Edutainment Radio, Women’s Status and PrimarySchool Participation:

Evidence from Cambodia∗

Maria Cheung†

April 10, 2012

Abstract

This paper investigates whether exposure to ”edutainment” (education - en-tertainment) radio leads to improved women’s status and primary schoolparticipation. Specifically, I examine a popular radio station focusing ongender issues in Cambodia. To identify the effect, I exploit plausible exoge-nous variation in over-the-air signal strength between radio transmitters andvillages within a district, as well as the variation across time and space in ex-posure. Using individual data, both approaches show that the exposure had asignificant impact on behavior by raising the women’s decision-making powerwithin the household and increasing children’s primary school attendance.The latter impact is also reflected by higher primary school enrollment threeyears after exposure. The impact was found in both poor and rural house-holds confirming that radio is an effective vehicle to transmit information inthe more marginalized areas. Suggestive evidence shows that the exposurealso affected attitudes towards domestic violence and the prevalence of sonpreference which is a stepping stone towards changing socially constructedgender norms.

Keywords: Radio exposure, information, gender, women’s status, school-ing, enrollment, attitude, behavior, edutainment, CambodiaJEL: D83, I25, J16

∗Thanks to the Lars Hierta Memorial Foundation providing a travel stipend to the Women’s MediaCenter of Cambodia (WMC) in Phnom Penh and to the Wallander Foundation financing this project.Special thanks to Mao Piseth at the Women’s Media Center. Jakob Svensson, Anne Boschini, Maria Per-rotta, Kristofer Jervinge and all the seminar participants at the Development Study Group of StockholmUniversity provided helpful comments. The opinions expressed herein are those of the author and do notnecessarily reflect the views of WMC or any financial supporters of the project.†Department of Economics, Stockholm University. Contact: maria.cheung@ne.su.se.

1 Introduction

Promoting gender equality and empowering women, which is one of theUnited Nations Millennium Development Goals, is not only an end in it-self; it has also proven to have large development gains. Several authorssuch as Qian (2008), King and Mason (2001), Duflo (2005), Jensen and Os-ter (2009) have found that when women obtain a higher status within thehousehold, investments in children’s human capital and well-being are en-hanced. Although efforts to fulfill gender equality are gaining momentum inmany developing countries, strong traditional norms that marginalize womencan undermine the implementation and enforcement of legislation and policyprovisions.1 Media can play a crucial role in either reinforcing these tra-ditional values or changing them, depending on how the media format isdesigned and implemented.

The World Bank has acknowledged the role of effective communicationstrategies in the development processes.2 ”Edutainment” (education - en-tertainment) is one such strategy, using entertainment media for educationalpurposes aiming to change attitudes, behaviors and social norms. It has beenconsciously applied to HIV prevention and control in the form of popular ra-dio and television soap operas. Research evaluations of such interventionsare mostly found within the sociology field showing in general that they havemeasurable effects on changing HIV prevention behavior.3There are howeverno robust empirical studies evaluating the causal effects and, in particular,on other outcomes than HIV prevention.

This paper investigates whether exposure to edutainment radio can af-fect attitudes and behavior leading to higher status for women and increasedprimary school participation. Specifically, I examine a popular radio sta-tion focusing on women’s issues in Cambodia, the Women’s Station FM 102.Their mission is to educate and inform the Cambodians on various social andwomen’s issues by designing and implementing edutainment programs. Toidentify the effect, I use two complementing identification strategies. The firstone exploits the within district variation in over-the-air signal strength whilethe second one uses variation across time and space, exploiting the fact thatthe radio coverage was gradually expanded over time and across regions. Rec-ognizing the multifaceted nature of women’s status, I use several indicatorsincluding the degree of son preference, women’s attitude towards wife beat-ing and their degree of participating in household decision-making, includinghaving the sole final say in decisions regarding the children’s schooling. To

1UNIFEM and ADB (2004)2Khalil et al. (2009)3For example, Singhal and Rogers 1999, Vaughan et al 2000, Valente 1997

1

estimate the effect on children’s school participation, I use school attendancefor children aged 6-12 as well as enrollment records for grade one to six.

Independent of the identification strategies, I find significant impact ontypes of behavior associated with women’s status such as participating inhousehold decision-making as well as investments in children’s human capital.Results from a within school fixed effects estimation show an increasing effectover time and there are no trends prior to the exposure that could havedriven the results in the exposed areas. In terms of impact, the averageschool enrollment increased by 5 and 7 percent respectively, two and threeyears after exposure. I also conduct placebo tests for both strategies and donot find that there were any confounding trends in the outcome variablesin the absence of exposure. Given the limitation of pre-exposure data, atechnique adopted by La Ferrara et al. (2008), Jensen and Oster (2009) andOster and Millett (2010), is employed to test the impact of future radiocoverage. I do not find that future coverage had a significant impact onthe outcome variables, confirming that these variables were not progressingfaster in coverage areas than in non-coverage areas.

I also find some suggestive evidence of edutainment radio on the attitudestowards domestic violence and the degree of son preference. These findingsimply that it might be more difficult to change structural inequalities, par-ticularly in societies where gender discrimination is enforced by social norms,than persuading individuals to adopt behaviors that are perceived as havingimmediate personal benefits. Another aspect is whether these findings actu-ally reflect personal or perceived attitude since the latter could be stronglyinfluenced by the survey context (answering what you believe is expected bythe interviewer) and the media priming from the radio station.4 However, ifbehavior is determined by the perception of what is normal or desirable in agiven context (the social norm) then, even if it is not internalized with thepersonal belief or attitude, changes in these attitudes may be an importantstepping stone towards changing behavior.

The contribution of this paper is fourth folded. Firstly, to my knowledgethis is the first paper investigating an edutainment intervention disseminat-ing gender information. Most of the literature on edutainment interventionsis found within the sociological field using more descriptive data. This is anattempt to establish a causal relationship using quantitative empirical meth-ods on a country-wide dataset. Secondly, the paper relates to the literatureon gender focusing on the role of women’s status on children’s well-being.

4A person exposed to a radio program designed to alter individuals perceptions aboutwhich attitudes and behaviors are typical or desirable in their community is normally awarethat many other members of the community are being exposed to the same program andwill believe that this is endorsed by the community members.

2

Thirdly, the paper also adds to the literature on media effects on attitudeand behavior outcomes. Contributions in this field include that of Gentzkowand Shapiro (2004) showing that television viewing in the Muslim worldaffects attitudes towards the West, Della Vigna and Kaplan (2007) findinglarge effects of the Fox News channel on voting patterns in the United States,La Ferrara et al. (2008) showing significant impact of Brazilian soap operason fertility, and Jensen and Oster (2009) finding an impact on women’s statusand fertility in rural India after introducing cable-TV.

Fourthly, this paper confirms that radio is a powerful tool in more marginal-ized areas, where access to other information technologies might be moreunequally distributed. The exact number of radios in developing countries isa matter of debate, but international agencies seem to agree that there are atleast ten times as many radios as televisions. This is mainly because radio hassome crucial advantages in these countries such as having both a wide-reachand a low production cost, an affordable and easy-to-use technology that isaccessible to an audience with limited reading skills. The power of radiohas already been studied by for example Stromberg (2004) who stresses therole of radio as mechanisms of accountability for politicians. Studies basedon data from developing countries include that of Olken (2009) who findsthat access to radio and television decreases participation in social organiza-tions in Indonesia and that of Svensson and Yanagizawa (2009) showing thatexposure to local FM radio made farmers better informed about prices.

Finally, the findings may have important policy implications for devel-oping countries. In contrast to resource-intensive interventions such as pro-viding scholarship to girls in order to increase the proportion of girls in avillage attending school, interventions using an edutainment communicationstrategy with radio as the driving force, attempt to change the communitymembers’ perception of a social norm which in turn, down the line, influ-ence actual behavior. The major difference here is that radio broadcastingprograms are in general relatively cheaper to produce (higher impact-per-dollar spent) and have the potential of reaching an overwhelming number ofpeople compared to education programs delivered in classrooms. If properlydesigned, edutainment radio can be an effective complement to these edu-cation programs for conveying important social contents and could thus besuccessfully employed in a broader set of development policies.

This paper is organized as follows. The next section provides the back-ground. Section 3 presents the details of the Women’s Radio Station FM102 and section 4 describes the data and sample used. Section 5 describesthe two empirical strategies. In section 6, the results of the estimations arepresented including some robustness checks and the final section concludesthe paper.

3

2 Background

2.1 Context

Despite a process of recovery, Cambodia remains one of the poorest and leastdeveloped countries in Asia. The society is characterized by a patriarchal andhierarchical social structure with strong traditional norms. According to thereport by UNIFEM in 20045, there are moral codes for a woman which isstated as part of the primary school curriculum (called Chba’p): “The womanis to be silent and walk so softly that one cannot hear the sound of her silkskirt rustling. She is shy and naive and must be protected. Before hermarriage, she has ideally never left the company of family members”.6 Suchnorms have traditionally reflected in a strong bias against girls’ access toeducation. Although the government of Cambodia is making huge efforts inachieving universal primary enrolment, convincing parents to invest in theirdaughters’ education is central to attain gender parity in education. Not onlydo the traditional values assign women a lower status, women also tend tohave a heavier work burden, work longer hours, have lower decision-makingpower, less control over resources, lower education and literacy rates andlower access to public services compared to men.7

The media sector in Cambodia plays a central role in reinforcing theseattitudes and social norms when portraying women with traditional genderroles. This sector is characterized as being one of the most unregulated inSoutheast Asia and is argued to have some close links with politics.8 Thedevelopment of the media sector is partly explained by the rapid changes ofpolitical regimes since the 1970s. During the Khmer Rouge period (1975-1979), more than a million people were killed, many of them were highlyeducated and journalists. Under the subsequent Vietnamese rule, media wascontrolled by the state and acted basically as a government tool for dissemi-nating information and party propaganda. After the Vietnamese occupationand prior to the UN-supported general election in 1993, the media sectorbecame highly deregulated and expanded rapidly.9

Print media is not a popular or reachable option in Cambodia due to

5See UNIFEM and ADB (2004).6Ledgerwood (1996)7See Dahlerup (2010) and UNIFEM and ADB (2004).8Most of the television and radio broadcast platforms are owned by or have a close af-

filiation with the ruling Cambodian People Party (CPP) as described by Deutsch Karlekar(2011). According to Edman (2000) and Rosette (2010), there are only two radio stationsthat are regarded as independent, one of which is the Women’s Station FM 102.

9According to CCIM (2009), there were some 341 newspapers, 119 magazines, 22 radiostations and seven television outfits registered in 2009.

4

problems of illiteracy while television is not an affordable alternative for peo-ple in areas where poverty rates are high. Radio remains the most accessibleform of mass media, particularly in rural areas. For example, in 1986 therewere about 200,000 radio receivers in the country but no daily newspaperwhile 10 years later there were an estimated one million radios and only100,000 televisions.10 Realizing its wide-reach, the Women’s Media Centerof Cambodia (WMC) founded and registered as an Cambodian NGO in 1995,launched their own radio station a few years later, the Women’s Station FM102, in order to change the stereotypes of women in traditional gender rolesusing an edutainment approach.

2.2 Edutainment as a Development Communication

The field of development communication is dominated by two conceptualmodels: diffusion and participation. The edutainment approach draws uponthe former model11 that uses a top-down or vertical approach to persuadeindividuals to change their behavior by providing them with information.In contrast to the participation model12 that uses a horizontal approach ofcommunication where dialogues are exchanged at the community level, thediffusion model is more outcome oriented. The standard formulation of thediffusion model is Knowledge - Attitudes - Practice: information providesKnowledge, which leads to a change in Attitudes, which in turn leads toPractice - the desired behavior change. Historically, many international de-velopment agencies in the health sector have relied on this model for messagedissemination. Expected outcomes such as increased child vaccination rates,use of prenatal care or family planning, and HIV/Aids prevention, are gen-erally easier to achieve than initiatives aimed at changing more structuralinequalities.13

Edutainment as an effective communication strategy has been embracedby many development communication practitioners.14 The purpose of edu-tainment media is to increase audiences’ knowledge about an educational is-sue using drama, music or other communication formats that engage the emo-tions to change attitudes, behaviors and social norms. This can be achieved

10Mehta (1997)11Rogers (1995)12Morris (2001)13Efforts to accomplish more structural changes would perhaps be better employed

with a participatory oriented intervention, or a combination of both approaches. Theparticipatory model stresses the role of an empowerment process, where individuals obtaingreater control over decisions that affect them, rather than on direct outcomes. See Izettand Toubia (1999).

14See, for example, Singhal and Rogers (2001) and Piotrow et al. (1990).

5

by either influencing the audience towards a socially desirable end or influ-encing their external environment to help create the necessary condition fora social change at the group or system level.15

There are several examples of successful edutainment interventions in-cluding televised serials, soap operas, radio dramas, children’s cartoons, videogames etc. The most famous example in recent years is the ”Soul City”television show broadcast by the Soul City Institute for Health and Devel-opment Communication, an NGO in Johannesburg, South Africa. Soul Cityaddresses a wide range of key issues including HIV/Aids prevention, landreform, child health, domestic violence etc.16 The series reached 16 mil-lion people in South Africa, achieving top audience ratings and winning sixawards, including the prize for South Africa’s Best Television Drama, andis therefore regarded to have a very high level of impact-per-dollar spent inchanging attitudes and behavior.

3 Radio Coverage

3.1 Women’s Station FM 102

This paper focuses on a popular radio station, the Women’s Station FM102, which is the only one focusing on gender issues in Cambodia. It isowned and operated by the WMC, a donor-driven17, non-profit and localnon-governmental organization, aiming to educate and inform Cambodianson women’s rights and health, domestic violence, HIV/AIDS awareness, traf-ficking, but also on elections, decentralization and poverty alleviation.18 Theradio station is daily broadcasting programs that consist of live and self-produced shows, daily news bulletins, as well as international news takenfrom the Australian Broadcasting Corporation service. Their radio programsmap into four areas: women’s issues such as law, domestic violence andhealth; news programs; teenagers’ programs and children’s programs.

15For example, Singhal et al. (2006) describes how the radio soap opera ”Tinka TinkaSukh” (Happines Lies in Small Things) led to an entire village in India rejecting traditionaldowry practices following the broadcast of this program.

16Singhal and Rogers (2004)17Some of the donors include UNESCO, UNICEF, foreign aid agencies from Sweden,

Denmark, United Kingdom, Australia, Japan etc.18The WMC consists of five departments that conduct various media-related activities

with a gender approach in all their programs. The WMC also has a TV unit that producesprograms (dramas, comedies, documentaries etc) with an edutainment approach but this isbroadcast through other television broadcast platforms (such as TV3). The radio stationFM 102 is their largest activity.

6

Several reports and surveys have confirmed that this is one of the leadingstations in Cambodia19. Basically because their reporting is considered to beneutral and unbiased20, and their programs are perceived as having a posi-tive influence on the audiences’ attitude and behavior.21 According to WMC(2006), equally many men and women tune into FM 102, mainly because it isinformative and educational. The station’s objective is to be ”recognized forhigh quality radio programs, and to be respected nationally and internation-ally for professionalism and ethical leadership in Cambodia’s media sector”.In 2009, the WMC was nominated by the One World Broadcasting Trust fora Special Award for Development Media.22

An example of an edutainment program is the popular call-in show foryoung people, called ”Os Tos Mhong!” (Cool) and broadcasted live every Sun-day morning in 2004, presented by two youths sharing their own experienceswith whom the listeners can identify.23 The program format included talkshows, information, advice, music and competition and the communicationstrategy was to create a safe but lively and entertaining space to question,share ideas, and voice opinions. Young people were encouraged to activelyshare their views with other listeners live on air as well as through e-mailsand letters. In addition to entertainment, the show provided educationalinformation about a particular topic such as youth culture, music, genderequality, education etc since each week had its particular focus. Other pro-grams are having a particular focus on women’s issues such as property andinheritance rights, health, education, and social roles. In general, all theirprograms are designed after an edutainment strategy where the educationalinformation is made more interesting by interacting it with entertainmentsuch as music, comedy, dramas etc.24

3.2 Coverage Expansion across Time and Space

The variation across space and time in coverage will be used in the identifi-cation strategy. In 1999 when the WMC launched the radio station, it wasfirst aired by a 200 W radio transmitter placed in Phnom Penh. Two yearslater, after receiving more funding, WMC upgraded this transmitter to a

19See for Human Rights (2003), Alison Mee and Savage (2003), Rosette (2010) andWMC (2006).

20Marseille (2008)21WMC (2006)22http://oneworldmedia.org.uk/.23WMC (2009b)24Another successful program is called ”The Road of Law” which tries to engage the

Cambodians about their legal rights through quiz shows, where people can call in andanswer questions about their rights, and a popular drama series about the legislation.

7

more powerful 10 kW transmitter. The prediction in coverage for the trans-mitter in Phnom Penh in each year of 1999 and 2001 is illustrated in Figure1 (hereafter called PP1999 and PP2001). The old transmitter was placed inanother province (Kampong Thom) as a relay station. They expanded fur-ther with an additional relay station in Svay Rieng province to reach 11 (of24) provinces covering about 60 percent of the population in Cambodia. Atthe end of 2006, the Kampong Thom transmitter was upgraded to a power of1 kW.25 Figures 2 and 3 illustrate the predicted coverage map for the Kam-pong Thom transmitter in 2001 and 2006 (hereafter KT2001 and KT2006)and that for the Svay Rieng transmitter in 2001 (SR2001), respectively.

In these maps, the red colored area closest to the transmitter locationindicates a strong radio signal, referring to signal 1, while the next signallevel is marked with a pink colored area indicating a medium strong signal,signal 2. In these signal levels or area covered, the simplest radio receivercan easily be tuned in to the specific frequency with clarity and with nointerference from any other radio broadcasts. The third signal level, signal 3,which is in the outer area of the coverage, indicates a medium/weak signal.In this area, one can still comfortably listen to the specific frequency but withsome static noise and a little interference from other radio broadcasts. Anysignal level below 45 dBV/m (not marked on the map), is highly affectedby the atmospheric and terrain conditions and is hence unreliable and veryweak and these areas are not considered as being covered.26

In order to predict the radio coverage for each transmitter, I use dataprovided from the WMC about each transmitter’s location and technical pa-rameters such as longitude and latitude, transmitter height above ground(m), transmission power (W), frequency (MHz), antenna gain (dBi) and po-larization. The coverage maps are created in COVWEB, a free web based ap-plication created by the Communications Research Centre Canada (CRC)27,where the calculation of the coverage for a single transmitter is based onthe Irregular Terrain Model (ITM) propagation prediction model, developedby the US Institute for Telecommunications Science. The ITM model as-sumes that the strength of electromagnetic signals declines proportionallywith the inverse square of the distance between the transmitting and receiv-ing location in the absence of mountains.28 Areas with direct line-of-sight to

25See WMC (2004), WMC (2008), WMC (2009a) and WMC (2009b).26At the lowest signal levels around 45 dBV/m, i.e. at the threshold of being exposed,

only the radios with high sensitivity are able to pick up the signals (typically higher qualityradio receivers).

27CRC is the government’s primary research lab for communications technologies inCanada.

28This model, which is a modified version of the Longley-Rice standard model that

8

the transmitter have the strongest signal.29 Information about the surfacerefractivity (N-units), dielectric constant of ground, conductivity of ground(Siemens/m) and climatic zone are also necessary parameters and taken intoaccount when predicting the coverage.

The coverage prediction maps are then imported into ArcGIS, software forworking with maps and geographical information (GIS)30, where new mapsare created from integrating the software with other mapped geographicalinformation. For example, the right hand side of Figure 4 shows an admin-istrative map over Cambodia with district boundaries combined with GISinformation from the Cambodian Demographic Health Survey (DHS) foreach cluster village (the white dots on the map) and the left hand side ofFigure 4 is a topology map combined with DHS clusters. Based on thesemaps, I can identify which clusters31 in the DHS that are located in coveredand non-covered areas.

3.3 Placement of Transmitters

Based on a qualitative analysis from an interview conducted by the authorin January 2011 in Phnom Penh with the communications specialist at theWMC, the placement of radio transmitters was not random and neither wasit driven by the demand. The availability of electricity was one of the funda-mental determinants to place the transmitters at their current locations. Forexample, the transmitter in Phnom Penh has its own generators at daytimebut at nighttime it uses the government electricity providers to supply thestation with electricity. 32 The other fundamental part of the decision wasthe operational costs which include expenses for office rental, staffs, equip-ments, technicians, transportation etc. The placement of the transmitterswas therefore largely driven by these operational costs rather than demand-driven. Although ideal in principle for targeting the audience, placing them

predicts long-term median transmission loss over irregular terrain relative to free-spacetransmission loss, has nowadays been adopted as a standard to predict radio coverage forcommercial use as it is accurate over a wide range of frequencies and over various distances.

29However, if a mountain blocks the signal’s path the signal pattern becomes morecomplex as it can diffract around and between mountains depending on its frequency(the diffraction is higher the lower the frequency). See Hufford (1995), Chamberlin andLuebbers (1982) and Longley (1968).

30www.esri.com31One cluster in the DHS is representing a whole village.32The reason for using the government providers during the night time is that they had

major issues with their own generators as the electricity was often cut off. During the daytime they are able to monitor their own generators which is more difficult and expensiveduring the night time.

9

according to demand would be an impossible task given their budget con-straints.

In order to quantitatively test this, Table 7 presents an OLS with dis-trict fixed effects analysis where an indicator of being exposed to the radiosignals of either the KT2001, PP2001 or SR2001 transmitter is regressed onvarious determinants, both before exposure using individual level data fromthe DHS 2000 and after exposure using data from the DHS 2005. I find thatthe only significant determinant before exposure is the distance to nearestWMC transmitter: the shorter the distance from the transmitter the higherprobability to receive radio coverage. Using post-exposure data, I find thatthe distance to nearest province capital and the elevation of each DHS clusterare also significant determinants. None of the other variables such as age,years of education and income wealth, reflecting some audience characteris-tics, are significant confirming that the placement of the transmitters werenot driven by these demand related characteristics.

However, even if the endogeneity of the placement is not problematic an-other concern is that there is no data available about the actual transmissionpatterns in order to verify the accuracy of the model prediction. Showingthe prediction map to the staff at the WMC during the field survey, theydid indeed acknowledge that the coverage map seemed accurate but this canhowever not be quantitatively verified.

4 Data

The data used in the paper are mainly from the DHS 2000 and 2005, con-ducted by the National Institute of Statistics (NIS), and include informationon demography, family planning, maternal mortality, infant and child mor-tality, domestic violence, women’s status and health related information suchas breastfeeding, antenatal care, children immunization, childhood diseases,and HIV/AIDS. The DHS 2000 is the first standard DHS survey while DHS2005 is the second survey conducted in Cambodia, where the latter is usingthe same methodology as its predecessor.33 Both surveys are based on anationally representative sample across all 24 provinces in the country, in-cluding 15,351 women aged 15-49 in 12,236 households and 16,823 women

33DHS 2000 has a stratified sample based on the 1998 Cambodia General PopulationCensus and consists of 600 villages selected with probability proportional to the numberof households within the village. The stratification was achieved by separating everyreporting domain into urban and rural areas and the sample was selected independentlyin every stratum. DHS 2000 was carried out over a six-month period, from February toJuly 2000, while DHS 2005 was conducted from 9 September 2005 to 7 March 2006. Formore technical information, see DHS (2001) and DHS (2006).

10

and 6,731 men aged 15-49 in 14,243 households interviewed in DHS 2000and DHS 2005, respectively. The women’s status module was implementedin one-fourth of the households included in these samples, a subsample of alleligible women (aged 15-49) yielding a total of 3,771 and 4,174 completedinterviews for the DHS 2000 and DHS 2005, respectively.

From data, I can identify the following five outcome variables defined inTable 1: an indicator of son preference (Sonpref); the number of situationsin which wife beating is justified (Wifebeat); the number of household deci-sions where the woman respondent has the final say (FinalSay); an indicatorof having the sole final say in decisions about children’s schooling (FS Sch)and an indicator of attending school (SchAtt). The sample size is differentdepending on which dependent variable is analyzed due to some necessaryrestrictions. Moreover, any observation that is located within the coveragearea being exposed to the PP1999 transmitter (see map on right-hand sidein Figure 1) are excluded from the analysis as they are exposed to the radiosignals in both surveys (i.e. both before and after) as well as any observa-tion located just at the boarder of having coverage or not. A more detailedexplanation of the sampling procedure is found in Table 2, appendix. Table3 shows that the number of observations used in the analysis with Sonpref asthe dependent variable is 10,865 and 11,820 (or 71% and 70% of all womeninterviewed) from DHS 2000 and 2005, respectively. The other three vari-ables Wifebeat, FinalSay and FS Sch are part of the women’s status moduleand only include a subsample of women aged 15-49 that are eligible for thewomen’s status module. The last DHS variable SchAtt is taken from thehousehold recode containing information about schooling on all children inthe household aged 6-12.

Tables 4 and 5 present the summary statistics for the control variables.The elevation is measured as the height above a fixed reference point ofthe Earth’s sea level in kilometers provided from Shuttle Radar TopographyMission. The topographic slope, or percent gradient, describes the steepnessof the topological location of a DHS cluster - a higher slope value indicatesa steeper incline.34 It is computed in ArcGIS using the change in elevationof the target raster cell, a DHS cluster, to its neighboring cells in the digitaltopographic map. Other geographic village-level control variables includethe distance from each DHS cluster to the nearest WMC transmitter andthe distance from each DHS cluster to the nearest Province Capital.35 The

34Slope can also be represented in degrees. A 100 percent slope has a 45 degree slopeand an increasingly vertical line (90 degrees) has a percentage slope approaching infinitywhile an decreasingly horizontal line (0 degrees) has a percentage slope close to zero.

35Data information about the former is provided by the WMC and the latter is basedon GIS information about each of the 24 province capitals.

11

individual or household control variables, including a dummy for living inurban areas, a dummy for whether the household have electricity, dummiesfor household’s wealth quintiles, respondent’s age and years of education, areprovided by the DHS.

School (gross) enrollment data, provided by the Education ManagementInformation System (EMIS) for school years 1999/00 up to 2003/04, is usedto verify the effect on children’s schooling at the school level. This seconddataset is collected by the Ministry of Education, Youth and Sports (MoEYS)who annually distributes school census forms to all schools in Cambodia, cov-ering 8,443 public schools of which 5,250 are primary schools. Since the schooldata is lacking the GIS (latitude and longitude) information, identificationof an exposed school is made at the lowest possible administrative level, thecommune level. Integrating provincial administrative commune boundariesmaps in ArcGIS (digitalized by the author) with the coverage prediction mapseach commune is mapped as non-covered, semi-covered or fully-covered, andlater merged with the school data. Any school within a fully-covered andnon-covered commune is then defined as an exposed or non-exposed school,respectively, while schools located in semi-covered communes are excludedin the analysis as I do not know whether they are exposed or not. Exclud-ing these schools is necessary to minimize potential biases where you haveunder- (or over) estimates of the true effect, as well as schools within cov-erage area of the PP1999 transmitter. Finally, I construct a balanced panelconsisting of 5,180 primary schools across 5 school years leading to 25,900observations, all summarized in Table 6.36 Moreover, in a falsification test Iuse a school census from 1998/1999 that contains information about grade1-6 for 5,086 schools of which 5,067 are used in the analysis, excluding a fewschools in semi-covered communes located at the boarder of having coverageand non-coverage.

5 Empirical Strategy

5.1 Cross-Section Identification

The first identification strategy uses a cross-section of individual data fromthe DHS 2005 to exploit plausible exogenous variation in over-the-air sig-nal strength between radio transmitters and villages within a district. Theidea is that the mountains in Cambodia and in particular the CardamomMountain range in the south west of Cambodia with the highest elevationat Phnom Aural, create variation in the radio coverage that is unrelated to

36Observations in the table refers to the number of schools, not pupils.

12

other village characteristics that might affect women’s status and school en-rollment of children. An important first step here is to investigate whetherthis variation is exogenous.37 The identification hinges on the assumptionthat the radio coverage is orthogonal to, i.e. uncorrelated with, the chosendependent variables before the exposure to WMC radio signals.

Table 8 illustrates this empirically where the dependent variable in PanelA, using DHS 2000 data, is an indicator of being exposed to the radio signalsof either the KT2001, PP2001 or SR2001 transmitter and that in Panel B,using DHS 2005 data, is an indicator of being exposed to the radio signals ofKT2006. Both panels in the table are estimated at the individual level witheach of Sonpref, Wifebeat, FinalSay, FS Sch and SchAtt as the independentvariables. The estimates show that the indicators for women’s status andchildren’s schooling observed before the exposure did not imply a significanthigher probability to receive radio coverage a year later. The following em-pirical model for the first strategy is estimated at the individual level usingOLS regression with district fixed effects:

wivd = βCvd +Xivdγ +Gvdδ + ζd + µm + εivd (1)

where the dependent variable wivd is one of the outcome variables Sonpref,Wifebeat, FinalSay, FS Sch or SchAtt for individual i in village v of district d.An indicator of being exposed to radio signals of either KT2001, PP2001 andSR2001 transmitter is denoted Cvd and measured at the village level for eachdistrict, Xivd is a vector of individual or household control variables includinga dummy for living in urban areas, a dummy for whether the household haselectricity, dummies for household’s wealth quintiles, respondent’s age andyears of education, Gvd is a vector of geographical control variables measuredat the village level and includes the elevation (per thousand meters), theslope (percent), distance to the nearest WMC transmitter, and distance tonearest province capital.

The reason for including these controls is to avoid different sources ofestimation biases arising when, if omitted in the specification, a number offactors are systematically correlated with the radio coverage indicator andat the same time determine the outcome variables. For example, Figure 5shows that the share of individuals that frequently listens to radio is largestin coverage areas referring to signal 1. One way to interpret the figure is

37A similar approach is found in Olken (2009), who compares villages that, withinthe same district, are exposed or not exposed to the television and radio signal due tosome geomorphologic characteristics of the territory. An illustration of the physics ofbroadcasting is found in page 26, Figure 2 in Olken (2009), which is modified from Ellingtonet al. (1980).

13

that exposure to WMC’s broadcasting induced more radio listening which iscaptured by the DHS questionnaire, but another way to interpret it relatesto the potential endogenetity problem: individuals within a closer distanceto the radio station tend to listen to radio more often compared to individ-uals located further away from the larger cites, or the fact that the radiostation is located in the suburbs of major big cities such as Phnom Penh andKampong Thom. In order to avoid such omitted variable biases, I control forthe distance to the transmitting WMC radio location as well as the distanceto the nearest major city such as the province capital. Another concern in-volves both individual and geographic differences. For example, if there aresystematically differences in the women’s status of women in mountain ar-eas than that of women in the surrounding lowland areas, then these factorscould drive and confound the estimated effect rather than the intended effectfrom radio exposure, suggesting that the individual as well as the other ge-ographical controls such as elevation and slope are also relevant and shouldbe added into the specification.

In all the specifications when using the DHS data, there are district fixedeffects ζd to account for time-invariant regional characteristics that mightinfluence the outcome as well as month fixed effects µm controlling for sea-sonality factors such as dry and wet seasons, school holidays, winter andsummer terms etc that might affect children’s schooling as well as labor mar-ket outcomes which in turn might affect the women’s status. Throughoutthe analysis using the individual level data, the standard errors are clusteredat the village level since this is the level determining whether a village, rep-resented by a cluster in the DHS, is exposed or not exposed to the radiosignals. I also conduct placebo regressions using the same specification (1)but with pre-exposure data and here I do not expect to find any effect onthe outcomes.

5.2 Across Time and Space Identification

The second strategy exploits the variation across time and geographical lo-cation in exposure by comparing individuals in villages within a district thatare and are not exposed, as well as before and after the exposure.38 I esti-mate the following OLS regression using district fixed effects with clusteredstandard errors at the village level and the same set of individual and geo-graphical controls as in the first strategy:

38The impact of the radio exposure on the outcome is estimated by computing thedouble difference, over time (before-after) and across subjects (between individuals incovered areas and individuals in non-coverage areas).

14

wivdt = αCvdtPvdt + βCvd +Xivdtγ +Gvdtδ + ζd + ηt + µm + εivdt (2)

where t=2000, 2005 denote the time of the DHS survey and the indicatorof having radio coverage Cvdt in village v of district d in time t is interactedwith a post P indicator which is a dummy for post exposure. The coeffi-cient α is interpreted as a difference-in-difference estimate where treatmentis exposure to the radio signals. In addition to the district and month fixedeffects, as well as the same set of geographical and individual factors as de-scribed under specification (1), the vector Xivdt also includes an interactionterm between the post indicator Pvdt and a dummy for living in urban areasto account for any differential trends in the outcome by urban and rural in-dividuals. For example, as the transmitters are placed in the suburbs of themajor cities of Cambodia as shown by the coverage maps in Figure 1 - 3 onemight suspect that it is some urban area characteristics driving the results.The interaction term thus allows each year to differ in the outcome by urbanand rural areas.

The validity of this strategy depends on one crucial assumption: thatin the absence of an intervention (here the exposure to edutainment radio),the trend in the outcomes of villagers (or schools when analyzing the schoolenrollment with school level data) in coverage areas would have been similar,or parallel, to that of the villagers (or schools) in non-coverage areas. Inorder to directly test the hypothesis that the growth paths were the same inthe two groups in the absence of intervention, one needs data on outcome forthe pre-exposure time period. However, with this limited dataset consistingonly of two pooled cross-sections it is not possible to directly verify theassumption but a technique which has been used before by, for example,La Ferrara et al. (2008), Jensen and Oster (2009) and Oster and Millett(2010) could be an alternative option. Because if we are concerned with thepossibility that the results are driven by time-variant unobservable factorswhich influence both the placement of the transmitters and the outcome,and whether the outcome is progressing faster in coverage areas than in non-coverage areas, I can test whether future radio coverage has an impact on theoutcome variables. By introducing an indicator for ”future” radio exposurefrom the KT2006 transmitter, denoted Fvdt in village v of district d in yeart, I can address these issues in the following specification:

wivdt = αCvdtPvdt + θFvdt + βCvd +Xivdtγ +Gvdtδ + ζd + ηt + µm + εivdt (3)

Apart from the future coverage indicator, this specification is identicalto specification (2). If the concern is true, a positive θ coefficient would

15

indicate that individuals who will become exposed in 2006 have relative fasterimprovement in status and school attendance between 2000 and 2005. Thiswould be an indication that the transmitters were placed in areas such thatthe outcome in coverage areas are progressing faster and thus the paralleltrend assumption is invalid. Albeit not perfect, an insignificant θ would bean indication of the validity of this strategy.

5.2.1 School Fixed Effect

The analysis of the children’s school participation is complemented usingschool-level data, a balanced panel with school fixed effects (FE) dating from1999/00 up to 2003/04 school year. The following equation is estimated:

ln(enrol)st =2003∑

t=2001

βtCstYst + ζs + ηt + εst (4)

where ln(enrol)st is the log of enrollment in school s at time t=1999, 2000,2001, 2002, 2003, representing the beginning of each school year; Cst is anindicator of having radio coverage in a given school and time interacted witha year dummy Yst, ζs is a vector of school fixed effects and ηt represents yearfixed effects. The coefficient βtis a FE within-school estimator for 2001 andeach year after exposure, respectively, and is expected to be insignificant in2001 and then increasing with time. This is because household’s enrollmentdecision is in most cased made at the beginning of the school year, so anyeffect of radio exposure on school enrollment is expected to happen withsome lag and probably not in the same period as they become exposed.The reference is average log enrollment in years 1999 and 2000, i.e. beforeexposure. As a placebo test, I run the same specification on pre-exposuredata for school years 1998/1999 and 1999/2000.

Although there are school fixed effects controlling for school character-istics that are constant over time (such as the distance from school to thetransmitter, the location of the school etc.), and year fixed effects controllingfor factors that are constant across schools but vary over time (such as infla-tion etc) in the specification, a concern arises if schools in the coverage areasare progressing faster than schools in non-coverage areas. In such case, thecoefficients would capture simply different time trends and not the effect ofradio exposure per se. This is the advantage of using the school data where Ihave several years of observations before radio exposure. Figure 6 illustratesthe hypothesis that the growth path is the same in two sets of schools in theabsence of exposure. It shows that the average log enrollment of a schoolin a coverage area, within a district, is slightly below that of a school in a

16

non-coverage area during the 1998-2001 periods but after exposure in 2001the log enrollment of schools in coverage areas increases at a faster rate thanthose in non-coverage areas. Each estimate is conducted from a district fixedeffect regression for each year using log enrollment as the dependent vari-able and standard errors clustered at the district level. The trend is clearlyparallel during the pre-exposure period and starts to diverge after 2001 illus-trating the validity of this identification. To formally test this hypothesis ina regression, I re-estimate equation (4) but add an interaction between Fst

which is an indicator for future coverage, i.e. if exposed in year 2006 by theKT2006 transmitter, and a year dummy Yst for school s in time t :

ln(enrol)st =2003∑

t=2001

[βtCstYst + θtFstYst] + ζs + ηt + εst (5)

The coefficient θt is expected to be insignificant for all t suggesting thatthe identification is valid.

6 Results

6.1 Results based on Cross Section Identification

The baseline results based on the cross section identification is presented inTable 9 where the first four columns represent the effect on women’s status(Sonpref, Wifebeat, FinalSay, FS Sch) and the last column the effect onschool attendance (SchAtt). The first panel presents the placebo test for thefive outcome variables before exposure, using DSH 2000, where the estimatesare as expected insignificant for all the outcomes. In the next panel, PanelB, the estimates of same specification but using post-exposure data, DHS2005, are significant and have the expected sign for three of the outcomevariables, while insignificant for the first two attitude related outcomes. Thehypothesis that exposure to edutainment radio will affect son preference andattitudes towards domestic violence is rejected here.

The remainder of the results from Panel B show that exposure to edu-tainment radio is associated with a 0.987 increase in the number of situationsin which a married woman with children has the sole final say in a numberof household decisions and a 26.6 percentage points higher probability tohave the final say in decisions regarding the children’s schooling comparedto women in non-exposed areas. Reference figures for these effects are theaverage means from the non-exposed areas showed by Table 5, i.e. withoutexposure a married woman with children has a final say in on average 2.26situations and the average share of married women with children having a

17

final say in schooling decisions is 6.64 percent. The last column shows thata child aged 6-12 in exposed area has on average a 11.7 percentage pointshigher probability to attend school relative to a child of the same age in anon-exposed area. These effects seems quite large; it is therefore a good ideato compare them with those based on the across time and space identifi-cation strategy in order to see whether the estimates are robust and if thestrategies produce similar results. However, before turning to the next set ofresults, it is important to notice the difference between the both strategies.The advantage with the second strategy, compared with this first strategy,is that the former uses an additional variation (across time) as well as takesinto account more data. On the other hand, the second strategy relies onthe assumption of parallel trend which can only be directly verified using theschool enrollment data.

6.2 Results based on Across Time and Space Identifi-cation

Table 10 presents the district fixed effect estimates from specification 2, with-out and with controls in Panel A.I and Panel A.II, respectively. Panel Bruns the specification 3 with all control variables and district fixed effectsbut adding the future coverage as a test for potential pre-trends. In con-trast to the cross-section results, I find significant effect of radio exposureon all outcome variables and the estimates are overall much lower.39 Theseestimates do not change much when adding the individual and geographicalcontrol variables. The coefficient for the year dummy is significant reflectingthat there is a significant time trend in the outcome variables. This is onlyproblematic if the time trend before exposure was stronger in the exposedareas such that the outcomes in these areas were progressing faster thanthose in non-exposed areas, which, as mentioned, cannot be directly verifiedgiven the data. On the other hand, adding the future coverage indicator inPanel B, I find that its coefficient is insignificant confirming the hypothesis ofequal growth path in the outcomes. In addition, the difference-in-differenceestimates are relatively unchanged throughout this analysis.

The direction of the effect is as expected when interpreting the estimatesfrom Panel A.II: exposure to edutainment radio leads to a 1.1 percentagepoints lower share of desired sons in the households (or a decrease of 2.35percent from the control mean of 47.8 percent), a 0.296 decrease in the num-ber of situations in which wife beating is justified (or a decrease of 29.0

39In general, regressions using both more variation and more data are yielding estimationwith more precisions.

18

percent from the control mean of 1.318), a 0.461 increase in the number ofsituations of having the sole final say (or an increase of 16.9 percent fromthe control mean of 2.263) and a 5.6 percentage points higher probability forhaving the sole final say in children’s schooling decision (or an increase of44.6 percent from the control mean of 6.64) relative to that of the controlgroup, i.e. women in non-exposed areas. Lastly, the probability to attendschool is 6 percentage points higher for children aged 6-12 (or an increase of7.4 percent from the control mean of 75.4 percent) in exposed areas relativeto children aged 6-12 in non-exposed areas.

6.2.1 Heterogeneous Impact

Table 11 investigates the impact on school attendance by gender and ageusing specification (2). Both Panel A and Panel B divide the sample intoa subsample of girls (column 2) and a subsample of boys (column 3). Theformer panel is based on a sample of children aged 6-12 while the latter panelis based on a sample of older children, aged 13-17. I find that both girls andboys in primary school age have on average a 5 and 7 percentage points higherprobability to attend school when located in a coverage area compared withchildren in primary school age located in non-coverage area, respectively.There is no significant difference in the effect between the genders and bothhave quite similar levels in the control areas i.e. girls and boys have onaverage 6.2 and 8.6 percent higher probability to attend school comparedwith the control means of 76.3 and 74.4 percent, respectively. However, Ido not find any significant effect on older children. Given that one of themajor goals of the WMC is to educate the public about children’s right toschooling as well as to inform them about issues related to the drop-out ratein Cambodia, these results suggest that there is a direct impact on bothyounger girls and boys. Additionally, some of these estimated effects mightalso reflect improvements in the women’s status itself.

An interesting part is therefore the mechanism behind the effect: howmuch of the effect was actually caused on account of increased status of themothers or caused by changed attitudes of both parents in the household afterlistening to the Women’s Station FM 102. Using a similar specification as inTable 12, I interact the difference-in-difference estimator Coverage*Year2005with mutually exclusive events of the FS Sch categorical variable to estimatethe total treatment effect of each event on children’s school attendance.40 Ifind that children in coverage areas, after the exposure, living in a householdwhere the woman has the sole final say in schooling decisions do not have a

40Results are not shown here, but available upon request.

19

significant higher probability to attend school compared to children in house-holds where women do not has the sole final say in these decisions. However,these estimations are based on a much smaller sample41, if compared withthat of column 1, Panel A of table 11, suggesting that much of the effecton school attendance are left unexplored in terms of finding the mechanismbehind it.

To see whether radio is a pro-poor tool reaching the most marginalized in-dividuals, I interact the difference-in-difference estimator Coverage*Year2005with indicators of household income quintiles. The results presented in Table12 are based on the same sample as in Panel A of Table 10 where q1 is adummy for a household that belongs to the poorest income quintile and q5 isa dummy for a household that belongs to the richest income quintile. I findan impact on the poorest quintile, except on attitudes towards wife beatingand son preference.

6.2.2 Robustness Checks

To rule out any selective movements of individuals between exposed and non-exposed villages after the time of exposure that could be driving the results, arobustness check is carried out and presented in Table13. However, the DHSdata include only information about how long an individual has lived in theirplace of residence but lacks information about where and which village theyhave moved from. This limitation implies that I cannot observe whetherindividuals have switched “treatment” assignment between the survey years.I therefore re-estimate specification (2) on a subset of individuals who havereported that they have lived at least 5 years in the place of residence toinvestigate how the estimates change. Compared with Panel A.II of Table10, the difference-in-difference estimates do not change much showing thatthe scope for bias due to selective mobility between villages is small. Theones that are notably smaller in magnitude and less significant include theeffect on the number of situations that married women has the sole final sayand the probability that they have the final say in schooling decisions.

Since the coverage areas are around major cities, one concern would bethat it is some characteristics typical for urban areas that are driving theresults and not the radio exposure per se. To prove that this is not thecase, Table 14 which runs specification (2) but adding the interaction termCoverage*Year2005*Urban to test for any significant differences between ur-ban and rural areas. The table shows no evidence of significant differences

41The indicator for final say in schooling decisions, FS Sch, is only available for 3,535observations.

20

between the areas given treatment, suggesting that it is not the individu-als in urban areas driving this. On contrary, most of the effect comes fromindividuals in rural areas.

Although I think it is very unlikely, I cannot rule out that there are stillsome unobservable characteristics that are relevant for the outcome variableswhich have not been considered in this paper. There could still be some im-portant pre-trends in other variables that affect both the coverage and theoutcome variables. For example, the exposed areas could have a relativefaster improvement in the women’s status due to a faster trend in “moder-nity” as these areas are closer to major big cities compared to non-exposedareas. Controlling for an indicator of “modernity”, a dummy for daily tele-vision watching, in all the specifications does not change the results.42

If individuals in covered areas have a higher probability to watch moretelevised contents then a related concern would be that it is this indicator for“modernity” that drives the results. The first column of Table 15 investigatesthis caveat by running the cross-sectional specification (1) while the secondcolumn runs the difference-in-difference specification (2). Both specificationsare using a binary variable TV that takes the value one if the respondentwatches TV almost every day and zero otherwise as the dependent variable.I do no find that the radio exposure increased the daily TV watching. If therehad been some other unobserved factors related to increased TV watchingsuch as trends in modernity that are driving the results, some of the effectwould have been captured by these estimate. As they are insignificant andsmall, it is very unlikely that the effect was driven by such factors.

6.3 Impact on School Enrollment

As already mentioned, a major threat to the validity of the results is if thetransmitters were placed in areas that already had a higher growth path inenrollment, such that the changes we observe are not caused by the interven-tion and would have happened even in the absence of the intervention. Toovercome this major concern, Figure 6 illustrated visually that there wereno pre-existing differential trends in covered and non-covered schools beforeexposure. Moreover, the placebo test in Panel A of Table 16, based on pre-exposure data, shows insignificant within school FE estimates, for both girls’and boys’ log enrollments, implying that the average log enrollment was notsignificantly different between schools in covered and non-covered areas. Thecoefficients of Coverage*Year 1999 which are the FE estimates of interest arealso small in size reassuring the validity of the identification.

42Not shown here but available upon request.

21

To deal with time-constant unobserved heterogeneity problems, the re-sults from estimating specification 4 and 5 are presented in Panel B of Table16 where the dependent variable is the natural logarithm of primary enroll-ment, all grades 1-6 pooled together, and the standard errors are clusteredat the school level. The first three columns show the effects of the currentcoverage and the last three columns show the effect including that of futurecoverage. Consistent with Figure 6, the estimates do not reveal any signif-icant effect in year 2001 but becomes significant and increasing after thatyear. The within-school FE estimate in column (1) indicates that primaryenrollment increased with 5.9 percent in coverage areas in 2002 comparedwith the reference period (i.e. the average log enrollment in 1999 and 2000).A year later in 2003 the average increase in enrollment is 7.3 percent higherthan the reference period. The effect on the same sample of schools but forgirls’ enrollment (column 2) and boys’ enrollment (column 3) is similar forboth genders: about 4-5 percent increase in 2002 and 5-6 percent increasein 2003 compared with the reference period. When adding the Future cov-erage dummy interacted with year dummies, the FE estimates of interest donot change much (column 4-6 of Table 16) and the coefficients of the Futurecoverage and year interaction term are, as expected, insignificant.

These results confirm that there has been a significant effect on the school-ing of children, consistent to what have been observed when analyzing theprobability to attend school at the individual level using the DHS. The maindifference between these two indicators of schooling is that enrollment figuresare reported by the schools themselves i.e. the principal in most cases, andcan thus be subject to over-reporting, while the individual school attendanceis self-reported by the households.43 Thus, even when using different identi-fication strategies I find effects on both girls’ and boys’ schooling indicatingthat the results on school participation are fairly robust.

7 Conclusions

This paper uses two types of identification strategies to analyze the impact ofbeing exposed to edutainment radio on various indicators of women’s statusas well as children’s school participation. Both strategies show that theseradio programs had a significant impact in improving the school attendanceof the younger children. The effect is found both looking at individual at-

43Cheung and Perrotta (2011) stress the importance of considering different measuresof school participation as the individual school attendance does not only capture theenrollment but also actual attendance since some children in Cambodia might be unableto enroll due to incomplete school records but still attend school.

22

tendance reports and at school-level enrollment data. The empirical investi-gation aims at dealing with possible trends in other factors that might drivethe results and potential correlation between the outcome variables and thefuture radio coverage, coming to the conclusion that these issues are not aconcern for the identification strategy.

The estimated effect from the school fixed effect results suggest that theenrollment of children in primary school increased, on average, 5.9 percentmore the first year and then 7.3 percent more the second year after exposureto the radio station compared to the years before exposure. In terms of im-pact the estimated effect is relatively large. Based on the school census, therewere in total 458,863 and 455,710 children enrolled in the coverage areas in1998 and 1999, respectively. Taking the average of the two and multiplying itwith the estimated effects of 5.9 percent and 7.3 percent, respectively, yieldsan approximate impact of 26,979 more children enrolled in primary schoolin year 2002 and 33,381 more children enrolled the following year. This canbe related to the program impact of, for example, a Cambodian scholar-ship program in 2004-2006 given to poor girls in 6th grade, which yieldedapproximately 830 more children in school44, and a large-scale CambodianFood-For-Education (FFE) program in 2002-2003 that induced 54,883 morechildren in school due to the school feeding and deworming program.45 Inaddition to its wide reach impact, this paper also showed that the radio sta-tion had a significant impact on the school participation for both rural andthe poorest households suggesting that radio may also be a pro-poor choice.

However, comparisons between the different interventions should be madewith caution as they do not have exactly the same purposes. For example, theFFE program does not only promote children’s schooling, it also contributesto the nutrition and health of the children as well as to their learning inclassrooms. But if the main purpose is to get children into school then, froma policy perspective, using school interventions such as scholarship or schoolfeeding programs may have a much more effective impact if complementedwith edutainment radio informing the public about their advantages.46 Ifproperly designed, edutainment radio could not only enhance and acceleratebehavior changes (such as increased investment in children’s human capital);

44Filmer and Schady (2008)45See more details in Cheung and Perrotta (2011). Unfortunately, I am unable to

compare the costs of interventions, as I do not have information about the radio station’sproduction cost.

46This is because pure didactic interventions or programs are usually incapable to retainthe audiences’ interest, even if well produced, and could thus successfully be accompaniedby an edutainment strategy as part of an overall development communication approach.See Barker et al. (2005)

23

they also attempt to promote changes in attitudes and social norms relatedto human capital investments (such as information about the importance ofeducation as well as the returns to education). Changing such attitudes couldtherefore promote both a higher school entry rate as well as a lower drop-outrate if the radio station managed to affect the social norms leading to a longerstay in school. Moreover, if attitudes and norms change, the correspondingchange in behavior might be sustained over time beyond the duration ofthe scholarship or school feeding intervention. However, this paper did notinvestigate the direct changes in attitude towards human capital investmentsbut on the other hand showed that part of the change in schooling-relatedbehavior is driven by deeper changes in attitudes and status of women.

Finally, suggestive evidence indicates that exposure to edutainment ra-dio also had affected gender related attitudes. This is interesting in itselfas many women’s stations are today starting up with funds from interna-tional organizations such as the UN. Compared with the impact on primaryschool enrollment, the effect was not immediate as households might takethe decision on investing in children’s human capital the school year after.The impact on enrollment was greatest some two to three years after theexposure to these radio programs. In relation to changing gender relatedattitudes, convincing households to invest in children’s schooling might bean easier task by enlightening the society about the importance of schoolingfor their future accumulated earnings. But trying to make structural adjust-ments by changing gender related attitudes in a society that is characterizedby a patriarchal system might take a much longer time.

24

ReferencesAlison Mee, Graham Haylor, S.V., Savage, W., 2003. ”Information Access Survey, Cambodia 2003”.

Survey. STREAM.

Barker, K., Sabido, M., Population Media Center (Shelburne, V., 2005. Soap Operas for Social Change toPrevent HIV/AIDs: A training guide for journalists and media personnel. Population Media Center.

CCIM, 2009. ”Enhancing Independent Media in Cambodia: An Ethics Perspective”. Technical Report.Cambodian Center for Independent Media (CCIM).

Chamberlin, K., Luebbers, R., 1982. An evaluation of longley-rice and gtd propagation models. Antennasand Propagation, IEEE Transactions on 30, 1093–1098.

Cheung, M., Perrotta, M., 2011. The impact of a food for education program on schooling in cambodia.Institute for International Economic Studies Seminar paper No. 766 .

Dahlerup, D., 2010. How to achieve the cambodian millennium development goal 3 on eliminating genderdisparity in public institutions by 2015 .

Della Vigna, S., Kaplan, E., 2007. The fox news effect: Media bias and voting. Quarterly Journal ofEconomics 122.

Deutsch Karlekar, K., 2011. ”Freedom of the Press 2011: Signs of change amid repression”. TechnicalReport. Freedom House.

DHS, 2001. ”Cambodia: DHS, 2000 - Final Report (English)”. DHS Final Reports. MEASURE DHS.

DHS, 2006. ”Cambodia: DHS, 2005 - Final Report (English)”. DHS Final Reports. MEASURE DHS.

Duflo, E., 2005. Gender equality in development. BREAD Policy paper 11.

Edman, B., 2000. Trapped in the past, seeking out a futurea study on the cambodian media sector, in: Astudy commissioned by Sida (Swedish International Development Agency) and Forum Syd.

Ellington, H., Addinall, E., Hately, M., 1980. The physics of television broadcasting. Physics Education15, 222.

Filmer, D., Schady, N., 2008. Getting girls into school: Evidence from a scholarship program in cambodia.Economic Development and Cultural Change 56, 581–617.

Gentzkow, M., Shapiro, J., 2004. Media, education and anti-americanism in the muslim world. Journalof Economic Perspectives 18, 117–134.

Hufford, G., 1995. ”The ITS Irregular Terrain Model, version 1.2. 2 The Algorithm”. Technical Report.Report of Institute for Telecommunication Sciences, Boulder USA.

for Human Rights, D.I., 2003. ”Annual Report 2003”. Report. The Danish Institute for Human Rights.

Izett, S., Toubia, N., 1999. Learning about social change: a research and evaluation guidebook usingfemale circumcision as a case study. Research Action & Information Network for Bodily Integrity ofWomen.

Jensen, R., Oster, E., 2009. ”the power of tv: cable television and womens status in india.”. QuarterlyJournal of Economics 124, 1057–1094.

Khalil, M., Dongier, P., Qiang, C., 2009. Information and communications for development 2009 report:Extending reach and increasing impact. Washington, DC: World Bank .

King, E., Mason, A., 2001. Engendering development: Through gender equality in rights, resources, andvoice. Washington, DC and New York: The World Bank and Oxford University Press. A World BankPolicy Research Report .

25

La Ferrara, E., Chong, A., Duryea, S., Dept, I.A.D.B.R., commerciale Luigi Bocconi, U., for Economic Re-search, I.G.I., 2008. Soap Operas and Fertility: Evidence from Brazil. Centre for Economic PolicyResearch.

Ledgerwood, J., 1996. Women in development: Cambodia. Asian Development Bank, Manila .

Longley, A., 1968. ”Prediction of Tropospheric Radio Transmission Loss Over Irregular Terrain. A Com-puter Method-1968”. Technical Report. DTIC Document.

Marseille, A., 2008. Cambodian radio and corruption: Quantitative and qualitative monitoring oncorruption-related stories in 4 khmer-language national radio programs .

Mehta, H., 1997. Cambodia silenced: The press under six regimes. White Lotus.

Morris, N., 2001. Bridging the gap: An examination of diffusion and participatory approaches in devel-opment communication. The CHANGE Project/USAID .

Olken, B., 2009. Do television and radio destroy social capital? evidence from indonesian villages. Amer-ican Economic Journal: Applied Economics, forthcoming .

Oster, E., Millett, M., 2010. ”do call centers promote school enrollment? evidence from india”.

Piotrow, P., Rimon, J., Winnard, K., Kincaid, D., Huntington, D., Convisser, J., 1990. Mass media familyplanning promotion in three nigerian cities. Studies in Family Planning 21, 265–274.

Qian, N., 2008. Missing women and the price of tea in china: The effect of sex-specific earnings on seximbalance. The Quarterly Journal of Economics 123, 1251.

Rogers, E., 1995. Diffusion of innovations. Free Pr.

Rosette, J., 2010. An overview of ee (entertainment-education) and its application in the cambodiancontext .

Singhal, A., Papa, M., Sharma, D., Pant, S., Worrell, T., Muthuswamy, N., Witte, K., 2006. Entertain-ment education and social change the communicative dynamics of social capital. Journal of CreativeCommunications 1, 1.

Singhal, A., Rogers, E., 2001. The entertainment-education strategy in communication campaigns. Publiccommunication campaigns 3, 343–356.

Singhal, A., Rogers, E., 2004. The status of entertainment-education worldwide. Entertainment-educationand social change: History, research, and practice , 3–20.

Stromberg, D., 2004. Radio’s impact on public spending. Quarterly Journal of Economics 119, 189–221.

Svensson, J., Yanagizawa, D., 2009. Getting prices right: The impact of the market information servicein uganda. Journal of the European Economic Association 7, 435–445.

UNIFEM, W., ADB, U., 2004. Dfid/uk. 2004. A Fair Share for Women: Cambodia Gender Assessment .

WMC, 2004. ”Women’s Media Centre of Cambodia - WMC - Strategic Plan 2005-2009”. Report. Women’sMedia Centre of Cambodia.

WMC, 2006. ”FM102 National Radio Survey: Final Report, 2006”. Survey. Women’s Media Centre ofCambodia.

WMC, 2008. ”Annual Report 2008”. Report. Women’s Media Centre of Cambodia.

WMC, 2009a. ”Annual Report 2009”. Report. Women’s Media Centre of Cambodia.

WMC, 2009b. ”Four Years Impact Evaluation Report 2005-2009”. Report. Women’s Media Centre ofCambodia.

26

Fig

ure

1:R

adio

Cov

erag

efr

omth

eP

hnom

Pen

hA

nte

nna

(PP

1999

and

PP

2001

)

Cov

erag

efr

omP

P19

99C

over

age

from

PP

2001

Note

:T

he

tria

ngle

sare

the

loca

tion

of

the

thre

era

dio

tran

smit

ters

:K

T,

PP

an

dS

R.

On

lyth

eco

ver

age

from

the

Ph

nom

Pen

htr

an

smit

ter

issh

ow

nh

ere

wh

ere

red

colo

ron

the

map

refe

rto

sign

al

1(7

5-1

00

dB

myV

/m

),p

ink

refe

rto

sign

al

2(6

0-7

5d

Bm

yV

/m

)an

db

lue

refe

rto

sign

al

3(4

5-6

0d

Bm

yV

/m

).S

ou

rce:

Th

era

dio

cover

age

iscr

eate

din

CR

C-C

OV

LA

Ban

dim

port

edin

Arc

GIS

wit

had

min

istr

ati

ve

dis

tric

tb

ou

nd

ari

esfr

om

Glo

bal

Adm

inis

trati

ve

Un

itL

ayer

s(G

AU

L).

27

Fig

ure

2:R

adio

Cov

erag

efr

omth

eK

amp

ong

Thom

Ante

nna

(KT

2001

and

KT

2006

)

Cov

erag

efr

omK

T20

01C

over

age

from

KT

2006

Note

:T

he

tria

ngle

sare

the

loca

tion

of

the

thre

era

dio

tran

smit

ters

:K

T,

PP

an

dS

R.

On

lyth

eco

ver

age

from

the

Kam

pon

gT

hom

tran

smit

ter

issh

ow

nh

ere

wh

ere

red

colo

ron

the

map

refe

rto

sign

al

1(7

5-1

00

dB

myV

/m

),p

ink

refe

rto

sign

al

2(6

0-7

5d

Bm

yV

/m

)an

db

lue

refe

rto

sign

al

3(4

5-6

0d

Bm

yV

/m

).S

ou

rce:

Th

era

dio

cover

age

iscr

eate

din

CR

C-C

OV

LA

Ban

dim

port

edin

Arc

GIS

wit

had

min

istr

ati

ve

dis

tric

tb

ou

nd

ari

esfr

om

Glo

bal

Ad

min

istr

ati

ve

Un

itL

ayer

s(G

AU

L).

28

Fig

ure

3:R

adio

Cov

erag

efr

omth

eSva

yR

ieng

Ante

nna

(SR

2001

)an

dA

dm

inis

trat

ive

Dis

tric

tB

oundar

ies

Map

Cov

erag

efr

omSR

2001

Adm

inis

trat

ive

Bou

ndar

ies

Map

Note

:T

he

tria

ngle

sare

the

loca

tion

of

the

thre

era

dio

tran

smit

ters

:K

T,

PP

an

dS

R.

On

lyth

eco

ver

age

from

the

Svey

Rie

ng

issh

ow

nh

ere

wh

ere

red

colo

ron

the

map

refe

rto

sign

al

1(7

5-1

00

dB

myV

/m

),p

ink

refe

rto

sign

al

2(6

0-7

5d

Bm

yV

/m

)an

db

lue

refe

rto

sign

al

3(4

5-6

0d

Bm

yV

/m

).S

ou

rce:

Th

era

dio

cover

age

iscr

eate

din

CR

C-C

OV

LA

Ban

dim

port

edin

Arc

GIS

wit

had

min

istr

ati

ve

dis

tric

tb

ou

nd

ari

esfr

om

Glo

bal

Ad

min

istr

ati

ve

Un

itL

ayer

s(G

AU

L)

(lef

tfi

gu

re).

Map

from

Glo

bal

Ad

min

istr

ati

ve

Un

itL

ayer

s(G

AU

L)

wit

hd

istr

ict

bou

nd

ari

esan

dG

PS

data

from

DH

Sim

port

edto

Arc

GIS

(rig

ht

figu

re).

29

Fig

ure

4:E

leva

tion

and

Slo

pe

Map

sof

Cam

bodia

Ele

vati

onSlo

pe

Note

:A

wh

ite

dot

isa

DH

S2005

clu

ster

s/villa

ge,

the

tria

ngle

sare

the

loca

tion

of

the

thre

era

dio

ante

nn

as:

KT

2001,

PP

2001

an

dS

R2001.

Sou

rce:

Shu

ttle

Rad

ar

Top

ogra

phy

Mis

sion

an

dG

PS

data

from

DH

S.

30

Figure 5: Pattern of Radio Listening by Signal Strength

0.53

0.47

0.40

0.60

0.46

0.540.59

0.41

0.2

.4.6

.8S

hare

No Coverage Coverage

0 1 2 3

Never/Less than once a week

Every day/At least every week

Note: Recoded from DHS 2005, the survey question is “Do you listen to the radioalmost every day, at least once a week, less than once a week or not at all?”. Thenumber 0 on the x-axis refer to no signal, 1 refer to signal 1, 2 refer to signal 2 and3 refer to signal 3.

31

Figure 6: Trends in Primary Enrollment

5180

5180

5180

5180

51805180

−.2

−.1

0.1

Cha

nge

in lo

g−en

rollm

ent

1998−99 1999−00 2000−01 2001−02 2002−03 2003−04

Coverage No Coverage

Note: The dotted lines show the 95% confidence interval.The labels report the number of observations.

(a) All

5180

51805180

5180

5180

5180

−.2

−.1

0.1

.2C

hang

e in

log−

enro

llmen

t

1998−99 1999−00 2000−01 2001−02 2002−03 2003−04

Coverage No Coverage

Note: The dotted lines show the 95% confidence interval.The labels report the number of observations.

(b) Girls

5180

5180

5180

5180

5180 5180

−.2

−.1

0.1

Cha

nge

in lo

g−en

rollm

ent

1998−99 1999−00 2000−01 2001−02 2002−03 2003−04

Coverage No Coverage

Note: The dotted lines show the 95% confidence interval.The labels report the number of observations.

(c) Boys

Note: Schools in non coverage areas are never exposed to the radio signals. Each estimate is a crosssection regression in that year including district fixed effects. Robust standard errors clustered at thedistrict level. Source: EMIS

32

Tab

le1:

Defi

nit

ion

ofO

utc

ome

Var

iable

s

Defi

nit

ion

Rec

od

edfr

om

surv

eyqu

esti

on

Son

pre

fT

he

des

ired

shar

eof

son

sin

the

hou

seh

old

”H

owm

any

of

thes

ech

ild

ren

(liv

ing

chil

dre

n)

wou

ldyo

uli

ke

tob

eb

oys,

how

many

wou

ldyo

uli

keto

be

gir

lsan

dfo

rh

owm

any

wou

ldth

ese

xn

ot

matt

er?”

Wif

ebea

tT

he

nu

mb

erof

situ

atio

ns

inw

hic

hw

ife

bea

t-in

gis

just

ified

”S

om

etim

esa

hu

sban

dis

an

noy

edor

an

ger

edby

thin

gs

wh

ich

his

wif

ed

oes

.In

you

rop

inio

n,

isa

hu

sban

dju

stifi

edin

hit

tin

gor

bea

tin

gh

isw

ife

inth

efo

llow

ing

situ

ati

on

s:a)

ifsh

egoes

ou

tw

ith

ou

tte

llin

gh

im,

b)

ifsh

en

egle

cts

chil

dre

n,

c)if

she

arg

ues

wit

hh

im,

d)

ifsh

ere

fuse

sh

ave

sex

wit

hh

im,

or

e)if

food

isla

teor

not

wel

lp

rep

are

d?”

Fin

alS

ayT

he

nu

mb

erof

dec

isio

ns

inw

hic

hth

ere

spon

-d

ent

has

aso

lefi

nal

say

”W

ho

inyo

ur

fam

ily

usu

all

yh

as

the

fin

al

say

on

”..

.a)

makin

gla

rge

hou

se-

hold

pu

rch

ase

s?a)

makin

gla

rge

hou

seh

old

purc

hase

s,b

)m

akin

gh

ou

seh

old

pu

r-ch

ase

sfo

rd

aily

nee

ds,

b)

wh

eth

eryou

shou

ldd

ow

ork

toea

rnm

on

ey,

c)yo

ur

own

hea

lth

care

,d)

wh

eth

erto

use

contr

ace

pti

on

s,vis

its

tofa

mil

y,fr

iend

s,or

rela

tive

s,e)

any

dec

isio

ns

ab

ou

tch

ild

ren

’ssc

hooli

ng

an

df)

dec

idin

gw

hat

tod

ow

ith

mon

eyhu

sban

dea

rns?

FS

Sch

Ad

um

my

for

hav

ing

the

sole

fin

al

say

ind

e-ci

sion

sab

out

chil

dre

n’s

sch

ool

ing

”W

ho

inyou

rfa

mil

yu

sual

lyh

as

the

fin

al

say

on

any

dec

isio

ns

ab

ou

tch

ild

ren’s

sch

ooli

ng?”

Sch

Att

Ad

um

my

ifa

chil

dag

ed6-

12is

att

end

ing

sch

ool

”D

id(N

AM

E)

att

end

sch

ool

at

any

tim

ed

uri

ng

the

pre

vio

us

sch

ool

year?

”

33

Table 2: Sampling Procedure for the Individual Level (DHS) Data

The whole DHS survey is considered in the analysis with the exception of following restric-tions:

i) Observations with missing values or answers such as ”don’t know” or ”decision notmade/not applicable” in either the dependent variables or in any of the control vari-ables as these are dropped from the analysis.

ii) I also exclude observations when I do not know whether they are exposed to the radiosignals or not. Such observations include visitors or individuals that have given anincomplete answer to whether they are visitors or not (80 and 153 observations inDHS 2000 and DHS 2005, respectively) and are dropped from the analysis as theycould potentially be visitor to the household for just one day, the survey day, and livein an unexposed area leading to an underestimation of the effect.

iii) Nine villages (310 observations) from DHS 2005 that lack geographical informationabout their location are also excluded since without this information I am unable toidentify whether they are exposed or not. This is however not a restriction in theDHS 2000 as the geographical information is complete in this survey.

iv) Moreover, observations that are within coverage area from PP1999 station, (concern-ing 1,303 observations from DHS 2000 and 1,394 from DHS 2005) are also droppedas these individuals are exposed to radio signals in both surveys;

v) As well as 23 and 30 villages (or 755 and 853 observations) from DHS 2000 and DHS2005, respectively, that are located just at the boarder of having coverage or not, i.e.they are located in communes with semi-coverage.

An additional sample restriction is to only include individuals that are aged 18 or above,i.e. excluding children, when analyzing the impact on women’s status; and to only includechildren aged 6-12 when analyzing the impact on primary school attendance. Furthermore,as the FinalSay variable is based on the FS Sch variable and the survey question defining thelatter variable is only addressed to women with children, these samples only include marriedwomen with children (1,787 and 1,748 observations from DHS 2000 and 2005, respectively).The inclusion of only married women, i.e. excluding widows and single mothers, is importantas I want to estimate the women’s probability to have the sole final say in schooling decisionsabout their children in relation to that of their living husband.

34

Tab

le3:

Sum

mar

ySta

tist

ics

for

the

Dep

enden

tV

aria

ble

s

(a)

Bef

ore

Exp

osu

re

mea

nsd

min

max

ob

sN

oC

over

age

Son

pre

f0.

490

0.13

50

17014

Wif

eBea

t0.

804

1.33

60

51796

Fin

alS

ay2.

080

1.89

00

81139

FS

Sch

0.10

50.

307

01

1139

Sch

Att

0.82

60.

379

01

5153

Sch

Att

f0.

824

0.38

10

12452

Sch

Att

m0.

829

0.37

60

12702

Cov

erag

eS

onp

ref

0.49

90.

136

01

3851

Wif

eBea

t1.

009

1.37

70

5978

Fin

alS

ay1.

986

1.79

70

8648

FS

Sch

0.08

280.

276

01

648

Sch

Att

0.76

20.

426

01

3214

Sch

Att

f0.

771

0.42

00

11590

Sch

Att

m0.

753

0.43

20

11624

Tot

alS

onp

ref

0.49

40.

136

01

10865

Wif

eBea

t0.

897

1.35

80

52774

Fin

alS

ay2.

037

1.84

80

81787

FS

Sch

0.09

490.

293

01

1787

Sch

Att

0.79

60.

403

01

8367

Sch

Att

f0.

798

0.40

10

14042

Sch

Att

m0.

794

0.40

40

14326

Not

e:D

HS

2000

.

(b)

Aft

erE

xp

osu

re

mea

nsd

min

max

ob

sN

oC

over

age

Son

pre

f0.4

78

0.1

33

01

8001

Wif

eBea

t1.3

18

1.3

12

04

2037

Fin

alS

ay2.2

63

1.6

68

08

1182

FS

Sch

0.0

664

0.2

49

01

1182

Sch

Att

0.7

54

0.4

31

01

6478

Sch

Att

f0.7

63

0.4

25

01

3256

Sch

Att

m0.7

44

0.4

37

01

3222

Cov

erage

Son

pre

f0.4

77

0.1

31

01

3819

Wif

eBea

t1.2

37

1.3

55

04

979

Fin

alS

ay2.6

26

1.7

20

08

566

FS

Sch

0.0

906

0.2

87

01

566

Sch

Att

0.7

80

0.4

15

01

3296

Sch

Att

f0.7

83

0.4

12

01

1666

Sch

Att

m0.7

76

0.4

17

01

1630

Tota

lS

on

pre

f0.4

77

0.1

32

01

11820

Wif

eBea

t1.2

79

1.3

33

04

3016

Fin

alS

ay2.4

38

1.7

03

08

1748

FS

Sch

0.0

780

0.2

68

01

1748

Sch

Att

0.7

66

0.4

23

01

9774

Sch

Att

f0.7

73

0.4

19

01

4922

Sch

Att

m0.7

60

0.4

27

01

4852

Note

:D

HS

2005.

35

Table 4: Summary Statistics for the Control Variables

mean sd min max obsNo CoverageElevation (1000 m) 0.0305 0.0502 0 0.675 7404Slope (percent) 0.400 0.829 0 10.32 7404Distance to nearest WMC transmitter (km) 147.8 81.11 1.201 328.6 7404Distance to nearest Province Capital (km) 28.56 22.45 0.357 105.2 7404Urban areas (0/1) 0.245 0.430 0 1 7404Dummy for having electricity (0/1) 0.229 0.420 0 1 7404Age 31.76 9.036 18 49 7404Years of education 2.124 1.962 0 6 7404Wealth quintile, poorest (0/1) 0.226 0.418 0 1 7404Wealth quintile, poorer (0/1) 0.206 0.404 0 1 7404Wealth quintile, middle (0/1) 0.172 0.377 0 1 7404Wealth quintile, richer (0/1) 0.159 0.365 0 1 7404Wealth quintile, richest (0/1) 0.238 0.426 0 1 7404CoverageElevation (1000 m) 0.0170 0.0173 0.00300 0.112 3999Slope (percent) 0.196 0.255 0 1.607 3999Distance to nearest WMC transmitter (km) 44.42 17.51 1.571 85.89 3999Distance to nearest Province Capital (km) 27.35 13.58 0.550 60.12 3999Urban areas (0/1) 0.0386 0.193 0 1 3999Dummy for having electricity (0/1) 0.0361 0.186 0 1 3999Age 32.33 8.914 18 49 3999Years of education 2.278 1.922 0 6 3999Wealth quintile, poorest (0/1) 0.246 0.431 0 1 3999Wealth quintile, poorer (0/1) 0.222 0.415 0 1 3999Wealth quintile, middle (0/1) 0.242 0.429 0 1 3999Wealth quintile, richer (0/1) 0.220 0.415 0 1 3999Wealth quintile, richest (0/1) 0.0695 0.254 0 1 3999TotalElevation (1000 m) 0.0245 0.0397 0 0.675 11403Slope (percent) 0.309 0.648 0 10.32 11403Distance to nearest WMC transmitter (km) 101.8 80.18 1.201 328.6 11403Distance to nearest Province Capital (km) 28.02 19.03 0.357 105.2 11403Urban areas (0/1) 0.153 0.360 0 1 11403Dummy for having electricity (0/1) 0.143 0.350 0 1 11403Age 32.01 8.986 18 49 11403Years of education 2.192 1.946 0 6 11403Wealth quintile, poorest (0/1) 0.235 0.424 0 1 11403Wealth quintile, poorer (0/1) 0.213 0.409 0 1 11403Wealth quintile, middle (0/1) 0.203 0.402 0 1 11403Wealth quintile, richer (0/1) 0.186 0.389 0 1 11403Wealth quintile, richest (0/1) 0.163 0.369 0 1 11403

Note: DHS 2000.

36

Table 5: Summary Statistics for the Control Variables

mean sd min max obsNo CoverageElevation (1000 m) 0.0373 0.0565 0 0.691 8254Slope (percent) 0.526 1.148 0 15.30 8254Distance to nearest WMC transmitter (km) 163.7 70.07 35.52 331.3 8254Distance to nearest Province Capital (km) 32.91 36.54 0.363 269.8 8254Urban areas (0/1) 0.205 0.404 0 1 8254Dummy for having electricity (0/1) 0.203 0.403 0 1 8254Age 31.87 9.457 18 49 8254Years of education 2.278 1.976 0 11 8254Wealth quintile, poorest (0/1) 0.218 0.413 0 1 8254Wealth quintile, poorer (0/1) 0.224 0.417 0 1 8254Wealth quintile, middle (0/1) 0.192 0.394 0 1 8254Wealth quintile, richer (0/1) 0.183 0.387 0 1 8254Wealth quintile, richest (0/1) 0.183 0.387 0 1 8254CoverageElevation (1000 m) 0.0167 0.0164 0.00200 0.0960 3951Slope (percent) 0.200 0.254 0 2.119 3951Distance to nearest WMC transmitter (km) 43.13 18.76 1.387 84.45 3951Distance to nearest Province Capital (km) 28.90 14.71 0.703 62.76 3951Urban areas (0/1) 0.0530 0.224 0 1 3951Dummy for having electricity (0/1) 0.112 0.316 0 1 3951Age 32.53 9.237 18 49 3951Years of education 2.614 1.880 0 12 3951Wealth quintile, poorest (0/1) 0.188 0.390 0 1 3951Wealth quintile, poorer (0/1) 0.213 0.410 0 1 3951Wealth quintile, middle (0/1) 0.249 0.432 0 1 3951Wealth quintile, richer (0/1) 0.230 0.421 0 1 3951Wealth quintile, richest (0/1) 0.120 0.325 0 1 3951TotalElevation (1000 m) 0.0275 0.0437 0 0.691 12205Slope (percent) 0.371 0.865 0 15.30 12205Distance to nearest WMC transmitter (km) 106.3 79.77 1.387 331.3 12205Distance to nearest Province Capital (km) 31.00 28.40 0.363 269.8 12205Urban areas (0/1) 0.133 0.339 0 1 12205Dummy for having electricity (0/1) 0.160 0.367 0 1 12205Age 32.19 9.358 18 49 12205Years of education 2.438 1.938 0 12 12205Wealth quintile, poorest (0/1) 0.204 0.403 0 1 12205Wealth quintile, poorer (0/1) 0.219 0.413 0 1 12205Wealth quintile, middle (0/1) 0.219 0.414 0 1 12205Wealth quintile, richer (0/1) 0.205 0.404 0 1 12205Wealth quintile, richest (0/1) 0.153 0.360 0 1 12205

Note: DHS 2005.

37

Table 6: Summary Statistics of Schools

Year No Coverage Coverage Total1998 mean sd obs mean sd obs mean sd obsTotal enrollment 335.8 399.0 4024 439.9 474.9 1043 357.2 417.8 5067Girl enrollment 153.2 186.3 4024 199.6 221.1 1043 162.7 194.8 5067Boy enrollment 182.6 214.4 4024 240.3 255.5 1043 194.5 224.7 5067

Year No Coverage Coverage Total1999 mean sd obs mean sd obs mean sd obsTotal enrollment 376.5 409.4 4212 470.8 359.5 968 394.1 402.2 5180Girl enrollment 172.5 190.0 4212 215.0 166.1 968 180.5 186.5 5180Boy enrollment 204.0 220.9 4212 255.8 195.7 968 213.7 217.3 5180

Year No Coverage Coverage Total2000 mean sd obs mean sd obs mean sd obsTotal enrollment 402.2 417.7 4212 509.0 372.2 968 422.1 411.7 5180Girl enrollment 187.7 196.5 4212 236.7 173.3 968 196.9 193.3 5180Boy enrollment 214.4 222.9 4212 272.4 201.3 968 225.2 220.2 5180

Year No Coverage Coverage Total2001 mean sd obs mean sd obs mean sd obsTotal enrollment 432.7 424.7 4212 539.6 375.4 968 452.7 418.0 5180Girl enrollment 201.3 199.3 4212 250.3 177.8 968 210.4 196.4 5180Boy enrollment 231.5 227.1 4212 289.3 199.9 968 242.3 223.4 5180

Year No Coverage Coverage Total2002 mean sd obs mean sd obs mean sd obsTotal enrollment 435.5 415.5 4212 548.0 373.8 968 456.5 410.4 5180Girl enrollment 203.7 196.3 4212 255.1 174.9 968 213.3 193.5 5180Boy enrollment 231.8 220.9 4212 292.8 201.2 968 243.2 218.6 5180

Year No Coverage Coverage Total2003 mean sd obs mean sd obs mean sd obsTotal enrollment 434.2 400.5 4212 545.1 357.6 968 455.0 395.1 5180Girl enrollment 204.6 190.9 4212 257.3 170.7 968 214.5 188.4 5180Boy enrollment 229.6 211.2 4212 287.8 188.8 968 240.5 208.4 5180

Note: EMIS.

38

Table 7: Potential Determinants of Radio Coverage(1) (2)

Before Exposure After ExposureElevation (1000 m) -0.294 -0.988∗∗

(0.408) (0.473)

Slope (percent) -0.012 -0.010(0.014) (0.011)

Dummy for having electricity (0/1) 0.002 -0.002(0.005) (0.004)

Distance to nearest WMC transmitter (km) -0.003∗ -0.001∗∗∗

(0.001) (0.000)

Distance to nearest Province Capital (km) -0.001 -0.003∗∗∗

(0.001) (0.000)

Age -0.000 0.000(0.000) (0.000)

Years of education 0.001 0.000(0.001) (0.000)

Wealth quintile, poorer (0/1) 0.001 0.002(0.003) (0.005)

Wealth quintile, middle (0/1) 0.005 0.004(0.003) (0.007)

Wealth quintile, richer (0/1) 0.007 0.004(0.005) (0.006)

Wealth quintile, richest (0/1) 0.004 0.009(0.005) (0.007)

Observations 11403 12205R2 0.962 0.986Districts 154 164Clusters 418 480

Note: The dependent variable is an indicator of being exposed to the radio signals of either the KT2001,PP2001 or SR2001 transmitter. Column (1) is based on DHS 2000, before the exposure, and column (2)is based on DHS 2005, after the exposure. Standard errors clustered at the village level in parentheses. ∗

p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. All the regressions include district and month fixed effects.

39

Table 8: Orthogonality of the Dependent Variables Before ExposurePanel A: Results using DHS 2000

Indep. var: Sonpref Wifebeat FinalSay FS Sch SchAtt0.009 0.002 0.002 0.006 -0.004

(0.008) (0.002) (0.001) (0.009) (0.004)Observations 10865 2774 1787 1787 8367R2 0.955 0.957 0.958 0.958 0.960Districts 154 154 154 154 151Clusters 418 418 413 413 407

Panel B: Results using DHS 2005

Indep. var: Sonpref Wifebeat FinalSay FS Sch SchAtt(1) (2) (3) (4) (5)

0.007 0.004 0.001 -0.000 -0.001(0.007) (0.002) (0.001) (0.006) (0.003)

Observations 11820 3016 1748 1748 9774R2 0.689 0.725 0.731 0.731 0.719Districts 164 164 162 162 163Clusters 480 480 473 473 478

Note: The dependent variable in Panel A is an indicator of being exposed to the radiosignals of either KT2001, PP2001 and SR2001 transmitter using DHS 2000 data. Thedependent variable in Panel B is an indicator of being exposed to the radio signals ofKT2006 transmitter using DHS 2005 data. Standard errors clustered at the villagelevel in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. All the regressions includedistrict and month fixed effect, individual and geographical controls.

Table 9: Results from Cross Section IdentificationPanel A: Placebo using DHS 2000

(1) (2) (3) (4) (5)Sonpref Wifebeat FinalSay FS Sch SchAtt

Coverage 0.020 0.437 0.620 0.030 -0.024(0.013) (0.273) (0.380) (0.068) (0.061)

Observations 10865 2774 1787 1787 8367R2 0.031 0.299 0.207 0.125 0.272Districts 154 154 154 154 151Clusters 418 418 413 413 407

Panel B: Results using DHS 2005

(1) (2) (3) (4) (5)Sonpref Wifebeat FinalSay FS Sch SchAtt

Coverage 0.009 -0.164 0.987∗ 0.266∗∗∗ 0.117∗∗

(0.016) (0.298) (0.561) (0.089) (0.045)Observations 11820 3016 1748 1748 9774R2 0.035 0.179 0.315 0.178 0.368Districts 164 164 162 162 163Clusters 480 480 473 473 478

Note: Standard errors clustered at the village level in parentheses. ∗ p < 0.10, ∗∗

p < 0.05, ∗∗∗ p < 0.01. All the regressions include district and month fixed effects,individual and geographical controls.

40

Table 10: Difference-in-Difference EstimatesPanel A.I: Simple DD with District FE

(1) (2) (3) (4) (5)Sonpref Wifebeat FinalSay FS Sch SchAtt

Coverage * Year 2005 -0.010∗∗ -0.303∗∗∗ 0.459∗∗∗ 0.062∗∗∗ 0.067∗∗∗

(0.005) (0.100) (0.144) (0.021) (0.019)

Year 2005 -0.012∗∗∗ 0.550∗∗∗ 0.169∗ -0.038∗∗∗ -0.052∗∗∗

(0.003) (0.064) (0.087) (0.013) (0.012)

Coverage 0.003 -0.253 -0.547∗ -0.033 -0.060(0.008) (0.224) (0.287) (0.041) (0.051)

Observations 22685 5790 3535 3535 18141R2 0.025 0.156 0.160 0.085 0.055Districts 174 174 173 173 174Clusters 538 538 536 536 536

Panel A.II: Effect of Current Radio Coverage

(1) (2) (3) (4) (5)Sonpref Wifebeat FinalSay FS Sch SchAtt

Coverage * Year 2005 -0.011∗∗ -0.296∗∗∗ 0.461∗∗∗ 0.056∗∗∗ 0.060∗∗∗

(0.005) (0.106) (0.146) (0.021) (0.016)

Year 2005 -0.015∗∗ 0.563∗∗∗ -0.361 -0.079∗∗∗ -0.208∗∗∗

(0.006) (0.129) (0.224) (0.026) (0.024)

Coverage 0.003 -0.278 -0.442 -0.050 -0.064∗

(0.008) (0.247) (0.311) (0.044) (0.037)Observations 22685 5790 3535 3535 18141R2 0.026 0.163 0.180 0.097 0.305Districts 174 174 173 173 174Clusters 538 538 536 536 536

Panel B: Effect of Future Radio Coverage

(1) (2) (3) (4) (5)Sonpref Wifebeat FinalSay FS Sch SchAtt

Coverage * Year 2005 -0.011∗∗ -0.287∗∗∗ 0.475∗∗∗ 0.057∗∗∗ 0.061∗∗∗

(0.005) (0.107) (0.148) (0.021) (0.016)

Future 0.012 0.433 0.616 0.064 0.049(0.014) (0.429) (0.496) (0.079) (0.058)

Year 2005 -0.015∗∗ 0.568∗∗∗ -0.364 -0.080∗∗∗ -0.207∗∗∗

(0.006) (0.128) (0.225) (0.026) (0.024)

Coverage 0.007 -0.117 -0.226 -0.027 -0.045(0.011) (0.282) (0.405) (0.054) (0.048)

Observations 22685 5790 3535 3535 18141R2 0.026 0.163 0.180 0.097 0.305Districts 174 174 173 173 174Clusters 538 538 536 536 536

Note: Panel A.II provides the simple difference in difference estimates with districtsfixed effects without any individual or geographical controls while Panel A.II includesboth district and month fixed effects, individual and geographical controls. PanelB tests for impact of future coverage adding a dummy for villages located in futurecoverage areas with the same set of controls as in Panel A.II. Standard errors clusteredat the village level in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

41

Table 11: Impact on School Attendance by Gender and AgePanel A: Children aged 6-12

(1) (2) (3)All Girls Boys

Coverage * Year 2005 0.060∗∗∗ 0.050∗∗ 0.070∗∗∗

(0.016) (0.020) (0.021)

Year 2005 -0.208∗∗∗ -0.186∗∗∗ -0.223∗∗∗

(0.024) (0.030) (0.031)

Coverage -0.064∗ -0.047 -0.080∗∗

(0.037) (0.046) (0.040)Observations 18141 8964 9178R2 0.305 0.304 0.324Districts 174 174 174Clusters 536 535 536

Panel B: Children aged 13-17

(1) (2) (3)All Girls Boys

Coverage * Year 2005 0.012 0.027 -0.004(0.020) (0.026) (0.024)

Year 2005 0.113∗∗∗ 0.151∗∗∗ 0.082∗∗

(0.031) (0.043) (0.036)

Coverage -0.008 0.052 -0.039(0.050) (0.068) (0.049)

Observations 14073 6760 7313R2 0.218 0.280 0.218Districts 174 174 174Clusters 536 534 536

Note: Dependent variable is school attendance (0/1), DHS 2000 and2005. Standard errors clustered at the village level in parentheses. ∗

p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. All the regressions include districtand month fixed effect, individual and geographical controls.

42

Table 12: Effects by Household Income(1) (2) (3) (4) (5)

Sonpref Wifebeat FinalSay FS Sch SchAttCoverage * Year 2005 * q1 -0.007 -0.070 0.635∗∗∗ 0.108∗∗∗ 0.071∗∗∗

(0.008) (0.157) (0.234) (0.039) (0.023)

Coverage * Year 2005 * q2 -0.010 -0.196 0.429∗ 0.020 0.098∗∗∗

(0.007) (0.146) (0.227) (0.038) (0.020)

Coverage * Year 2005 * q3 -0.008 -0.369∗∗ 0.186 0.016 0.053∗∗∗

(0.007) (0.147) (0.203) (0.028) (0.020)

Coverage * Year 2005 * q4 -0.018∗∗∗ -0.463∗∗∗ 0.589∗∗∗ 0.078∗∗ 0.025(0.006) (0.156) (0.211) (0.033) (0.021)

Coverage * Year 2005 * q5 -0.014∗ -0.488∗∗∗ 0.487 0.061 0.034(0.008) (0.162) (0.296) (0.050) (0.022)

Observations 22685 5790 3535 3535 18141R2 0.027 0.165 0.181 0.100 0.306Districts 174 174 173 173 174Clusters 538 538 536 536 536

Note: q1 and q5 are a dummies for a housholds belonging to the poorest and richest quintiles, respectively.Standard errors clustered at the village level in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Allthe regressions include district and month fixed effect, individual and geographical controls.

Table 13: Selective Migration between Exposed and Non-Exposed Villages(1) (2) (3) (4) (5)

Sonpref Wifebeat FinalSay FS Sch SchAttCoverage * Year 2005 -0.010∗ -0.291∗∗∗ 0.375∗∗ 0.037∗ 0.067∗∗∗

(0.005) (0.109) (0.152) (0.022) (0.016)

Year 2005 -0.014∗∗ 0.542∗∗∗ -0.380∗ -0.069∗∗∗ -0.209∗∗∗

(0.006) (0.133) (0.223) (0.026) (0.024)

Coverage 0.000 -0.241 -0.404 -0.044 -0.077∗∗

(0.009) (0.239) (0.342) (0.050) (0.031)Observations 20518 5261 3251 3251 16719R2 0.028 0.164 0.183 0.094 0.311Districts 173 173 172 172 174Clusters 538 537 533 533 544

Note: These regressions include only respondents who lived in the same place for at least 5years. Standard errors clustered at the village level in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗

p < 0.01. All the regressions include district and month fixed effect, individual and geographicalcontrols.

43

Table 14: Testing for differences between urban and rural areas(1) (2) (3) (4) (5)

Sonpref Wifebeat FinalSay FS Sch SchAttCoverage * Year 2005 * Urban 0.009 -0.027 0.148 0.050 0.024

(0.014) (0.361) (0.548) (0.065) (0.037)

Coverage * Year 2005 -0.012∗∗ -0.294∗∗∗ 0.450∗∗∗ 0.052∗∗ 0.058∗∗∗

(0.005) (0.108) (0.148) (0.022) (0.017)Observations 22685 5790 3535 3535 18141R2 0.026 0.163 0.180 0.097 0.305Districts 174 174 173 173 174Clusters 538 538 536 536 536

Note: Standard errors clustered at the village level in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.All the regressions include district and month fixed effect, individual and geographical controls.

Table 15: Testing for a ”Modernity” Factor - Daily TV Watching(1) (2)

Cross-section DDCoverage * Year 2005 -0.031

(0.021)

Year 2005 0.089∗∗∗

(0.031)

Coverage -0.046 -0.002(0.074) (0.038)

Observations 11815 22675R2 0.268 0.261Districts 164 174Clusters 480 538

Note: The dependent variable TV is defined as a dummy if the repondent watchesTV almost every day. Column (1) is based on the cross-section specification (1) andcolumn (2) is based on the difference-in-difference specification (2). Standard errorsclustered at the village level in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Allthe regressions include district and month fixed effect, individual and geographicalcontrols.

44

Table 16: Effects on School EnrollmentPanel A: Effects Before Expoure

(1) (2) (3)All Girls Boys

Coverage * Year 1999 -0.002 -0.005 -0.003(0.007) (0.008) (0.008)

Observations 10134 10134 10134R2 0.047 0.050 0.030Schools 5067 5067 5067

Panel B: Effects of Current and Future Radio Exposure

(1) (2) (3) (4) (5) (6)All Girls Boys All Girls Boys

Coverage * Year 2001 0.026 0.016 0.021 0.026 0.016 0.021(0.017) (0.016) (0.016) (0.017) (0.016) (0.016)

Coverage * Year 2002 0.059∗∗∗ 0.044∗∗ 0.049∗∗ 0.058∗∗∗ 0.044∗∗ 0.049∗∗

(0.021) (0.019) (0.019) (0.021) (0.019) (0.019)

Coverage * Year 2003 0.073∗∗∗ 0.059∗∗∗ 0.055∗∗∗ 0.073∗∗∗ 0.060∗∗∗ 0.056∗∗∗

(0.022) (0.021) (0.021) (0.022) (0.021) (0.021)

Future * Year 2001 0.031 0.013 0.034(0.040) (0.037) (0.037)

Future * Year 2002 0.013 -0.002 0.014(0.038) (0.034) (0.033)

Future * Year 2003 0.034 0.018 0.051(0.064) (0.056) (0.058)

Observations 25900 25900 25900 25900 25900 25900R2 0.006 0.012 0.008 0.006 0.012 0.008Schools 5180 5180 5180 5180 5180 5180

Note: Standard errors clustered at the school in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Thedependent variable is the natural logarithm of primary enrollment. All the regressions include school andyear fixed effects. Panel A include 5067 primary schools regressed over two years and Panel B include5180 primary schools regressed over 5 years.

45