Risk factors for vaginal fistula symptoms in Sub-Saharan Africa: A pooled analysis of national household survey data

Background: Vaginal fistula (VF) is one of the most severe maternal morbidities with the immediate consequence of chronic urinary and/or fecal incontinence. The epidemiological evidence regarding risk factors for VF is dominated by facility-based studies. Our aim is to estimate the effect size of selected risk factors for VF using population-based survey data. Methods: We pooled all available Demographic and Health Surveys and Multiple Indicators Cluster Surveys carried out in sub-Saharan Africa that collected information on VF symptoms. Bayesian matched logistic regression models that accounted for the imperfect sensitivity and specificity of self-reports of VF symptoms were used for effect size estimation. Results: Up to 27 surveys were pooled, including responses from 332,889 women. Being able to read decreased the odds of VF by 13 % (95 % Credible Intervals (CrI): 1 % to 23 %), while higher odds of VF symptoms were observed for women of short stature (<150 cm) (Odds Ratio (OR) = 1.31; 95 % CrI: 1.02-1.68), those that had experienced intimate partner sexual violence (OR = 2.13; 95 % CrI: 1.60-2.86), those that reported sexual debut before the age of 14 (OR = 1.41; 95 % CrI: 1.16-1.71), and those that reported a first birth before the age of 14 (OR = 1.39; 95 % CrI: 1.04-1.82). The effect of post-primary education, female genital mutilation, and having problems obtaining permission to seek health care were not statistically significant. Conclusions: Increasing literacy, delaying age at first sex/birth, and preventing sexual violence could contribute to the elimination of obstetric fistula. Concomitant improvements in access to quality sexual and reproductive healthcare are, however, required to end fistula in sub-Saharan Africa. DHS and MICS surveys conducted in sub-Saharan Africa that included questions about VF symptoms were considered for this analysis. A comprehensive overview of DHS and MICS surveys can be found elsewhere [34]. Briefly, both DHS and MICS are household-based surveys that use a multistage stratified cluster sampling design to select a nationally representative sample of women of reproductive age (15-49 years old). Socio-demographic characteristics and information on selected health indicators are collected through face-to-face interviews by trained personnel and recorded in standard questionnaires. The majority of surveys administered the VF questions to all women of reproductive age but some restricted it to women that were ever married (Mauritania MICS 2011), ever pregnant (Swaziland MICS 2010 and Guinea-Bissau MICS 2010), or that had a live birth in the previous five years (Rwanda DHS 2005). The specific questions related to vaginal fistula symptoms varied slightly from survey to survey and a contingency question about knowledge of vaginal fistula was sometime incorporated. A full description of the VF and contingency questions (if any), their probes, and the coding of the outcome can be found elsewhere [10]. Based on previous studies and the information available from DHS/MICS surveys, we estimated the effect of the following risk factors: illiteracy, education level, whether the respondent has experienced female genital mutilation (FGM), short stature, experience of intimate partner sexual violence, young age at first sexual intercourse, young age at first birth, and women’s difficulty to get permission to access health care. Literacy status was ascertained in the surveys by asking the interviewee to read a sentence on a card that was handed out to her. If the woman was able to read only part of the sentence, she was considered not being able to read properly. Women who reported having had some secondary education or higher were de facto assumed to be literate. For genital mutilation, we did not stratify our analysis by FGM type as a validation study of the DHS FGM questions in Sierra Leone demonstrated that they were accurate to determine FGM prevalence but inaccurate for determining cutting extent [35]. Not all surveys recorded information for these risk factors and the list of countries for which such data was collected is presented in Tables 1 and ​and2.2. As for women’s anthropometric measurements, this information is not collected by MICS surveys and the women’s height was recorded from a sub-sample of participants in most DHS surveys. Similarly, questions on domestic violence were often administered to a subsample of women, depending on the survey, and the questions about ever having experienced intimate partner sexual violence were only asked to ever married women (or those in a union). As for age at first sexual intercourse, inconsistent responses were disregarded and considered as missing (e.g., a women reporting never having had sexual intercourse but having given birth). Finally, most DHS surveys asked women if getting permission to seek health care was a problem. Those who responded that it was a big problem were considered as having limited ability to seek the care they need. Number of vaginal fistula (VF) by survey datasets for the following risk factors: literacy status, education level, female genital mutilation (FGM), and short stature (<150 cm) VF = Vaginal Fistula; FGM = Female genital mutilation; DHS = Demographic and Health Survey; MICS = Multiple Indicators Cluster Survey The survey-specific total sample sizes can vary by risk factor depending on the number of missing observations and eligibility criteria Number of vaginal fistula (VF) by survey datasets for the following risk factors: experience of intimate partner sexual violence (IPSV), young age at first sex (<14 years old), young age at first birth (<14 years old), and permission to seek health care VF = Vaginal Fistula; IPSV = Intimate Partner Sexual Violence; DHS = Demographic and Health Survey; MICS = Multiple Indicators Cluster Survey aAmong married and/or ever married women (or those in a union) bAmong sexually active women cAmong primi/multiparous women The survey-specific total sample sizes can vary by risk factor depending on the number of missing observations and eligibility criteria The principal threat to the internal validity of our analyses is confounding of the exposure-outcome relationship. The main potential confounders for which information was collected by the survey questionnaires are age, literacy status, location of residence (rural versus urban), gravidity status, and religion. Socio-economic status and marital status were not considered in this analysis because these variables are likely both a cause and an effect of VF. That is, due to the cross-sectional nature of data collection, we do not have information on the temporal sequence in which changes in socio-economic status or marital status occurred. Three surveys (Chad MICS 2010, Mauritania MICS 2011, and Togo MICS 2010) did not record information on gravidity status and we assumed that all nulliparous women were also nulligravid – a reasonable assumption giving the high correlation observed between these two variables. Finally, four surveys did not record information on religion and these were coded using a missing variable indicator to retain them in the analyses (Mauritania MICS 2011, Niger DHS 2012, Swaziland MICS 2010, and Tanzania DHS 2010). To circumvent the lack of balance and overlap for some of the covariates, matching was used to make the group with the selected risk factor (i.e., exposed) as similar as possible to the group without (i.e., unexposed). By reducing model dependency through this semi-parametric data preprocessing, we aim to produce more robust inferences that are less sensitive to modeling assumptions [36]. Three of our risk factors are continuous and were dichotomized. Respondents with a height less than 150 cm, a commonly used threshold [12, 15], were defined as having a short stature. For age at first birth, visual inspection of the exposure-outcome relationship suggested that this variable could be dichotomized at less than 14 years of age at first delivery. This corresponds roughly to the 4th percentile of the distribution of age at first birth. The same threshold of less than 14 years was used to define young age at first sexual intercourse. All country datasets were pooled together as the low number of VF cases precludes data analysis at the country level for many surveys (i.e., all cases were either exposed or unexposed in these surveys). For the selected risk factors, a nearest neighbor algorithm was used to match women on sampling weight (for sexual violence, the sampling weight from the domestic violence questionnaire was used), age (continuous), and survey identifier. For this latter variable, exact matching was used for risk factors that consistently had more unexposed than exposed observations across surveys: short stature, intimate partner sexual violence, young age at first sexual intercourse, young age at first birth, and problem obtaining permission to seek care (otherwise, nearest neighbor matching was used). The matching ratio of exposed to unexposed units varied for each risk factor and was chosen as to minimize unbalance and maximize statistical power. Matching was implemented using the ‘MatchIt’ package [37] in R. Unmatched women were excluded from the analyses. Logistic regression models were used on the matched data to estimate the effect of the selected risk factors on lifetime prevalence of VF. Missing values for the selected risk factors and covariates were always less than 1 %, except for height (2.0 %) and age at first sexual intercourse (6.1 % of inconsistent or missing values). Observations with missing values were excluded from the analyses (with the exception of those for religion which were retained using a missing indicator). To provide for additional control of potential confounders, we adjusted for the following covariates: age (15-19, 20-29, and 30-49 years), literacy status (this covariate was not included for literacy status and education level risk factors), gravidity status (not included for age at first birth), location of residence (urban/rural), religion (Christian, Muslim, others, missing), and the survey’s country. Such analyses have been described as doubly-robust because statistically consistent inferences can be made “if either the matching analysis or the analysis model is correct (but not necessarily both)” [37]. Surveys that had a different population denominator were included in the analysis since we matched on survey identifier and country fixed effects were included in the parametric analyses. These logistic regressions did not account for the clustered design of surveys as our preliminary analyses have shown that clustering the standard errors had no impact on our conclusions (also discussed in [10]). Importantly, women’s self-report of vaginal fistula symptoms do not have perfect sensitivity and specificity, as compared to the gold standard of a pelvic examination. In order to account for non-differential misclassification of the self-reported outcome, we used a latent-class Bayesian statistical model [10, 38, 39]. The underlying assumption being that all surveys have a common sensitivity and specificity (see [10] for details). This model takes the following form: Because of our very large sample sizes and the computing-intensive nature of Bayesian calculations, we grouped observations with the same covariate patterns and used a binomial likelihood instead of the standard Bernoulli (i.e., grouped logistic regression). In this model, yi is the total number of women reporting VF symptoms with covariate pattern i; Ni is the total number of women with covariate pattern i; pi is the observed probability of reporting VF symptoms, πi is the true probability of women having ever had VF symptoms; Se and Sp are the sensitivity and specificity of the survey instrument, respectively; α is the model’s intercept; β is a vector of coefficients for the covariates included in Xi. The model’s specification is completed using the following prior distributions: Both α and β are given non-informative priors that follow a normal distribution with a mean of zero and standard deviation of 20. For sensitivity and specificity, we used uniform distributions that match the 95 % credible intervals of the posterior distributions of these quantities, as estimated previously [10]. Posterior distributions were obtained using Markov Chain Monte Carlo sampling, implemented in R using the ‘rstan’ package [40]. Samples are obtained using the no-U-turn sampler, a computationally efficient variant of Hamiltonian Monte Carlo [41]. Inferences were based on three chains of 30,000 samples after an initial warm-up period of 2,500 samples per chain (total of 90,000 iterations used for inferences). Convergence was examined using traceplots and ensuring that the potential scale reduction factor was equal to one. All analyses were performed using the R statistical software [42].