Socio‐economic gradients in prevalent tuberculosis in Zambia and the Western Cape of South Africa

Abstract Objective To describe the associations between socio‐economic position and prevalent tuberculosis in the 2010 ZAMSTAR Tuberculosis Prevalence Survey, one of the first large tuberculosis prevalence surveys in Southern Africa in the HIV era. Methods The main analyses used data on 34 446 individuals in Zambia and 30 017 individuals in South Africa with evaluable tuberculosis culture results. Logistic regression was used to estimate adjusted odds ratios for prevalent TB by two measures of socio‐economic position: household wealth, derived from data on assets using principal components analysis, and individual educational attainment. Mediation analysis was used to evaluate potential mechanisms for the observed social gradients. Results The quartile with highest household wealth index in Zambia and South Africa had, respectively, 0.55 (95% CI 0.33–0.92) times and 0.70 (95% CI 0.54–0.93) times the adjusted odds of prevalent TB of the bottom quartile. College or university‐educated individuals in Zambia and South Africa had, respectively, 0.25 (95% CI 0.12–0.54) and 0.42 (95% CI 0.25–0.70) times the adjusted odds of prevalent TB of individuals who had received only primary education. We found little evidence that these associations were mediated via several key proximal risk factors for TB, including HIV status. Conclusion These data suggest that social determinants of TB remain important even in the context of generalised HIV epidemics.


Introduction
Socio-economic gradients in access to health care mean the association between tuberculosis (TB) diagnosis and socio-economic position (SEP) may not reflect social gradients in communities [1][2][3]. Prevalence surveys enable more accurate estimation of associations between SEP and TB.
ZAMSTAR [13][14][15] was a large community-randomised trial in Zambia and the Western Cape of South Africa. Using a 2 9 2 factorial design, it tested case-finding interventions, one delivered in the community, one in households. In 2010, after these interventions, a TB prevalence survey was conducted. Data were captured concurrently on SEP, socio-demographic characteristics, proximal risk factors for TB, plus current TB and HIV treatment. HIV testing was offered. The household (but not the community) intervention may have reduced TB prevalence (adjusted prevalence ratio 0.82; 95% CI 0.64-1.04) [15].
Here, we report associations between both individual educational attainment and household wealth, two measures of SEP [16], and prevalent TB in ZAMSTAR. We calculate population attributable fractions (PAFs) for SEP by each measure. We use mediation analysis [17] to evaluate potential mechanisms for social gradients and describe differences in social gradients between diagnosed and prevalent disease.

Ethics
The protocol was approved by ethics committees at Stellenbosch University, the University of Zambia and London School of Hygiene and Tropical Medicine. Prevalence survey participants provided written informed consent.

Population
The survey was conducted in 16 communities in Zambia and eight in the Western Cape of South Africa. The communities, both urban and rural, had TB notification rates >400 per 100 000 per annum, high HIV prevalence and were the catchment populations of clinics offering TB diagnostics.
In each community, standard enumeration areas (SEA) were identified from census maps and visited in random order. Once 4000 adults were enrolled in a community, no further SEAs were included; for each SEA included, all households were visited. Up to three visits were made to each household.

Measurement
Data on household-level exposures were obtained from a responsible adult. Other data were obtained from individuals.
Participants were asked whether they had ever tested for HIV and, if so, whether they were willing to report their status. All were offered point-of-care HIV testing, regardless of self-reported status. Blood sugar measurement was offered concurrently. These tests were performed in households in Zambia and at mobile centres in South Africa.
Measures of household crowding, exposure to indoor air pollution and migration were derived from answers to other questions ( Table 1). The exact wording of these questions is detailed in Appendix 1.
A single respiratory specimen was collected from each participant and cultured in duplicate in liquid culture. When exploring the association between SEP and prevalent TB and in the mediation analysis, we included only individuals with an 'evaluable' sputum sample. This meant a non-contaminated sample which passed quality controls [15]. For the main analyses, prevalent TB was defined as culture positivity.

Conceptual framework
Proximal determinants of TB infection or progression from infection to disease were considered potential mediators of the association between SEP and prevalent tuberculosis ( Figure 1). Age, sex and community were considered potential confounding variables. No adjustment for previous TB was made asgiven it may be similarly associated with SEPthis might artificially diminish any association between prevalent TB and SEP. To assess for social gradients in access to TB treatment, the primary analysis was repeated with self-report of current TB treatment as the outcome. *For 16% of individuals in Zambia and 39% in SA, there was no information from either serology or self-report. This was because these individuals did not give a blood sample for HIV testing, and either reported they had never tested for HIV or that they had tested but did not know/did not wish to disclose the result of their last HIV test.

Analysis plan
Given their different socio-economic landscapes, separate analyses were conducted for each country. Household wealth indices were generated for each country by principal components analysis [18] (PCA) using data from all consented participants, irrespective of whether their sputum sample was evaluable. The variables included were household ownership of a set of assets, dwelling type, the material used to construct the floor, available sanitation facilities and the household's source of drinking water. We considered the first principal component only, with scores calculated as the sum of the factor weights for each variable. Individuals were assigned to wealth index quartiles.
Analyses of TB prevalence and current TB treatment were restricted to individuals with an evaluable sputum sample. Logistic regression models were fitted, adjusting for age group and gender, allowing the pattern by age to differ by gender.
To control for confounding by community, community was included as a fixed effect. In Zambia, communities were aggregated into four 'regions', each containing four communities because, in eight communities, fewer than 10 cases of prevalent TB were found. Aggregation considered force of TB infection (high or low), from a baseline survey [19], then divided communities into rural, urban (non-Lusaka) and urban (Lusaka). Communities in the same 'region' were not necessarily geographically close. We accounted for clustering by SEA using robust standard errors.
PAFs were calculated for each measure of SEP in each country. We estimated the prevalent TB that would be avoided if all individuals had the same prevalence as those in the highest household wealth quartile. We then   TB that would be avoided if individuals with no upper secondary education had the same  prevalence as individuals with some upper secondary edu-cation, leaving the prevalence in college and universityeducated individuals unchanged. We used the approach of Valeri and VanderWeele [17] to assess how much of the association between SEP and prevalent TB might be mediated via each of a set of proximal risk factors (Table 1). This permits decomposition of total effects into that explained by (indirect effect) and that not explained by the putative mediator (direct effect). Age, gender and community or region were held constant and clustering by SEA disregarded (in earlier analyses, it had minimal impact upon standard errors).
Missing data 21 843 individuals in Zambia and 9793 in South Africa had complete data for all variables used in these analyses. There were no missing data on educational attainment or household wealth, meaning the main analyses excluded only 401 (Zambia) and 19 individuals (South Africa) with missing age data.
For the mediation analyses, we excluded individuals, with missing data on age, migration or household crowding -2410 in Zambia and 961 in South Africa. A composite measure of diabetes, incorporating self-report, was used to eliminate missingness in this variable (Table 1).
For missing HIV status, we explored two approaches. First, we reduced missingness by generating a measure incorporating self-reported status ( Table 1) then performed mediation analysis excluding those still having missing HIV status. In the second approach, we repeated the HIV mediation analysis imputing missing HIV test results assuming missing at random (MAR), i.e. that the value of the missing data, after accounting for measured predictors of HIV status, was not predicted by unobserved data. The imputation model included data on selfreported HIV status and all variables included in the analytical model or thought to predict either HIV status or missingness of HIV status [20].

Tools
Most analysis was conducted in Stata 13. The mediation analysis using HIV status imputed under MAR was performed in R.

Sensitivity analyses
We repeated our main analyses stratified by gender; excluding individuals who reported previous TB; using a simple asset count as the measure of household wealth; and, for Zambia, adjusting for community rather than region as a fixed effect. We also repeated the PCA and the main analysis in Zambia using only data from the 12 urban communities.
We also repeated our main analyses, stratifying individuals who tested and/or self-reported HIV-positive according to whether they self-reported that they were on antiretroviral therapy (ART), as follows: those who selfreported they were HIV-positive and that they were taking ART; those who self-reported that they were HIVpositive and that they were not taking ART; and those who tested or self-reported HIV positive, but for whom, we had no data about whether they were taking ART. The latter group included people who self-reported that they had never previously tested, self-reported that the last time they tested the result was HIV-negative, or declined to discuss prior HIV testing.
We restricted the diabetes and HIV mediation analyses to individuals with test results from a blood sample. Given early symptoms might alter tobacco and alcohol intake, we repeated these mediation analyses including only current vs. never smokers and heavy vs. never drinkers. We repeated the mediation analysis for malnutrition using a different measure of food security. That question asked 'During the past 3 months, did it happen even once that you or any member of your family experienced hunger because you did not have any food to eat?' For key mediators, HIV and IAP, we tested the sensitivity of our mediation analyses to choices made regards the level at which to fix age group, gender and community/region.

Survey participation
There were 57 809 individuals in Zambia and 32 792 in South Africa who consented to participate in the study, representing 71% and 78% of those approached. Men were under-represented in both countries.
There were 34 446 evaluable culture results in Zambia and 30 017 in South Africa meaning outcome data were available for 60% and 92% of individuals who consented. The proportion of unevaluable culture results differed by community. Within communities, there was little association between whether cultures were evaluable and individual characteristics. Characteristics of individuals included in the primary analysis are presented in Table 2.
‡HIV test result was missing in 32% in Zambia and in 66% in South Africa, respectively, because these individuals did not consent to give a blood sample for HIV testing.
§A derived variable incorporating test result plus self-report (see Table 1).
[18] explained 17.2% and 20.1% of total variation in Zambia and South Africa. The distributions of wealth score by country and by community or region are shown in Figures S1-S2. There was little evidence of clumping or significant truncation. Examination of household assets associated with high and low wealth scores suggested the circumstances of individuals in these households were qualitatively different. Household wealth scores correlated closely with individual educational attainment, particularly in Zambia ( Figures S3-S4).

Primary analyses
We observed associations between low SEP and prevalent TB in both countries, by both measures (Tables 3-4

Population attributable fractions
Were everyone to have the TB prevalence of those in the highest quartile of household wealth, then 23.5% (95% CI À10.7-47.1%) and 13.5% (À0.6-25.6%) of prevalent TB might be avoided in Zambia and South Africa, respectively (Table S2). Were individuals with no upper secondary education to have the rates of TB of individuals with some upper secondary education, then 19.3% (95% CI À3.1-36.9%) and 15.1% (95% CI 7.7-21.9%) of prevalent TB might be avoided in Zambia and South Africa, respectively (Table S2).

Mediation analyses
The associations between SEP and HIV status are shown in There was little evidence, after accounting for covariates, that the observed associations between prevalent TB and low SEP were mediated by any proximal risk factors considered (Table 5); that is, the conditional natural indirect effects were all approximately equal to one, with conditional natural direct and indirect effects presented on the odds ratio scale [17].

Social gradients in receipt of TB treatment
Social gradients in prevalent TB observed in Zambia were stronger than for diagnosed TB, but the trend was similar. In South Africa, the strength of the association between education and current TB treatment was similar to that between education and prevalent TB. There was also an association between wealth index and current TB treatment, but we did not observe that the odds of diagnosed TB fell with every increment in SEP, as with prevalent TB. These data are in Table 4.

Sensitivity analyses
Our findings were robust to the sensitivity analyses undertaken. Importantly, the social gradient in prevalent TB using a simple asset count was similar to that seen using the household wealth index (Table 4).
However, there were insufficient data to permit exploration of ART use, in addition to HIV status. Among 24 440 Zambians for whom we had information on HIV status, 842 self-reported they were HIVpositive and not taking ART, 1611 self-reported they were HIV-positive and taking ART (6.6% of the total population, and 32.4% of those who tested or selfreported HIV-positive), and 2521 were HIV-positive based on survey testing, but they did not self-report they were HIV-positive and so there was no information on ART use. Among 11 340 South Africans for whom we had information on HIV status, 767 selfreported they were HIV-positive and not taking ART, 1022 self-reported they were HIV-positive and taking ART (9.0% of the total population, and 34.5% of those who tested or self-reported HIV-positive), and 1176 were HIV-positive based on survey testing, but they did not self-report that they were HIV-positive and so there was no information on ART use.

Main findings
We observed strong social gradients in prevalent TB in two very different settings in Southern Africa. These were steeper in Zambia. These gradients were not explained by a number of putative mediating variables. The association between SEP and being on TB treatment was less clear. As observed in the Zambian ZAMSTAR pilot prevalence survey [21] (and in a study of diagnosed TB in Brazil [22]), a substantial proportion of TB might be avoided if people with low SEP suffered the same burden as those with high SEP.
Poor educational attainment was more strongly associated with prevalent TB than household wealth. Individual-level SEP may be more important than householdlevel SEP; absolute measures of SEP may better predict TB than relative measures; human capital (knowledge or skills) might be more protective than wealth or living conditions; or longer-term disadvantage might be important with education fixed early in adulthood whereas asset ownership may fluctuate. Alternatively, assets included in our index may not fully explain variation in SEP in these communities.

Limitations
The size of the study and the consistency of our findings in two settings and by two measures of SEP suggest this is not a chance finding. However, a number of biases might affect our estimates.
Within study communities, it is possible sickness or employment affected recruitment into the study. We would expect any bias introduced to be modest.
The weighting of components of the wealth indices broadly agreed with our beliefs about which assets were associated with relative wealth. However, choices made in the construction of wealth indices can bias estimates of the association between SEP and TB.
Inclusion of assets directly associated with the outcome can lead to overestimation of the health inequalities [23]. For TB, it has been argued that measures of household construction should not be included, given they may affect ventilation [24]. That we obtain similar results when using a simple asset count, without measures of household construction, suggests including them did not substantially affect our estimates.
The inclusion of 'urban' assets in wealth indices can theoretically attenuate the association between low SEP and TB, given urban areas tend to be wealthier and have a higher burden of TB [16,24]. However, we obtained very similar results in Zambia when both the PCA and the main analysis were undertaken using only data from the 12 urban communities. A previous study in Zambia suggested excluding urban variables did not alter the association between SEP and TB [24]. Additional discriminatory power obtained by including urban assets may offset any attenuation. Many people in 'rural' communities in this study were living in peri-urban areas.
The association between SEP and prevalent TB did not appear to be mediated by any of the proximal risk factors measured. An important caveat here is that many of these risk factors were imperfectly measured. Our measure of diabetes was insensitive [25]. We did not measure protein intake, which an earlier study from Zambia suggested might be the component of malnutrition that is associated with TB [11]. We had no data on recent migration.
Three variables were dichotomised to enable them to be included in the mediation analysis. Of these, diabetes was associated with higher SEP, so could not explain the association between low SEP and prevalent TB observed. Furthermore, in these communities, the association between diabetes and prevalent TB is weak and diabetes only thought to explain around 1.0% of prevalent TB (95% CI 0.1-1.9%) [26]. No association was observed between household crowding and prevalent TB, even when household crowding was more finely categorised. Time spent outside the community appeared protective, at least in Zambia, and was associated with higher SEP in Zambia and lower SEP in South Africa. However, in Zambia, the majority of people (91%) had either never lived outside the community or had done so for more than 10 yearsthat is, it was already essentially a binary variable. A finer categorisation of the migration variable was not informative, as there were too few cases of TB among individuals who had lived outside the community for between 1 and 10 years. Many participants declined to test for HIV. The inclusion of self-reported status in one measure of HIV will have resulted in some misclassification, given HIV remains stigmatised and a proportion of individuals will have become HIV-positive since their last test. However, the odds ratios for the association between HIV and prevalent TB were consistent with previous studies, and similar in Zambia and South Africa.     There was little evidence to show that HIV mediated the association between SEP and prevalent TB. This was surprising given lower SEP was associated with HIV positivity in complete case analysis and HIV infection clearly associated with prevalent TB. This was true for both men and women and among both younger and older individuals.
In an analysis imputing missing HIV serology data assuming MAR, we also did not find evidence of mediation. However, HIV status in population based surveys is often missing not at random (MNAR) [27,28]. Individuals who know themselves to be HIV positive are more likely to decline testing. The imputation methods we used are not valid if there is substantial departure from MAR. We have explored the extent to which MNAR might affect our conclusions in a sensitivity analysis, finding little evidence of mediation by HIV status across a range of plausible MNAR assumptions [20].
The association between SEP and HIV may be both complex and dynamic [29]. Given HIV is a key risk factor for TB, in settings with a stronger social gradient in HIV positivity, we would expect HIV to, at least partially, mediate any association between SEP and prevalent TB. However, improvements in HIV care, including the earlier initiation of ART, may attenuate this effect.
ART modifies the association between HIV and TB [30], but we were unable to examine the effect of ART due to data limitations. An increasing proportion of HIVpositive people are taking ART, and with WHO now recommending ART initiation regardless of CD4 count, this trend is likely to continue. The impact of ART on the social patterning of TB should be examined in future analyses. Other potential mediators of the social gradient in prevalent TB were also not examined. These include social contact carrying a TB risk and structural barriers to accessing TB treatment.
Our analysis accounts for within-community and not between-community social gradients in prevalent TB. A brief exploratory analysis of the Zambian data suggested that modest between-community social gradients also existed with higher TB prevalence observed in poorer and less well-educated communities.

Strength of evidence for a causal association
A key assumption behind the PAFs that we present is that the association between SEP and prevalent TB that we describe is causal.
TB disease is a cause of impoverishment [31]. As educational attainment is usually fixed in early adult life, reverse causality is unlikely to explain its association with prevalent TB. As a measure of SEP, household assets are *2410 Zambians with missing data on age, household crowding or migration excluded from these mediation analyses. †961 South Africans with missing data on age, household crowding or migration excluded from these mediation analyses.
‡The measure incorporating self-report (see Table 1). §Missing HIV status imputed under the Missing at Random assumption.
considered relatively stable to short-term economic shocks [32,33], such as illness. However, individuals with TB may sell assets to fund hospital visits or when they became too unwell to work. This would result in overestimates of the association between household wealth and prevalent TB. However, prevalent TB or current receipt of TB treatment was not strongly associated with household sale of assets in either country. In Zambia, 8.2% of individuals included in the primary analysis lived in households reporting sale of assets in the preceding 18 months. The equivalent figures for individuals with prevalent TB and individuals in receipt of TB treatment were 12.5% and 11.3%, respectively. Sale of assets was reported more frequently in less assetrich households. In South Africa, 3.1% of individuals in the primary analysis, 3.2% of individuals with prevalent TB and 4.3% of individuals on TB treatment lived in households reporting sale of assets in the preceding 18 months.
An alternative explanation for the association between SEP and prevalent TB that we describe is residual confounding. Under our conceptual framework, proximal determinants of prevalent TB would be considered to be on the causal pathway rather than putative confounding variables. However, we cannot exclude the possibility that some of the observed association is explained by confounding by upstream factors, such as a healthier environment or better governance. Including community or region as a fixed effect might not control for such upstream factors if, for example, they were operating at a different scale or if political constituencies and study communities did not overlap. The extent to which this matters depends on whether one wishes to isolate the pure effect of SEP from its environmental and contextual determinants.

Results in context
The social gradient in diagnosed disease was less clear than for prevalent disease. The effects were subtle, but this might suggest some social patterning in access to treatment, as noted elsewhere [1][2][3]. Note the ZAMSTAR communities had their diagnostic capacity strengthened through participation in this trial.
A prevalence survey in Myanmar found prevalent TB was not associated with SEP after stratifying by rurality [3]. The recent prevalence survey in Zambia observed prevalent TB was associated with lower SEP in urban areas but no association between prevalent TB and SEP in rural areas [8]. However, our results are consistent with large TB prevalence surveys from South India [1], the Philippines [5], Vietnam [5], Bangladesh [2,5,34], Shandong Province in China [7], Kenya [5] and Tanzania [9] which all found prevalent TB to be associated with lower SEP. They are also consistent with a prevalence survey in Zimbabwe which found a non-significant reduction in the odds of prevalent TB per asset owned in univariable analysis [4].
ZAMSTAR pilot prevalence surveys in Zambia [11] and the Western Cape [12] both reported associations between lower SEP and prevalent TB, with some evidence that this was mediated by poor protein intake [11].
Studies of the association between diagnosed disease and SEP in Southern Africa [10,[35][36][37] have yielded divergent results, perhaps due to differing social gradients in access to health care.
Odone reported interesting differences in the associations between various measures of SEP and diagnosed TB in a large study from Northern Malawi [10]. Whilst ownership of assets appeared protective, better household construction and working in the cash rather than the subsistence economy were associated with higher rates of TB [10].
There are plausible reasons why aspects of higher SEP might place one at greater risk of TB infection [10,38]. Employed individuals may have greater exposure to other people whilst commuting; in the workplace; and, perhaps, via more frequent attendance at social or commercial venues, made possible by their earnings. Alternatively, better constructed homes may be less well ventilated [39]. This might explain the association between better household construction and diagnosed TB observed by Odone [10]. However, a growing body of evidence suggests that most Mycobacterium tuberculosis transmission in Southern Africa occurs outside the household [40][41][42][43].

Conclusions
We have shown that steep socio-economic gradients in prevalent TB persist in Southern African communities with high HIV prevalence. These associations are probably causal. If so, low SEP is responsible for a substantial proportion of prevalent TB in these communities. We were unable to identify any mediating factors that explained these associations. Confirmation of the previously noted [11] association between poor protein intake and prevalent TB would be valuable. Future studies of the association between SEP and TB must consider differences by SEP in access to TB treatment as part of the explanation for any observed associations. Longitudinal studies would be valuable in establishing causality and, potentially, in measuring the effect of interventions to reduce poverty.
That previous studies from HIV-endemic areas of Kenya [5], Zimbabwe [4], Tanzania [9], the Western Cape [12], and (at least urban) Zambia [8,11] have also found an association between low SEP and prevalent TB suggests this may be the case more generally. These findings lend support to the inclusion of poverty alleviation and social protection as 'key actions' under Pillar 2 of WHO's End TB Strategy [44]. National Treatment Programmes in HIV-endemic settings, as elsewhere, must ensure that their services can be accessed by individuals with little education, members of asset-poor households, and other less advantaged members of the community.

Supporting Information
Additional Supporting Information may be found in the online version of this article: Figure S1. The distribution of household wealth index overall and by community in Zambia. Figure S2. The distribution of household wealth index overall and by community in South Africa. Figure S3. Household wealth score by educational attainment for individuals in Zambia. Figure S4. Household wealth score by educational attainment for individuals in South Africa. Table S1. By country, frequency of asset ownership, and the weights assigned to them in construction of household wealth index. Table S2. Population attributable fractions of prevalent TB for household wealth and educational attainment in both countries .  Table S3. The adjusted associations between HIV status and measures of socioeconomic position. Table S4. The adjusted associations, in Zambia, between measures of socio-economic position and putative mediating factors. Table S5. The adjusted associations, in South Africa, between measures of socio-economic position and putative mediating factors. Table S6. The adjusted associations between putative mediators and prevalent TB in Zambia. Table S7. The adjusted assocations between putative mediators and prevalent TB in South Africa.