 -0.014 -0.014 -0.014 -0.021** (0.009) (0.009) (0.009) (0.009) (0.011) Household size 0.000 0.000 0.000 0.000 0.001 (0.001) (0.001) (0.001) (0.001) (0.002) Wealth index 0.008 0.007 0.007 0.007 0.010 (0.005) (0.005) (0.005) (0.005) (0.006) Home owner -0.004 -0.004 -0.004 -0.004 -0.015* (0.007) (0.007) (0.007) (0.007) (0.008) Father’s education > secondary 0.010** 0.009** 0.009** 0.009** 0.009 (0.005) (0.005) (0.005) (0.005) (0.007) Mother’s education > secondary 0.005 0.005 0.004 0.004 0.004 (0.004) (0.004) (0.004) (0.004) (0.007) Muslim 0.012* 0.012** 0.012** 0.012** 0.016** (0.006) (0.006) (0.006) (0.006) (0.008) Yoruba 0.018 0.018 0.018 0.018 0.006 (0.013) (0.013) (0.013) (0.013) (0.014) Constant 0.948*** 0.949*** 0.963*** 0.956*** 0.914*** (0.041) (0.042) (0.043) (0.043) (0.049) 5,166 5,166 5,166 5,166 3,246 R-squared 0.094 0.094 0.094 0.093 0.080 Controls Yes Yes Yes Yes Yes City FE Yes Yes Yes Yes Yes Mean Dep. Var. (Control) 0.966 0.966 0.966 0.966 0.966 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. The dependent variable is an indicator for whether the individual has been interviewed at follow-up. xii Table A6: HIV indexes Sign with which Loading variable enters factor index HIV knowledge # of correct sources of contagion listed + 0.338 Can get HIV through intercourse + 0.164 Has mentioned drugs to live longer with HIV (not ARVs) + 0.155 Has mentioned ARV + 0.206 Recognizes ARV when mentioned by enumerator + 0.288 # of correct ways to avoid contracting HIV listed + 0.312 Avoid HIV knowing your/your partner’s status + 0.048 Window period 3 months + 0.343 Knows that an early negative test is no guarantee of no HIV + 0.364 Second test necessary + 0.396 HIV transmitted during pregnancy + 0.235 Says exist drugs to reduce transmission risk to baby + 0.250 Says HIV can be transmitted from mother to baby during delivery + 0.237 Says HIV can be transmitted from mother to baby by breastfeeding + 0.167 HIV attitudes Would not prefer to keep HIV of family member a secret + -0.066 Would reveal HIV status + 0.228 Would buy from an HIV+ shopkeeper + 0.389 An HIV+ boy should play footbal + 0.406 If a young person get tested for HIV, he has been sleeping around + 0.441 People with HIV should be blamed + 0.490 HIV/AIDS is a punishment for sleeping around. + 0.439 HIV testing Tested for HIV at least once + 0.501 Tested last 12 months (self reported) + 0.467 Tested last 6 months (self-reported) + 0.371 Asked him(her)self for the test + 0.345 Tested and picked up results + 0.493 Knows a place to get HIV test + 0.174 Notes: Shaded cells refer to ordinal variables. The variables are coded so that higher values correspond to higher levels of disagreement. xiii Table A7: Indexes of risky sexual behavior Attitudes towards risky sex Sign with which Loading variable enters factor index “Men who are not married should not only have sex with one partner” + 0.227 “Women who are not married should not only have sex with one partner” + 0.231 “OK for a young girl to date an older married man if he offers to pay for her educaton” + 0.555 “OK for a young girl to date an older married man if her family needs financial support” + 0.557 “OK for a young girl to date an older married man if he offers to take her out” + 0.513 xiv If a woman brings a condom it does not mean she’s no serious + 0.113 Risky sexual behavior Not multiple concurrent sexual partners + 0.506 In the last 6 months had only one partner in the same month + 0.464 Number of current sexual partners - -0.475 Used condom last time he/she had sex + -0.070 Has a main partner + 0.198 Has not an “other”partner + 0.510 Notes: Shaded cells refer to ordinal variable. The variables are coded so that higher values correspond to higher levels of disagreement. Table A8: Correlates of Health Camp attendance Full sample Treated Control Dep. Var=1 if attended health camp (1) (2) (3) (4) (5) (6) (7) (8) (9) Treated 0.000 -0.001 0.005 (0.015) (0.015) (0.019) HIV knowledge -0.001 -0.000 -0.003* (0.001) (0.001) (0.002) HIV attitudes 0.003 0.002 0.003 (0.002) (0.002) (0.004) HIV testing 0.001 0.002 -0.002 (0.001) (0.002) (0.003) Attitudes towards risky sexual behavior 0.003 0.003 0.004 (0.002) (0.002) (0.004) Risky sexual behavior (sexually active) 0.001 0.003 -0.003 (0.002) (0.003) (0.002) Female 0.000 0.001 -0.029 0.016 0.015 -0.019 -0.028 -0.026 -0.040 (0.015) (0.015) (0.018) (0.019) (0.019) (0.022) (0.028) (0.028) (0.033) Age 0.003 0.002 0.005 0.003 0.002 0.005 0.003 0.003 0.004 (0.003) (0.003) (0.003) (0.004) (0.003) (0.004) (0.004) (0.005) (0.006) Currently attending school -0.057*** -0.058*** -0.063*** -0.058*** -0.060*** -0.057*** -0.058** -0.057** -0.081** (0.014) (0.014) (0.016) (0.017) (0.017) (0.018) (0.023) (0.024) (0.036) Years of education -0.013** -0.013** -0.005 -0.009 -0.010 -0.002 -0.020* -0.019* -0.012 (0.006) (0.006) (0.007) (0.008) (0.008) (0.009) (0.010) (0.010) (0.011) Speaks English -0.012 -0.009 -0.033 -0.023 -0.024 -0.046 -0.004 0.003 -0.016 (0.032) (0.032) (0.031) (0.044) (0.044) (0.043) (0.040) (0.040) (0.038) Single -0.024 -0.024 -0.059*** -0.026 -0.027 -0.058** -0.020 -0.020 -0.061 (0.015) (0.015) (0.021) (0.020) (0.020) (0.026) (0.024) (0.024) (0.038) Does not live with the family -0.153*** -0.151*** -0.144*** -0.155*** -0.154*** -0.157*** -0.150*** -0.150*** -0.115*** (0.021) (0.022) (0.022) (0.027) (0.027) (0.029) (0.036) (0.036) (0.033) Household size 0.013*** 0.013*** 0.014*** 0.011*** 0.011*** 0.011** 0.015*** 0.014*** 0.018** (0.003) (0.003) (0.004) (0.004) (0.004) (0.005) (0.005) (0.005) (0.007) Wealth index -0.006 -0.006 -0.007 0.004 0.002 0.006 -0.030 -0.026 -0.040* (0.014) (0.014) (0.015) (0.018) (0.018) (0.019) (0.021) (0.021) (0.023) Owns his/her house 0.011 0.012 0.013 0.012 0.013 0.010 0.011 0.013 0.018 (0.015) (0.015) (0.018) (0.019) (0.019) (0.022) (0.024) (0.024) (0.030) Father obtained education higher then sec. -0.032* -0.032* -0.024 -0.043** -0.044** -0.035 -0.012 -0.010 0.001 (0.017) (0.017) (0.019) (0.021) (0.021) (0.024) (0.030) (0.031) (0.031) Mother obtained education higher then sec. 0.024 0.024 0.028 0.021 0.019 0.014 0.036 0.038 0.058 (0.018) (0.018) (0.022) (0.020) (0.020) (0.023) (0.035) (0.035) (0.049) Muslim 0.024* 0.023* 0.027* 0.021 0.021 0.025 0.022 0.019 0.022 (0.013) (0.013) (0.015) (0.015) (0.015) (0.016) (0.031) (0.030) (0.036) Native language Yoruba 0.069*** 0.068*** 0.075*** 0.042 0.042 0.039 0.125*** 0.124*** 0.139** (0.026) (0.026) (0.027) (0.034) (0.034) (0.034) (0.040) (0.040) (0.051) Sample All All Sexually active All All Sexually active All All Sexually active Observations 4,986 4,986 3,618 3,402 3,402 2,487 1,584 1,584 1,131 R-squared 0.067 0.068 0.071 0.064 0.065 0.068 0.085 0.088 0.100 Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean Dep. Var 0.773 0.767 0.765 0.763 0.773 0.767 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level . 80 clusters in the Full Sample, 54 clusters in the Treated Sample, 26 clusters in the Control Sample. ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. All regressors are measured at baseline. xv Table A9: Exogeneity of treatment assignment Dep. Var. =1 if Treated. Panel A: Controls Coeff. Std. Err. Female 0.003 (0.013) Age -0.002 (0.004) Currently attending school 0.019 (0.022) Years of education 0.000 (0.008) English Spoken 0.034 (0.038) Single -0.005 (0.022) Does not live with the family -0.038 (0.032) Household size -0.005 (0.004) Wealth index -0.013 (0.027) Home owner -0.089** (0.039) Father obtained education higher then sec. -0.048*** (0.017) Mother obtained education higher then sec. -0.019 (0.021) Muslim 0.008 (0.032) Yoruba Native 0.012 (0.052) Constant 0.784*** (0.172) Observations 5,166 R-squared 0.021 P-val F-test of joint significance 0.147 Panel B: Outcomes (1) (2) (3) (4) (5) HIV knowledge 0.002 (0.001) HIV attitudes -0.000 (0.003) HIV testing 0.001 (0.002) Attitudes towards risky sexual behavior -0.000 (0.002) Risky sexual behavior (for sexually active) 0.002 (0.002) Constant 0.820*** 0.782*** 0.796*** 0.784*** 0.745*** (0.168) (0.171) (0.168) (0.171) (0.176) Observations 5,166 5,166 5,166 5,166 3,246 R-squared 0.022 0.021 0.021 0.021 0.025 Controls (a) Yes Yes Yes Yes Yes P-val F-test of joint significance 0.156 0.159 0.188 0.142 0.304 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. The dependent variable takes value one if the individual has been assigned to treatment. (a) Controls in each regression of panel B include: female, age, currently attending school, years of education, english spoken, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. xvi Table A10: Impact on Indexes calculated with principal component method Dep. Var. (Yt ): HIV knowledge HIV attitudes HIV testing (1) (2) (3) (4) (5) (6) (7) (8) (9) Full Sample Females Males Full Sample Females Males Full Sample Females Males Treated 0.184*** 0.302*** 0.089 0.177*** 0.156*** 0.207*** 0.146** 0.161* 0.131 (0.055) (0.079) (0.083) (0.042) (0.059) (0.055) (0.057) (0.085) (0.083) [0.002] [0.001] [0.288] [0.000] [0.021] [0.001] [0.011] [0.069] [0.225] Treated*Yt−1 -0.066* -0.016 -0.114** -0.011 -0.006 -0.014 -0.028 -0.039 -0.035 (0.039) (0.049) (0.045) (0.040) (0.051) (0.053) (0.033) (0.042) (0.050) [0.257] [0.932] [0.043] [0.787] [0.932] [0.798] [0.630] [0.725] [0.731] Yt−1 0.433*** 0.420*** 0.442*** 0.436*** 0.487*** 0.391*** 0.543*** 0.591*** 0.492*** (0.031) (0.041) (0.033) (0.032) (0.040) (0.046) (0.026) (0.034) (0.038) Observations 4,986 2,323 2,663 4,986 2,323 2,663 4,971 2,320 2,651 R-squared 0.208 0.240 0.204 0.188 0.247 0.157 0.279 0.378 0.195 Controls(a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean dep var (Control) 0.189 0.177 0.199 0.119 0.188 0.0594 0.0895 0.236 -0.0391 P-value test joint sig 0.000 0.000 0.138 0.000 0.012 0.001 0.008 0.064 0.059 Dep. Var. (Yt ): Attitudes towards risky sex- Risky sexual behavior (for sexu- ual behavior ally active) (1) (2) (3) (4) (5) (6) Full Sample Females Males Full Sample Females Males Treated 0.042 -0.025 0.112** 0.060 0.151 0.029 (0.036) (0.056) (0.056) (0.075) (0.166) (0.140) [0.432] [0.661] [0.096] [0.432] [0.622] [0.836] Treated*Yt−1 0.017 0.007 0.022 -0.166** -0.263* -0.150** (0.031) (0.046) (0.042) (0.064) (0.154) (0.075) [0.571] [0.887] [0.607] [0.026] [0.236] [0.098] Yt−1 0.278*** 0.295*** 0.260*** 0.396*** 0.386*** 0.405*** (0.024) (0.037) (0.036) (0.054) (0.145) (0.064) Observations 4,973 2,320 2,653 1,682 760 922 R-squared 0.113 0.147 0.101 0.265 0.146 0.125 Controls (a) Yes Yes Yes Yes Yes Yes Mean Dep. Var. (Control) 0.0858 0.144 0.0346 -0.0488 0.837 -0.753 P-value test joint sig 0.254 0.647 0.0418 0.370 0.509 0.339 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. P-values in square brackets corrected for multiple hypothesis testing using FWER. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. xvii Table A11: Impact on outcome Indexes, ANCOVA Dep. Var. (Yt ): HIV knowledge HIV attitudes HIV testing Attitudes towards Risky sexual behavior risky sexual behav- (for sexually active) ior (1) (2) (3) (4) (5) Treated 0.781*** 0.339*** 0.336** 0.148 0.150 (0.214) (0.103) (0.128) (0.090) (0.146) Yt−1 0.346*** 0.366*** 0.449*** 0.292*** 0.241*** (0.020) (0.019) (0.016) (0.017) (0.028) xviii Observations 4,986 4,986 4,986 4,986 3,070 R-squared 0.197 0.179 0.261 0.099 0.149 Controls(a) Yes Yes Yes Yes Yes Mean Dep. Var. (Control) 0.0400 0.0439 -0.139 0.00186 -0.0631 Std Dev Dep. Var (Control) 6.106 3.474 4.145 3.452 3.625 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. FWER adjusted p-values in square brackets, based on 10,000 replications. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. Table A12: Impact on outcome indexes, by gender Dep. Var. (Yt ): HIV knowledge HIV attitudes HIV testing Attitudes towards Risky sexual behavior (for risky sexual behavior sexually active) Females Males Females Males Females Males Females Males Females Males (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Cross-section Treated 1.191*** 0.674** 0.201 0.480*** 0.440* 0.283 -0.133 0.426*** -0.170 0.335 (0.361) (0.316) (0.183) (0.172) (0.248) (0.178) (0.167) (0.161) (0.121) (0.220) [0.005] [0.076] [0.265] [0.021] [0.152] [0.120] [0.422] [0.022] [0.296] [0.136] Observations 2323 2663 2323 2663 2323 2663 2323 2663 1526 2092 R-squared 0.097 0.098 0.078 0.044 0.144 0.057 0.046 0.022 0.037 0.018 Panel B: Conditional specification Treated 1.174*** 0.462 0.275* 0.408*** 0.421** 0.260 -0.074 0.363** -0.118 0.293 xix (0.313) (0.285) (0.140) (0.137) (0.199) (0.179) (0.138) (0.152) (0.199) (0.244) [0.001] [0.204] [0.074] [0.012] [0.074] [0.204] [0.811] [0.037] [0.811] [0.230] Treated*Yt−1 -0.023 -0.106*** 0.020 -0.002 -0.053 -0.033 -0.024 0.012 -0.060 -0.104 (0.049) (0.039) (0.039) (0.052) (0.040) (0.046) (0.048) (0.048) (0.096) (0.067) [0.854] [0.025] [0.854] [0.964] [0.469] [0.723] [0.792] [0.798] [0.792] [0.233] Yt−1 0.379*** 0.399*** 0.389*** 0.334*** 0.523*** 0.425*** 0.311*** 0.279*** 0.228*** 0.326*** (0.042) (0.028) (0.029) (0.046) (0.033) (0.036) (0.039) (0.041) (0.085) (0.057) R-squared 0.225 0.194 0.238 0.146 0.361 0.181 0.126 0.094 0.080 0.079 Controls(a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 2323 2663 2323 2663 2323 2663 2323 2663 1287 1783 Mean Dep. Var. (Control) 0.001 0.0741 0.00715 0.0760 0.182 -0.420 0.208 -0.178 1.353 -1.078 Std dev Dep. Var. (Control) 5.944 6.248 3.222 3.683 4.186 4.091 3.304 3.568 2.090 4.119 ı. FWER Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively` adjusted p-values in square brackets, based on 10,000 replication. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. Table A13: Impact on selected HIV outcomes CROSS-SECTION CONDITIONAL SPECIFICATION Coefficient on: Treated Treated Treated*Yt−1 Coeff p-value Coeff p-value Coeff p-value p-value No. Mean in (Std.Err) FWER (Std.Err) FWER (Std.Err) FWER (3)+(5) Obs control group Dep. Var. (Yt ): (1) (2) (3) (4) (5) (6) (7) (8) (9) Tested for HIV (observed) (a) 0.031** 3828 0.033 (0.013) [0.014] Tested last 6 months (self-reported) 0.025*** 0.026*** -0.038 0.005 4971 0.086 (0.009) [0.014] (0.009) [0.004] (0.063) [0.550] HIV transmitted during pregnancy 0.069*** 0.127*** -0.098*** 0.000 4986 0.621 (0.019) [0.002] (0.033) [0.002] (0.032) [0.025] Has mentioned ARV drugs spontaneously or 0.037** 0.032** -0.019 0.486 4986 0.047 when mentioned by enumerator (0.017) [0.041] (0.015) [0.045] (0.036) [0.623] Second test necessary 0.047*** 0.055 -0.039 0.003 4986 0.343 (0.015) [0.013] (0.018) [0.014] (0.037) [0.726] Window period 3 months 0.049*** 0.045*** -0.009 0.000 4986 0.129 xx (0.012) [0.001] (0.012) [0.003] (0.062) [0.955] Can get HIV through intercourse 0.01* -0.037 0.050 0.285 4986 0.969 (0.006) [0.136] (0.034) [0.500] (0.034) [0.552] Would buy from an HIV+ shopkeeper 0.047*** 0.082*** -0.094*** 0.000 4986 0.487 (0.015) [0.013] (0.020) [0.000] (0.032) [0.024] An HIV+ boy should play football 0.051*** 0.080*** -0.045 0.006 4986 0.662 (0.016) [0.012] (0.028) [0.022] (0.036) [0.581] People HIV+ should not be blamed 0.022 0.045 -0.031 0.240 4986 0.676 (0.018) [0.422] (0.037) [0.417] (0.041) [0.701] HIV is not punishment for sleeping around 0.045** 0.027 0.021 0.063 4986 0.486 (0.023) [0.137] (0.028) [0.417] (0.034) [0.701] Would reveal HIV status to partner 0.015 0.049* -0.044 0.098 4986 0.713 (0.013) [0.422] (0.029) [0.263] (0.034) [0.581] Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. FWER adjusted p-values in square brackets, adjusted within family of variables (e.g., knowledge, attitudes, testing) and based on 10,000 replications. ”P-Value (3)+(5)” tests the null hypothesis Treated + Treated*Yt−1 m =0, where Yt−1 m denote the sample mean of Yt−1 at baseline. Estimates in cols. 1-2 come from the cross sectional model; estimates in cols. 3-5 from the model that includes the lagged dependent variable and its interaction with treatment. Col. 9 reports the mean of the dependent variable at follow-up. All regressions include the following controls: female, age, currently attending school, years of education, English spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as native language. (a) For this outcome we cannot estimate the model including the lag of the dependent variable because teasting through health-camps was only offered at follow-up Table A14: Treatment Effect on the probability of HIV ] Dep. Var: Prob. of being currently Prob. that partner is HIV Prob. that a person your age HIV positive positive in the community has HIV (1) (2) (3) (4) (5) (6) (7) (8) (9) Treated 0.000 0.001 0.000 -0.006 -0.005 -0.002 0.027** 0.024** 0.034** (0.002) (0.002) (0.003) (0.005) (0.005) (0.005) (0.011) (0.011) (0.013) [0.852] [0.721] [0.947] [0.410] [0.528] [0.928] [0.055] [0.070] [0.035] Yt−1 0.088*** 0.073** 0.110*** 0.142*** 0.116*** 0.137*** (0.017) (0.028) (0.019) (0.027) (0.016) (0.030) Treated*Yt−1 0.025 -0.047 -0.031 xxi (0.034) (0.037) (0.037) [0.654] [0.531] [0.654] Observations 4.986 4.986 4.986 4.066 3.452 3.452 4.986 4.986 4.986 R-squared 0.006 0.015 0.016 0.018 0.031 0.031 0.021 0.034 0.034 Controls(a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean Dep. Var. (Control) 0.019 0.019 0.019 0.058 0.055 0.055 0.337 0.337 0.337 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. FWER adjusted p-values in square brackets, based on 10,000 replications. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. Table A15: Heterogeneous effects by stability of partnerr Dep. Var.: HIV knowledge HIV attitudes HIV testing Attitudes towards Risky sexual behav- risky sexual behav- ior (for sexually ac- ior tive) (1) (2) (3) (4) (5) Treated 0.690 0.702 0.999** 0.615 0.081 (0.708) (0.434) (0.395) (0.414) (0.561) [0.339] [0.208] [0.044] [0.276] [0.884] Treated * Stable Partnert-1 -0.045 -0.496 -0.927** -0.623 0.027 (0.757) (0.456) (0.401) (0.426) (0.614) [0.956] [0.485] [0.071] [0.281] [0.947] Stable Partnert-1 -0.489 0.332 0.879*** 0.565 -0.745 (0.610) (0.395) (0.243) (0.341) (0.549) Treated*Yt−1 -0.043 -0.016 -0.011 0.064 -0.110* (0.047) (0.045) (0.037) (0.046) (0.062) [0.750] [0.926] [0.926] [0.165] [0.162] Yt−1 0.371*** 0.365*** 0.436*** 0.260*** 0.339*** (0.039) (0.040) (0.031) (0.036) (0.050) Observations 2,827 2,827 2,827 2,827 2,777 R-squared 0.195 0.179 0.250 0.109 0.160 Controls(a) Yes Yes Yes Yes Yes Mean Dep. Var. (Control) 0.040 0.0439 -0.139 0.00186 -0.0631 Std Dev Dep. Var. (Control) 6.106 3.474 4.145 3.452 3.625 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. FWER adjusted p-values in square brackets, based on 10,000 replications (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks yoruba as a native language. The dummy ”Stable partner” is measured at baseline and takes value one if the respondent reports being married or living with someone or having a main partner. xxii Table A16: Impact on selected risky sexual behavior outcomes CROSS-SECTION CONDITIONAL SPECIFICATION Coefficient on: Treated Treated Treated*Yt−1 Coeff P-Value Coeff P- Coeff P-Value P-Value No. Mean in con- (Std.Err) FWER (Std.Err) Value (Std.Err) FWER (3)+(5) Obs trol group FWER Dep. Var. (Yt ): (1) (2) (3) (4) (5) (6) (7) (8) (9) Men should have one partner only 0.036*** 0.074** -0.048 0.048 4973 0.867 (0.012) [0.013] (0.037) [0.100] (0.036) [0.347] Women should have one partner only 0.027*** 0.103*** -0.088** 0.004 4986 0.906 (0.010) [0.021] (0.035) [0.014] (0.036) [0.051] Has not had multiple concurrent sexual partners 0.027* 0.113*** -0.103** 0.007 2623 0.780 (0.015) [0.232] (0.041) [0.018] (0.041) [0.028] Number of current sexual partners if sexually active -0.051 0.215** -0.222*** 0.014 1850 1.323 (0.034) [0.232] (0.086) [0.049] (0.07) [0.020] xxiii If a woman brings condom does not mean she’s not serious 0.014 0.012 -0.004 0.640 4986 0.613 (0.015) [0.347] (0.026) [0.636] (0.030) [0.901] Used condom the last time he/she had sex -0.003 -0.021 0.038 0.513 2308 0.494 (0.019) [0.844] (0.032) [0.506] (0.053) [0.574] Have you ever used a condom? 0.018 0 .069* -0.088** 0.054 3070 0.737 (0.014) [0.495] (0.035) [0.153] (0.039) [0.086] Are you currently using condoms? -0.014 -0.027 0.01 0.508 2990 0.775 (0.015) [0.588] (0.04) [0.759] (0.047) [0.972] The first time you will have sex, do you plan to use a condom? 0.012 0.003 0.006 0.875 1311 0.168 (0.02) [0.588] (0.019) [0.874] (0.077) [0.972] Notes: Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. FWER adjusted p-values in square brackets based on 10,000 replications. ”P-Value (5)+(6)” tests the null hypothesis Treated+ Treated*Yt−1 m =0, where by Yt−1 m we denote the sample mean of Yt−1 at baseline. Controls in each regression include: female, age, currently attending school, years of education, english spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language Table A17: Involvement in the narrative, loading factors Loading factor Transportation index You could easily imagine what was going on in the show. 0.3055 You were not distracted by activities in the room around you. 0.1387 You could imagine yourself being part of the story. 0.2800 You were really following the story. 0.3851 You wanted to learn how the story ended. 0.3673 It affected you emotionally. 0.3113 You were thinking of ways the story could have ended in a different way. 0.3146 While watching the show, you did not found your mind wandering. 0.0933 You found the story relevant to your everyday life. 0.2632 You had a clear picture of the characters in the story. 0.3504 You did not found it easy to forget about it. 0.1667 You feel the story has changed your life. 0.3183 Identification index While viewing the show, you felt as if you were part of the action 0.3174 While viewing the show, you forgot yourself and you were fully absorbed 0.2869 You were able to understand the events in the show like the characters understood them 0.3607 You have a good understanding of the characters 0.3688 You understand the reasons why the characters did what they did. 0.3510 While viewing the show you could feel the emotions the characters displayed. 0.3385 During the show, you felt you could read the characters? minds. 0.2808 At key moments in the show, you felt you had experienced the same thing that the 0.2117 characters were going through While viewing the show, you wanted the characters to succeed in achieving their goals. 0.3156 When the characters succeeded you felt joy, but when they failed, you felt sad. 0.2996 xxiv Table A18: Correlates of Transportation and Identification Dep. Var.: Transportation Index Identification Index Coeff. Std. Err. Coeff. Std. Err. Female -0.010 (0.088) -0.262*** (0.098) Age 0.016 (0.015) 0.018 (0.015) Currently attending school 0.065 (0.077) -0.048 (0.072) Years of education -0.060 (0.040) -0.072* (0.040) English spoken at home 0.237 (0.162) 0.374** (0.151) Single -0.036 (0.083) 0.060 (0.085) Does not live with the family -0.055 (0.105) 0.048 (0.120) Household size -0.006 (0.017) -0.016 (0.020) Wealth index 0.116* (0.063) 0.157** (0.074) Home owner -0.087 (0.104) -0.126 (0.111) Father obtained education higher then sec. -0.029 (0.087) 0.051 (0.087) Mother obtained education higher then sec. -0.084 (0.100) -0.123 (0.102) Muslim 0.015 (0.086) -0.100 (0.084) Yoruba 0.043 (0.128) 0.126 (0.146) Constant 0.009 (0.520) 0.026 (0.651) Observations 3,753 3,753 R-squared 0.019 0.030 P-value F-test of joint significance 0.391 0.001 Notes: Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. xxv Table A19: Transportation and Identification, cross sectional specification (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Dependent variable: HIV knowledge HIV attitudes HIV testing Attitudes towards risky sex- Risky sexual behavior (for ual behavior sexually active) Panel A: Transportation Index Treated*Transportation 0.491*** 0.492*** 0.204*** 0.207*** 0.140 0.140 0.351*** 0.357*** 0.090 0.082 (0.109) (0.110) (0.073) (0.072) (0.085) (0.086) (0.073) (0.074) (0.063) (0.065) [0.000] [0.000] [0.015] [0.013] [0.112] [0.098] [0.000] [0.000] [0.168] [0.208] Treated 1.145*** -7.852** 0.358** -1.990 0.332* -1.364 -0.012 -1.983 0.045 -2.862 (0.262) (3.824) (0.164) (1.975) (0.186) (2.349) (0.120) (1.982) (0.128) (1.988) [0.000] [0.121] [0.054] [0.549] [0.073] [0.562] [0.915] [0.319] [0.915] [0.269] Transportation -0.047 -0.051 0.001 -0.003 -0.011 -0.015 -0.021 -0.021 -0.057 -0.050 (0.077) (0.077) (0.046) (0.045) (0.056) (0.055) (0.052) (0.053) (0.043) (0.045) Observations 3,753 3,753 3,753 3,753 3,753 3,753 3,753 3,753 2,667 2,667 R-squared 0.103 0.107 0.056 0.058 0.092 0.097 0.041 0.047 0.102 0.107 Controls(a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls*Treated No Yes No Yes No Yes No Yes No Yes Mean Dep. Var. (Control) 0.0400 0.0400 0.0439 0.0439 -0.139 -0.139 0.00186 0.00186 -0.0631 -0.0631 Std Dev Dep. Var. (Control) 6.106 6.106 3.474 3.474 4.145 4.145 3.452 3.452 3.625 3.625 P-value test joint sig 3.77e-05 0.0434 0.0325 0.317 0.0781 0.563 0.919 0.320 0.727 0.154 xxvi Panel B: Identification Index Treated*Identification 0.310*** 0.322*** 0.124* 0.118* 0.140* 0.143* 0.181** 0.174** 0.028 0.014 (0.097) (0.096) (0.064) (0.064) (0.076) (0.077) (0.069) (0.069) (0.055) (0.058) [0.006] [0.006] [0.117] [0.139] [0.117] [0.139] [0.024] [0.033] [0.612] [0.823] Treated 1.172*** -7.976** 0.377** -2.060 0.333* -1.398 0.011 -2.108 0.043 -2.882 (0.256) (3.830) (0.161) (1.994) (0.190) (2.353) (0.121) (2.007) (0.130) (1.977) [0.000] [0.122] [0.042] [0.516] [0.083] [0.541] [0.935] [0.313] [0.935] [0.267] Identification -0.006 -0.014 0.012 0.014 -0.009 -0.015 0.018 0.026 -0.035 -0.022 (0.072) (0.072) (0.047) (0.047) (0.056) (0.056) (0.049) (0.050) (0.040) (0.043) Observations 3,753 3,753 3,753 3,753 3,753 3,753 3,753 3,753 2,667 2,667 R-squared 0.098 0.102 0.052 0.054 0.092 0.097 0.031 0.036 0.102 0.107 Controls(a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controls*Treated No Yes No Yes No Yes No Yes No Yes Mean Dep. Var. (Control) 0.0400 0.0400 0.0439 0.0439 -0.139 -0.139 0.00186 0.00186 -0.0631 -0.0631 Std Dev Dep. Var. (Control) 6.106 6.106 3.474 3.474 4.145 4.145 3.452 3.452 3.625 3.625 P-value test joint sig 1.70e-05 0.0405 0.0218 0.305 0.0838 0.554 0.928 0.297 0.741 0.149 Notes: Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. P-values in square brackets corrected for multiple hypothesis testing using FWER. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. Table A20: Impact of T2 on outcome indexes (1) (2) (3) (4) (5) HIV knowledge HIV attitudes HIV testing Attitudes towards Risky sexual behav- risky sexual behav- ior (for sexually ac- ior tive) Panel A: Cross-section T2 -0.089 -0.065 0.019 -0.075 0.057 (0.209) (0.137) (0.174) (0.122) (0.132) Observations 3,402 3,402 3,402 3,402 2,487 R-squared 0.091 0.057 0.093 0.029 0.084 Panel B: Conditional specification T2 -0.074 0.037 0.101 -0.049 -0.053 (0.195) (0.128) (0.156) (0.121) (0.148) T2*Yt−1 0.036 0.057 -0.028 -0.027 0.005 (0.044) (0.044) (0.037) (0.040) (0.067) Yt−1 0.301*** 0.334*** 0.447*** 0.304*** 0.209*** (0.032) (0.031) (0.027) (0.029) (0.050) Observations 3.402 3.402 3.402 3.402 2.117 R-squared 0.188 0.184 0.246 0.107 0.135 Controls(a) Yes Yes Yes Yes Yes Mean Dep. Var. (Control) 1.080 0.534 0.278 0.176 0.0740 Std Dev Dep. Var. (Control) 6.043 3.373 4.609 3.433 3.562 Notes: Standard errors in parentheses adjusted for clustering at the screening centre level (54 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. xxvii Table A21: Spillovers, cross sectional specification Dep. Var. (Yt ): HIV knowledge HIV attitudes HIV testing Attitudes towards Risky sexual be- risky sexual behav- havior (for sexu- ior ally active) (1) (2) (3) (4) (’5) (6) (7) (8) (9) (10) Friend of Treated 0.642* 0.587 -0.063 -0.080 -0.454 -0.520 -0.300 -0.253 -0.408 -0.366 (0.376) (0.375) (0.256) (0.258) (0.296) (0.313) (0.210) (0.220) (0.319) (0.324) [0.216] [0.263] [0.772] [0.749] [0.195] [0.263] [0.290] [0.439] [0.290] [0.439] Friend of Treated*Opposite Sex 1.388 0.636 1.719 -1.214 -1.181 (1.221) (1.028) (1.266) (1.069) (1.448) [0.541] [0.559] [0.541] [0.532] [0.532] Opposite Sex -0.799 0.187 -0.812 0.630 0.724 xxviii (0.738) (0.732) (0.984) (0.800) (0.756) Observations 924 924 924 924 924 924 924 924 680 680 R-squared 0.119 0.119 0.093 0.094 0.139 0.140 0.057 0.058 0.125 0.126 Controls (a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean Dep. Var. (Control) -0.185 -0.185 -0.316 -0.316 -0.164 -0.164 -0.107 -0.107 -0.213 -0.213 Std Dev Dep. Var. (Control) 6.063 6.063 3.601 3.601 4.410 4.410 3.512 3.512 3.862 3.862 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. Sample includes network friends that have non-missing Yt−1 . FWER adjusted p-values in square brackets, based on 10,000 replications. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. Table A22: Balance between main respondents and network members used for spillover analysis Mean main Mean Diff=0 Normalized No. Obs. respondents network (p-value) Diff.(a) members (1) (2) (3) (4) (5) Indexes HIV knowledge 0.141 0.300 0.456 0.018 6042 xxix HIV attitudes 0.007 -0.158 0.167 -0.033 6042 HIV testing -0.029 0.189 0.141 0.035 6042 Attitudes towards risky sexual behavior -0.022 -0.367 0.003 -0.071 6042 Risky sexual behavior (sex act) 0.023 0.107 0.573 0.017 3808 Notes: The table reports sample means at baseline. (a) Normalized difference is the difference in the sample means of main respondents and their network members divided by the square root of the sum of the sample variances. Table A23: Indexes of conformism, loading factors Loading factor Conformity index He/she believes that people should do what they’re told. He/she thinks people 0.4832 should follow rules at all times, even when no-one is watching. It is important to him/her to always behave properly. He/she wants to avoid 0.5541 doing anything people would say is wrong. He/she believes he/she should always show respect to his/her parents and to 0.4393 older people. It is important to him/her to be obedient. It is important to him/her to be polite to other people all the time. He/she 0.5162 tries never to disturb or irritate others. Tradition index He/she thinks it’s important not to ask for more than what you have. He/she 0.5239 believes that people should be satisfied with what they have. Religious belief is important to him/her. He/she tries hard to do what his/her 0.5127 religion requires. He/she thinks it is best to do things in traditional ways. It is important to 0.4236 him/her to keep up the customs he has learned. It is important to him/her to be humble and modest. He/she tries not to draw 0.5321 attention to himself/herself. Self-Direction index Thinking up new ideas and being creative is important to him/her. He/she 0.4765 likes to do things in his/her own original way. It is important to him/her to make his own decisions about what he/she does. 0.5445 He/she likes to be free to plan and to choose what to do himself/herself. He/she thinks it’s important to be interested in things. He/she likes to be 0.4678 curious and to try to understand all sorts of things. It is important to him/her to be independent. He likes to rely on himself. 0.5076 xxx Table A24: Treatment and conformism, cross section (1) (2) (3) (4) (5) Dependent Variable: HIV knowl- HIV atti- HIV testing Attitudes to- Risky sexual be- edge tudes wards risky havior (for sex- sexual behavior ually active) Panel A: Conformity Treated 0.890*** 0.346** 0.355** 0.147 0.126 (0.244) (0.138) (0.148) (0.090) (0.131) [0.001] [0.025] [0.025] [0.208] [0.341] Treated*Conformity 0.186 -0.147 -0.017 -0.057 -0.069 (0.149) (0.110) (0.099) (0.081) (0.087) [0.216] [0.463] [0.876] [0.669] [0.669] Conformity -0.001 0.188** -0.045 0.177*** 0.121* (0.128) (0.089) (0.070) (0.061) (0.071) R-squared 0.086 0.052 0.092 0.024 0.091 P-Value of joint signif. 0.000 0.014 0.018 0.113 0.348 Panel B: Tradition Treated 0.900*** 0.344** 0.354** 0.144 0.118 (0.244) (0.136) (0.147) (0.090) (0.131) [0.000] [0.024] [0.024] [0.216] [0.370] Treated*Tradition -0.031 -0.127* 0.017 0.007 -0.139 (0.151) (0.068) (0.088) (0.090) (0.106) [0.975] [0.174] [0.982] [0.926] [0.321] Tradition Index -0.012 0.145*** -0.055 0.137* 0.215** (0.114) (0.054) (0.060) (0.078) (0.094) R-squared 0.085 0.051 0.092 0.024 0.092 P-Value of joint signif. 0.000 0.014 0.018 0.114 0.386 Panel C: Self Direction Treated 0.865*** 0.334** 0.349** 0.144 0.125 (0.239) (0.136) (0.149) (0.093) (0.127) [0.000] [0.028] [0.028] [0.225] [0.326] Treated*Self-direction 0.134 0.005 0.125* 0.038 0.137* (0.121) (0.071) (0.068) (0.081) (0.078) [0.288] [0.937] [0.206] [0.640] [0.164] Self-direction -0.538*** -0.149*** -0.093** -0.077 -0.154** (0.100) (0.052) (0.046) (0.068) (0.059) R-squared 0.095 0.054 0.092 0.021 0.091 P-Value of joint signif. 0.000 0.016 0.021 0.123 0.324 Controls (a) Yes Yes Yes Yes Yes Observations 4,986 4,986 4,986 4,986 3,618 Mean Dep. Var. (Control) 0.04 0.044 -0.139 0.00186 -0.063 Std Dev Dep. Var. (Control) 0.000 0.0133 0.0186 0.113 0.383 Notes. Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. FWER adjusted p-values in square brackets, based on 10,000 replications. (a) Controls in each regression include: female, age, currently attending school, years of education, english spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language. xxxi Table A25: Heterogeneity by education and English Dep. Var. (Yt ): HIV knowledge HIV attitudes HIV testing Attitudes towards Risky sexual behavior risky sexual behavior (for sexually active) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Treated -3.648* -0.084 -1.808 -0.239 0.779 -0.762 -1.780 -0.315 -1.748 0.377 (2.173) (0.890) (1.147) (0.466) (1.129) (0.688) (1.088) (0.568) (1.059) (0.527) [0.285] [0.924] [0.285] [0.852] [0.492] [0.622] [0.225] [0.733] [0.225] [0.733] Treated*Years of Education 0.382** 0.185* -0.038 0.166* 0.163* (0.185) (0.096) (0.098) (0.094) (0.090) [0.135] [0.135] [0.697] [0.173] [0.173] Years of Education 0.007 0.050 0.062 -0.019 -0.100 (0.161) (0.087) (0.055) (0.084) (0.068) Treated*English spoken at home 0.903 0.604 1.144* 0.482 -0.242 (0.926) (0.479) (0.686) (0.591) (0.529) [0.387] [0.387] [0.284] [0.674] [0.674] English spoken at home 0.580 -0.528 -0.682 -0.376 -0.103 (0.749) (0.394) (0.620) (0.501) (0.454) xxxii Treated*Y lagged -0.077** -0.068* -0.002 0.007 -0.032 -0.035 -0.005 -0.001 -0.118** -0.117** (0.036) (0.036) (0.038) (0.037) (0.032) (0.032) (0.034) (0.033) (0.056) (0.056) [0.103] [0.177] [0.934] [0.861] [0.536] [0.483] [0.887] [0.985] [0.078] [0.087] Y lagged 0.399*** 0.393*** 0.367*** 0.361*** 0.471*** 0.473*** 0.296*** 0.293*** 0.322*** 0.321*** (0.028) (0.029) (0.031) (0.030) (0.027) (0.027) (0.028) (0.027) (0.046) (0.046) Observations 4.986 4,986 4.986 4,986 4.986 4,986 4.986 4,986 3.07 3,070 R-squared 0.199 0.198 0.180 0.179 0.261 0.262 0.099 0.099 0.153 0.153 Controls(a) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Mean Dep. Var .(Control) 0.0400 0.585 0.0439 0.297 -0.139 0.0925 0.0019 0.0849 -0.0631 0.0156 Std Dev Dep. Var. (Control) 6.106 6.127 3.474 3.476 4.145 4.441 3.452 3.417 3.625 3.546 P-value test joint sig 0.0972 0.925 0.119 0.609 0.492 0.272 0.106 0.581 0.103 0.476 Notes: Standard errors in parentheses adjusted for clustering at the screening centre level (80 clusters). ***, ** and * denote significance at the 1, 5 and 10 percent level, respectively. Increasing values of the dependent variable correspond to outcomes more consistent with the messages of Shuga. FWER adjusted p-values in square brackets, based on 10,000 replications (a) Controls in each regression include: female, age, currently attending school, years of education, English spoken at home, single, does not live with the family, household size, wealth index, home owner, father obtained higher than secondary education, mother obtained higher than secondary education, muslim, speaks Yoruba as a native language.