The Government Response to Informed Citizens: New Evidence on Media Access and the Distribution of Public Health Benefits in Africa Philip Keefer and Stuti Khemani We use a “natural experiment” in media markets in Benin to examine the impact of community radio on government responsiveness to citizens. Contrary to prior research on the impact of mass media, in this experiment government agents do not provide greater benefits to citizens whose exposure to community radio increased their demand for those benefits. Households with greater access to community radio were more likely to pay for government-provided bed nets to combat malaria than to receive them for free. Mass media changed the private behavior of citizens—they invested more of their own resources in the public health good of bed nets—but not citizens’ ability to extract greater benefits from government. While the welfare consequences of these results are ambiguous, the pattern of radio’s effects that we uncover has implications for policy strategies to use mass media for development objectives. JEL codes: D72, D73, D83, H51, I18 This paper examines the impact of community radio in Africa on government responsiveness to citizens. Community radio has proliferated as donors and gov- ernments have turned to media to advance development objectives in remote rural areas where poor households tend to live (Buckley et al. 2008). Community radio stations are, by definition, licensed as noncommercial radio organized to Philip Keefer (corresponding author) is Principal Economic Advisor, Inter-American Development Bank (IADB); his email address is: pkeefer@iadb.org. Stuti Khemani is Senior Economist, World Bank; his email address is: skhemani@worldbank.org. We thank Peter Lanjouw and the two anonymous referees. We also thank Tim Besley, Alberto Diaz-Cayeros, Quy-Toan Do, Esther Duflo, Pascaline Dupas, Lawrence Katz, and participants at the World Bank research seminar, Stanford University conference on public goods provision, and the Annual Bank Conference of Development Economics for very useful comments. We thank Anne-Katrin Arnold, Tony Lambino, and Sina Odugbemi for very useful references and discussion. We are extremely grateful to Ayite-Fily D’ Almedia and Njara Rakotonirina for their generosity in providing detailed information about the malaria prevention programs in Benin. We thank Illenin Kondo and Quynh Nguyen for excellent research assistance. We are indebted to Leonard Wantchekon and the team at the Institute for Empirical Research in Political Economy (IREEP) in Cotonou, Benin, for the expert survey work and assistance with survey design. We are grateful to generous financing from the Knowledge for Change Program, which made this research possible. The opinions and conclusions expressed here are those of the authors and not those of either the World Bank or the Inter- American Development Bank, nor their respective Executive Directors. THE WORLD BANK ECONOMIC REVIEW, VOL. 30, NO. 2, pp. 233– 267 doi:10.1093/wber/lhv040 Advance Access Publication August 6, 2015 # The Author 2015. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com. 233 234 THE WORLD BANK ECONOMIC REVIEW serve the public interest of the communities in which they are based (Fraser and Restrepo-Estrada 2002; Buckley et al. 2008). Contrary to prior research on the impact of mass media, we find that government agents do not provide greater benefits to citizens who, because of their greater exposure to community radio, exhibit greater demand for those benefits. A “natural experiment” in radio markets in Benin reveals that households with greater access to community radio were more likely to pay for government-provided bed nets to combat malaria than to receive them for free. In Benin, as in most of Africa, donors and government ministries distribute free anti-malaria bed nets through local public health personnel and sponsor pro- gramming on community radios to persuade households to take up public health services. This policy context allows us to examine the potentially powerful role of mass media in shaping development outcomes through two different channels: citizens’ media exposure could shift the behavior of government towards citizens, a governance issue, and it could change private household behavior towards public health. We find that radio changed the private behavior of citizens; they invested more of their own resources in the public health good of bed nets. It did not increase citizens’ ability to extract greater benefits from government. The study area in Benin offers a natural experiment in radio access that we use, in combination with purposefully designed survey data, to identify the effect of variation in radio access on the government distribution of free bed nets. The data come from an original survey, undertaken in March 2009, soon after a massive wave of free bed-net distribution by the Benin government. More than 4000 households were surveyed, in 210 villages spread across 32 communes in northern Benin. These households are representative of the types of households that community radio is designed to reach, since this is the more rural and poorer region of Benin. The unique structure of radio markets in northern Benin gives rise to exoge- nous variation in radio access across villages. Twenty-one community radio sta- tions serve this region; all carry similar programming on public health messages sponsored by the government and donors on the value of bed nets in combatting malaria. The headquartering of a station in a particular commune can be endoge- nous to commune characteristics, so we rely only on within-commune differences in radio access. Intra-commune variation, though, is exogenous. The stations are dispersed, rather than concentrated around specific markets or densely populated areas, and each is small so that it can be community-owned and managed. This dispersion yields significant intra-commune variation in radio access across vil- lages. Because the signals of these low-powered stations are easily disrupted, even villages located close to each other and sharing the same economic, geo- graphic, and topographical features can differ significantly in the number of com- munity radio signals reaching them both from within and outside the commune. We drew our village sample by exploiting this potential for exogenous variation in intra-commune radio access across similarly situated villages. Keefer and Khemani 235 The National Malaria Control Program of Benin, in effect since 2006, three years before the survey, classifies anti-malaria bed nets as an “essential” com- modity that should be provided free by local public health personnel in order to achieve full population coverage against malaria and especially full coverage of poor and vulnerable groups (such as pregnant women and young children). Substantial donor assistance in the procurement and supply of bed nets sup- ports this objective, especially under the Roll Back Malaria Initiative. The policy of free distribution is well-known: 93 percent of household respondents in our survey correctly identify it.1 However, regardless of official policy, local staff has always had considerable de facto discretion about whether and whom to charge for government bed nets. Our results show that they are more likely to use their discretion to charge for bed nets in villages with greater radio access. On average, in villages with greater access to radio, households are more likely to pay the local public health providers to acquire government bed nets, com- pared to households in villages with lower radio access that are more likely to receive government bed nets for free. Radio access has no significant effect on the total quantity of bed nets that households report owning. This pattern of evi- dence is consistent with a radio-induced shift in the demand for bed nets without a corresponding increase in their supply. In contrast, earlier work has found that governments respond to citizens who have greater media access by increasing the supply of benefits such as disaster relief and welfare assistance (Besley and Burgess 2002; Eisensee and Stro ¨ mberg 2007). The different result in this study can be explained by what we know about the content of community radio programming and political institutions in Benin. Both of these influence the framing of the messages that radio stations broadcast. In the contexts examined in prior literature, media programs were likely to frame issues in terms of government accountability for delivering greater benefits (e.g., disaster relief ) to citizens.2 This is less likely in the case of community radio sta- tions, which depend upon sponsored programming for their revenues and func- tioning. We know from documents issued by government and donors that they directly sponsor programming that is framed to encourage households to acquire and use bed nets to combat malaria rather than to encourage households to query government about the adequacy of bed-net distribution. Community radio programs therefore emphasize changes in private household behavior towards health and education rather than exhort citizens to make greater demands upon government. Since this framing emerges from the programming that stations need to broadcast in order to stay afloat financially, the evidence we present sup- ports prior research that has concluded that incentives in media markets are key 1. The national policy prior to 2006 also emphasized free distribution to indigent and vulnerable groups, but left more discretion to local health staff on pricing practices. 2. Another strand of the literature focuses on media market effects on generalized corruption but not on specific public service delivery outcomes (Ferraz and Finan 2008; Besley and Prat 2006; Brunetti and Weder 2003; Adsera et al. 2003). 236 THE WORLD BANK ECONOMIC REVIEW determinants of the type of programming content that media broadcasts and how that content is framed to influence public policy (Prat and Stro ¨ mberg 2005; Eisensee and Stro¨ mberg 2007). Nevertheless, when radio programs inform households about the value of bed nets, increasing their demand for them, prior research predicts that government political incentives should lead to an increase in supply to more demanding citi- zens (Maskin and Tirole 2004). The political context in Benin explains why we do not observe such a supply response: politicians do not have an incentive to respond to increased demand with greater supply, in the case of bed nets. Catering to greater household demand by shifting the distribution of free bed nets from uninformed to informed households is unlikely to be an effective policy instrument to garner votes in the political context of Benin. Political competition in Benin is heavily clientelist (Wantchekon 2003), depending on the exchange of specific benefits for political support from narrow constituen- cies embedded in personalized networks. In such an environment, the political benefits of responding to more informed households (e.g., by providing more free bed nets) are less certain, since these households need not belong to the ap- propriate clientelist networks.3 Beyond improving our understanding of the effects of media on government responsiveness, the analysis also contributes to research on the determinants of household take-up of public health services (Cohen and Dupas 2010; Dupas 2009).4 Dupas (2009), for example, finds that an experiment with social market- ing had no effect on household sensitivity to the price of bed nets, but this and other prior research is silent on the effects of media on the demand for public health services. Our contribution suggests that community radio can be an im- portant medium to persuade households about the value of health goods. It is also consistent with related results from Benin showing that radio induces house- holds to invest more in the education of their children (Keefer and Khemani 2014).5 A growing literature deals with how to improve health and education out- comes by providing information to households, focusing on concerted campaigns ¨ rkman that often entail intensive, face-to-face interactions with households (Bjo and Svensson 2009; Banerjee et al. 2010; Jensen 2010; Pandey et al. 2009; Andrabi et al. 2008). Our finding of radio’s effect on household demand for a public health good, along with those of Keefer and Khemani (2014) on demand for education, is notable in this context, showing that media programming can 3. For a formal analysis of policy effects of this mismatch, see Keefer and Vlaicu (2008). 4. A large communications literature shows that focused media programming can influence household behavior in public health and family planning, though these studies do not control for the endogeneity of radio access (Arnold and Lambino 2009 provide a review). 5. This paper differs from Keefer and Khemani (2011) in the policy context it examines and in the media effects it identifies. The policy context of bed net distribution allows us to be the first to uncover a government price response to more informed citizens. Keefer and Khemani 237 provide a less transactions-intensive and potentially more cost-effective approach to information transmission to change household behaviors.6 These results add to research on media effects on private household behavior and social norms (Paluck, 2009). For example, Chong and La Ferrara (2009) report that expanding access to the broadcasts of soap operas in Brazil increased rates of divorce and separation; La Ferrara et al. (2008) find that it reduced fertility. Jensen and Oster (2009) find that programming on cable television im- proved the status of women in India. The potential adverse influence of radio access on household behavior, with tragic consequences, was recently quantified in Yanagizawa’s (2009) work on the impact of “hate radio” in the Rwandan genocide. In the next section we describe the policy and institutional context of Benin to show that media are more likely to influence private household demand for health goods rather than government accountability for delivering greater bene- fits. Section 3 describes the empirical identification strategy to test the impact of greater media access. Section 4 presents the main results, the reduced-form effects of radio access on bed net acquisition. A large battery of tests for the ro- bustness of our main results is presented in Section 5. Section 6 discusses the possible welfare consequences of the results. Section 7 concludes by describing the relevance of these findings for policies that use mass media to support devel- opment objectives, such as by changing household behavior and government accountability. I . T H E S T U DY CO N T E X T : B E D N E T D I S T R I B U T I O N AND MEDIA MARKETS IN BENIN Benin is a small country in Francophone West Africa. Malaria is endemic and the number one killer of children under five.7 The country has spent significant re- sources in successive national malaria control programs on the distribution of bed nets through government health facilities. To this end, it received substantial international aid under the Roll Back Malaria Partnership (sponsored most notably by UNICEF, the World Health Organization, and the World Bank), which contributed to a massive expansion in the distribution of free government bed nets in 2008, the year prior to our survey.8 6. A potential concern with interpreting the reduced form results as a consequence of a shift in household demand for bed nets is that households with greater media access do not acquire more bed nets from the private market and private sellers do not increase the prices they charge. We discuss further below that in this policy context—of large scale government distribution of free bed nets, a small private sector, and related public policies that restrict the penetration of the private market in selling bed nets—it is likely to be difficult for the private market to detect and respond to subtle changes in household demand in the way local, community-embedded public health providers do. 7. Consistent with this, we find that respondents with greater access to community radio and its malaria programming are more likely to be knowledgeable about child mortality rates in Benin. 8. According to the Plan Inte ´ des Activite ´ gre ´ s de Lutte Contre le Paludisme pour l’anne ´ e 2009 au Be´ nin, http://rollbackmalaria.org/countryaction/docs/warn/beninPlanPNLP2009.pdf, 8. 238 THE WORLD BANK ECONOMIC REVIEW Anti-malaria campaigns in the Roll Bank Malaria program also include com- munication efforts using local radio programming (USAID 2011, 7, Roll Back Malaria Partnership 2008). Greater exposure to such radio programs could there- fore lead to greater take-up, reflected in higher bed-net ownership in areas with more radio access, if government distributors increase the supply of bed nets or shift bed nets from areas with fewer informed households to those with more. Data collected for this study show, however, that government distributors respond to higher demand by raising prices, keeping quantity supplied unchanged. Two features of Benin are important for the empirical analysis that follows. First, the mechanics and oversight of bed-net distribution allow locally deter- mined, village-level variation in the prices charged for bed nets. Second, the nature of programming carried on radio is more likely to influence private household behavior than government accountability. This section discusses these in turn. The Distribution of Free Bed Nets and the Market for Bed Nets in Benin We find that in areas exposed to more community radio stations, household demand for bed nets is higher and households are more likely to report that they paid for their government-provided bed nets but not that they have more govern- ment bed nets. Policy and project documents verify four facts that are important for interpreting these findings and the fact that we observe a demand but not a supply response. First, the bulk of bed nets are distributed by the government, so the effects we observe are largely unaffected by the interaction of private and govern- ment markets. In 2007, 1,450,000 bed nets were distributed by the government (prior to the Roll Back Malaria campaign) and 800,000 in 2009, during the cam- paign. Donors and government also support the distribution of a small portion of total bed nets through social marketing by the nongovernment sector, at heavily subsidized prices (USAID 2011, 7 and 14). The most important private organiza- tion that directly distributes bed nets in Benin is an NGO called Population Services International (PSI). However, it reported selling only 75,000 bed nets in 2009. That is, the private market is not only thin, but more importantly, also domi- nated by nongovernmental organizations, with similar objectives of achieving full coverage against malaria through price subsidies, rather than servicing profit motives. Private suppliers of bed nets are therefore less likely—both because of their lack of profit incentives and their sparse distribution in the study area—to detect and respond to subtle, localized shifts in household demand. Second, households that paid for government bed nets must have done so in contravention of official government policy, since free distribution of bed nets has been an important element of government and donor policy at least since 2006, three years before our survey.9 Donors focus on free distribution, 9. Information on bed-net distribution rules over the earlier period, 2001–05, is sparse. However, it appears that even then, local authorities were asked to distribute free bed nets to poor and vulnerable households. (World Bank (2006). “Project Information Document (PID) Appraisal Stage”. Report No.: AB2224, Malaria Booster Project, 4.) Keefer and Khemani 239 especially aimed at vulnerable populations of pregnant women and young chil- dren, and not on pricing strategies. Targets for USAID contractors, for example, never refer to cost recovery, emphasizing instead aggressive goals for the cover- age of vulnerable populations. It is nevertheless possible that households that paid for bed nets prior to 2006 might have been more likely to keep them after being exposed to radio program- ming, confounding our interpretation of the effect of community radio access. Our multipurpose survey has no data on the timing of bed-net acquisition. Nevertheless, for numerous reasons, this is an unlikely explanation for the results we present below. From 2006 onwards, the government massively increased the supply of (free) bed nets. Bed nets acquired prior to 2006, even if households had retained them, would have been a small fraction of the total. In addition, bed nets more than three years old are in significantly worse condition, and the WHO recommends that they not be used.10 Households that are more exposed to radio programming promoting the proper use of bed nets should therefore be less likely to keep (and report) older bed nets.11 Moreover, if households exposed to radio were persuad- ed to retain bed nets longer, they should also report more total bed nets. However, they do not. Finally, the emphasis on distributing treated bed nets in- creased through the 2000s. Untreated bed nets are therefore more likely to have been purchased prior to 2006 than subsequently. However, our results only strengthen if we focus solely on treated bed nets or on households that reported only possessing treated bed nets. The third key characteristic of bed-net distribution in Benin is that, regardless of official pricing policies, local health staff enjoy significant discretion regarding the prices they charge. On the one hand, employees of the village health centers are permitted to charge fees for some health services under national guidelines, as long as they use those fees to finance the services they provide. The guidelines prohibit fees in certain cases, such as for indigent households and, as indicated above, for goods classified as “essential,” such as bed nets. On the other hand, though, the government has little capacity to monitor local adherence to these guidelines.12 Donors monitor the distribution of bed nets from receipt at the port of Cotonou down to the department. However, once bed nets reach the next level of government, the commune, the distribution of bed nets is managed by 10. The only two investigations of the condition of three-year-old bed nets (World Health Organization 2008, 2009) indicate that they had numerous holes and that the insecticide was insufficiently effective. 11. There is sparse evidence on how long households actually keep bed nets. However, the WHO also sought to evaluate bed nets that had been used for five years; it had difficulty finding households with five-year-old bed nets. In general, households are known to be less likely to use bed nets that are older than twelve months or not in good condition; bed nets that are not used are more likely to be repurposed or thrown out and to not be reported. 12. USAID (2011, 10 and 14) emphasizes the weakness of the country’s health infrastructure and lack of knowledge of government pricing policies. 240 THE WORLD BANK ECONOMIC REVIEW the Comite ´ s de Gestion des Centres de Sante ´ (COGECS), the Health Center Management Committees. They allocate bed nets to villages; the distribution to households is then undertaken as part of regular government health service deliv- ery by health workers at village-level health centers.13 Embedded as they are in the community, village health officials have local knowledge that allows them to assess whether households have a higher willing- ness to pay for bed nets. They also have an incentive to charge for bed nets since they may use revenues, in collaboration with community committees, to provide more services to citizens in their areas; they may also take advantage of weak fi- nancial controls to divert the fees to personal uses. The fourth important characteristic of bed net markets is that the quantity of bed nets available to local communities is a high level decision that would fall within the domain of politicians. In contrast to their control over bed-net pricing, village officials cannot unilaterally change the quantity of bed nets allo- cated to their health centers. Commune-level representatives of the central offices of the National Malaria Control Program fix the quantities village officials can distribute. Local officials can secure larger allocations for their health centers only if they persuade commune officials to respond to local information about household demand and reallocate bed nets across villages, or if they persuade commune officials to lobby national government officials to increase commune allotments at the expense of other communes. The survey instrument allowed us to elicit detailed information from 4,200 households about the mode of acquisition of each bed net currently in the posses- sion of the household—whether it was acquired from the government, private market sellers, charitable organizations or international NGO, friends or family, and whether households paid for it.14 Specifically, we use answers to the ques- tion, “How did you get this bed net?” (asked of each bed net the household lists as owning). The options read out to the household were: “1. I bought it from private sellers. 2. I bought it from the government. 3. It is a government donation. 4. It is a donation from a relative or a friend. 5. Other.” Our variable “paid for government bed nets” is the number of bed nets for which the house- hold response is category 2. 13. For example, in its 2009 BASICS contract (Basic Support for Institutionalizing Child Survival), USAID expected the contractor to ensure that bed net supply logistics were in place with partners down to the departmental level only (not the commune or community level) (http://www.fightingmalaria.gov/ funding/contracts/basics_task-order.pdf, 8). 14. While the survey has many questions on mosquito bed nets, it was not primarily a survey about malaria and bed nets. It contained questions, as well, about education policy, household knowledge, and political behavior. We do not, therefore, have reliable data on the actual usage of bed nets by households (who sleeps under the bed nets, how regularly, etc.), since this information is particularly difficult to collect through multi-purpose surveys. In contrast, the focused micro-empirical studies on bed-net usage (Cohen and Dupas 2010; Dupas 2009) relied on enumerators visiting households in the evenings to directly observe the practice of sleeping under bed nets. Our information on whether bed nets are treated with long-lasting insecticide is also noisy, relying on households to list which were treated and which were not. Keefer and Khemani 241 Both the success of these distribution programs in reaching far-flung house- holds and the ability of local officials to charge for bed nets are evident in our sample. Of the sample households, spread throughout rural and sparsely popu- lated northern Benin, 86 percent report having at least one bed net of any kind; 69 percent report having received at least one free bed net from the government; and 16 percent of households purchased at least one government bed net.15 The median household in our sample reports owning two bed nets in total. The average share of free bed nets in bed net–owning households is 66 percent, ranging from a minimum of zero (20 percent of the bed net–owning households) to a maximum of 100 percent (52 percent). The table in the appendix provides summary statistics for all the variables used in the analysis below. In principle, we should prefer to measure the demand effects of radio access using the actual prices that households pay. However, price effects are unlikely to be accurately observed, and price data is less informative, in a policy context where the majority of transactions are nonmarket and at zero price, as in the market for bed nets in Benin. The availability of data on prices is conditional upon paying for a net in the first place. Given the policy context of large-scale public distribution of free bed nets, there are few observations for which we observe price data and even fewer for prices paid to the government in the pur- chase of bed nets. Not surprisingly, we find no significant effect of radio on prices that households report paying for bed nets acquired through purchase.16 Exposure to Anti-malaria Media Campaigns in Benin The analysis that follows uses the fact that dispersed radio stations broadcast similar programming. We document this fact in this section: programming on the value of public health and the use of anti-mosquito bed nets is carried by all com- munity radio stations in northern Benin. We document, as well, that households listen to these stations. At the same time, we are able to show that radio stations could broadcast information that specifically encourages households to hold government officials accountable for the distribution of free bed nets, but they do not. Instead, they broadcast programming designed to increase household demand for bed nets. The most popular media in northern Benin are twenty-one small community radio stations scattered throughout our study area.17 We surveyed the managers of all sixty-eight radio stations operating in Benin as of March and April 2009, 15. 25 percent of the households report purchasing a bed net from private market sources. 16. These results are available upon request. We chose not to report them in the paper because price is not a useful statistic in a policy context where a market is dominated by free public distribution. 17. The southern region of Benin is more urbanized and densely settled. Most areas have access to multiple national and commercial radio stations, and there is little within-commune variation in access across villages. Since our identification strategy exploits within-commune differences across villages, as detailed in the next section, our sample is drawn from thirty-two northern communes of Benin. Only the northern commune of Parakou is excluded; it is the second largest city of Benin, after the capital region of Cotonou in the south. 242 THE WORLD BANK ECONOMIC REVIEW eliciting extensive information about radio characteristics, including ownership, licensing, and programming. According to the broadcast licensing policy, only private, noncommercial stations that satisfy certain criteria demonstrating that they are organized to serve the public interest of local communities can operate as “community radio” (Fraser and Restrepo-Estrada 2002; Buckley et al. 2008). Our survey interviews verify that community radio stations are owned and managed by local community organizations; established with the avowed objective of promoting public-interest and educational programming; and sup- ported and sponsored by foreign donors, government ministries, and charitable organizations. Because of the paucity of commercial advertising, community radio stations rely on support from donors and government to broadcast educational program- ming. Fourteen of the twenty-one community radio stations in the study area report depending on funds from foreign donors and three on funds from nongov- ernmental organizations. Regardless of funding source, respondents in each of the stations describe the provision of information on health and education as a “very important” objective. Donor representatives and station managers confirm that they support programming that broadcasts general information about the value of bed nets and sponsor announcements regarding the availability of bed nets and the timing of distribution. The Roll Back Malaria program, specifically, has used community radio to broadcast information about malaria control poli- cies and advocacy for public health practices. The remarkable feature of these radio markets in Africa, of which Benin is typical, is that people actually listen to them even though they carry more educa- tional, public-interest programming. These radio stations are the main type of media accessible to households living in relatively poor and remote regions. Of the 3,828 households in our sample that report listening to some radio, 64 percent report listening to at least one community radio station and 45 percent report listening to national public radio, compared to only 176 that listen to private commercial radio.18 These preferences are consistent with availability (few households have access to private commercial radio) and with claims of media experts that the programming of national broadcasters conforms less well to the tastes and linguistic preferences of poor households in rural Africa com- pared to local radio stations (Buckley et al. 2008). Our survey of radio stations indicates that they do attempt to match musical and cultural programming to their local area. The measure of radio access is the number of community radio signals accessi- ble at the village level. Enumerators first asked village-level key informants to list which radio stations they were able to receive. The enumerators also used their 18. “Listenership” data was gathered by simply asking households to name the station they listen to and then having our investigating team post-code the response for type of radio. We did not directly prompt the household to categorize what type of radio they like listening to. Even among these 176 that report listening to some private commercial radio, only sixty-seven report listening only to commercial radio; the rest listen as well to at least one other public or community radio station. Keefer and Khemani 243 own transistor radios to verify and expand on this, if they received additional signals. The radio codes from the village survey were then matched with the radio survey data to establish the types of radio available to the village. The next section discusses how the sample of villages was picked to exploit a natural ex- periment in northern Benin that results in significant intra-commune variation in this measure of radio access. Here, we document how this measure of radio access is closely correlated with exposure to programming on health and educa- tion issues. Table 1 shows the extensive reach of community radio in northern Benin. Since community radio stations have low signal strength, no single station is ac- cessible to more than a few villages in our sample. However, because the number of community broadcasters is large, 93 percent of the sampled villages receive at least one community signal. The number of community radios to which villages have access ranges from zero to seven stations, averaging 2.4 and with a standard deviation of 1.5. Only 23 percent of villages receive private commercial radio signals. Although private radio has greater signal strength, it is concentrated in the southern com- munes of Benin, outside the study area. In contrast, government-owned public radio has several relay transmitters that allow it to cover 92 percent of the villag- es. Nearly all villages (175) have access to both a community radio station and national public radio. About 34 percent of villages have access to religious radio stations. However, the information from religious stations is less reliable (for example, in the case of signal strength).19 We also have considerable information about radio programming choices dis- played in table 1. Station managers report broadcasting a large number of health programs (138) over the three months prior to the survey, on average, more than three times as many as public or private commercial radio. Hence, access to a higher number of community radio stations is likely to be associated with expo- sure to more information about health issues. On the one hand, households that are interested in health programs are more likely to find one that suits their taste or language preference. On the other, radio listenership is likely to be higher, since households are more likely to find a station that matches their tastes; even respondents who are not particularly interested in health programs will be more likely to nevertheless hear them if they are more likely to listen to the radio. As results in table 2 indicate, village exposure to health programming is signif- icantly and substantially greater with the reception of more community radio sta- tions. For each village in the sample, we aggregated the number of health programs that station managers reported broadcasting across all radio stations heard in the village. The first column shows that additional community radio sta- tions received by a village, over and above the average of the commune, are asso- ciated with receiving 129 additional health programs. The second two columns 19. However, only thirteen respondents identified a religious station as their most preferred broadcaster. 244 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . Access to Radio Stations in North Benin No. of Average no. of stations No. of sampled health programs accessible to villages covered Average signal broadcast (over the region (total ¼ 210) strength(watts) the past 3 months) Government-owned 2 200 2667 40 public radio Private ¼ noncommercial 21 195 231 138 or community radio Private commercial 10 49 476 27 Radio Religious radio 2 71 NA 36 Source: Survey data collected by authors, described in text. T A B L E 2 . Community Radio Effects on Village Exposure to Health Programs Including total Dependent variable: number of health programs Community number of other Disaggregating broadcast in the three months prior to the survey radios radios by radio type Community radio received by village (# stations) 129.14*** 126.27*** 124.66*** (43.00) (47.21) (39.67) All radio received by village (# stations) 2.77 19.26 (17.94) (0.00) Public radio received by village (# stations) 2 51.28 (52.52) Commercial radio received by village (# stations) 68.83* (34.61) Religious radio received by village (# stations) 2 63.98 (63.10) Observations 210 210 210 R2 0.80 0.80 0.81 Notes: OLS with commune fixed effects. Standard errors in parentheses, clustered at commune level. *** Significant at 1% level; ** significant at 5% level; * significant at 10% level. Dependent variable is the number of health programs received by a village over the three months prior to the survey team’s visit, as determined by information supplied by the managers of the radio stations whose signals reach the village. Source: Author analysis based on data described in text. compare the impact of having an additional community radio to that of an addi- tional radio of any other type, confirming again that community radio is a partic- ularly special media in this context, in terms of communicating information about health issues. We also find indications in our sample that other forms of mass media, other than radio, are unlikely to be sources of information about health. Especially in the study area in northern Benin, radio is essentially the only media to which Keefer and Khemani 245 most citizens have access. Twenty-five percent of respondents to the nationally representative Afrobarometer (2005) survey of Benin report that their household owns a television. In the sample here, of 4,200 households from northern Benin, only 8 percent own a television, but 84 percent own radios; television ownership in this sample is uncorrelated with access to good radio signals. Newspapers have little penetration in Benin as a whole. The largest has a circulation of seven thousand, and that almost entirely in the capital city of Cotonou (from author in- terviews of journalists in Cotonou). Interviews with radio station staff indicate, however, that their news reports about Benin are often prepared based on reports from the Cotonou newspapers. Community radio stations could, in principle, broadcast programs intended to raise citizen awareness of government performance and frame messages to influ- ence citizens to demand more benefits from government. Maskin and Tirole (2004) argue that media can increase government incentives to deliver benefits to more informed citizens by making them more demanding. In our survey inter- views with station managers, however, no station indicated that it monitors gov- ernment performance in implementing anti-malaria programs. This is unlikely to be because of significant government-imposed restrictions on media freedom. According to ratings of Reporters Without Borders from 2009, the year of our survey, Benin ranks 72nd in the world, out of 175, just behind Brazil (number 71), Botswana, Malawi, Tanzania, Liberia, and Togo but substantially ahead of most other African countries. Instead, media broadcasts about the specifics of government performance are less likely to the extent that those specifics are not politically salient, as in the case of bed-net distribution. Both the political context of Benin and the techni- calities of bed-net distribution reduce political salience. With respect to political context, when politicians have weak incentives to be responsive to well-informed constituents, as in more clientelist democracies, Khemani (2007) suggests that media markets are less likely to cover information about how to improve public goods performance. Clientelist political settings, such as Benin’s, are precisely those in which politicians have weak incentives to care about local-level pricing of some bed nets by health officials or excess demand for free bed nets among well-informed constituents Political parties are weak and political competition is personalized. Politicians cannot easily make credible commitments to voters outside of their private, clientelist networks (Keefer and Vlaicu 2008). As a consequence, politicians cannot easily coordinate among themselves to transfer benefits from uninformed households represented by one politician to informed households represented by another. The number of parties has fluctuated between twenty-seven and 129 since competitive elections were first held in 1991. Over one hundred tried to compete in the 2007 elections (Economist Intelligence Unit). None of the four major parties emerging from the 2007 elections had a programmatic identity—a basis for making credible com- mitments to large groups of citizens. Party fractionalization is correspondingly high: in 2006 the probability that two randomly selected legislators do not 246 THE WORLD BANK ECONOMIC REVIEW belong to the same party was 80 percent in Benin, compared to 52 percent in Ghana (Beck et al. 2001). As a consequence, politicians are likely to gain little by shifting the distribution of free bed nets from uninformed to informed households. Politician—and station manager—disinterest in the specifics of bed-net distri- bution is also related to the difficulties of assessing whether, in fact, government performance is deficient. As we observe in the discussion of our results at the end of the paper, the welfare consequences are ambiguous of responding to increased household demand for bed nets by charging for them rather than shifting sup- plies from low- to high-demand households. To work out those consequences re- quires information about household bed-net coverage, who pays for bed nets, and how bed net fees are used. However, at least currently, donors and the health ministry have not collected such data, making it difficult for station managers to broadcast programming meant to increase government accountability for its bed-net distribution performance. We have shown that community radio stations carry significantly more health programming, and that households listen to them. Since people listen to one radio station at a time, however, why, should access to a larger number of com- munity radio stations increase their exposure to health programming? Moreover, if people are indifferent to health programming and switch stations to find other, more preferred programming, the availability of multiple stations might imply that they are less likely to listen to health programming. In fact, our survey data suggest that the opposite is true. Households report that they listen to radio pre- cisely to hear news and information about public affairs, so the availability of more stations gives them greater opportunities to find a station that, at any given time, is broadcasting this programming. The survey asked respondents to indicate how important different types of radio programs are to them (music, sports, news and public affairs, etc.). Perhaps not surprisingly, since radio is by and large their only source of information about the outside world, households attach the highest level of importance to news and public affairs (“very important/I listen a lot every day”): 56 and 48 percent of respondents cite these as very important, compared to 40 percent and 25 percent, respectively, who say the same for music and sports. Furthermore, re- spondents who indicate that they prefer news and public affairs programming are also significantly more likely to listen to community radio stations. If people strongly prefer a particular type of programming, if different stations (e.g., community radio stations) are more likely to broadcast that programming, and if the timing of public interest programming varies across stations, then more stations give more households the opportunity to listen to the public- interest programs that they prefer. At any given time, the likelihood that at least one station is broadcasting the type of public-interest programming that a house- hold in the village finds valuable is higher in villages with access to more commu- nity radio stations. Keefer and Khemani 247 I I . I D E N T I F I CAT I O N : T H E N AT U R A L E X P E R I M E N T IN NORTHERN BENIN Unobserved factors could influence both media access and outcomes of interest, such as whether households have received bed nets from the government and whether they have paid for those nets or received them for free. Northern Benin provides a natural experiment in village-level radio access that allows us to use a novel identification strategy to estimate media effects. As table 1 illustrates, al- though 93 percent of the sampled villages can access signals from at least one local community radio, variation in the number of community radios to which they have access is large, ranging from zero to seven stations. The average number is 2.4; the standard deviation is 1.5, and exposure to a greater number of signals should be correlated with greater exposure to programming on the value of bed nets. In contrast, villages are homogeneous in their access to the relay sta- tions of the national public radio station: all but twenty villages in the sample receive strong signals from one or more relays. Community radio proliferated in this region when donors supported multiple small stations, spread out across many communes, in order to extend the reach of public interest programming to remote, rural areas. Donors chose to support such small radio stations in order to communicate public interest messages more effectively, with the expectation that community-based station managers could “translate” messages more persuasively, taking into account local cultures and tastes. This history gave rise to substantial village-level variation in radio access that is driven by accidental features that determine where signals from many, small and dispersed radio stations happen to end their reach and that are exoge- nous to village characteristics. Our preferred specification controls for commune fixed effects, identifying radio effects from intra-commune village-level variation, since commune charac- teristics that are associated with bed net distribution can be endogenous to radio placement. For example, donors needed local partners to establish community radio stations. Idiosyncratic conditions affected which communes yielded such partners and, therefore, the communes in which radio stations were estab- lished.20 We control for these unobserved characteristics using commune fixed effects. Nevertheless, reliance on within-commune variation in access to community radio could fail to identify the effects of radio access if the market consisted of only a few community radio stations, to which all villages in communes had 20. In a detailed case study of one of these community radios, Radio Tanguieta located in Tanguieta commune in Atacora department, Gra ¨ tz (1999) reports on the political contestation over the locally elected committee to oversee the radio and manage its funds. Control of the community radio rested in the hands of local politicians and Gra¨ tz indicates concerns about embezzlement and over-spending on salaries of numerous presenters representing different ethnic groups. Community radio projects are subject, then, to the same political risks of local elite capture and clientelist and ethnic politics as community-based projects in other sectors. 248 THE WORLD BANK ECONOMIC REVIEW similar access. This is not the case: within-commune variation in access is sub- stantial. Unobserved village characteristics that affect bed-net acquisition might also influence within-commune variation in village access to community radio sta- tions. For several reasons, however, all related to the number and dispersion of small radio stations in northern Benin, this is unlikely to be the case. That is, it is implausible that any particular unobserved village characteristics could lead to systematic bias in observed village-level variation in radio access, after account- ing for commune fixed effects. First, we use a sampling strategy where villages are closely matched in location and access to infrastructure and still exhibit post-sampling variation in radio access. The 210 villages in the study were selected from maps showing the loca- tion of all villages relative to radio towers and major road networks. We chose villages that were equidistant from towers and roads, as well as from commune boundaries, all of which are factors that could yield spurious correlations between radio access and bed-net distribution.21 As a consequence, it is not the case that variation is driven by the clustering of some villages in a commune near a neighboring commune with a large number of community radio stations, nor that variation is driven by clustering of some villages around community radios located within the commune. Second, this variation exists because the community radio stations are small, dispersed, and numerous: villages that are close to each other and not separated by any remarkable topographical features can nevertheless exhibit significant dif- ferences in access. Short distances and small changes in elevation are enough to degrade the signal received by some of them. These differences are unlikely to affect bed-net distribution except through their influence on radio access. In con- trast, identification in previous research on media effects in developing countries has relied on more significant geological features that are sufficient to obstruct broadcasts from one or a few large stations (Stro ¨ mberg 2004; Olken 2009; Yanagizawa 2009).22 21. Unfortunately, no data are available on the precise GPS locations of stations and villages to directly control for this in our analysis. We therefore relied on pictorial maps provided by our local consultants to identify neighboring villages that were located equally distant from radio towers. As we discuss below, we are able to verify post-survey that our sampled villages exhibit no correlation between radio access and observable characteristics of location that could independently impact bed-net distribution. 22. US research on media effects has taken advantage of richer data to reduce reliance on geographic instruments altogether. Gentzkow et al. (2011), for example, have many decades of fine-grained data on newspaper markets in the US and can control for trends before newspaper entry and exit in their variable of interest (voter turnout). In contrast to that work, ours does not benefit from time series data in either the dependent or independent variables of interest. However, ours is the first research to consider media effects on public services in a young democracy and to document the presence of behavioral but not accountability effects of media. We are able to uncover these effects precisely because we focus on a developing-country setting where media effects are likely to be different than in a richer country but where comprehensive panel data are less common. Keefer and Khemani 249 Third, most of the intra-commune variation in access to radio signals is ex- plained by differences in access to out-of-commune radio stations. The average household has access to .63 within-commune community radio stations, with a standard deviation of .51. It has access to 1.78 out-of-commune stations, with a standard deviation of 1.45. Exposure to signals from neighboring communes is particularly likely to be exogenous, since the location decisions of out-of- commune community broadcasters are unrelated to the characteristics of villages in neighboring communes. The fact that variation in access is driven significantly by out-of-commune radio stations does not attenuate the ability of our tests to detect whether govern- ment caters to more informed citizens. The health programming of out-of- commune stations should also boost demand for bed nets in neighboring communes, increasing the incentives of politicians in those communes to deliver bed nets. Out-of-commune stations do not inform them about the compliance of their local officials with centrally mandated rules regarding pricing, but house- holds’ personal observation of local official behavior makes this irrelevant. In sum, the fragmentation of the Benin radio market and a sampling procedure that exploits village-level variation in access to community radio stations allow us to include more homogeneous villages, with differential radio access, than has been possible in previous research. Most studies of media impact are based on variation across jurisdictions or households in their access to the same large broadcasting or publishing outlets. In contrast, the analysis here is based on within-commune variation in village access to small community radio stations. Unobserved differences between commune villages that can access out-of- commune radio broadcasts and those that cannot are likely to be insignificant compared to the differences between villages that can access centrally broadcast media and those that cannot. Observable village characteristics and the number of community radio sta- tions to which villages have access are largely uncorrelated. This further supports the argument that variation in village radio access is exogenous. These correla- tions, after controlling for commune fixed effects, are reported in table 3. Access to other types of radio may provide summary proxies of a variety of village socio-economic conditions that can drive spurious correlation between community radio and bed-net distribution. The number of community radio signals available to a village is uncorrelated with the number of signals received from private commercial, public, and religious radio. This supports our argu- ment that the effects of radio access on bed-net distribution are driven by particu- lar characteristics of the programming carried on community radio and not by other village characteristics that vary with general radio access. These results are also consistent with historical accounts that donors preferred to place communi- ty radios in more remote areas, contrary to the usual practice of radio entrepre- neurs. We show in the section below that radio effects on bed nets are driven by community radio rather than access to any other type of radio. 250 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Correlation of Village Characteristics with Community Radio Access Dependent variable: number of community radio Multivariate Bivariate stations received by village Coefficient P-value Coefficient P-value Private commercial stations received by village 0.108 0.59 0.156 0.52 Religious stations received by village 0.173 0.51 0.182 0.45 Public stations received by village 2 0.210 0.47 2 0.123 0.69 Village population (1,000s) 0.007 0.90 0.00005 0.29 Does village have a paved road? 2 0.119 0.63 0.327 0.26 # of potables water sources built in 2007 or 2008 2 0.003 0.93 2 0.010 0.74 Secondary school dummy 0.307 0.09 0.258 0.07 Number of functional private schools 2 0.481 0.02 2 0.159 0.32 Literacy center dummy 0.004 0.99 0.047 0.79 Health center or maternity dummy 2 0.064 0.71 0.116 0.29 Village chief has primary schooling 0.114 0.57 0.048 0.79 Village chief has secondary education 0.279 0.19 0.287 0.21 Distance of village to nearest urban center 2 0.008 0.08 2 0.008 0.09 Distance of village to nearest bus or train stop 2 0.005 0.15 2 0.006 0.25 Share of surveyed households that report income less 2 0.520 0.16 2 0.569 0.10 than 27,500 CFA Most common language in village and commune is 0.162 0.68 0.253 0.31 the same Probability that main language in any two 2 0.439 0.43 2 0.308 0.57 households in the village is the same Observations, R2 199, 0.80 208 þ , varies Notes: OLS with commune fixed effects. Standard errors in parentheses, clustered at commune level. The multivariate specification also controls for the fraction of village respondents that come from each of eight ethnic groups, the fraction that is Muslim and the fraction that is Catholic. None of these are significant, either in the multivariate nor bivariate specifications. Source: Author analysis based on data described in text. An important issue is whether radio access simply captures village remoteness. This is unlikely, since, given the fragmented radio market, the small size of com- munity radios, and the way in which villages were selected, our measure of com- munity radio access is not driven by remoteness but incidental features of terrain and geography. Moreover, our results are insensitive to controls for noncommu- nity radio stations, which are intended to reach populated, not remote areas. In addition, though, we control for survey indicators of village distance from urban centers and to a train or bus stop. There is no significant correlation between radio access and distance to the nearest bus or train stop. Distance to the nearest urban center is marginally significant, but this is because of a single outlying village that reports a distance that is 50 percent higher than the next most remote village in the sample. Once this outlier is dropped, there is no significant correla- tion of distance with radio access. The results of this study are robust to control- ling for distance and to estimating radio effects in sub-samples that omit outlying observations (in both remoteness and closeness to urban centers). Keefer and Khemani 251 Population may also reflect remoteness. The most recent census in Benin, from 2002, has information on village population for nearly all of the villages in the sample. Radio access is not correlated with village population, however. Poorer villages could have lower radio access and simultaneously be targeted with more free bed nets. There are no census or statistically representative data of incomes and poverty rates at the village-level in Benin. The survey offers one proxy for village income, which could be correlated with both radio and bed net access. We asked our sampled households to select from a set of different income brackets to which their household belonged. From a sample of twenty house- holds in each village, we calculated the village-level variable of the share of re- spondents that reported belonging to the lowest income bracket. This variable is insignificantly correlated with radio access in the multivariate regression reported in column 1 of table 3 but marginally significant in the bivariate specification re- ported in column 2. The results of this study are robust to controlling for this household-reported income variable. In addition, our results are robust across two subsamples, one in which all households belong to the lowest income bracket and one in which no households belong to this bracket. Other proxies of village wealth, such as availability of various facilities and infrastructure—private primary schools, secondary schools, paved roads, potable water sources, literacy center, and health center—are not systematically correlated with radio access.23 Among all of these, only the presence of a second- ary school is positively correlated in both specifications, but the magnitude of the correlation is small: the 17 percent of villages with secondary schools are exposed to fewer than 0.3 (0.20 standard deviations) additional community radio station signals. At the same time, the number of functional private schools is negatively correlated with radio access in the multivariate specification. Richer villages could have more educated village leaders who can influence bed-net dis- tribution; the survey gathered data on village chiefs’ schooling. There is no corre- lation of this with radio access. Village linguistic characteristics could be associated both with their inclusion in specific radio markets and independently with cultural practices that influence adoption of health technologies. Based on the household survey, it is possible to compare villages according to, first, the probability that any two of these house- holds would speak the same language at home and, second, an indicator variable for whether the most common language among the sample of twenty households in a village is different from the most common language among all of the house- holds surveyed in the commune in which that village resides. These village char- acteristics are uncorrelated with radio access. 23. Radio access could independently influence both public infrastructure and the presence of private market facilities. Keefer and Khemani (2011) show that radio access in Benin increases the demand for education and parental investment in children’s education. We discuss further in the sections below why the reduced form effect of radio on bed-net distribution and pricing can be attributable to the persuasive power of radio programming, rather than driven by other mechanisms such as radio’s impact on incomes and/or village infrastructure. 252 THE WORLD BANK ECONOMIC REVIEW Village ethnic and religious characteristics, similarly inferred from the household data, could also affect village access to community radio. None of the controls for eight ethnic groups and two religions (Catholic and Islam) are significant in either the multivariate or bivariate specifications. I I I . R E S U L T S : V I L L A G E - L E V E L VA R I A T I O N I N R A D I O A C C E S S AND BED-NET DISTRIBUTION The impact of radio on government responsiveness and household behavior is estimated with regressions specified according to (1): Bed netsijk ¼ b0 þ b1 Number of community radio jk þ X ijk B þ eij þ mk ; ð1Þ where the dependent variable is, in turn, the total number of bed nets (from all sources), the number of free bed nets provided by the government, and the number of bed nets purchased from the government. The units of observation are household i in village j and commune k; mk is a commune fixed effect. Given that the number of community radio signals accessible to a village is exog- enous to village characteristics once we control for commune fixed effects, we can identify b1, the reduced form effect of community radio access on bed nets, without including any further controls. Control variables are not likely to be exogenous; in- cluding them could induce bias in our coefficient of interest. The bivariate, commune fixed effects specification is therefore our preferred one. Nevertheless, in order to support our argument that the reduced form impact of radio is driven by the persuasive power of radio programming content, rather than through radio’s impact on other correlates of bed net acquisition, we provide additional results on a rich set of covariates X ijk on which the survey gathered data. The appendix contains a table with the summary statistics of all the variables used in the analysis. The first panel of table 4 reports the central result, from the preferred bivari- ate, fixed effects specification, that greater access to community radio signifi- cantly increases the number of bed nets that households report purchasing from the government. The average number of purchased government bed nets per household is .27. The regression results indicate that every additional community radio station increases the average number of purchased government bed nets per household by .07 bed nets, or approximately 25 percent of the mean. In contrast, radio has a negative effect on the number of free bed nets received, though the estimated coefficient is only marginally significant. There is no signifi- cant effect on the total number of bed nets owned by the household. That is, greater radio access increases the likelihood that households pay for the bed nets acquired from the government, rather than receiving them free (as per official policy).24 The last two columns report the results for the share of free and 24. There is also no separate radio effect on bed nets purchased from private market sellers (which is included, of course, in the total number of bed nets). T A B L E 4 . Impact of Community Radio on Household Bed Net Acquisition Dependent variable: Total bed nets Free gov’t. bed nets Paid gov’t. bed nets Share of free govt. nets Share of paid govt. nets Number of community radio stations received 0.051 (0.040) 2 0.041 (0.027) 0.074*** (0.021) 2 0.023* (0.010) 0.023*** (0.007) by village Observations, R2 4200, 0.06 4200, 0.05 4200, 0.10 3606, 0.06 3606, 0.10 Number of community radio stations received 0.041 (0.040) 2 0.045* (0.027) 0.073*** (0.022) 2 0.022** (0.010) 0.024*** (0.007) by village Number of public radio stations received by 2 0.100 (0.106) 2 0.093 (0.065) 2 0.006 (0.043) 2 0.002 (0.029) 0.006 (0.017) village Number of commercial radio stations received 0.025 (0.056) 0.007 (0.036) 2 0.021 (0.030) 2 0.004 (0.015) 2 0.006 (0.011) by village Number of religious radio stations received by 0.217** (0.088) 0.070 (0.069) 0.079 (0.051) 2 0.031 (0.028) 0.018 (0.020) village Observations, R2 4200, 0.060 4200, 0.05 4200, 0.10 3607, 0.05 3607, 0.09 Notes: Each panel of regressions is estimated at the household level using OLS with commune fixed effects. Standard errors in parentheses, clustered at village level. *** Significant at 1% level; ** significant at 5% level; * significant at 10% level. Source: Author analysis based on data described in text. Keefer and Khemani 253 254 THE WORLD BANK ECONOMIC REVIEW purchased government bed nets in the total number of bed nets owned by a household. These show more explicitly that radio access changes how house- holds acquire bed nets: they pay the government for bed nets that they could have received for free. The next panel in table 4 reports the results of adding other types of radio to the specification. The results for the impact of community radio are unchanged. No other radio type is significantly correlated with paying for government- provided bed nets. This gives further confidence that the central results in the first panel are driven by particular characteristics of community radio and not by other mechanisms that could be correlated with radio or media access generally.25 Table 4A reports estimates after controlling for a host of household and village characteristics that could be correlated with radio access and bed net dis- tribution. The results for radio’s impact remain intact and are robust to alternate specifications, including different sets of controls. Household demographic com- position can be important since bed nets are supposed to be targeted to pregnant women and young children, who are particularly vulnerable to complications from malaria. The number of bed nets owned by a household, both free and pur- chased, is significantly correlated with the total number of household residents, adults and children. Controlling for the total numbers of children and adults, it is not correlated with the number of children reported as younger than five. However, as shown in column 4, the share of free government bed nets in the total owned by the household is significantly greater when there are more young chil- dren, suggestive evidence that the targeting policy may be working as intended. Household income and education can also be important. General policy guidelines target benefits to indigent households. Consistent with this, we find that households reporting their income to be in the lowest bracket also report fewer bed nets in total but a significantly larger share of free bed nets from the government. Households in homes with more expensive cement floors report more bed nets in total: the share of free bed nets is significantly lower in these households, but the number and share of paid bed nets is significantly higher. Education can serve as a proxy for income or poverty but can also be indepen- dently correlated with knowledge of and demand for health goods. Households where the respondents have greater education, especially secondary education, report more total bed nets, both free and purchased. Controlling for the total number of bed nets, these households are also more likely to purchase bed nets from governments, as shown in the last column. The size of the radio effects is comparable to these other, important correlates of bed-net acquisition. A one standard deviation increase in community radio 25. An additional result here is that religious radio is correlated with more bed nets in total and marginally correlated with more free bed nets from government. We cannot strictly interpret this correlation because we do not have information on the possibly endogenous location of religious radio. It is unlikely to be driven by the persuasive power of religious radio to take-up bed nets, however, because households don’t report listening to radio that we document as religious, and religious radio station managers report carrying few health programs. T A B L E 4 A . Community Radio Access and Household Acquisition of Bed Nets—Multivariate Free gov’t. bed Share of free govt. Share of paid govt. Dependent variables Total bed nets nets Paid gov’t. bed nets nets nets Number of community radio stations received 0.038 (0.039) 2 0.036 (0.027) 0.066*** (0.023) 2 0.020* (0.010) 0.019** (0.008) by village Secondary education—Respondent 0.277*** (0.070) 0.113* (0.063) 0.129** (0.053) 2 0.030 (0.021) 0.043** (0.020) Higher education—Respondent 0.523** (0.236) 0.004 (0.226) 0.079 (0.140) 2 0.140** (0.055) 0.025 (0.047) Number of adult household members listed 0.250*** (0.025) 0.157*** (0.020) 0.0320* (0.017) 0.0045 (0.006) 2 0.006 (0.004) Number of children listed 0.162*** (0.018) 0.109*** (0.015) 0.013 (0.009) 0.006 (0.004) 2 0.002 (0.003) Number of children listed aged 0– 5 2 0.025 (0.031) 0.023 (0.024) 2 0.001 (0.017) 0.028*** (0.008) 2 0.008 (0.006) Respondent single 2 0.144 (0.093) 2 0.133 (0.010) 0.095* (0.049) 2 0.103** (0.043) 0.122*** (0.035) Income is below 27.5 K CFA 2 0.122** (0.047) 2 0.012 (0.041) 2 0.103*** (0.023) 0.045*** (0.017) 2 0.041*** (0.012) Catholic religion (Respondent) 0.113 (0.071) 0.037 (0.056) 0.112** (0.047) 2 0.025 (0.023) 0.041* (0.021) Floor—cement 0.120** (0.054) 0.007 (0.052) 0.066* (0.038) 2 0.046** (0.019) 0.033** (0.016) Rooms—five or more 0.367*** (0.070) 0.140** (0.066) 0.061 (0.043) 2 0.027 (0.020) 2 0.006 (0.015) Owns a TV 0.502*** (0.118) 0.147 (0.094) 0.067 (0.066) 2 0.061** (0.028) 2 0.002 (0.021) Distance to closest bus/train Stop 0.006*** (0.001) 0.00490*** (0.001) 0.001 (0.001) 0.0003 (0.0005) 0.0001 (0.0003) Village chief with secondaryeducation 0.187** (0.081) 0.152** (0.061) 0.089* (0.048) 0.026 (0.026) 0.033* (0.019) # of potable water sources built in 2007 2 0.002 (0.011) 2 0.00603 (0.009) 0.015** (0.008) 2 0.001 (0.003) 0.0059** (0.003) or 2008 Observations 3,663 3,663 3,663 3,131 3,131 R-squared 0.29 0.19 0.13 0.10 0.12 Notes: Each panel of regressions is estimated at the household level using OLS with commune fixed effects. Standard errors in parentheses, clustered at village level. *** Significant at 1% level; ** significant at 5% level; * significant at 10% level. The following coefficients were never or rarely significant and are not reported: gender and age of household respondent; respondent elementary education; all ethnic variables (respondent was Adja, Fon, Bariba, Dendi, Yoa/Lokpa, Peulh, Goua/Otamari, Yoruba); number of household members over sixty; number of children under five; gender/age of household head; whether respondent is polygamous; respondent is Muslim; has brick walls; owns a mobile phone; village population and distance to nearest urban center; # of functional private schools; village chief has primary education; paved road in village; presence of secondary school, literacy center, or health/maternity Keefer and Khemani center; most common language in village is most common in commune; village linguistic fractionalization. Source: Author analysis based on data described in text. 255 256 THE WORLD BANK ECONOMIC REVIEW access (1.5 stations) is associated with .10 additional paid government bed nets (column three). This effect is comparable to that of secondary education: house- holds where the respondent had completed secondary education had .13 more bed nets than households with only primary education. It is also comparable to income effects. Households with income reported in the lowest bracket had .103 fewer purchased bed nets than other households. Ownership of mobile phones and television is significantly correlated with owning more bed nets (column 1, phone results not reported) but particularly so from private market sellers. However, neither television nor mobile phone own- ership is significantly correlated with receiving more bed nets from the govern- ment, either free or purchased.26 On the contrary, television ownership is negatively correlated with free government bed nets as a share of all household bed nets. We also test for correlations with other village-level characteristics reported in table 3 previously. Where village chiefs are more educated, households report owning more bed nets, both free and purchased from the government. Distance from a train or bus stop is associated with more bed net ownership and with more free bed nets but not with more bed nets purchased from the government nor with a higher share of government bed nets generally, whether purchased or free. Households in villages with more potable water sources appear to have pur- chased more bed nets from the government. Other village characteristics are not systematically or significantly correlated with household bed net acquisition. Given these effects, a natural question is whether radio access undermines the goals of bed net policy. Chief among these goals is to ensure bed-net provision among the most vulnerable households by making sure that poorer households and households with young children receive free bed nets. Radio access under- mines these goals if local health officials are more likely to charge families with young children or poorer households for bed nets when they have greater access to radio. Table 5 reports tests of this hypothesis, interacting the radio access vari- able with the number of young children and with the indicator variable for whether a household belongs to the lowest income bracket. If local officials’ pricing actions undermine policy, then these households would be significantly less likely to receive free bed nets when they have greater radio access: the interaction terms would be significant and negative. The esti- mates reject the hypothesis: the coefficients on the interaction terms with the number or share of free bed nets, though negative, are statistically indistinguish- able from zero. Two estimates provide support for the contrary hypothesis that government officials respond to informed households in a manner consistent 26. The regressions also control for a number of dichotomous ethnicity and religion variables, in case particular ethnic or religious groups are better mobilized to take advantage of government services or, in contrast, more likely to be excluded from it. Controls also include a variety of other measures of household income (quality of housing construction), the marital status of the household head (including single or polygamous), and the gender and age of both the respondent and household head. None of these are systematically correlated with household ownership of the different kinds of bed nets. T A B L E 5 . Radio’s Effects on Targeting Bed Nets to Vulnerable Households Free gov’t. bed Paid gov’t. bed Share of free govt. Share of paid govt. Dependent variable: Total bed nets nets nets nets nets No. of community radio stations received by 0.050 (0.042) 2 0.038 (0.032) 0.083*** (0.026) 2 0.029** (0.012) 0.029*** (0.01) village Community radio X Number of children 0.002 (0.016) 0.003 (0.011) 2 0.007 (0.012) 0.004 (0.004) 2 0.007** (0.003) under 5 Community radio X Self-reported income in 2 0.042 (0.031) 2 0.003 (0.026) 2 0.030* (0.018) 0.014 (0.012) 2 0.008 (0.003) lowest bracket Number of children under 5 2 0.032 (0.049) 0.014 (0.033) 0.015 (0.032) 0.018 (0.013) 0.009 (0.009) Indicator for self-reported income in lowest 2 0.023 (0.095) 2 0.005 (0.077) 2 0.032 (0.044) 0.012 (0.031) 2 0.022 (0.022) bracket Observations 3663 3663 3663 3131 3131 R2 0.29 0.19 0.12 0.10 0.13 Notes: Each regression is estimated at the household level, using OLS with commune fixed effects, and all the controls included in table 4A specifications (coefficients not reported). Standard errors in parentheses, clustered at village level. *** Significant at 1% level; ** significant at 5% level; * significant at 10% level. Significance of the linear terms in each interaction is calculated assuming the other linear term in the interaction is zero. Source: Author analysis based on data described in text. Keefer and Khemani 257 258 THE WORLD BANK ECONOMIC REVIEW with government policy. The interaction with young children is significantly negatively correlated with the share of paid bed nets and the interaction with the lowest income bracket is significantly negatively correlated with the number of paid bed nets. That is, greater radio access significantly reduces the share of paid government bed nets reported by households with more young children, and reduces the number of paid bed nets reported by the poorest households. This is suggestive evidence that the radio-induced diversion of free bed nets to the paid market by local officials is likely to be concentrated among less vulnerable households. I V. R O B U S T N E S S The results in tables 4 and 4A are robust to a large number of other specifica- tions. Table 6 summarizes robustness tests taking the multivariate specification in table 4A as the base, though the bivariate results are equally robust. First, the estimates in tables 4 and 4A are based on household observations. They are clustered and therefore not assumed to have independently distributed errors at the village level. However, since bed-net distribution could be focused on villages, and radio access varies by village, the village could be the more appro- priate unit of observation. As the first row of table 6 reports, the results are robust to estimating radio’s effects by aggregating the household variables up to the village level. It is also possible that the effects of radio access should be properly viewed as influencing whether households acquire any bed nets at all. Results are unchanged after substituting dichotomous variables for the number of bed nets, as in table 4A, and using a logit specification with commune fixed effects (results not reported). A key part of public health policy in Benin, as in other poor countries, is out- reach by health workers. We check that our results are robust to including out- reach visits by health workers, which could be independently correlated with radio access or with underlying household demand.27 The survey asked respon- dents if they had received a visit by a health worker to discuss malaria preven- tion. These visits are common: 41 percent of respondents report that a health worker called on their household. Health worker visits are significantly correlat- ed with the total number of bed nets that households report, and with the number and share of free bed nets, but not with the number or share of paid bed nets. This is consistent with the arguments advanced here: where household demand is weak (for example, when household acquisition of bed nets depends on a visit by health workers) households are less likely to pay for bed nets. However, as the second row of table 6 reports, estimates of the effects of radio access on bed-net acquisition are entirely robust to controls for health worker visits. 27. We also directly test with our data whether village radio access has an effect on outreach visits to households by health workers and find none. T A B L E 6 . Robustness Checks: Coefficient on Number and Share of Community Radios with Different Specifications/Samples Free gov’t. bed Paid gov’t. bed Share of free govt. Share of paid govt. Dependent variable: Total bed nets nets nets nets nets Specification change: Village-level estimates .038 (.044) 2 .043 (0.035) .052** (0.025) 2 0.024* (0.014) .013* (0.10) Add control for health worker visits to discuss .040 (.038) 2 0.032 (0.027) .067*** (0.003) 2 0.018* (0.010) .019** (0.008) malaria Include only HHs with no free government bed .121** (.057) NA 0.137*** (0.006) NA 0.046* (0.026) nets (N ¼ 1,136) Jointly estimate govt./own-acquired with NA 2 .036 (0.025) .066*** (0.017) NA NA seemingly unrelated regressions Include only HHs in villages lacking paved roads .035 (.039) 2 0.038 (0.027) 0.058*** (0.022) 2 0.020* (0.010) 0.017** (0.008) Exclude 25 percent of sample closest to urban .041 (.038) 2 0.047 (0.035) 0.085*** (0.021) 2 0.026 (0.135) 0.027** (0.009) center Exclude 25 percent of sample furthest from .089 (.040) 2 .019 (0.029) 0.077*** (0.025) 2 0.022** (0.010) 0.015* (0.008) urban center Include only HHs in lowest income category .004 (.936) 2 .044 (.262) .048* (.095) 2 .030* (.072) .028** (.034) Include only HHs not in lowest income category .048 (.239) 2 .036 (.353) .078*** (.004) 2 .014 (.240) .017* (.076) Replace single community radio variable with two variables: community radios inside and outside of commune Stations inside commune 2 0.025 (0.141) 2 0.213 (0.139) 0.300*** (0.078) 2 0.028 (0.075) 0.101** (.044) Stations outside commune 0.092** (0.041) 2 0.014 (0.029) 0.071*** (0.025) 2 .023** (0.010) .013 (.009) Notes: Each coefficient is the estimated effect of the number of community radios on the indicated dependent variable, from the corresponding specifica- tion of table 5, as modified according to the indication in the first column above. In the last two rows, the number of community radio stations in the specifi- cation of table 5 is replaced with two variables: the number of stations broadcasting from inside the commune and the number from outside. Both coefficients from the corresponding regression are reported. *** Significant at 1% level; ** significant at 5% level; * significant at 10% level. Source: Author analysis based on data described in text. Keefer and Khemani 259 260 THE WORLD BANK ECONOMIC REVIEW The estimates in tables 4 and 4A ignore the possible interdependence of demand for government bed nets and own-investment in bed nets. The third and fourth rows of table 6 present results of specifications that account for this inter- dependence. The third row takes a naı ¨ve approach to this issue and looks only at the sub-sample of households that should have the strongest demand for own- investments in bed nets—those that receive no free government bed nets at all. The magnitude of the effect of community radio access is much larger here than in table 4A, as one would expect. An alternative approach, controlling for the number of government bed nets in the own-acquired bed-net specification, yields the same result: the more free government bed nets a household has, the fewer bed nets it acquires using its own resources (every additional government bed net is associated with .21 fewer nongovernment bed nets, results not reported). The fourth row explicitly allows for the possibility that unobserved household characteristics simultaneously influence household access to free government nets and households’ private investments in bed nets. Seemingly unrelated regres- sions (SUR) correct for this possibility, allowing for the error terms across the free and paid government bed nets regressions to be correlated. Estimating equa- tion (1) using SUR again yields a significant effect of access to community radio on the acquisition of bed nets using own resources. The next three rows look at sub-samples that are more homogeneous with regard to remoteness. The seventh row examines only villages that lack a paved road. The eighth excludes the 25 percent of the sample that is nearest to a bus stop and the eighth excludes the 25 percent that is furthest away. Again, the results mimic those in table 4A; access to noncommercial private radio has a strong, positive effect on the number and share of paid government bed nets that households report. We also examine whether the results are robust when examining only house- holds that are in the lowest income bracket, and only households that are not in this bracket. Among the 1,437 households that are in the lowest income bracket, access to an additional community radio increases the number of paid bed nets by .05 ( p ¼ .095). Consistent with the earlier discussion of targeting, showing that greater access to community radio was more likely to induce richer house- holds to purchase bed nets, the effects for the richer subsample of households are larger and more significant. Among the 2,226 households that are not in the lowest income bracket, access to an additional community radio increases the number of paid government bed nets by .08 ( p ¼ .004). The specifications in tables 4 and 4A include all community radios to which villages have access, both within and outside their commune. Given the small size of community radio stations and the way in which the sample of villages was selected, village access to both types of radio stations is likely to be exogenous. The arguments underlying identification are strongest, however, for within- commune variation in village access to out of-commune radio stations. An im- portant robustness issue is therefore whether the main results of the paper—the lack of effect of radio access on the number of free bed nets and the positive Keefer and Khemani 261 effect on the number of paid government bed nets—persist when looking at within- and out-of-commune stations separately. The last two rows of table 6 therefore examine a specification in which the number of out-of-commune and of within-commune stations replace the single community radio variable in the table 4 specification. Access to out-of-commune stations has a highly significant positive effect on the number of paid government bed nets that households report and no effect on the number of free bed nets; co- efficient magnitudes are about the same as those in table 4A. While the effect of outside stations on the share of paid bed nets is not quite significant ( p ¼ .14), they significantly reduce the share of free bed nets, consistent with the argument that government officials respond to radio-induced increased demand and shift bed nets from the free to the paid channel. V. D I S C U S S I O N OF POSSIBLE WELFARE CONSEQUENCES Prior work has implied that greater provision of government benefits to media- driven citizen demand is likely to be welfare-enhancing, based on models in which the lack of citizen information gives governments greater discretion to divert resources to themselves or to exert less effort in servicing citizens. We are conservative in recognizing, however, that shifting bed nets from low-demand villages to high-demand villages could come at the expense of optimal bed net targeting to serve public health goals. The objective of malaria-control policy is to encourage households to take up a particular health technology, bed nets, using price subsidies as needed for this purpose to address public health goals (Cohen and Dupas 2010). If the lack of supply responsiveness is due to the scarcity of free government-provided bed nets, then shifting supply to where there is greater demand might reduce welfare by lowering bed-net provision to other areas where they are more needed to serve public health goals. If the public health goal is to achieve population coverage by targeting free nets to those with high price elasticity and greater vulnerability, who would not acquire or use bed nets in the absence of the subsidy, then an optimal response might even be to reduce the provision of free nets in those areas where greater media access, rather than price subsidies, has persuaded house- holds to acquire nets. However, we do not observe any significant effect of radio on inter-village variation in the number of bed nets, free or paid. Instead, we only observe a shift in the composition of bed nets, from free to paid, in villages with greater radio access. Our evidence does support the policy conclusion that government programs to increase demand for bed nets should be accompanied by greater efforts to expand aggregate supply, either free or paid. Villages with higher demand (those exposed to radio) exhibit no greater bed-net ownership than villages with lower demand, suggesting that supply is inelastic. Moreover, as much as 14 percent of the sample reports not owning any bed nets, despite our survey being undertaken in the aftermath of a massive wave of distribution. The median household 262 THE WORLD BANK ECONOMIC REVIEW reports one bed net for every three household members, numbers that include old bed nets that were either never treated with insecticides or where the treatment has lost its efficacy (Dupas 2009, discusses the importance of regular treatments with insecticide). The benefits of additional bed-net usage are likely to be large. Bed-net ownership by households explains as much as 58 percent of the recent decline in infant mortality in the region (Demombynes and Trommlerova 2012). One reason for this, found in the discussion and references in Cohen and Dupas (2008), is the large positive externalities that bed-net usage has on disease trans- mission. A quantity response to households with greater radio-induced demand for bed nets may therefore increase welfare; we observe no such government re- sponse in the data. At the same time, however, the data on which we base our analysis are insufficient to offer guidance as to the best strategy to expand supply. Just as we cannot be sure that the responsiveness of government supply to radio-driven demand is welfare-enhancing, so also it is not possible to assess the welfare effects of the fact that radio induces health workers to charge for bed nets or to derive policy implications about whether they should be charging. Stronger welfare conclusions require data that are not available on how health officials use the revenues from selling bed nets. We are therefore careful not to draw the conclusion that local deviation from nationally set policies is necessarily a sign of costly corruption in our study context. However, we note that prior work has drawn precisely such conclusions in a different context. Reinikka and Svensson (2004) interpret local deviations from national budget allocations for schools as evidence of “capture” or corruption but are not able to provide direct evidence of how those funds were used by local governments. Community radio exposure increased household demand for bed nets in an environment in which the supply of bed nets was inelastic to increases in demand, both in the aggregate and locally. Local officials, facing higher demand, were compelled to use some criterion to ration bed nets. Our evidence indicates that they chose to use pricing to ration. The fact that local officials chose this method of rationing is insufficient to assess welfare consequences since some price-charging by local officials at the margin may or may not be an efficient way to target subsidies where to best achieve population coverage. On the one hand, officials may divert bed nets away from priority targets (e.g., young children) in order to sell nets to low price-elasticity households. On the other hand, they may or may not use the proceeds from the sale of bed nets to improve the health care of disadvantaged households. We can exclude one potential negative welfare consequence of charging for bed nets. The results suggest that radio access has no effect, or a slight positive effect, on the targeting of bed nets to vulnerable households. Vulnerable house- holds, those with young children or low incomes, are less likely to pay for bed nets, and not less likely to receive free bed nets, when they enjoy greater radio access. This means that price discrimination by local officials is not concentrated among households targeted by the free distribution program. However, data are Keefer and Khemani 263 not available on how health officials use the revenues from selling bed nets and other health services and how they perform in delivering public health as a public good. VI. CONCLUSION The new evidence provided here of a government price response to more in- formed citizens underlines the importance of taking policy context and the struc- ture of media markets into account when analyzing the effects of media on government responsiveness. We provide the first rigorous evidence on the impact of a particular type of media—community radio—that is proliferating across Africa and other parts of the developing world and is used by donors, ministries, and civil society organizations to promote public interest programming to support development goals. These results have implications for the new thrust in international aid to use information and transparency to promote government accountability in difficult political economy environments (Zoellick 2011). The analysis here indicates that, in contrast to the findings of previous research, the media dissemination of general policy information (e.g., on the availability of a government benefit) is not, by itself, sufficient to enable citizens in less developed democracies to extract greater benefits from government. Instead, information can have important de- velopment effects by changing private household behavior. This, of course, leaves open the questions of how to use transparency initia- tives to improve government behavior, what information content is likely to be effective in this regard, and what media institutions would have incentives to carry such content. Evidence from Brazil suggests that greater access to radio played a facilitating role in reducing local government corruption, but in this case media content was shaped by a campaign of public disclosure of audit find- ings of irregularities in municipal government spending (Ferraz and Finan 2008)). Tendler (1997) describes how a reform-minded state governor in Brazil successfully used radio broadcasts to spread messages specifically aimed at un- dercutting municipal patronage politics, while also increasing citizen demand for municipal public health services. However, it is difficult to disentangle the effects of the radio broadcasts from other actions taken by this governor. More recently, and specifically for Benin, Wantchekon (2009) examines how “expert” information about the quality of public policies, provided in town- hall style meetings, can reduce the appeal of clientelist political promises to voters; these results raise important questions about how the broadcast of such information can be introduced into different media markets. Our work under- lines the need for more research along these lines, focusing both on identifying optimal information content, but also on effective ways to deliver that informa- tion, including the exploitation of local media with considerable power to per- suade poor citizens. 264 THE WORLD BANK ECONOMIC REVIEW AP P E N D I X . S U M M A RY S TAT I S T I C S Variable Obs Mean Std. Dev. Min Max Number of bed nets Total 4200 1.92 1.48 0 10 Free from government 4200 1.21 1.20 0 10 Bought from government 4200 .27 .79 0 10 Number of radio Number of community radios 4200 2.41 1.50 0 7 Number of private commercial radio 4200 0.472 0.996 0 6 Number of public radio 4200 1.038 0.36 0 2 Number of religious radio 4200 0.372 0.522 0 2 Respondent characteristics Age of respondent 4179 40.06 13.36 17 90 Respondent female 4200 .18 .38 0 1 Respondent education Primary 4197 .18 .38 0 1 Secondary 4197 .11 .31 0 1 Higher 4197 .01 .12 0 1 Respondent ethnicity Adja 4197 .01 .09 0 1 Fon 4197 .07 .26 0 1 Bariba 4197 .32 .47 0 1 Dendi 4197 .08 .27 0 1 Yom 4197 .12 .33 0 1 Peulh 4197 .07 .25 0 1 Ditamari 4197 .18 .38 0 1 Yoruba 4197 .15 .36 0 1 Respondent single 4195 .07 .26 0 1 Respondent polygamous 4195 .20 .40 0 1 Respondent religion Muslim 4198 .52 .50 0 1 Catholic 4198 .24 .43 0 1 In household, number of: Adults 4200 2.99 1.61 0 15 Adults over 60 4200 .14 .40 0 3 Children 4200 2.63 2.07 0 14 Children 5 and younger 4200 .85 1.02 0 8 Age of household head 4138 42.17 13.13 18 90 Gender of household head (1 ¼ male) 4149 1.06 .24 1 2 Income less than 27,500 CFA 4074 .39 .49 0 1 Characteristics of home Brick walls 4175 .20 .40 0 1 Cement floor 4172 .29 .45 0 1 More than 5 rooms 4120 .19 .39 0 1 Owns TV 4200 .08 .27 0 1 Owns mobile phone 4200 .27 .44 0 1 Characteristics of household villages Population 4120 2055.91 1386.06 115 8205 Distance to nearest urban center 4160 23.39 18.71 0 145 (Continued ) Keefer and Khemani 265 APPENDIX. Continued Variable Obs Mean Std. Dev. Min Max Distance to nearest bus or train stop 4120 23.31 29.31 0 150 Functional private school 4180 .07 .26 0 1 Chief has at least primary education 4200 .25 .43 0 1 Chief has at least secondary education 4200 .20 .40 0 1 Paved road 4200 .10 .30 0 1 Secondary school 4200 .17 .38 0 1 Literacy center 4180 .48 .50 0 1 Health center 4200 .54 .50 0 1 Potable water source 4200 1.14 2.32 0 17 Most common language in village is most common 4200 .85 .35 0 1 in commune Probability that any two households in village speak 4200 .75 .23 .27 1 the same language at home. Source: Survey data collected by authors, described in text. REFERENCES Adsera, A., C. Boix, and M. Payne. 2003. “Are You Being Served? Political Accountability and the Quality of Government.” Journal of Law, Economics and Organization 19 (2): 445– 90. Andrabi, T., J. Das, and A. I. Khwaja. 2008. “Report Cards: The Impact of Providing School and Child Test Scores on Education Markets.” Working Paper, Kennedy School of Government, Harvard University, http://www.hks.harvard.edu/fs/akhwaja/papers/RC_14Nov08.pdf. Banerjee, A., R. Banerji, E. Duflo, R. Glennerster, and S. Khemani. 2010. “Pitfalls of Participatory Programs: Evidence from a Randomized Evaluation in Education in India.” American Economic Journal: Economic Policy 2 (1): 1 –30. Beck, T., G. Clarke, A. Groff, P. Keefer, and P. Walsh. 2001. “New Tools in Comparative Political Economy: The Database of Political Institutions.” World Bank Economic Review 15 (1): 165–76. Besley, T., and R. Burgess. 2002. “The Political Economy of Government Responsiveness: Theory and Evidence from India.” The Quarterly Journal of Economics 117 (4): 1415– 51. Besley, T., and A. Prat. 2006. “Handcuffs for the Grabbing Hand?: Media Capture and Government Accountability.” American Economic Review 96 (3): 720–36. Bjorkman, M., and J. Svensson. 2009. “Power to the People: Evidence from a Randomized Field Experiment on Community-Based Monitoring in Uganda.” Quarterly Journal of Economics 124 (2): 735–69. Brunetti, A., and B. Weder. 2003. “A Free Press is Bad News for Corruption.” Journal of Public Economics 87 (7 –8): 1801– 24. ´ . Siochru Buckley, S., K. Duer, T. Mendel, S. O ´ , M. E. Price, and Marc Raboy. 2008. “Broadcasting, Voice, and Accountability.” World Bank Institute, The World Bank: Washington DC. Chong, A., and E. La Ferrara. 2009. “Television and Divorce: Evidence from Brazilian Novelas.” Journal of the European Economic Association Papers and Proceedings 7: 2–3, 458– 68. Cohen, J., and P. Dupas. 2010. “Free Distribution or Cost-Sharing? Evidence from a Malaria Prevention Experiment.” Quarterly Journal of Economics 125 (1): 1–45. Demombynes, G., and S. K. Trommlerova. 2012. “What Has Driven the Decline of Infant Mortality in Kenya?” World Bank Policy Research Working Paper No. 6057, Washington, DC. Dupas, P. 2009. “What Matters (and What Does Not) in Households’ Decision to Invest in Malaria Prevention?” American Economic Review 99 (2): 224– 30. Economist Intelligence Unit. 2006. Benin: Country Profile. London: Economist Intelligence Unit. 266 THE WORLD BANK ECONOMIC REVIEW Eisensee, T., and D. Stro¨ mberg. 2007. “News Floods, News Droughts, and U.S. Disaster Relief.” Quarterly Journal of Economics 122 (2): 693–728. Ferraz, F., and C. Finan. 2008. “Exposing Corrupt Politicians: The Effect of Brazil’s Publicly Released Audits on Electoral Outcomes.” Quarterly Journal of Economics 123: 2, 703–45. Fraser, C., and S. Restrepo-Estrada. 2002. “Community Radio for Change and Development.” Development 45 (4), 69 –73. Gentzkow, M., J. M. Shapiro, and M. Sinkinson. 2011. “The Effect of Newspaper Entry and Exit on Electoral Politics.” American Economic Review 101 (7): 2980– 3018. ¨ tz, T. 2000. “Local Radio Stations in African Languages and the Process of Political Transformation: Gra The case of Radio Rurale Locale Tanguieta in Northern Benin.” In Richard Fardon, and Graham Furniss, eds., African Broadcast Cultures. London: James Currey, 110 –27. Jensen, R. 2010. “The (Perceived) Returns to Education and the Demand for Schooling.” The Quarterly Journal of Economics 125 (2): 515– 48. Jensen, R., and E. Oster. 2009. “The Power of TV: Cable Television and Women’s Status in India.” The Quarterly Journal of Economics 124 (3): 1057– 94. Keefer, P., and S. Khemani. 2005. “Democracy, Public Expenditures, and the Poor: Understanding Political Incentives for Providing Public Services.” World Bank Research Observer. 20 (1): 1–28. ———. (2014). “Mass Media and Public Services: The Effects of Radio Access on Public Education in Benin.” Journal of Development Economics XXXXX. Keefer, P., and R. Vlaicu. 2008. “Democracy, Credibility and Clientelism.” Journal of Law, Economics and Organization 24 (2): 371– 406. Khemani, S. 2007. “Can Information Campaigns Overcome Political Obstacles to Serving the Poor?” In S. DevarajanI. Widlund, eds., The Politics of Service Delivery in Democracies: Better Access for the Poor. Expert Group on Development Issues, Ministry for Foreign Affairs, Stockholm, Sweden. La Ferrara, E., A. Chong, and Suzanne Duryea. 2008. “Soap Operas and Fertility: Evidence from Brazil.” Working Paper, Department of Economics, University of Bocconi, Milan, Italy http://didattica. unibocconi.it/mypage/upload/49273_20090112_123523_SOAPNOV08.PDF. Olken, B. 2009. “Do TV and Radio Destroy Social Capital? Evidence from Indonesian Villages.” American Economic Journal: Applied Economics 1(4): 1– 33. Pandey, P., S. Goyal, and V. Sundararaman. 2009. “Community Participation in Public Schools: Impact of Information Campaigns in Three Indian States.” Education Economics 13 (3): 355– 75. ¨ mberg. 2005. “Commercial Television and Voter Information.” CEPR Discussion Prat, A., and D. Stro Paper 4989. Reinikka, R., and J. Svensson. 2004. “Local Capture: Evidence from a Central Government Transfer Program in Uganda.” The Quarterly Journal of Economics 119 (2): 679–705. Roll Back Malaria Partnership. 2008. Global Malaria Action Plan. http://www.rbm.who.int/gmap/ gmap.pdf ¨ mberg, D. 2004. “Radio’s Impact on Public Spending.” The Quarterly Journal of Economics 117 (4): Stro 1415– 51. Tendler, J. 1997. Good Government in the Tropics. Johns Hopkins University Press, Baltimore, MD. Wantchekon, L. 2003. “Clientelism and Voting Behavior: Evidence from a Field Experiment in Benin.” World Politics, Vol. 55, 399–422. ———. 2009. “Can Informed Public Deliberation Overcome Clientelism? Experimental Evidence from Benin.” Working paper, Department of Politics, Princeton University. World Health Organization. 2008. “Report of the Twelfth WHOPES Working Group Meeting. WHO/ HQ, Genevea, 8– 11 December 2008. Review of: Bioflashw GR, Permanetw 2.0, Permanetw 3.0, Permanetw 2.5, Lambda-cyhalothrin LN.” http://www.who.int/whopes/Long_lasting_insecticidal_ nets_Jul_2011.pdf Keefer and Khemani 267 ———. 2009. “Report of the Thirteenth WHOPES Working Group Meeting. WHO/HQ, Geneva, 28– 30 July 2009. Review of Olysetw LN, Dawaplusw 2.0 LN, Tianjin Yorkoolw LN.” http://whqlibdoc.who. int/publications/2009/9789241598712_eng.pdf. The World Bank. 2007. The World Bank Booster Program for Malaria Control in Africa: Scaling up for Impact, A Two-Year Progress Report, Washington, DC http://siteresources.worldbank.org/ EXTAFRBOOPRO/Resources/MALARIAREPORTfinalLOWRES.pdf. USAID. 2011. President’s Malaria Initiative, Benin: Malaria Operational Plan FY 2011. Yanagizawa, D. 2009. “Propaganda and Conflict: Theory and Evidence from the Rwandan Genocide.” Mimeo, Stockholm University. Zoellick, R. B. 2011. “The Middle East and North Africa: A New Social Contract for Development.” Speech delivered to The Peterson Institute for International Economics, April 6.