The demographic and socioeconomic correlates of behavior and HIV infection status across sub-Saharan Africa.

Background: Predisposition to become HIV positive (HIV + ) is influenced by a wide range of correlated economic, environmental, demographic, social, and behavioral factors. While evidence among a candidate handful have strong evidence, there is lack of a consensus among the vast array of variables measured in large surveys.
Methods: We performed a comprehensive data-driven search for correlates of HIV positivity in >600,000 participants of the Demographic and Health Survey across 29 sub-Saharan African countries from 2003 to 2017. We associated a total of 7251 and of 6,288 unique variables with HIV positivity in females and males respectively in each of the 50 surveys. We performed a meta-analysis within countries to attain 29 country-specific associations.
Results: Here we identify 344 (5.4% out possible) and 373 (5.1%) associations with HIV + in males and females, respectively, with robust statistical support. The associations are consistent in directionality across countries and sexes. The association sizes among individual correlates and their predictive capability were low to modest, but comparable to established factors. Among the identified associations, variables identifying being head of household among females was identified in 17 countries with a mean odds ratio (OR) of 2.5 (OR range: 1.1-3.5, R2 = 0.01). Other common associations were identified, including marital status, education, age, and ownership of land or livestock. Conclusions: Our continent-wide search for variables has identified under-recognized variables associated with being HIV + that are consistent across the continent and sex. Many of the association sizes are as high as established risk factors for HIV positivity, including male circumcision.

Introduction

The number of new HIV infections in Sub-Saharan Africa has been declining for over a decade, but the prevalence of HIV has stagnated, around 1 million people are newly infected every year, and over 5 million are unaware of being HIV-positive[1,2]. The global strategy for improving the population burden of HIV calls for intensive identification of HIV-infected individuals in order to pursue high treatment coverage, which, in turn, is anticipated to reduce HIV transmission and incidence[1]. Sub-Saharan Africa–where the burden of HIV is highest – remains short of these major targets, including identifying at least 90% of those with HIV[1]. One challenge hindering progress at identifying those with HIV infection is the heterogeneity of HIV risk in terms of modes of transmission and prevalence. In other words, it is hard to know where and whom to target for testing and public health interventions when aiming for very high (>90%) identification rates. This heterogeneity is manifest in the proliferation of inquiries into HIV risk factors, without unifying approaches or systematic studies into HIV risk[3-7]. At the time of this writing, no systematic reviews of HIV risk factors in Africa have been published. In this paper we describe a systematic approach to the identification of a broad array of behavior, economic, and social factors in HIV risk and across multiple samples and geographic regions of Africa. The goal of this analysis is to help with identifying groups that may benefit from HIV-specific public health efforts.

The epidemiology of risk factor identification commonly proceeds by identifying candidate risk factors. These approaches are grounded in real-world observations but suffer from several limitations: generalizing to populations other than those directly examined is often tenuous; testing of one or several candidate risk factors may spuriously identify candidate risk factors;[8,9] and limited number of candidate risk factors may leave important relationships unobserved[10].

We address these shortcomings using a large-scale risk factor testing approach across the entire sample of individuals that had an HIV test as part of the Demographic and Health Surveys (DHS) between 2003 and 2017 in Sub-Saharan Africa, a sample of over 700,000 individuals extending our previous work to identify factors associated with HIV in Zambian females[11]. We test every variable or candidate risk factor available in DHS against individual-level HIV outcomes, a total of 32,353 variables (27,506 in females and 24,713 in males). This approach allows us to systematically identify variables that include behavioral, social, and economic factors and have a high degree of generalizability as risk factors for HIV, that play an important role in explaining who does and does not have HIV, over multiple survey waves across multiple countries in Sub-Saharan Africa, in men and women separately.

Methods

Data sources

The primary data for this work comes from the Demographic and Health Surveys (DHS). The DHS are nationally representative surveys conducted approximately every 5 years in many low- and middle-income countries. We used data from every individual who had consented and tested positive or negative for HIV-1[12]. There were a total of 50 surveys available for analysis for females and males each across 29 countries. We followed the STROBE statement and checklist in preparation and analysis of the data and approval for the use of the data was approved by the DHS program.

According to the DHS Program, procedures and questionnaires for standard DHS surveys have been reviewed and approved by ICF (the administrator of the DHS Survey) Institutional Review Board (IRB). Second, country-specific DHS survey protocols are reviewed by the ICF IRB as well as country-specific IRB. ICF IRB ensures that the survey complies with the U.S. Department of Health and Human Services regulations for the protection of human subjects (45 CFR 46). Further, before each interview, an informed consent statement is read to the participant, who can then decline or accept the interview. The informed consent statement emphasizes that participation is voluntary; that the respondent may refuse to answer any question, and information will be kept confidential. More information can be found on the DHS Program website: https://dhsprogram.com/Methodology/Protecting-the-Privacy-of-DHS-Survey-Respondents.cfm.

Appending the women and men’s recode

While appending the individual and men’s recode datasets (IR and MR), we also systematically identified common variables to map between MR and IR. The following prefix changes were made to variables in MR in this sequence: removing “_”, “mv” to “v”, removing “s”, “mcase” to “case”, “m” to “v”, “sm” to “mm” or “s” if no “mm” candidates exist, “dm” to “d”, and “sm6” to “s8” and “mv304a” to “v304a_”. These changes allowed us to harmonize variables with similar content across men's and women's surveys.

Additionally, IR variables that had at least a 70% syntactic similarity to MR variables (computed using Levenshtein distance) were mapped using Amazon Mechanical Turk (mTurk) Master Workers. These workers were trained briefly and presented with an MR variable and a set of potential IR candidates. They selected the IR candidate that they thought was the best match, or indicated that the MR variable was an individual category of the question header label, or that none of the IR candidates were appropriate. We validated the mTurk responses by excluding any workers who only selected the same positional answer (or semantically similar answers), or whose responses were not submitted through the mTurk interface. A majority vote rule (70%) was used to select the MR-IR variable mapping, and any candidate mappings that were indeterminate remained unmapped.

Harmonizing the household, male, and female surveys

Each DHS contains several parts: a census of all household members, in-depth interviews with reproductive-age (15–49 years old) women, men (15–59 years old), a ledger of household information, and HIV test results for each tested individual. We harmonized the surveys to match individuals in the household with available in-depth data, household data, and HIV test results. Variables that represented the same concept were identified by similarities in their variable name or the variable description, and merged together. Amazon Mechanical Turk (mTurk) was used to match men’s and women’s variables when the relationship between the variable names was intermediate[13]. A harmonized dataset was created for each country-wave survey, and each of these merged surveys was then analyzed.

We analyzed the harmonized surveys from all sub-Saharan African countries and survey years that performed the DHS HIV testing protocol. All surveys that reported an HIV prevalence ≥ 0.01 were included in our analysis (see Table 1).

Table 1

Sample sizes across sub-Saharan Africa.

Country	Sample Size	# Cases	Prevalence (SE)	Prevalence Female (SE)	Prevalence Male (SE)	Mean Age (SE)	% Female (SE)	% Urban (SE)
Angola	12260	271	1.95 (0.2)	2.57 (0.28)	1.21 (0.22)	27.71 (0.16)	54.21 (0.57)	70.77 (2.41)
Burkina_Faso	23434	295	1.17 (0.09)	1.29 (0.11)	1.04 (0.13)	29.1 (0.1)	53.93 (0.33)	26.43 (1.85)
Burundi	8790	167	1.43 (0.15)	1.72 (0.19)	1.11 (0.17)	27.75 (0.15)	52.78 (0.54)	12.55 (1.62)
Cameroon	10348	565	5.35 (0.3)	6.75 (0.44)	3.91 (0.31)	28.63 (0.12)	50.5 (0.49)	56.23 (2.68)
Chad	11034	172	1.56 (0.17)	1.8 (0.23)	1.29 (0.21)	28.11 (0.16)	52.73 (0.53)	25.74 (2.36)
Congo_Democratic_Republic	27530	313	1.17 (0.13)	1.62 (0.2)	0.7 (0.11)	28.96 (0.13)	51.76 (0.39)	39.74 (2.36)
Congo	12382	379	3.16 (0.3)	4.12 (0.46)	2.06 (0.28)	28.95 (0.18)	53.16 (0.42)	63.07 (3.91)
Cote_d'Ivoire	18070	721	4.33 (0.24)	5.49 (0.34)	3.11 (0.28)	27.82 (0.16)	51.15 (0.52)	49.44 (3.19)
Ethiopia	41165	787	1.43 (0.12)	1.86 (0.17)	0.98 (0.1)	28.31 (0.09)	51.48 (0.28)	20.79 (1.97)
Gabon	11264	502	4.24 (0.38)	5.81 (0.6)	2.69 (0.36)	28.65 (0.18)	49.67 (0.64)	87.65 (1.74)
Ghana	18795	351	1.96 (0.13)	2.56 (0.2)	1.31 (0.15)	29.89 (0.11)	51.79 (0.38)	49.62 (2.04)
Guinea	15407	272	1.69 (0.14)	2.03 (0.2)	1.27 (0.14)	29.71 (0.13)	55.12 (0.39)	36.11 (2.42)
Kenya	13359	902	6.54 (0.39)	8.32 (0.48)	4.63 (0.4)	28.8 (0.12)	51.78 (0.5)	24.71 (2.15)
Lesotho	18730	4308	23.67 (0.47)	27.57 (0.59)	18.99 (0.57)	28.66 (0.1)	54.52 (0.45)	29.57 (1.76)
Liberia	20362	320	1.65 (0.15)	1.88 (0.18)	1.39 (0.2)	29.37 (0.15)	54.04 (0.42)	48.69 (2.74)
Malawi	34409	3514	10.25 (0.31)	12.13 (0.39)	8.24 (0.33)	28.37 (0.08)	51.61 (0.28)	18.92 (1.18)
Mali	17978	199	1.16 (0.12)	1.36 (0.15)	0.93 (0.14)	29.58 (0.12)	53.25 (0.38)	30.21 (2.33)
Mozambique	16976	1466	7.48 (0.47)	8.75 (0.56)	5.97 (0.47)	32.49 (0.22)	54.43 (0.4)	31.03 (3.53)
Namibia	9309	1297	14.33 (0.6)	16.84 (0.75)	11.45 (0.81)	32.69 (0.23)	53.51 (0.72)	54.62 (2.64)
Niger	7897	70	0.63 (0.1)	0.58 (0.12)	0.7 (0.15)	30.31 (0.16)	57.32 (0.6)	22.55 (2.38)
Rwanda	41044	1206	2.8 (0.11)	3.38 (0.15)	2.17 (0.11)	28.8 (0.07)	52.28 (0.21)	17.15 (1.24)
Sao_Tome_and_Principe	4860	78	1.54 (0.25)	1.29 (0.29)	1.79 (0.35)	29.77 (0.21)	50.49 (0.92)	52.15 (5.62)
Senegal	18122	123	0.52 (0.07)	0.57 (0.11)	0.43 (0.12)	28.37 (0.11)	53.66 (0.73)	52.62 (2.53)
Sierra_Leone	21619	328	1.46 (0.12)	1.68 (0.15)	1.22 (0.16)	29.69 (0.13)	52.87 (0.29)	36.29 (2.36)
Swaziland	13008	2481	18.89 (0.58)	22.25 (0.69)	15 (0.67)	26.89 (0.12)	53.69 (0.53)	23.02 (2.47)
Tanzania	44896	2180	5.8 (0.19)	6.69 (0.24)	4.75 (0.22)	28.45 (0.06)	54.18 (0.26)	26.84 (1.54)
Togo	9321	200	2.5 (0.22)	3.11 (0.3)	1.85 (0.24)	29.34 (0.14)	51.65 (0.52)	44.81 (3.05)
Uganda	31644	1592	5.2 (0.21)	6.02 (0.27)	4.26 (0.22)	30.65 (0.11)	53.4 (0.31)	18.05 (2.08)
Zambia	40821	5634	13.68 (0.37)	15.34 (0.45)	11.97 (0.39)	28.68 (0.07)	50.7 (0.26)	45.19 (2.04)
Zimbabwe	43835	7158	15.71 (0.3)	18.35 (0.37)	12.76 (0.33)	28.01 (0.06)	52.8 (0.31)	35.71 (1.61)

Resolving value-label conflicts

When mapping categorical variables together in IR and MR, we resolved conflicting value labels by: 1) obtaining a complete list of all possible values of categorical variables, 2) merging all labels to the values if value-labels are available in the native Stata data dictionary, 3) appending both complete lists of value-labels from IR and MR, 4) removing duplicates from this new value-label dictionary, 5) finding conflicting value-labels (where the same value in a variable has two different labels, e.g. 1 = “Yes” in IR but 1 = “Definitely” in MR) and re-levelling one of the conflicting value-labels. If a conflict is found, then the value without a label receives priority (and the labelled conflict is relevelled); otherwise the IR value keeps its level while the MR value gets relevelled to the next larger level number in the set of levels in that variable. All value-labels are prefixed with their original value in square brackets in the label (e.g. the relevelled value of 2 might have the label “[1] Yes” because the value of 1 had two conflicting labels).

Merging recode datasets

The household member recode (PR) is then left-joined with the appended IR/MR file (with variables mapped and levels relevelled) in a 1:1 merge using (1) the cluster ID (v001), (2) household ID (v002), (3) respondent line number (v003), and occasionally, (4) structure ID (sstruct, sconces, svivi, sqnumber, or snumber) if it is needed to uniquely identify a respondent. Not all PR records are merged to a record in the appended IR/MR file, but are retained in the dataset with empty cells for the IR/MR variables. The HR dataset is left-joined to the PR/IR/MR dataset in a 1:m merge using (1) the cluster ID, and (2) household ID. The AR dataset (which contains the HIV test result of a subset of respondents) is left-joined to the PR/IR/MR/HR dataset in a 1:1 merge and left as missing if not available. In the AR file, v001 = hivclust, v002 = hivnumb and v003 = hivline. The GE dataset is left-joined in a 1:m merge using cluster ID only.

Handling missing recode datasets

The IR dataset is the minimum file needed for merging. If the IR dataset is not present, the datasets are not merged. Alerts are given if <75% of the IR/MR records are merged to PR, or if >25% of the IR/MR records are not mergeable to PR. Similarly, alerts are given if <75% of HR records are merged to PR, or if >25% of HR records are not mergeable to PR. Alerts are given if <50% of HIV test result data are merged to PR. The final merged dataset of PR, IR/MR, HR, and AR was saved as a Stata 15 file with the embedded data dictionary and named after the country-survey code and given the “flattened” suffix.

Processing of duplicate and categorical variables

We removed variables that exhibited no variation or were “duplicates”. We then computed pairwise Pearson correlations for all variables in a given survey across all 50 surveys. We identified groups of variables whose pairwise Pearson correlation was 0.9 or greater. For each of these groups of variables that were highly correlated with one another, we selected one to represent the group that had the largest average sample size for the variable across all surveys. These decision rules eliminated redundant variables, preserved most meaningfully continuous variables, and discretized most non-ordinal variables. All continuous variables were scaled and centered.

Selection and preparation of variables per survey

We retained all variables that had at least 90% complete data, a total of 29,092 and 25,980 unique variables in female surveys and male surveys, respectively, and 33,729 unique variables overall in both males and females. All variables with 30 or fewer levels were treated as binary variables. Variables with 30 or more levels were treated as continuous. We retained 7251 and 6288 unique variables in females and males, respectively and 8980 unique variables overall. The number of unique variables per country ranged from 337 (Congo) to 1659 (Malawi) for females and from 348 (Congo) to 1219 (Malawi) for males.

Number of surveys available per country

Five countries had 3 surveys available for analysis (Lesotho, Malawi, Rwanda, Tanzania, Zimbabwe), 11 had 2 surveys (Burkina Faso, Democratic Republic of Congo, Cote d'Ivoire, Ethiopia, Ghana, Guinea, Kenya, Liberia, Mali, Sierra Leone, Zambia). The 13 remaining countries had 1 survey (Angola, Burundi, Cameroon, Chad, Congo, Gabon, Mozambique, Namibia, Sao Tome and Principe, Senegal, eSwatini, Togo, and Uganda).

Pan Sub-Saharan sex-specific systematic meta-analyses across all years and countries

For each variable, we combined associations across all of the surveys (year and country combination) for males and females with a random effects meta-analysis procedure. Specifically, given an association between a variable (e.g., marital status, call it X1) and HIV + in a survey for a sex (e.g., Zambia, 2013-2014, females denoted by ; Zimbabwe, 2005, females, denoted by ), we estimated the overall association for each of the variables and measures of their heterogeneity, including the I2 for males and females using a DerSimonian-Laird random effects meta-analysis model, arriving at an overall estimate of association (e.g., for our example,)[14]. Out of the 7,251 and 6,288 variables unique available for females and males, 2,830 and 2,307 variables for females and males were measured in greater than one survey and thus available for a meta-analysis (Supplementary Fig. 1A, B). As examples, 4421 and 3921 variables were only available in 1 survey for males and females respectively. 79 and 57 variables were available across all 50 surveys for females and males respectively. For the 4421 and 3921 variables that appeared only in one survey (for females and males), we retained the survey-specific estimate. We used the metafor package in R to compute the meta-analytic association estimates[15].

Survey-specific association models and meta-analysis across sub-Saharan Africa

We associated each of the variables with HIV status using a weighted logistic regression model:Where s indexes the survey (including country, year, and female or male survey; e.g., Zimbabwe, 2005, female survey), j indicates the individual observation (all analyses are person-level analyses), and i indexes the variables for that survey. We estimated the Nagelkerke pseudo-R2 to assess the improved goodness of fit from a logistic model with zero variables (equivalent to the prevalence of HIV) to a model with X[16]. We used survey-weighted logistic regression model to account for the probability-based sampling of DHS, implemented in the survey package in R[17]. In a senstivity analysis, we tested the associations in a multivariate model with adjustments for age, place of residence (urban or rural), and the DHS wealth index (5 point scale of relative wealth). Last, for each variable, we combined associations across all of the surveys (year and country combination) for males and females with a random effects meta-analysis procedure, estimating the average association and heterogeneity of the associations across the surveys and countries.

We prioritized variables that were the most explanatory and whose associations were statistically non-zero across the surveys after correction for multiple hypotheses. Specifically, we report variables across pan-Saharan-Africa meta-analytic p-value lower than a conservative DHS-wide Bonferroni threshold of 1 × 10−6 (for 7,251 plus 6,288 variables, a Bonferroni threshold would be 0.05/13,539, or p < 3.7 × 10−6) and whose average R2 across the surveys were the top 25% of all R2 for males and females, which was equivalent to a R2 of 0.001.

Sex-specific meta-analyses within countries

For each variable, we combined associations of all of the surveys within a country (e.g., Zambia, 2007 and Zambia 2013-2014) to estimate a country-specific association for each sex. Given an association between a variable (i.e, marital status, call it X1) and HIV+ in a survey for a sex (i.e, Zambia, 2013–2014, females denoted by in the equation; Zambia, 2007, females, denoted by ), we estimated the overall association for each of the variables for males and females using a DerSimonian-Laird random effects meta-analysis model, arriving at an overall estimate of association ().

Comparison of associations assayed across multiple countries

To facilitate comparison of associations that are measured in multiple countries, we binned by the number of countries in which the variables appear, including 1 country, 2–10 countries, 11–19 countries, and 20–29 countries (Fig. 1, Fig. S2A–C, Table 2). Variables that were identified and assessed across a larger number of countries exhibited similar Nagelkerke R2 distributions (Fig. 1A, Fig. S2A); however, their odds ratios were attenuated (Fig. 1B, Fig. S2B).

Fig. 1

Volcano plots of -log10(p value) versus R2 and Odds Ratios.

A Nagelkerke R2 (B) Odds Ratios (C) Heterogeneity I2 of identified variables across all surveys. OR capped at 0.01 and 100 for visualization purposes. In A and B, Dotted line denotes Bonferroni-level of significance (1 x 10−6). In C, dotted line corresponds to I2 of 50%. f females. m males.

Table 2

Distributions of odds ratios, Nagelkerke R2, and I2 (heterogeneity) estimates across countries.

gender/number	Num Assoc.	25th OR	Median OR	75th OR	25th R2	Median R2	75th R2	25th I2	Median I2	75th I2
1 Country
Female	4543	1.2	1.62	3.69	8.56 × 10−5	3.58 × 10−4	1.17 × 10−3	.	.	.
Male	4089	1.24	1.78	12.67	8.39 × 10−5	3.29 × 10−4	1.12 × 10−3	.	.	.
2–10 Countries
Female	1702	1.19	1.67	136.4	2.56 × 10−4	6.30 × 10−4	1.39 × 10−3	0	62.1	99.01
Male	1488	1.24	2.3	526.3	2.42 × 10−4	5.91 × 10−4	1.26 × 10−3	0	64.85	99.14
11–20 Countries
Female	467	1.15	3.93	110.95	4.10 × 10−4	7.67 × 10−4	1.22 × 10−3	49.8	98.59	99.39
Male	368	1.17	3.44	189.43	4.09 × 10−4	7.70 × 10−4	1.57 × 10−3	47.49	98.38	99.41
20–29 Countries
Female	539	1.17	1.65	9.88	4.77 × 10−4	8.42 × 10−4	1.42 × 10−3	56.04	97.57	99.35
Male	343	1.26	3.68	56.47	4.71 × 10−4	7.10 × 10−4	1.22 × 10−3	67.7	99.14	99.48

Num Assoc. denotes the number of associations for that category.

Multivariate association study

The primary analyses in this project are univariate. This was done to maintain the tractability of the analysis, as multivariate analyses can increase the dimensionality and reduce sample sizes of the analysis substantially. However, we recognize that univariate associations, even those with very tight associations and meaningful effect sizes, can be confounded. We thus present the findings of a multivariate analysis with a fixed choice of covariates, replicating only the basic association study where all variables are examined and assessed as in the primary survey-specific analysis.

We run the following model on all candidate variables:

The indexing is identical to the model described, and we additionally include 3 variables: wealth (a 5-point relative wealth based on household assets, indexed by k), age in years, and place of residence (using an indicator for living in a rural area). As with the univariate analyses, the are combined across surveys and countries using a random-effects meta-analysis.

The results of the meta-analyzed multivariate analysis are presented in Supplementary Data 5-6, and a histogram comparing the OR values of the univariate and multivariate models are in Figure S3.

Prediction of HIV+ across aub-Saharan Africa

We calculated the predicted probability of HIV as a function of the variables that were identified in common in 29 African countries. First, we identified the latest survey with HIV status for each country and then built a logistic regression model for the top 10 identified factors (had p value less than 1 × 10−6 and R2 greater than 0.001) present in all 29 countries. We removed DHS participants who did not have all 10 variables available. Using this model, we estimated the area under the curve and the predicted probability for each individual.

The Area Under the Curve (AUC) estimates were calculated by varying the threshold at which a person was considered “HIV+” based on their predicted probability, estimating the sensitivity and specificity based on these thresholds, and calculating the area under the receiver operator curve (AUC). Precision-recall curves are preferred when the proportion of cases are much lower than non-cases by focusing on positive cases and are estimated by computing the area under the precision (positive-predictive value) versus recall (sensitivity) curve (PRAUC). For random predictors, the AUC will be equal to approximately 0.5 and the PRAUC will be equal to the prevalence of HIV+. Therefore, a predictor that achieves an AUC greater than 0.5 or a PRAUC greater than the prevalence is considered to be better than random. A perfect AUC or PRAUC is 1.

Finally, we assess the concentration of HIV risk. We do this by calculating the proportion of the population that carries the predicted probability of HIV from our model. This allows us to calculate a “gini” coefficient of HIV risk by estimating the cumulative HIV risk distribution that is accounted for by the portion of the sample population using the ineq package in R[18]. For example, the gini coefficient allows us to see if 20% of HIV risk is accounted for by 80% of the population as ranked by predicted risk[19].

Statistics and reproducibility

Data preparation was performed using Stata 15 MP on a MacBook Pro with a 2.7 GHz Intel Core i7 processor and 16GB of RAM. In Stata 15, setting the maxvar to 18000 (or higher) is necessary given the high number of variables. The codebase for data preparation is available on Zenodo (10.5281/zenodo.6819777)[20].

The source code for meta-analysis across surveys is available here: https://github.com/chiragjp/dhs_hiv_meta. A website of the findings is here: https://www.chiragjpgroup.org/dhs_hiv_meta/. We have placed all of our summary statistics, including the overall and country-specific sample sizes and meta-analytic odds ratios, R2, I2, standard errors, and p-values in Supplementary Data 1–4. All summary statistics for all surveys can be found on Zenodo (see reference[20]).