The Gradient in Sub-Saharan Africa: Socioeconomic Status and HIV/AIDS

Abstract

Using data from the Demographic and Health Surveys (DHS) for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003), I investigate the cross-sectional relationship between HIV status and socioeconomic status. I find evidence of a robust positive education gradient in HIV infection, showing that, up to very high levels of education, better-educated respondents are more likely to be HIV-positive. Adults with six years of schooling are as much as three percentage points more likely to be infected with HIV than adults with no schooling. This gradient is not an artifact of age, sector of residence, or region of residence. With controls for sex, age, sector of residence, and region of residence, adults with six years of schooling are as much as 50% more likely to be infected with HIV than those with no schooling. Education is positively related to certain risk factors for HIV, including the likelihood of having premarital sex. Estimates of the wealth gradient in HIV, by contrast, vary substantially across countries and are sensitive to the choice of measure of wealth.

Considerable research documents the link between health and socioeconomic status in developed countries (see, e.g., Case, Lubotsky, and Paxson 2002). Almost universally, these studies show that those with higher socioeconomic status (often measured by income and education) are in better health. Because of data constraints, comparatively little work has assessed the relationship between socioeconomic status and health in developing countries, but existing evidence suggests that they are positively linked in developing countries, as they are elsewhere (e.g., Wagstaff 2000). In sub-Saharan Africa today, HIV status is an important component of health among adults. However, because the prevalence and spread of HIV in sub-Saharan Africa has been distinctive, the relationship between socioeconomic status and HIV may be quite unlike that of socioeconomic status and health more generally.

Available empirical evidence on the link between HIV infection and socioeconomic status in sub-Saharan Africa is mixed. Several studies found that HIV infection, unlike other measures of health, was positively correlated with socioeconomic status. Early work by Over and Piot (1993), using data from the late 1980s, showed higher HIV infection rates among the highly educated and among higher occupational classes. This link was corroborated by cross-country work by Gregson, Waddell, and Chandiwana (2001) that documented a positive relationship between national HIV prevalence and literacy rates. Using data from 2003, Shelton, Cassell, and Adetunji (2005) drew attention to the positive unadjusted relationship between HIV and wealth in individual-level data from Tanzania. However, the observed positive relationship between HIV and socioeconomic status may very well be a remnant of other characteristics that are correlated with both HIV and socioeconomic status, such as age, sex, or sector of residence (urban or rural). Hargreaves and Glynn (2002) performed a meta-analysis of published literature that took into account both age and sex, and nevertheless found that individual HIV infection and education were positively linked in countries across Africa.

However, work by Glynn et al. (2004) in several sub-Saharan African cities reached different conclusions. The authors looked separately at men and women, controlling for age, religion, and ethnicity. They found that in two cities (in Kenya and Zambia), HIV and schooling were unrelated. In another city (in Cameroon), more-educated women were less likely to be HIV-positive; for men, the relationship was not statistically significant. In a fourth city (in Benin), education and HIV were negatively related among men, whereas the relationship between education and HIV was not statistically significant for women. A study by de Walque (2006), using data from five countries, concluded that wealth—but not education—was positively related to HIV infection after controlling for individual characteristics. Using these data and data from three additional countries, Mishra et al. (2007) likewise found that wealth was positively related to HIV infection. And in an earlier longitudinal study in Uganda, de Walque (2002) found that among young people, education was negatively correlated with HIV.

Although the empirical evidence documenting the relationship between socioeconomic status and HIV infection lacks consensus, the variance could reflect true heterogeneity in the relationship, both across space and over time. For instance, because people with different levels of education may respond differently to information about avoiding HIV infection (as suggested by de Walque 2002 and Gregson et al. 2001), the gradient may be changing over time. Furthermore, the differing conclusions of these studies may stem from the fact that they used a range of markers for socioeconomic status. Different indicators of socioeconomic status may provide different information about an individual’s social and economic resources (Daly et al. 2002). Despite these challenges, understanding the relationship between HIV and socioeconomic status may be important for designing policies to both reduce the spread of HIV and minimize its adverse effects.

In this paper, I investigate this relationship further using a broad sample from the Demographic and Health Surveys (DHS). Using HIV test results linked to demographic microdata from the DHS for Burkina Faso, Cameroon, Ghana, Kenya, and Tanzania, I estimate the relationship between HIV and socioeconomic status. My analysis improves upon past work because I produce estimates that both cover a broad sample of countries and control for individual characteristics, such as age and sector of residence. Furthermore, I relax the parametric assumptions of earlier work to better pinpoint the gradient and perform a series of robustness checks. I find that for all but the most-educated respondents, education is positively related to HIV, even with controls for age and sector and region of residence. Education is positively related to several risk factors for HIV, including the probability of having premarital sex (for both men and women). Although wealth and HIV are correlated, the relationship between wealth and HIV is highly sensitive to the choice of the measure of wealth.

DATA

The DHS provides nationally representative, repeated cross sections of demographic, economic, and fertility microdata from countries across the world; for several countries, recent waves of the survey include the results of an HIV test.1 My analysis uses the 2003 cross sections for Burkina Faso, Ghana, Kenya, and Tanzania, and the 2004 cross section for Cameroon. Because earlier waves do not have HIV testing results, I do not use data from previous cross sections.2 In households selected for the survey, all women ages 15–49 were eligible to be interviewed. In Burkina Faso, Cameroon, and Kenya, men were interviewed in only a subsample of households. In Burkina Faso, Cameroon, and Ghana, men ages 15–59 were eligible to be interviewed. In Kenya, the sample included men up to age 54; and in Tanzania, the sample included men up to age 49.

In Ghana and Tanzania, all households were eligible for HIV testing; in Burkina Faso, Cameroon, and Kenya, only a subsample of households was eligible for HIV testing.3 Respondents were asked to provide blood samples, by finger prick, for HIV testing. Response rates for the HIV portion of the survey were 89.3% in Burkina Faso, 91.0% in Cameroon, 84.9% in Ghana, 73.4% in Kenya, and 80.5% in Tanzania, with refusal making up most of the nonresponse. Response rates for the household questionnaire exceeded 96% in all five countries.

Using years of schooling and wealth as proxies for socioeconomic status, I study the relationship between socioeconomic status and HIV status. The average number of years of schooling varies considerably across countries, as shown in Table 1. In Burkina Faso, the majority of respondents did not complete any schooling, whereas average schooling in Cameroon, Ghana, Kenya, and Tanzania exceeds five years. Within country, average years of schooling is consistently greater for men than for women.4

Table 1.

Sample Means

Variable	Burkina Faso	Cameroon	Ghana	Kenya	Tanzania
A. Women
Full sample
Years of schooling	1.380	5.618	5.898	7.122	5.363
Years of schooling = 0	0.803	0.227	0.284	0.128	0.222
Wealth index	0.296	0.338	0.337	0.300	0.268
Rural	0.784	0.452	0.516	0.749	0.691
Age	29.119	27.775	29.511	28.546	28.367
Number of observations	12,477	10,656	5,691	8,195	6,863
HIV subsample
Years of schooling	1.434	5.606	5.866	7.068	5.367
Years of schooling = 0	0.794	0.224	0.284	0.128	0.220
Wealth index	0.299	0.340	0.335	0.300	0.272
Rural	0.772	0.450	0.516	0.753	0.692
Age	29.076	27.787	29.486	28.477	28.446
HIV-positive	0.018	0.066	0.027	0.087	0.077
Number of observations	4,189	5,154	5,289	3,271	5,969
B. Men
Full sample
Years of schooling	2.621	7.067	7.751	7.938	6.202
Years of schooling = 0	0.652	0.119	0.177	0.067	0.112
Wealth index	0.301	0.348	0.344	0.304	0.277
Rural	0.759	0.427	0.551	0.746	0.697
Age	31.310	30.355	31.627	29.565	28.606
Number of observations	3,605	5,280	5,015	3,578	5,659
HIV subsample
Years of schooling	2.590	7.033	7.725	7.905	6.212
Years of schooling = 0	0.646	0.118	0.175	0.065	0.108
Wealth index	0.299	0.345	0.341	0.301	0.279
Rural	0.759	0.429	0.551	0.749	0.699
Age	31.155	30.301	31.365	29.517	28.622
HIV-positive	0.019	0.039	0.016	0.046	0.063
Number of observations	3,341	5,041	4,265	2,917	4,774

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: The table gives the mean value of each variable by country; all results are weighted using provided sample weights. The full sample includes women ages 15–49 (all countries), and men ages 15–59 (Burkina Faso, Cameroon, and Ghana), ages 15–54 (Kenya), and ages 15–49 (Tanzania). The HIV subsample includes women and men ages 15–49 who were tested for HIV. The wealth index is the fraction of nine assets or amenities that the respondent’s household has; the assets and amenities are radio, television, refrigerator, bicycle, motorcycle, car, telephone, electricity, and a flush toilet or pit latrine (must have nonmissing values for at least five components to be included). Rural is an indicator for whether the respondent currently lives in a rural area.

Wealth, the other proxy for socioeconomic status, is based on household assets or amenities. The wealth index is calculated as the fraction of nine assets or amenities (radio, television, refrigerator, bicycle, motorcycle, car, telephone, electricity, and a flush toilet or pit latrine) that the respondent’s household has.5 A precedent exists for using such an index with these data (e.g., Case, Paxson, and Ableidinger 2004). There are several advantages of using this index rather than the DHS-provided wealth index. First, interpreting the coefficient on this wealth index is easier because the units have significance; for example, a 0.11 increase in the index is equivalent to the household gaining one asset or amenity. Second, this index is directly comparable across countries. Although the index is not a perfect measure of wealth, the sample means are consistent with gross domestic product (GDP) per capita estimates of relative wealth across countries.6 As shown in Table 1, estimates of wealth based on this index are quite similar for men and women, largely because wealth is estimated based on household amenities or assets. Nevertheless, household wealth is somewhat greater for men than for women.

HIV infection rates for the sample, shown in Table 1, are generally higher for women than for men and vary considerably across countries. These estimates show that HIV is more prevalent in Cameroon, Kenya, and Tanzania than in Burkina Faso and Ghana.

RESULTS

Previous research has shown that in other contexts, socioeconomic status and health status are approximately linearly related (e.g., Case et al. 2002). Nonetheless, I first study nonparametric estimates of the relationship between socioeconomic status and HIV status. For years of schooling, I calculate the percentage of respondents who are HIV-positive at each educational level and then plot these values. Figures 1 (women) and 2 (men) show these results. The size of the circle indicates the weight, which takes into account both the sampling probability and the number of respondents at that educational level. These results show that HIV infection rates tend to be higher among the more highly educated. However, this relationship appears to be nonlinear, with respondents who have very high levels of schooling (e.g., more than 12 years) showing lower rates of HIV than those respondents with less schooling (e.g., between 6 and 12 years).

Figure 1.

Education Gradient in HIV Infection: Women

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: The sample includes women ages 15–49 who were tested for HIV. The figure shows the fraction of women who tested positive for HIV at each educational level between 0 and 16 years. The size of the circle indicates the number of women at that educational level who were tested for HIV, weighted using provided HIV sample weights.

To nonparametrically estimate the link between wealth and HIV, I use Fan (1992) regression to plot the relationship between the wealth index and HIV status. In particular, I estimate Eq. (1):

$$Y_i=f(W_i)+{\epsilon }_i,$$ (1)

where Y_i is HIV status, and W_i is the wealth index. I estimate this regression country by country, separately by sex. Figures 3 (women) and 4 (men) show the results, plotted over the middle 80% of wealth index values.7 These results reveal little systematic relationship between wealth and HIV. For some groups, HIV infection appears to be increasing with wealth, particularly in Burkina Faso (among women) and Cameroon and Tanzania (among men). That is, the evidence shows that in these countries, the rich are more likely to be HIV-positive than those with less wealth, as measured by the wealth index.

Figure 3.

Wealth Gradient in HIV Infection: Women

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: The sample includes women ages 15–49 who were tested for HIV. The wealth index is the fraction of nine assets or amenities that the respondent’s household has; the assets and amenities are radio, television, refrigerator, bicycle, motorcycle, car, telephone, electricity, and a flush toilet or pit latrine (must have nonmissing values for at least five components to be included). Results are from Fan locally weighted regressions with a biweight quartic kernel with a halfwidth of 0.35, and are weighted using provided HIV sample weights.

Although these nonparametric results suggest some general trends, there is considerable variation in these relationships across countries, perhaps reflecting differences in HIV epidemic start dates or cultural norms. I test whether a linear specification differs across countries, separately for education and wealth (results not shown but available upon request). An F test of equality for country-specific education coefficients suggests that I can reject a pooled linear model at the 10% significance level. For wealth, there is also evidence of cross-country differences in the relationship with HIV. Therefore, in the results that follow, I present estimates separately by country for both education and wealth.

Next I turn to parametric estimates of these relationships. Because the nonparametric estimates reveal nonlinearities in the data, I estimate numerous nonlinear specifications for both education and wealth. (These estimates are not shown but are available upon request.) Although there is variability across countries, I find some evidence of quadratic relationships between years of schooling and HIV and between wealth and HIV. Because education and wealth are highly collinear and are likely to be measured with error, including both as regressors could be problematic for interpreting my coefficients. Therefore, in all specifications, I look at education and wealth separately.

In Table 2, I show coefficients for linear and quadratic education specifications and linear and quadratic wealth specifications, controlling for sex.8 In Tables 3 and 4, I investigate whether these relationships are artifacts of age, sector of residence (urban or rural), or region of residence. In particular, in Table 3, I present country-by-country estimates of the quadratic education gradient in HIV, controlling for five-year age group (all panels), sector of residence (Panels B and C), and region of residence (Panel C). In Table 4, I look at the relationship between wealth and HIV, including linear and quadratic terms for wealth. These specifications include controls for five-year age group (all panels), sector of residence (Panels B and C), and region of residence (Panel C).9

Table 2.

Economic Gradient in HIV

HIV-Positive	Burkina Faso	Cameroon	Ghana	Kenya	Tanzania
A. Education, Linear
Years of schooling	0.0013 (0.0009)	0.0033* (0.0006)	0.0002 (0.0003)	0.0021* (0.0009)	0.0030* (0.0009)
Number of observations	7,142	9,732	9,146	5,991	10,743
B. Education, Quadratic
Years of schooling	−0.0006 (0.0027)	0.0078* (0.0016)	0.0011 (0.0009)	0.0062* (0.0022)	0.0056* (0.0021)
Years of schooling, squared	0.0002 (0.0003)	−0.0003* (0.0001)	−0.0001 (0.0001)	−0.0003 (0.0001)	−0.0002 (0.0002)
F statistic	1.22	19.87	0.70	5.72	7.38
p value	.30	.00	.50	.00	.00
Number of observations	7,142	9,732	9,146	5,991	10,743
C. Wealth, Linear
Wealth index	0.0369* (0.0150)	0.0197 (0.0147)	−0.0153* (0.0076)	0.0138 (0.0216)	0.0202 (0.0172)
Number of observations	7,137	9,750	9,142	5,988	10,742
D. Wealth, Quadratic
Wealth index	0.0312 (0.0349)	0.0879 (0.0494)	0.0095 (0.0268)	0.0320 (0.0625)	0.1637* (0.0489)
Wealth index, squared	0.0063 (0.0449)	−0.0876 (0.0613)	−0.0304 (0.0296)	−0.0243 (0.0784)	−0.1839* (0.0574)
F statistic	4.04	2.00	3.10	0.25	5.62
p value	.02	.14	.04	.78	.00
Number of observations	7,137	9,750	9,142	5,988	10,742

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: All regressions are weighted least squares regressions, weighted using provided HIV sample weights with clustering on the household. The sample includes women and men ages 15–49 who were tested for HIV. The dependent variable is an indicator for whether the respondent is HIV-positive. The F statistic tests the joint significance of years of schooling and years of schooling, squared (Panel B), and the wealth index and the wealth index, squared (Panel D). The wealth index is the fraction of nine assets or amenities that the respondent’s household has; the assets and amenities are radio, television, refrigerator, bicycle, motorcycle, car, telephone, electricity, and a flush toilet or pit latrine (must have non missing values for at least five components to be included). All regressions include a constant and an indicator for whether the respondent is female. Huber-White standard errors are in parentheses.

^*p < .05

Table 3.

Education Gradient in HIV

HIV-Positive	Burkina Faso	Cameroon	Ghana	Kenya	Tanzania
A. Quadratic With Age Group Controls
Years of schooling	0.0008 (0.0028)	0.0116* (0.0017)	0.0035* (0.0011)	0.0086* (0.0022)	0.0067* (0.0021)
Years of schooling, squared	0.0001 (0.0003)	−0.0006* (0.0001)	−0.0002* (0.0001)	−0.0005* (0.0001)	−0.0003 (0.0002)
F statistic	1.87	26.98	5.38	7.64	8.32
p value	.16	.00	.00	.00	.00
B. Quadratic With Age Group Controls, Rural Indicator
Years of schooling	−0.0014 (0.0027)	0.0105* (0.0017)	0.0034* (0.0011)	0.0091* (0.0022)	0.0063* (0.0021)
Years of schooling, squared	0.0001 (0.0003)	−0.0006* (0.0001)	−0.0002* (0.0001)	−0.0006* (0.0002)	−0.0005* (0.0002)
F statistic	0.28	18.95	5.16	8.29	4.79
p value	.75	.00	.01	.00	.01
C. Quadratic With Age Group Controls, Rural Indicator, Region Fixed Effects
Years of schooling	−0.0015 (0.0027)	0.0061* (0.0022)	0.0025* (0.0012)	0.0060* (0.0024)	0.0058* (0.0020)
Years of schooling, squared	0.0001 (0.0003)	−0.0004* (0.0001)	−0.0002* (0.0001)	−0.0004* (0.0002)	−0.0006* (0.0002)
F statistic	0.43	4.88	3.14	3.95	4.53
p value	.65	.01	.04	.02	.01
Number of Observations	7,142	9,732	9,146	5,991	10,743

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: All regressions are weighted least squares regressions, weighted using provided HIV sample weights with clustering on the household. The sample includes women and men ages 15–49 who were tested for HIV. The dependent variable is an indicator for whether the respondent is HIV-positive. The F statistic tests the joint significance of years of schooling and years of schooling, squared. All regressions include a constant and an indicator for whether the respondent is female. Age group controls are five-year age groups. Huber-White standard errors are in parentheses.

^*p < .05

Table 4.

Wealth Gradient in HIV

HIV-Positive	Burkina Faso	Cameroon	Ghana	Kenya	Tanzania
A. Quadratic With Age Group Controls
Wealth index	0.0266 (0.0349)	0.0861 (0.0493)	0.0105 (0.0268)	0.0389 (0.0622)	0.1613* (0.0486)
Wealth index, squared	0.0132 (0.0450)	−0.0747 (0.0610)	−0.0289 (0.0296)	−0.0378 (0.0785)	−0.1659* (0.0570)
F statistic	4.13	2.62	2.44	0.24	5.66
p value	.02	.07	.09	.78	.00
B. Quadratic With Age Group Controls, Rural Indicator
Wealth index	0.0038 (0.0321)	0.0258 (0.0504)	−0.0026 (0.0279)	0.0350 (0.0621)	0.1196* (0.0486)
Wealth index, squared	0.0072 (0.0453)	−0.0477 (0.0610)	−0.0254 (0.0296)	−0.0790 (0.0800)	−0.2095* (0.0584)
F statistic	0.25	0.53	3.72	0.88	8.40
p value	.78	.59	.02	.41	.00
C. Quadratic With Age Group Controls, Rural Indicator, Region Fixed Effects
Wealth index	−0.0014 (0.0345)	−0.0032 (0.0516)	−0.0286 (0.0288)	−0.0177 (0.0689)	0.0986* (0.0492)
Wealth index, squared	0.0096 (0.0462)	−0.0124 (0.0626)	−0.0033 (0.0306)	−0.0169 (0.0860)	−0.2042* (0.0605)
F statistic	0.09	0.28	4.48	0.64	8.76
p value	.91	.76	.01	.53	.00
Number of Observations	7,137	9,750	9,142	5,988	10,742

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: All regressions are weighted least squares regressions, weighted using provided HIV sample weights with clustering on the household. The dependent variable is an indicator for whether the respondent is HIV-positive. The F statistic tests the joint significance of the wealth index and the wealth index, squared. The wealth index is the fraction of nine assets or amenities that the respondent’s household has; the assets and amenities are radio, television, refrigerator, bicycle, motorcycle, car, telephone, electricity, and a flush toilet or pit latrine (must have nonmissing values for at least five components to be included). All regressions include a constant and an indicator for whether the respondent is female. Age group controls are five-year age groups. Huber-White standard errors are in parentheses.

^*p < .05

The results for schooling in Table 2 show that in Burkina Faso and Ghana education is unrelated to HIV infection. However, in Cameroon, Kenya, and Tanzania, there is evidence of a positive, nonlinear education gradient. This gradient, although concave, is positive for all but the most-educated respondents. These estimates imply that adults who completed six years of schooling are 3.5 (Cameroon), 2.7 (Kenya), and 2.5 (Tanzania) percentage points more likely to be infected with HIV than adults with no schooling. In other terms, compared with those with no schooling, adults with six years of schooling are 141% (Cameroon), 68% (Kenya), and 51% (Tanzania) more likely to be infected with HIV. Estimates in Table 3 show that the education gradient is robust to the inclusion of controls for five-year age group, sector of residence, and region of residence. These results show a significant positive and concave education gradient in HIV in Cameroon, Ghana, Kenya, and Tanzania. In particular, with the full set of controls, the results suggest that adults with six years of schooling are 2.1 (Cameroon), 0.8 (Ghana), 2.0 (Kenya), and 1.5 (Tanzania) percentage points more likely to be HIV-positive than adults with no schooling. These are quite large differences in HIV infection rates: infection rates among adults with six years of schooling are 51% (Cameroon), 44% (Ghana), 36% (Kenya), and 24% (Tanzania) higher than infection rates among those with no schooling. In Burkina Faso, the relationship between education and HIV is not significant, perhaps reflecting the fact that comparatively few Burkinabe have any formal education.

Tables 2 and 4 show mixed evidence of a wealth gradient in HIV. In Table 2, where regressors include linear and quadratic wealth terms, the relationship between wealth and HIV is significant in Burkina Faso, Ghana, and Tanzania (see Panel D). The estimated effects from this specification reveal that the rich are more likely to be infected with HIV in Burkina Faso, consistent with the Fan regression results in Figure 3. In Ghana, the rich are generally less likely to be infected with HIV. In Tanzania, the estimated effects suggest that the wealth gradient is positive but concave. In Table 4, I investigate whether these relationships are robust to the inclusion of controls for age group and sector and region of residence. Adding controls for five-year age group has little effect on the estimates in Burkina Faso, Ghana, and Tanzania, and reveals a positive, concave gradient in Cameroon that is significant at the 10% level. With controls for sector of residence (Panel B of Table 4), the gradient disappears in Burkina Faso and Cameroon. However, the wealth-HIV link in Ghana and Tanzania is not explained by sector or region of residence, as shown in Panels B and C of Table 4: an F test shows that the linear and quadratic terms are jointly significant after the full set of controls are added. Nevertheless, wealth and HIV are significantly related in only two of five countries (Ghana and Tanzania) after I account for region and sector of residence.

ROBUSTNESS CHECKS

The results in Panel C of Tables 3 and 4, which include region fixed effects, allow HIV prevalence to vary across areas. However, one might also be concerned that the gradient varies across high- and low-HIV areas. For instance, in areas with low levels of HIV, there might be less scope for variation in infection across educational groups. Therefore, one might expect that the true gradient is actually higher in high-HIV areas and much lower in low-HIV areas. I test whether this is the case by separately excluding low-HIV areas (regions with prevalence less than 1%) and high-HIV areas (regions with prevalence greater than 7%). I find that the education gradient is robust to the exclusion of low-HIV regions and to the exclusion of high-HIV regions (which are only in Cameroon, Kenya, and Tanzania). The wealth gradient, by contrast, is robust to the exclusion of low-HIV regions, but not to the exclusion of high-HIV regions. In particular, the wealth gradient in Tanzania disappears with the exclusion of high-HIV regions, leaving the wealth terms jointly significant only in Ghana.

In Tables 5 and 6, I further test the robustness of these results by estimating these relationships using alternate measures of education and wealth. Table 5 shows the relationship between HIV and education with controls for age, region, and sector of residence. Rather than including linear and quadratic terms for years of schooling, I include indicators for completed levels of schooling (primary, secondary, or higher). As in Table 3, the education terms are at least marginally significant in Cameroon, Ghana, Kenya, and Tanzania, and are consistent with a positive but concave gradient.

Table 5.

Education Gradient in HIV With Alternate Measures

HIV-Positive	Burkina Faso	Cameroon	Ghana	Kenya	Tanzania
Completed Educational Levels (ref. = no education)
Primary education	−0.0084 (0.0063)	0.0109 (0.0063)	0.0032 (0.0043)	0.0037 (0.0083)	0.0060 (0.0060)
Secondary education	0.0007 (0.0198)	0.0106 (0.0222)	−0.0083 (0.0060)	−0.0166 (0.0121)	−0.0555* (0.0241)
Higher education	0.0289 (0.0482)	−0.0483* (0.0241)	−0.0079 (0.0083)	−0.0215 (0.0154)	−0.0697* (0.0247)
Rural	−0.0207* (0.0066)	−0.0236* (0.0066)	−0.0037 (0.0041)	−0.0506* (0.0139)	−0.0597* (0.0092)
Age (ref. = 15–19)
20–24	0.0043 (0.0047)	0.0381* (0.0060)	0.0088* (0.0034)	0.0453* (0.0087)	0.0304* (0.0069)
25–29	0.0173* (0.0065)	0.0685* (0.0076)	0.0215* (0.0044)	0.0905* (0.0124)	0.0625* (0.0085)
30–34	0.0208* (0.0077)	0.0785* (0.0090)	0.0337* (0.0060)	0.0827* (0.0125)	0.0900* (0.0101)
35–39	0.0247* (0.0080)	0.0682* (0.0096)	0.0377* (0.0068)	0.0896* (0.0139)	0.0895* (0.0109)
40–44	0.0057 (0.0052)	0.0498* (0.0089)	0.0319* (0.0075)	0.0806* (0.0143)	0.0937* (0.0128)
45–49	0.0062 (0.0057)	0.0394* (0.0090)	0.0202* (0.0056)	0.0337* (0.0127)	0.0478* (0.0105)
Female	−0.0011 (0.0036)	0.0272* (0.0047)	0.0106* (0.0030)	0.0369* (0.0065)	0.0129* (0.0055)
Constant	0.0143 (0.0081)	0.0219* (0.0111)	−0.0068 (0.0066)	−0.0409* (0.0127)	0.0201 (0.0126)
F Statistic	0.69	4.16	2.02	2.12	33.97
p Value	.56	.01	.11	.10	.00
Number of Observations	7,143	9,751	9,146	5,994	10,743

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: All regressions are weighted least squares regressions, weighted using provided HIV sample weights with clustering on the household. All regressions include region fixed effects, though the coefficients are not reported. The dependent variable is an indicator for whether the respondent is HIV-positive. The sample includes women and men ages 15–49 who were tested for HIV. The F statistic tests the joint significance of primary education, secondary education, and higher education. These education variables are indicators for whether the respondent completed each level of schooling; for example, respondents who completed secondary school are coded as having completed both primary and secondary school. Rural is an indicator for whether the respondent currently lives in a rural area. Huber-White standard errors are in parentheses.

^*p < .05

Table 6.

Wealth Gradient in HIV With Alternate Measures

HIV-Positive	Burkina Faso	Cameroon	Ghana	Kenya	Tanzania
A. DHS Wealth Index Quintiles
Wealth quintile = 2	0.0052 (0.0059)	0.0031 (0.0074)	0.0033 (0.0053)	0.0200 (0.0122)	0.0049 (0.0072)
Wealth quintile = 3	0.0007 (0.0048)	0.0317* (0.0085)	0.0113* (0.0057)	0.0167 (0.0112)	0.0152 (0.0078)
Wealth quintile = 4	−0.0045 (0.0053)	0.0371* (0.0100)	−0.0007 (0.0066)	0.0410* (0.0123)	0.0468* (0.0093)
Wealth quintile = 5	−0.0054 (0.0080)	0.0279* (0.0111)	−0.0073 (0.0078)	0.0473* (0.0185)	0.0381* (0.0131)
F statistic	0.94	5.62	2.00	3.27	7.24
p value	.44	.00	.09	.01	.00
Number of observations	7,143	9,751	9,146	5,994	10,743
B. Wealth Index Components
Radio	0.0074 (0.0039)	−0.0110 (0.0065)	−0.0028 (0.0047)	−0.0088 (0.0105)	0.0084 (0.0068)
Television	0.0160 (0.0106)	0.0019 (0.0092)	−0.0123* (0.0056)	0.0030 (0.0114)	−0.0413 (0.0230)
Refrigerator	0.0100 (0.0194)	0.0070 (0.0105)	−0.0066 (0.0056)	−0.0198 (0.0224)	−0.0055 (0.0231)
Bicycle	−0.0074 (0.0074)	−0.0057 (0.0067)	−0.0034 (0.0041)	0.0085 (0.0087)	−0.0032 (0.0065)
Motorcycle	−0.0073 (0.0049)	−0.0072 (0.0077)	0.0126 (0.0103)	0.0110 (0.0382)	−0.0197 (0.0184)
Car	−0.0154 (0.0173)	0.0035 (0.0127)	0.0063 (0.0090)	−0.0128 (0.0172)	−0.0167 (0.0209)
Telephone	0.0143 (0.0206)	−0.0433* (0.0128)	−0.0061 (0.0067)	−0.0157 (0.0138)	−0.0111 (0.0158)
Electricity	−0.0077 (0.0128)	0.0097 (0.0081)	0.0019 (0.0055)	0.0109 (0.0174)	0.0007 (0.0168)
Toilet	−0.0081 (0.0047)	0.0072 (0.0113)	0.0013 (0.0055)	0.0026 (0.0129)	0.0163* (0.0078)
F statistic	1.20	2.08	1.71	0.84	3.09
p value	.29	.03	.08	.58	.00
Number of observations	7,109	9,696	9,099	5,913	10,664

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: The table reports results from two different specifications. In the first, the explanatory variables of interest are quintiles of the DHS wealth index. In the second, the explanatory variables of interest are the wealth index components included separately. All regressions are weighted least squares regressions, weighted using provided HIV sample weights with clustering on the household. Coefficients on the following regressors are not reported: rural; age groups 20–24, 25–29, 30–34, 35–39, 40–44, and 45–49; region fixed effects; female; and a constant. The sample includes women and men ages 15–49 who were tested for HIV. The dependent variable is an indicator for whether the respondent is HIV-positive. The F statistic in Panel A tests the joint significance of wealth quintiles 2–5. The F statistic in Panel B tests the joint significance of the nine assets/amenities that make up the wealth index. Wealth quintiles are indicators for quintiles of the DHS-provided wealth index. Radio, television, refrigerator, bicycle, motorcycle, car, telephone, and electricity are indicators for whether the respondent’s household owns or has these amenities. Toilet is an indicator for whether the respondent’s home has a flush toilet or pit latrine. Rural is an indicator for whether the respondent currently lives in a rural area. Huber-White standard errors are in parentheses.

^*p < .05

In Table 6, I estimate the relationship between wealth and HIV using two alternate measures of wealth, controlling for age and sector and region of residence. In Panel A, I include indicators for quintiles of the DHS-provided wealth index, where the first quintile is omitted. These results show that the quintiles are jointly significant in Cameroon, Kenya, and Tanzania, showing that wealth is positively related to HIV infection. In Ghana, the terms are marginally significant and indicate a positive but concave gradient. In contrast, the constructed wealth index (based on the fraction of nine assets or amenities in the household) in Table 4 shows significant results in only Ghana and Tanzania after the inclusion of controls for age, region, and sector of residence. In Panel B, I instead include each of the nine components of the wealth index separately and report results from an F test of their joint significance. These results show that wealth and HIV are related only in Cameroon and Tanzania, with marginally significant results in Ghana. These results and results in Tables 2 and 4 suggest that estimates of the wealth gradient in HIV are sensitive to the choice of proxy and specification.

These results suggest that there is indeed a link between HIV and socioeconomic status. Years of schooling is positively correlated with HIV for all but the elite, even taking into consideration age, sector of residence, and region of residence. Although the positive, concave education gradient is robust, the wealth gradient varies considerably across measures of wealth—perhaps because wealth is measured with error.

Even if the household characteristics studied by the DHS were perfect indicators of household wealth, we might still expect to find differences between the education gradient in HIV and the wealth gradient in HIV. Although wealth and education are highly correlated, they may signal different things about socioeconomic status. And although education among adults is fixed for most respondents, wealth may change over the life cycle; thus, current wealth may not be a good indication of lifetime wealth. Also, individuals’ access to household wealth may not be uniform across households, making household wealth a poor proxy for individual resources. Furthermore, although individual HIV infection is not likely to affect completed schooling, it could very well have an impact on household wealth. Therefore, to the extent that I am interested in the characteristics of the HIV-infected prior to infection, education may be a better reflection of socioeconomic status.

Other researchers using these data (e.g., de Walque 2006; Mishra et al. 2007) argued that wealth was positively related to HIV infection. These studies used quintiles of the DHS-provided wealth index, developed in Filmer and Pritchett (2001), as I do in Panel A of Table 6. Despite some methodological differences, the results are not dissimilar. De Walque (2006) found that wealth was positively related to HIV infection in Cameroon, Kenya, and Tanzania, consistent with my results in Panel A of Table 6. Likewise, Mishra et al. (2007) found a positive wealth gradient in HIV infection in these countries, as well as additional countries not under study here. But because I find that the results are not robust to using alternate measures of wealth (such as those in Table 4), I interpret my results differently.

Using these same data, de Walque (2006) concluded that education was not related to the probability of HIV infection. His main specification included a linear education term and quintiles of the DHS-provided wealth index. Here, I study education and wealth separately because they are highly collinear and measured with error. However, if I control for wealth using quintiles of the DHS-provided wealth index (as de Walque 2006 did), the education gradient continues to be positive and significant in Cameroon, Kenya, and Tanzania when education is included as a quadratic function. In Ghana, the gradient is positive and significant at the 10% level.10 However, if I include education only linearly with wealth quintile controls, education is not statistically significant for any of the five countries. This suggests that nonlinearities may be quite important in estimating the education gradient.

DISCUSSION

In this paper, I have shown evidence that more-educated respondents are more likely to be HIV-positive. But why is this the case? It could be that education causes certain behaviors that increase the probability of becoming infected with HIV. Alternatively, the same people who receive more schooling possibly also have other characteristics that make them more likely to become infected with HIV. Here, I am not able to distinguish between causality (the former) and omitted variable bias (the latter). Nevertheless, I investigate how education is related to proximate determinants of HIV.

One possible explanation for these results is that varying modes of transmission produce an education gradient. Anecdotal evidence suggests that intravenous drug use is quite low in Africa, so this is unlikely to drive the results. Likewise, transfusions are unlikely to play a role in explaining the positive relationship between HIV and education: in Tanzania, less than 1% of respondents received transfusions in the previous year. Furthermore, transfusions are unrelated to HIV and to education. Mother-to-child (vertical) transmission of HIV could affect only the youngest respondents because HIV did not become widespread until the mid-1980s (after most respondents were born).11 The positive education gradient is robust to the exclusion of those younger than age 20, which suggests that this gradient in HIV infection represents a gradient in sexually transmitted HIV infection.

To shed light on the mechanisms, I look at the association between education and risky sexual behavior. Previous studies found that the relationship between education and sexual behavior varies with the measure of behavior. For instance, Glynn et al. (2004) found that in four African cities, the more educated generally reported less risky sexual behavior, although some variability was reported by city, sex, and behavior. Using data from eight African countries, Glick and Sahn (2008) found that although education and wealth were associated with increased condom use among both men and women, education and wealth increased the demand for sexual partners among men. Education and these risk factors may or may not be causally related. Because any effect of education on HIV is mediated by sexual activity, public policies aimed at reducing HIV center on influencing behavior. Although I cannot determine here whether risky behavior is caused by or correlated with education, better understanding the links between risky sexual behavior and education could nevertheless help inform these policies.

In the estimates that follow, I look at the relationship between education and risk factors separately for women and men because past estimates have shown that these relationships differ for men and women (e.g., Glick and Sahn 2008). In Tables 7 (women) and 8 (men), I estimate weighted least squares regressions of HIV risk factors on linear and quadratic terms for years of schooling, five-year age group indicators, a sector of residence (rural) indicator, and region of residence indicators. The dependent variable is years since first intercourse (Panel A), partners in the past year (Panel B), and an indicator for whether the respondent had premarital sex (Panel C).12 Table 8 also includes an indicator for whether the respondent is not circumcised (Panel D), which has been shown to increase the risk of HIV infection (Auvert et al. 2005).

Table 7.

Education Gradient in Risk Factors: Women

	Burkina Faso	Cameroon	Ghana	Kenya	Tanzania
A. Years Since First Intercourse
Years of schooling	0.0042 (0.0307)	−0.0889* (0.0284)	0.0184 (0.0342)	−0.1226* (0.0323)	0.0218 (0.0331)
Years of schooling, squared	−0.0119* (0.0032)	−0.0081* (0.0019)	−0.0115* (0.0028)	−0.0147* (0.0021)	−0.0242* (0.0034)
F statistic	48.64	178.18	52.23	357.92	143.69
p value	.00	.00	.00	.00	.00
Number of observations	11,841	10,257	5,355	7,822	6,855
B. Partners in the Past Year
Years of schooling	0.0005 (0.0056)	0.0276* (0.0102)	0.0061 (0.0043)	0.0035 (0.0133)	−0.0006 (0.0046)
Years of schooling, squared	−0.0002 (0.0005)	−0.0018* (0.0006)	−0.0007* (0.0003)	−0.0010 (0.0006)	−0.0010* (0.0004)
F statistic	0.28	4.33	3.32	21.87	15.91
p value	.76	.01	.04	.00	.00
Number of observations	12,460	10,614	5,688	8,165	6,863
C. Had Premarital Sex
Years of schooling	0.0156* (0.0055)	0.0224* (0.0044)	0.0150* (0.0038)	0.0456* (0.0039)	0.0093* (0.0044)
Years of schooling, squared	0.0002 (0.0005)	−0.0009* (0.0003)	−0.0007* (0.0003)	−0.0028* (0.0002)	−0.0007 (0.0004)
F statistic	35.05	22.03	10.67	72.48	2.45
p value	.00	.00	.00	.00	.09
Number of observations	12,440	10,622	5,685	8,171	6,861

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: All regressions are weighted least squares regressions, weighted using provided sample weights with clustering on the household. Coefficients on the following regressors are not reported: rural; age groups 20–24, 25–29, 30–34, 35–39, 40–44, and 45–49; region fixed effects; and a constant. The sample includes women ages 15–49. The dependent variables are years since first intercourse (Panel A), partners in the past year (Panel B), and had premarital sex (Panel C). The F statistic tests the joint significance of years of schooling and years of schooling, squared. Rural is an indicator for whether the respondent currently lives in a rural area. Huber-White standard errors are in parentheses.

^*p < .05

Table 8.

Education Gradient in Risk Factors: Men

	Burkina Faso	Cameroon	Ghana	Kenya	Tanzania
A. Years Since First Intercourse
Years of schooling	0.0220 (0.0629)	0.2833* (0.0495)	0.0393 (0.0388)	0.2093* (0.0658)	−0.0429 (0.0447)
Years of schooling, squared	0.0024 (0.0056)	−0.0145* (0.0029)	−0.0008 (0.0025)	−0.0128* (0.0038)	−0.0016 (0.0036)
F statistic	2.02	17.12	2.05	5.76	5.66
p value	.13	.00	.13	.00	.00
Number of observations	3,597	5,248	4,987	3,555	5,647
B. Partners in the Past Year
Years of schooling	0.0204 (0.0206)	0.0349 (0.0357)	−0.0024 (0.0080)	−0.0082 (0.0117)	0.0073 (0.0124)
Years of schooling, squared	0.0005 (0.0020)	−0.0010 (0.0020)	0.0008 (0.0006)	0.0004 (0.0007)	−0.0009 (0.0010)
F statistic	7.40	1.74	4.62	0.24	0.40
p value	.00	.18	.01	.78	.67
Number of observations	3,598	5,258	5,010	3,569	5,650
C. Had Premarital Sex
Years of schooling	0.0153* (0.0052)	0.0185* (0.0046)	0.0096* (0.0034)	0.0172* (0.0049)	0.0036 (0.0040)
Years of schooling, squared	−0.0007 (0.0004)	−0.0009* (0.0003)	−0.0002 (0.0002)	−0.0008* (0.0003)	−0.0001 (0.0003)
F statistic	6.43	9.06	14.63	6.46	1.02
p value	.00	.00	.00	.00	.36
Number of observations	3,597	5,248	5,009	3,562	5,647
D. Not Circumcised
Years of schooling	−0.0086* (0.0041)	0.0031 (0.0043)	−0.0117* (0.0020)	−0.0153* (0.0052)	−0.0068 (0.0041)
Years of schooling, squared	0.0006 (0.0003)	−0.0002 (0.0002)	0.0005* (0.0001)	0.0004 (0.0003)	−0.0009* (0.0003)
F statistic	2.24	0.27	23.05	11.63	53.83
p value	.11	.76	.00	.00	.00
Number of observations	3,604	5,261	5,014	3,567	5,652

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: All regressions are weighted least squares regressions, weighted using provided sample weights with clustering on the household. Coefficients on the following regressors are not reported: rural; age groups 20–24, 25–29, 30–34, 35–39, 40–44, 45–49, 50–54 (where applicable), and 55–59 (where applicable); region fixed effects; and a constant. The sample includes men ages 15–59 (Burkina Faso, Cameroon, and Ghana), ages 15–54 (Kenya), and ages 15–49 (Tanzania). The dependent variables are years since first intercourse (Panel A), partners in the past year (Panel B), had premarital sex (Panel C), and not circumcised (Panel D). The F statistic tests the joint significance of years of schooling and years of schooling, squared. Rural is an indicator for whether the respondent currently lives in a rural area. Huber-White standard errors are in parentheses.

^*p < .05

Table 7 shows that, in all five countries, more highly educated women become sexually active at a later age; all else equal, this should reduce their risk of HIV infection. However, more highly educated women are also more likely to have had premarital sex. More highly educated men are similarly more likely to have had premarital sex (Burkina Faso, Cameroon, Ghana, and Kenya), as shown in Table 8. For men, years of schooling is also positively related to years since first intercourse (Cameroon and Kenya), which is a measure of sexual activity. However, men with more education are more likely to be circumcised (Burkina Faso, Ghana, Kenya, and Tanzania), which should reduce their risk of contracting HIV. For both men and women, education is not consistently positively or negatively associated with the number of partners in the past year.13

A related but distinct hypothesis is that more-educated respondents may be more likely to have concurrent relationships. Epstein (2007) argued that concurrent relationships may be partly responsible for HIV’s rapid spread in sub-Saharan Africa. Although the DHS data do not provide direct measures of concurrency, the surveys for Burkina Faso, Ghana, and Kenya contain information about the number of times that male respondents were away from home in the previous year. Because concurrency is thought to be related to migration for work, this may be a good proxy for concurrency. Indeed, results (not shown) suggest that moreeducated men are more likely to have been away from home during the previous year.

These results indicate an education gradient in most risk factors for HIV. However, there are many risk factors for HIV, including infection with other sexually transmitted diseases, that are not measured by these data. Although education is positively related to some risk factors and negatively related to others, there is strong evidence that education is related to behavior, which is consistent with the idea that differences in sexual behavior may explain the positive education gradient in HIV.

THREATS TO VALIDITY

The education gradient in HIV infection is robust across samples and measures and is echoed for risky sexual behavior. Despite this, one concern is that the gradient measured here is driven by nonresponse bias, rather than a true positive relationship between education and HIV. Mishra et al. (2006) showed that national estimates of prevalence based on these data are not biased by nonresponse. Nevertheless, one might be concerned about bias at the individual level.

The observed education gradient could be an artifact of HIV test nonresponse if nonresponse is correlated with both HIV status and years of schooling. Using survey information on HIV test nonrespondents and respondents, I investigate the importance of nonresponse. With controls for sex, five-year age group, region, and sector of residence, more-educated respondents are slightly more likely to have HIV test results in these data. However, I can reweight my results to adjust for nonresponse by educational category. In particular, I recalculate the education gradient in HIV infection (in Panel C of Table 3), multiplying each respondent’s weight by the inverse of the probability of having a test result in these data. The results are unchanged: adjusting for compositional differences between respondents and nonrespondents does not affect the estimated education gradient in HIV infection.

Nevertheless, if respondents and nonrespondents face different probabilities of HIV infection, this could bias my results. If nonrespondents are uniformly more likely to be infected with HIV, or if less-educated nonrespondents are more likely to be infected with HIV than more-educated nonrespondents, nonresponse bias could drive the estimated positive education gradient in HIV infection. To assess whether this is the case, I test whether there are differing education gradients in risky behavior that is often associated with HIV infection.

In Tables 7 and 8, I find a consistent education gradient in the probability of having premarital sex. Premarital sex is highly correlated with HIV infection, and the survey gives information about premarital sex for both HIV test respondents and nonrespondents. Therefore, I test whether the probability of having premarital sex differs across respondents and nonrespondents, as well as whether that difference varies with education.14 I find no significant difference in the probability of having premarital sex between respondents and nonrespondents, nor do I find a different education gradient in premarital sex. These results suggest that nonresponse bias is not driving the education gradient in HIV infection.

Differential survival could also be part of the reason that more-educated people are more likely to have HIV—that is, that the positive gradient reveals a greater likelihood of survival, rather than a greater likelihood of infection. I test this possibility by looking at those in the youngest age cohort. Respondents in this cohort may have contracted HIV, but it is unlikely that they would have died from the disease because survival after infection (in the absence of treatment) is generally thought to be roughly 7–10 years. Even within this cohort (ages 15–24) there is a positive education gradient, which suggests that differential survival does not explain these results.

CONCLUSION

My results show evidence of a robust positive education gradient in HIV infection that is not driven by age, sector of residence, or region of residence. Adults with six years of schooling are up to three percentage points more likely to be infected with HIV than adults with no schooling. With controls for sex, age group, region, and sector of residence, adults with six years of schooling are as much as 50% more likely to be infected with HIV than those with no schooling. There is likewise a positive education gradient in certain risk factors for HIV, including the likelihood of having had premarital sex. For instance, in Cameroon, men with six years of schooling are eight percentage points (or 12%) more likely to have had premarital sex than men with no schooling; Cameroonian women with six years of schooling are 10 percentage points (or 16%) more likely to have had premarital sex than women with no schooling. In principle, even small behavioral differences could produce a large gradient when compounded by assortative mating.15

Although my results indicate that wealth and HIV are interrelated, the wealth gradient in HIV is not robust to alternate specifications or measures of wealth. In light of this, existing studies showing that wealth and HIV are positively correlated should be interpreted with caution. This is not to say, however, that wealth and HIV are unrelated. However, because DHS wealth measures are constructed from questions about a limited number of household assets and amenities, wealth may be poorly measured in these data.

These results suggest that links between socioeconomic status and health may be more heterogeneous than previously thought. Although past research has suggested that socioeconomic status and health are positively related, I find evidence that the more educated are more likely to be infected with HIV in sub-Saharan Africa.

Figure 2.

Education Gradient in HIV Infection: Men

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: The sample includes men ages 15–59 (Burkina Faso, Cameroon, and Ghana), 15–54 (Kenya), and 15–49 (Tanzania) who were tested for HIV. The figure shows the fraction of men who tested positive for HIV at each educational level between 0 and 16 years. The size of the circle indicates the number of men at that educational level who were tested for HIV, weighted using provided HIV sample weights.

Figure 4.

Wealth Gradient in HIV Infection: Men

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: The sample includes men ages 15–59 (Burkina Faso, Cameroon, and Ghana), 15–54 (Kenya), and 15–49 (Tanzania) who were tested for HIV. The wealth index is the fraction of nine assets or amenities that the respondent’s household has; the assets and amenities are radio, television, refrigerator, bicycle, motorcycle, car, telephone, electricity, and a flush toilet or pit latrine (must have nonmissing values for at least five components to be included). Results are from Fan locally weighted regressions with a biweight quartic kernel with a halfwidth of 0.35, and are weighted using provided HIV sample weights.

Appendix Table A1.

Descriptive Statistics

Variable	Number of Observations	Mean	SD	Minimum	Maximum
Years of Schooling	66,982	5.3	4.5	0	26.0
Wealth Index	66,975	.311	.193	0	1
Rural	67,019	.639	.480	0	1
Age	67,019	29.2	10.2	15.0	59.9
Female	67,019	.655	.475	0	1
Years Since First Intercourse	65,199	11.7	10.0	0	45.6
Partners in the Past Year	66,912	.850	1.109	0	95.0
Had Premarital Sex	66,879	.636	.481	0	1
Circumcised	23,116	.859	.348	0	1
Used Condom at Last Intercourse	47,148	.145	.352	0	1
Transfusion in the Past Year	12,514	.009	.092	0	1
Primary Education	67,019	.504	.500	0	1
Secondary Education	67,019	.078	.268	0	1
Higher Education	67,019	.031	.173	0	1
Radio	66,953	.720	.449	0	1
Television	66,919	.219	.413	0	1
Refrigerator	66,905	.114	.318	0	1
Bicycle	66,938	.449	.497	0	1
Motorcycle	66,901	.110	.313	0	1
Car	66,900	.061	.239	0	1
Telephone	66,905	.084	.277	0	1
Electricity	66,967	.294	.456	0	1
Toilet	66,816	.745	.436	0	1
HIV-Positive	44,210	.047	.211	0	1

Source: DHS for Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003).

Notes: All results are weighted using provided sample weights. The sample includes women ages 15–49 (all countries) and men ages 15–59 (Burkina Faso, Cameroon, and Ghana), men ages 15–54 (Kenya), and men ages 15–49 (Tanzania). The wealth index is the fraction of nine assets or amenities that the respondent’s household has; the assets and amenities are radio, television, refrigerator, bicycle, motorcycle, car, telephone, electricity, and a flush toilet or pit latrine (must have nonmissing values for at least five components to be included). Rural is an indicator for whether the respondent currently lives in a rural area. Had premarital sex is an indicator for whether the respondent has had premarital sex; respondents who have not had intercourse are assigned a value of zero. Primary education, secondary education, and higher education are indicators for whether the respondent completed each level of schooling. Radio, television, refrigerator, bicycle, motorcycle, car, telephone, and electricity are indicators for whether the respondent’s household owns or has these amenities. Toilet is an indicator for whether the respondent’s home has a flush toilet or pit latrine. The transfusion data are from Tanzania.