The Effects of Social Interactions on Female Genital Mutilation Evidence from Egypt

The Effects of Social Interactions on Female Genital Mutilation: Evidence from Egypt Abstract Female genital mutilation (FGM) is a traditional procedure of removing the whole or part of the female genitalia for non-medical reasons—typically as a signal of ‘quality’ in the marriage market. It has been found by the World Health Organization to be harmful to the health of women, and is internationally recognized as illegal. This paper attempts to identify the social effects of FGM and its medicalization—the shift from traditional practitioners to professional health providers—on a household’s decision to opt for FGM using instrumental variables based on spatial location. We find that FGM itself has a strong social effect: households are more likely to opt for FGM the more widely adopted it is adopted among their peers, while medicalization is found to have a significant negative effect in some areas: households are less likely to opt for FGM the more widely is medicalization utilized among their peers. 1. Introduction Female genital mutilation (FGM)1 is the traditional practice in some cultures of removing the whole or part of the external female genitalia of girls, from infancy to 15 years of age, for non-medical reasons. Some forms of the practice also seal the vaginal opening. It is mostly done as a rite of passage into female adulthood, an act of ‘cleansing’ in preparation for marriage, and/or a method of curbing sexuality to ensure virginal purity before marriage and fidelity after (WHO, 2010). It is prevalent to different degrees in western, eastern and northern Africa; its prevalence could be as high as 91% for women between the ages of 15 and 49, in countries like Egypt (WHO; El-Zanaty and Way, 2009). World-wide, between 100 to 140 million girls are estimated to have undergone this procedure (WHO, 2010). The FGM procedure has been found to cause a variety of health problems if carried out in 1 This is sometimes referred to with the less severe terms of “circumcision” and “cutting”. We use all these terms synonymously. an non-sanitary environment (by traditional circumcisers, for example), as well as long term problems and complications in childbirth Mackie (2003). No health benefits have been found. Additionally, it has been internationally recognized as a violation of the human rights of girls who are forced to undergo this procedure (WHO, 2010). It has been compared to footbinding in being a harmful traditional practice, ethically indefensible due to its permanent physical and psychological damage (Mackie, 1996). The World Health Organization (WHO) has directed its advocacy and research efforts towards the elimination of this practice, in conjunction with local governments that have attempted a variety of policy interventions (WHO, 2010). In addition to health-related or ethical objections to FGM, it has been estimated in a study conducted by the WHO, that the cost of obstetric complications caused by FGM to be $3.7 million (PPP) (Adam et al., 2010). Typically, the procedure of FGM is done by a traditional circumciser2 , but increasingly professional health providers, such as doctors or nurses, are also doing it. This is referred to as the medicalization of FGM, which has become a major concern for the WHO and many antiFGM activists. The WHO declares that under no circumstances should health professionals preform FGM, regarding it as violation of the medical ethic of “Do no harm”. There are also fears that medicalization might legitimize the practice, giving it the appearance of being beneficial, and hence rolling back the gains made in the elimination of FGM (OHCHR et al., 2008). A more amendable position views medicalization as a harm reducing temporary solution in societies where a sudden elimination of FGM is unlikely to take place. Such a view regards the resistance to medicalization as counterproductive and harmful to the young girls who would then have to suffer the painful procedure without anesthetics and proper sanitation and care (Shell-Duncan, 2001). Yet another view sees medicalization as helpful in eroding the traditional practice, as FGM is moved from the traditional community-level domain and marriage market to the domain of modern medicine. It would no longer be a moral issue, but a health issue, and hence would meet with weaker resistance to attempts to completely eliminate it on health grounds. Another contributing factor to this possible story is the unobservable nature of FGM before and after marriage, and hence the possible reliance on traditional circumcisers as certifiers of ‘quality’. Thus, as FGM is increasingly done in government clinics by professional health providers, who are less connected to the marriage market in local communities, it loses its effectiveness as a signal of ‘quality’. The practice of FGM can be viewed as an innovation that has gained wide acceptance and adoption in society, and the move to medicalization can be viewed as a form of reinvention of this practice. What is interesting about the changes we observe is that this form of re-invention is undermining the overall practice, slowly leading to its discontinuance. What appears to be the modernization of an entrenched practice that is resistant to policy, actually weakens it by changing its characteristics significantly—moving it from the marriage market to the health domain (Rogers, 2003). What makes FGM difficult to eliminate is that, while it is a form of physical violence against female children, it is not regarded as such by 2 In Egypt, mostly midwives (daya) and barbers 2 its practitioners. In fact, parents would be viewed as negligent were they not to have this procedure done. Medicalization can then be viewed as taking away this perceived benefit, whether by removing its benefit as a signal of quality in the marriage market, or by changing the method of benefit evaluation—it is evaluated by its health benefit as opposed to its social benefit. This paper investigates the social effects of both circumcision and medicalization. It attempts to shed light on the extent to which members of a social reference group influence each other, setting up convention. The aim of this is to aid policy makers in understanding the consequences of possible interventions targeting the banned practice. Recently, economists have taken a strong interest in studying such nonmarket interactions between agents. There is now a recognition of the need to go beyond the conventional model of homo economicus in order to respond to public policy questions on social behavior. Any attempt to investigate persistent social behavior, such as poverty or in our case FGM, without incorporating the influence of peers and family would be grievously incomplete (Durlauf and Young, 2001). Of the literature dealing with social/peer effects, this paper follows research on social networks and welfare use (Bertrand et al., 2000); family and neighborhood effects that influence criminal activity, drug and alcohol use, school dropout, and teenage behavior (Case and Katz, 1991; Crane, 1991; Evans et al., 1992). It should be noted that this work does not undertake to discover the mechanisms driving peer influence; it only attempts to show the causal influence of peer decisions (FGM and its medicalization) on households’ decisions. The possible stories mentioned above, possibly explaining our results, are not explicitly tested, and neither are some of the common theories in the literature on FGM3 . This paper follows Becker’s (1981) classical work on marriage markets. This work is particularly important in studying low-income countries, where marriage is a critical aspect of a woman’s life, in the face of low education and few employment opportunities outside the role of wife and mother. This paper was also inspired by Chesnokova and Vaithianathan’s (2010) work on the persistence of FGM as an equilibrium in society, using DHS data from Burkina Faso. They find that as long there exist some circumcised women and circumcision is viewed as a desirable quality by men, there will always be an incentive to have FGM done to girls to improve marriageability4 . The paper also follows the literature on pre-marital investment (Burdett and Coles, 2001; Peters and Siow, 2002). In terms of method, this work is inspired by Bramoullé et al. (2009) and Blume et al. (2010) which attempt to address some of the econometric challenges raised by Manski (1993) and Moffitt (2001). What is found is that the FGM decision is strongly influenced by the decision of a household’s peers. This is not unexpected intuitively and from previous investigations. What is an interesting contribution of this paper is the finding that the choice of medicalization by a household’s peers has a negative influence on the FGM decision in some areas—rural 3 For more on theory see Mackie (1996) and Mackie and LeJeune (2009). For a recent work on testing theory see Hayford (2005) and Shell-Duncan et al. (2011). 4 A finding that Shell-Duncan et al. (2011) find weak evidence in Senegambia. 3 areas, and urban areas with no problemin accessing medical help (due to distance). While we cannot identify the mechanism by which negative influence operates, it might be explained by the move of FGM from the traditional marriage market to the professional health domain or a change in information structure in the marriage market due to the unobservable nature of FGM. This paper is organized as follows: In section 2 the data used to produce our results is described, in section 3 our empirical strategy is presented, in section 4 our results are presented, and finally we conclude with section 5. 2. Data In this paper we utilize the Egyptian Demographic and Health Survey data for the year 2008 (EDHS 2008). This is the ninth such survey in Egypt, conducted every two or three years since 1988. The survey focuses on a wide set of population, health, fertility, and nutrition indicators. In 2008, 16,527 ever-married women of ages between 15 and 49 were interviewed, as well as a subsample of 5,430 men of ages between 15 to 59 who are residing in one in four of the womens’ households. Relevant to this paper, the survey also collected data on FGM status for the interviewed women and their daughters, by whom the procedure was done, at what age was it done, intentions to circumcise uncircumcised daughters, exposure to information on FGM, and attitudes towards FGM (El-Zanaty and Way, 2009). The DHS survey design is standardized across countries and surveys data is made available in a standard recode. The survey aims for national population and geographic coverage, representing the entire population across all domains, relying on random probability sampling. For EDHS 2008, surveyed households were clustered into 1,241 clusters with an average of 14.3 households per cluster (with a minimum of 1 and a maximum of 45). Observations were gathered from all the 27 governorates of Egypt in 2008 (DHS, 1996). Actual field work was conducted between 15 March 2008 and late May 2008. The primary data survey used in this paper is that of the unmarried daughters residing in the household. Each eligible woman was asked about the FGM status of her daughters and other related information. In our data, each observation corresponds to a daughter of a surveyed woman, and for each we have their FGM characteristics and their household’s characteristics: the parent’s education levels, occupation, wealth level, etc. In addition, for each cluster, we have GPS coordinates, which we use in our spatial analysis to infer social effects. The full survey produced 17,991 observations from 9,963 households. However, since the majority of girls are circumcised around the age of puberty, only those born before 1996 are considered (ages 12 and above). The upper plot in Figure 1 shows the density distribution of the daughter’s sample. The vertical line shows the cut off age for our sample in order not to bias our estimates with younger girls who have not yet reached the age at which they are risk of FGM. Using the entire sample, we find that only 5% of girls would undergo the FGM procedure after they have reached the age of 12. The subsample that we use for our analysis is composed of 6,563 observations in 4,619 households. Table 1 gives descriptive statistics for the considered sample, stratified by wealth level, urban/rural residence, and religion. We see a greater tendency for FGM among the poorer, 4 FGM Density Plot by Cohort 0.30 Cohort Density 0.25 1988 1989 1990 1991 1992 1993 1994 1995 1996 0.20 0.15 0.10 0.05 0.00 0 5 10 15 20 Age at FGM FGM Density Plot (Medicalization vs. Traditional) Density 0.4 0.3 Traditional Medical. 0.2 0.1 0.0 0 5 10 15 Age at FGM Figure 1: FGM Density Plot 5 Table 1: Summary Statistics Circumcised Medicalization1 All All N 100% 67.74% 71.88% 6253 By Wealth Index (Rich) 16.02% 40.22% 87.59% 1002 By Wealth Index (Poor) 83.98% 73.00% 70.23% 5251 By Education (None) 48.58% 75.61% 64.91% 3038 By Education (Primary) 16.33% 76.30% 73.81% 1021 By Education (Secondary) 29.46% 57.17% 83.67% 1842 By Education (Higher) 5.63% 30.40% 91.59% 352 Rural Residence 62.40% 76.68% 69.35% 3902 Urban Residence 37.60% 52.91% 77.97% 2351 By Religion (Christian) 4.77% 42.62% 29.87% 298 By Religion (Muslim) 95.30% 69.02% 71.92% 5959 1 Conditional on circumcision 6 Proportion of FGM Yes No Frequency 1.0 0.8 0.6 0.4 0.2 0.0 1988 1989 1990 1991 1992 1993 1994 1995 1996 1995 1996 Cohort Proportion of Medicalization Yes No Frequency 1.0 0.8 0.6 0.4 0.2 0.0 1988 1989 1990 1991 1992 1993 1994 Cohort Figure 2: FGM Proportions by Cohort rural and Muslim households. Also noticeable is a greater tendency for medicalization for richer, urban, and Muslim households. In Figure 2, we notice a decrease over cohorts in the rate of overall circumcision with increase in the rate of medicalization. 3. Empirical Strategy 3.1. The Model The linear-in-means model that we are interested in estimating is yihr = β · xi + δ · X xj j∈Pi ni +γ X yj j∈Pi ni + ηh + ρr + εihr , 7 E[εihr |x(1,h) , . . . , x(Rh ,h) , ηh , ρ1 , . . . , ρRh ] = 0 (1) 1 Wealth Index (Rich) Rural Residence Religion (Christian) Mother Circumcized Mother’s Marital Age 67.7% 38.0% 78.1% 97.8% 46.1% 79.1% 48.3% 56.6% 25.6% 57.7% 70.6% 65.6% 82.9% 74.7% 78.5% 9.9% 82.5% 60.3% 82.2% 42.6% 15.6% 93.7% 89.6% 79.3% 89.9% 68.0% 63.6% 48.7% 32.2% 31.0% 70.2% 35.5% 59.9% 33.0% 41.4% 21.8% 24.4% 53.3% 57.5% 73.0% 66.2% 74.7% 6.1% 66.7% 39.4% 54.8% 16.7% 7.3% 75.5% 81.2% 65.0% 53.7% 40.0% 46.1% 16.0% 47.3% 6.6% 12.4% 10.3% 6.4% 45.8% 12.1% 40.6% 2.7% 15.4% 32.6% 15.8% 4.4% 18.8% 7.6% 7.7% 5.6% 24.7% 15.7% 79.8% 10.3% 20.8% 4.5% 7.0% 48.0% 33.8% 62.4% 0.0% 72.0% 70.8% 81.9% 86.6% 0.0% 69.0% 57.1% 84.9% 73.5% 39.2% 51.3% 73.3% 58.6% 42.7% 76.5% 85.1% 42.5% 41.7% 0.0% 71.6% 0.0% 89.2% 81.9% 24.0% 0.0% 4.8% 4.4% 11.6% 3.4% 0.9% 2.2% 5.9% 0.3% 0.0% 1.4% 0.4% 5.1% 0.0% 0.9% 7.3% 0.0% 3.4% 12.3% 0.0% 0.0% 0.9% 6.3% 18.8% 1.6% 11.8% 8.0% 0.6% 96.0% 90.7% 97.0% 100.0% 97.5% 100.0% 96.9% 96.9% 95.5% 99.3% 96.7% 96.0% 100.0% 99.6% 100.0% 26.0% 98.3% 94.2% 100.0% 83.3% 95.4% 99.1% 100.0% 100.0% 99.5% 100.0% 98.1% 18.5 20.6 17.6 17.2 18.5 17.5 20.1 18.8 19.7 16.7 19.9 18.5 19.2 19.1 19.3 17.5 19.3 16.9 20.3 19.3 22.2 17.2 18.2 18.2 17.7 19.0 21.0 Standard deviation is reported between parenthesis 8 (3.8) (4.0) (3.3) (3.9) (3.6) (3.5) (4.2) (3.7) (3.9) (2.9) (3.7) (4.0) (3.3) (3.8) (3.7) (3.7) (3.8) (3.4) (4.3) (3.6) (3.5) (3.6) (3.9) (3.0) (3.4) (4.2) (3.9) Age of Circumcision Medicalization All Alexandria Assuit Aswan Behera Beni Suef Cairo Dakahlia Damietta Fayoum Gharbia Giza Ismailia Kafr El Sheikh Kalyubia Matrouh Menoufia Menya New Valley North Sinai Port Said Qena Red Sea Sharkia Souhag South Sinai Suez Circumcized Table 2: Governorate Summary Statistics N 9.2 (2.8) 10.4 (2.0) 8.1 (2.4) 5.7 (2.3) 10.8 (2.0) 10.9 (1.8) 9.6 (2.0) 10.5 (2.2) 10.9 (2.1) 11.2 (1.7) 9.9 (1.7) 10.4 (1.5) 10.6 (1.6) 10.7 (1.3) 9.5 (1.0) 10.2 (1.7) 9.9 (1.6) 10.5 (1.7) 9.5 (2.4) 10.9 (1.5) 10.6 (2.7) 5.8 (3.5) 8.3 (1.9) 10.5 (1.3) 7.4 (3.3) 9.9 (1.4) 10.3 (1.7) 6253 205 439 178 321 359 321 290 133 291 272 273 152 225 261 131 234 464 73 108 109 458 48 314 415 25 154 The subscript i = {1, . . . , N} identifies daughters (the observations of interest), h = {1, . . . , M} their household, and r = {1, . . . , Rh } their order of birth within their household, where Rh is the number of daughters in household h. The variable yihr is a indicator of whether a particular daughter has undergone FGM. Each daughter i has a peer group whose characteristics and behavior might influence i’s household’s FGM decision. In this specification, we model it as the set observations Pi (ni = |Pi | is the number of i’s neighbors). The dependent variable is regressed on • The K × 1 vector xi of a daughter’s individual characteristics, composed of daughterspecific regressors and household-invariant regressors • The mean of x of a daughter’s peer group Pi , and whose K × 1 vector of coefficients, δ, represents exogenous social effects • The mean of y of a daughter’s peer group Pi , and whose coefficient, γ, represents endogenous social effects • The household fixed effect ηh • The birth order fixed effect ρr We transform this model to use matrix notation to facilitate the use of an interaction matrix to represent peer groups (Bramoullé et al., 2009). We use the logical N × N matrix f W to indicate whether any two daughters are considered peers (we will further explain what defines peers for the purposes of this model below). (f Wij ) = 1{i and j are peers} We further normalize this matrix to the row stochastic W, where (Wij ) = f Wij ni Our model now becomes y = Xβ + WXδ + Wyγ + η + ρ + ε (2) where y is N × 1 vector of FGM status, X is a N × K matrix of daughters’ characteristics, η is a N ×1 vector of {η1 , . . . , η M }, each repeated Rh times, and ρ is a N ×1 vector of {ρ1 , . . . , ρRh }. 3.1.1. Reference Group In order to carry out our social effects analysis we need to define each daughter’s peer group Pi or the interaction matrix W. First we define the logical matrices 9 e : (A eij ) = 1{disti j ≤ 10 kilometers5 } A e : (C eij ) = 1{|agei − age j | ≤ 1 year} C eij ) = 1{householdi = household j } e : (C H from which we define the interaction matrix, excluding all same household daughters from the peer group, since we are ultimately interested in estimating the influence of peers on a household’s decision making. e ◦ ¬H f e◦C e W=A This means we define the peer group that would influence a household’s FGM decision for daughter i as all daughters i) not in the same household, ii) who are within a ten kilometer radius, iii) and whose absolute age difference does not exceed one year. While we do model the peer groups or ‘neighborhoods’ as a network (as is commonly done in the social networks literature), we do not mean to model explicit social links between households or daughters. A daughter’s peer group is used to estimate what is common practice in the area of residence6 . In other words, we decide to consider two daughters as network neighbors if all the conditions outlined above are true. It might be more intuitive to understand how this is modeled by viewing the space of observations as divided into planes, with each plane representing a cohort group, and on each plane, observations of that cohort group are placed according to their GPS coordinates. On each plane we then have overlapping circles, each centered on an observation (or more accurately, its cluster), and any observation within such a circle is considered a neighbor, which we represent by a network link. Figure 3 shows an example of such a network. 3.1.2. Identification So far, this is a typical linear-in-means model with all the identification challenges this entails: simultaneity (the reflection problem), endogeneity, and correlated effects (Manski, 1993; Moffitt, 2001; Blume et al., 2010; Bramoullé et al., 2009). The reduced form of this model would be y = (I − Wγ)−1 Xβ + (I − Wγ)−1 WXδ + (I − Wγ)−1 η + (I − Wγ)−1 ρ + (I − Wγ)−1 ε 5 (3) Any smaller range would be problematic according to the DHS because of location displacement, whereby GPS coordinates are randomly altered in order to preserve survey subjects’ privacy. This is also the reason we do not rely on the distance between clusters to weigh our interaction matrix (DHS, 2012). 6 Inference based on such estimated averages is further discussed below. 10 Figure 3: For node 1, nodes 2-4 are considered direct neighbors, while nodes 5-9 are indirect neighbors. 11 It is clear from equation (3) that we cannot separately identify exogenous and endogenous social effects (the parameters δ and γ). We therefore follow the identification strategy described by Bramoullé et al. (2009). First they show that by substituting in equation (3) P k k the expansion (I − Wγ)−1 = ∞ k=0 γ W , premultiplying by W, and taking expectation we get ∞ ∞ ∞ X X X k k+2 k k+1 E[Wy|X, η, ρ] = ( γ W )Xβ + ( γ W )Xδ + ( γk Wk+1 )(η + ρ) k=0 k=0 (4) k=0 According to Bramoullé et al. (2009), if the matrices I, W, and W2 are linearly independent the social effects are separately identified7 . What this entails is the existence of intransitive triads in the social network: daughters that have a common peer but are not each other’s peers (see Figure 3). This allows us to use (W2 X, W3 X, . . . ) as instrumental variables for Wy in equation (2). As shown in equation (4) these variables influence E[Wy|X, η, ρ] without having a direct effect on y. In other words, we can use the exogenous social effects of a daughter’s peers’ peers (their exogenous characteristics) as instruments for endogenous social effects (the peer households’ endogenous FGM decision), as long as there exists intransitive triads. Another identification problem in linear-in-means models is correlated effects: unobservable neighborhood/regional effects that would influence FGM decision making and are correlated with observed daughter and household characteristics. Bramoullé et al. (2009) address this problem by introducing either global network fixed effects, a fixed effect for each component in the network, or local network fixed effects, a fixed effect for each individual’s direct neighbors. They then use within differencing to eliminate these fixed effects in their estimation. In this analysis, we introduce a similar fixed effect to capture household hetergeneity. Since we observe possibly multiple daughters from the same household we rely on a similar within transformation. Bramoullé et al. (2009) show that identification in this case requires the stronger condition of having I, W, W2 , and W3 . Yet another identification challenge in linear-in-means models is peer group endogeneity: unobservable characteristics that influence FGM decision making and cause households to choose to associate with other households that make similar decisions (Moffitt, 2001). We find it improbable that household are going to be driven to select their location of residence based on the FGM decisions of their neighbors. Nonetheless, we attempt to forestall any possible problems by eliminating from our analysis any households that have not resided in their current location for at least ten years. 3.1.3. Medicalization So far, our model has not deviated much from Bramoullé et al. (2009), but since we are mainly interested in the effect of medicalization on a household’s decision we modify our model to be 7 Note that Wk provides an interaction matrix for all neighbors at a distance of k, where two nodes are said to be of k distance from each other if there exists a path between them passing through k other nodes. The concept of ‘distance’ here is based on network links, not spatial distance. 12 y = Xβ + WXδ + Wyγ + Wmλ + η + ρ + ε (5) where m is a N × 1 vector of binary variables indicating whether a household, conditional on having opted for FGM, decided to rely on a medical practitioner rather than a traditional circumciser. Hence, Wm results in the N × 1 vector of mean medicalization of a daughter’s peer group. While Wm is unlikely to suffer from the same simultaneity problems of Wy, since, having opted for FGM, a household is unlikely to be influenced peer’s FGM decision, Wy, in choosing miht , we nevertheless rely on the same instrumental variable approach used to identify Wy. 3.1.4. Generated Regressors One problem in linear-in-means models, raised by Manski (1993), is the use of sample peer means as if they are the actual unobserved social influence regressor of interest, effectively neglecting the fact that such sample means are estimates from a first stage estimation. This could cause problems for inference. In Wooldridge (2010) such regressors are referred to as generated regressors, which are shown to be consistent but the standard errors produced by OLS would be incorrect. In our case, another advantage of using instrumental variables for identification is that heteroskedasticity robust standard errors are sufficient in testing the significance of our estimates. 3.1.5. Proxies for Modernization One remaining concern is that our estimation might suffer from omitted variable bias because of possible unobservable influence of ‘modernization’ in medical care on the FGM decision: as different areas become more modernized they would perhaps become less likely to opt for a traditional practice such as FGM, and should they decide to have it done they would seek a medical practitioner rather than a traditional daya—it would be correlated with the Wm. In order to control for this unobservable modernization effect, we use some proxy variables related to the child delivery choices of neighbors, such as (a) the fraction of households in the peer group that decided to have their child delivered by a daya as opposed to a medical practitioner, (b) and the location of delivery. We calculate these averages for births during the same period the FGM decision is made8 . Recall that FGM decision is normally done around the age of twelve, so we are seeing different groups who are making FGM and child delivery decisions. In other words, focusing one such proxy, dayaihr , we make the following assumptions E(yihr |zihr , modernihr , dayaihr ) = E(yiht |zihr , modernihr ) L(modernihr |1, zihr , dayaihr ) = L(modernihr |1, dayaihr ) where zihr are all other regressors and modernihr is unobserved modernization. 8 Data is only available for child births preceding the time of survey by six years. 13 3.2. Implementation 3.2.1. Variables Of the household characteristics we are interested in we use: • Wealth level: ‘poor’ or ‘rich’, with ‘poor’ as the omitted level9 • Urban or rural residence, with rural residence as the omitted level • The mother’s marital age • The mother’s FGM status • Religion: Christian or Muslim, with Muslim as the omitted level • The sex of the head of household, with male as the omitted level • Whether seeking medical help is ‘not a problem’ or a ‘big problem’ due to distance or cost, with ‘not a problem’ as the omitted level • Highest level of mother’s education: no education, primary education, secondary education, or higher education, with no education as the omitted level 3.2.2. Instrumental Variables The first step in identifying possible instrumental variables for Wy from W2 X we run the first stage regression Wy = θ1 + W2 Xφ1 + ν1 and, similarly to instrument Wm Wm = θ2 + W2 Xφ2 + ν2 . Tables 3 and 4 show the results of these regressions . The first columns shows the regression using all of W2 X, while the second columns show the regression using the instrumental variables selected for our analysis. Using a heteroskedastic overidentification test, using three instruments for the two endogenous regressors in equation (5), we fail to reject the hypothesis that the instruments are valid (p-value = 0.8657648) (Wooldridge, 2010). 9 The DHS wealth index is used here, which is divided into quintiles: poorest, poorer, middle, richer, richest. Calculations of household wealth are based on standard of living measures derived from the survey questionnaires, based on such indicators as the ownership of assets, the characteristics of housing, and the source of drinking water. In this paper, the wealth levels are reduced to two groups: the poorest/poorer/middle are reclassified as poor, and the rest as rich. 14 Table 3: FGM Endogenous Social Effect First Stage Regression Estimate (Standard Error) (1) (2) -0.208*** (0.054) 0.087* (0.037) 0.022 (0.044) -0.078 (0.054) 0.603*** (0.127) 0.005 (0.005) 0.314*** (0.092) -0.320** (0.122) 0.142* (0.066) -0.038 (0.049) -0.004 (0.040) -0.007 (0.054) -0.217*** (0.050) F-test on instruments 7.664 17.629 N 6206 6206 R2 0.760 0.757 0.757 0.754 Wealth (Rich) Residence (Urban) Education (Primary) Education (Secondary) Education (Higher) Marital Age Mother FGM Religion (Christian) HH Head Sex (Female) Medical Help Dist. (Big Problem) Discussed FGM Received Info on FGM Adjusted R2 1 0.731*** (0.117) 0.352*** (0.084) Asterisks indicate significance levels: ‘·’ is 10 percent, ‘*’ is 5 percent, ‘**’ is 1 percent, and ‘***’ is 0.1 percent. 15 Table 4: Medicalization Endogenous Social Effect First Stage Regression Estimate (Standard Error) (1) Wealth (Rich) Residence (Urban) Education (Primary) Education (Secondary) Education (Higher) Marital Age Mother FGM Religion (Christian) HH Head Sex (Female) Medical Help Dist. (Big Problem) Discussed FGM Received Info on FGM -0.263*** (0.073) -0.018 (0.052) 0.025 (0.071) -0.034 (0.063) 0.372** (0.141) 0.015** (0.006) 0.149· (0.083) -0.426*** (0.111) 0.122 (0.095) 0.041 (0.061) 0.023 (0.050) -0.084* (0.035) (2) -0.444*** (0.108) F-test on instruments 5.659 17.006 N 6206 6206 R2 0.608 0.603 0.601 0.597 Adjusted R2 1 Asterisks indicate significance levels: ‘·’ is 10 percent, ‘*’ is 5 percent, ‘**’ is 1 percent, and ‘***’ is 0.1 percent. 16