Menu
Menu
Menu
Menu
Background: The Ebola virus disease (EVD) epidemic has threatened access to basic health services through facility closures, resource diversion, and decreased demand due to community fear and distrust. While modeling studies have attempted to estimate the impact of these disruptions, no studies have yet utilized population-based survey data. Methods and Findings: We conducted a two-stage, cluster-sample household survey in Rivercess County, Liberia, in March–April 2015, which included a maternal and reproductive health module. We constructed a retrospective cohort of births beginning 4 y before the first day of survey administration (beginning March 24, 2011). We then fit logistic regression models to estimate associations between our primary outcome, facility-based delivery (FBD), and time period, defined as the pre-EVD period (March 24, 2011–June 14, 2014) or EVD period (June 15, 2014–April 13, 2015). We fit both univariable and multivariable models, adjusted for known predictors of facility delivery, accounting for clustering using linearized standard errors. To strengthen causal inference, we also conducted stratified analyses to assess changes in FBD by whether respondents believed that health facility attendance was an EVD risk factor. A total of 1,298 women from 941 households completed the survey. Median age at the time of survey was 29 y, and over 80% had a primary education or less. There were 686 births reported in the pre-EVD period and 212 in the EVD period. The unadjusted odds ratio of facility-based delivery in the EVD period was 0.66 (95% confidence interval [CI] 0.48–0.90, p-value = 0.010). Adjustment for potential confounders did not change the observed association, either in the principal model (adjusted odds ratio [AOR] = 0.70, 95%CI 0.50–0.98, p = 0.037) or a fully adjusted model (AOR = 0.69, 95%CI 0.50–0.97, p = 0.033). The association was robust in sensitivity analyses. The reduction in FBD during the EVD period was observed among those reporting a belief that health facilities are or may be a source of Ebola transmission (AOR = 0.59, 95%CI 0.36–0.97, p = 0.038), but not those without such a belief (AOR = 0.90, 95%CI 0.59–1.37, p = 0.612). Limitations include the possibility of FBD secular trends coincident with the EVD period, recall errors, and social desirability bias. Conclusions: We detected a 30% decreased odds of FBD after the start of EVD in a rural Liberian county with relatively few cases. Because health facilities never closed in Rivercess County, this estimate may under-approximate the effect seen in the most heavily affected areas. These are the first population-based survey data to show collateral disruptions to facility-based delivery caused by the West African EVD epidemic, and they reinforce the need to consider the full spectrum of implications caused by public health emergencies.
This study was approved by the ethics review boards at the Liberian Institute for Biomedical Research, Georgetown University, and Partners Healthcare. All participants provided verbal informed consent prior to participation. Rivercess County is a rural county located in south-central Liberia with approximately 71,500 residents as of the 2008 national census [37]. It had limited Ebola transmission, with 34 confirmed or probable cases reported to the World Health Organization, principally linked to a single cluster in October–November 2014 [38–41]. Participants were sampled using a stratified, two-stage cluster-sample approach. The sample was stratified for purposes of the survey’s function as baseline for a stepped-wedge impact evaluation, with two strata corresponding to the intervention’s phased implementation. The third stratum included areas within 5 kilometers (km) of a health facility, where CHWs are not deployed by current Liberian national policy, and which were assessed to provide county-wide estimates for health officials. Prior to sampling, we enumerated all households in each village in the county. At the first stage, we sampled villages with probability proportionate to size within each stratum using the standard DHS approach: listing clusters with a running cumulative number of households, determining the sampling interval necessary to take the correct number of clusters, randomly determining a starting value, and then selecting each subsequent cluster that corresponded to the sampling interval [42–44]. At the second stage, 21 households were selected per cluster in compact segments by (1) spinning a laminated paper triangle on the ground in the village’s center, (2) using a random number generator to randomly select an initial house in the direction pointed between the center and margin of the village, and (3) continuing to the next closest dwelling until 21 households were sampled [45,46]. We surveyed all women aged 18 to 49 in each selected household. The survey (see S1 Survey) included questions on household wealth, including asset ownership, water and toilet facilities, and housing materials, and maternal health, drawn principally from the 2013 Liberian Demographic and Health Survey [27]. A section on Ebola knowledge, attitudes, and practices was produced by the research team. All questions were translated from American English to Liberian vernacular English and back-translated by bilingual staff to ensure accuracy. Because some respondents were expected to speak only Bassa, a local language without a commonly used written form, bilingual enumerators administered the survey. Prior to survey administration, all enumerators attended a five-day training, which included practice administering the survey as well as training on informed consent, proper use of the mobile platform, survey skip logic, and techniques to reduce bias [47]. Field supervisors (with one supervisor per three-enumerator field teams) observed implementation of surveys daily and ensured quality assurance at the point of survey implementation, and one additional supervisor oversaw all field teams to ensure consistency. Data were entered using Commcare, an Android-based mobile platform, and maintained in a MySQL database, with basic data cleaning conducted prior to exportation for analysis. Our primary outcome of interest, having a facility-based delivery, was recorded for all respondents with at least one prior pregnancy. Our primary predictor of interest, whether the delivery occurred during the EVD period, was generated from child dates of birth. June 15, 2014, was chosen to dichotomize the periods before and during the epidemic, because it was the approximate date on which Ebola re-emerged as a broadly perceived national threat and when the first media reports emerged that patients were avoiding health facilities [48–54], and because, in limited published hospital record data from elsewhere in the country, June is the earliest month in which there appears a reduction in facility visits [15,16]. The pre-EVD comparison period began 4 y prior to the beginning of survey administration. In our models, we included potential confounding variables that have been identified as determinants of FBD in prior studies [30,34,45]. These included whether the birth occurred in the rainy or dry season (rainy season is May to October [55]), maternal marital status, household language (Liberian English versus Bassa), birth order (categorized as first, second or third, and fourth or higher), and self-reported maternal education (categorized as none, primary school only, and any secondary school or higher). Maternal age at each birth was calculated by measuring elapsed time between date of birth and the mother’s reported age when surveyed. Because prior studies [30,34,45] have found no consistent relationship between maternal age and FBD to give us an a priori basis on how to include maternal age, we categorized it into quartiles. A household wealth index was constructed using the standard DHS approach of assessing household assets, housing quality, water source, and toilet facilities [56]. The index was constructed using principal components analysis and then assigning relative percentiles of wealth. Prior work has demonstrated that FBD increases consistently with greater wealth [27,28,57]; we investigated the association between FBD and wealth in the pre-Ebola period using locally weighted scatterplot smoothing (LOWESS), which confirmed that the relationship between wealth percentiles and FBD was logit-linear (see S1 Fig). As a result, we included wealth as a continuous variable, rescaling it by dividing by its own interquartile range for ease of interpretation. Road distance from the center of each cluster to the nearest health facility was measured by global positioning systems (GPS) devices (Garmin eTrex 10; Garmin Ltd.). Prior research has shown complex inverse relationships between distance and FBD [45,58–61], so we examined the relationship using LOWESS plots, which suggested distance appeared to have a logit-linear, splined relationship with nodes at 10 and 21 km (see supplemental S1 Fig). Finally, to assess the causality of a relationship between time period and FBD, we included a survey item about whether respondents believed health facilities to be a source of Ebola transmission. We dichotomized this variable by combining respondents who stated they believed health facilities posed a definite or uncertain risk versus those who stated facilities posed no Ebola risk. We used standard summary statistical methods to describe respondents’ demographic and socioeconomic characteristics. We tested differences in respondents’ characteristics before versus during the Ebola period using design-corrected chi-squared analysis for categorical variables. Differences in the distribution of continuous variables, which were not normally distributed, were tested using Somers’ D, an analogue to the Mann-Whitney U test that can accommodate complex sample survey data [62,63]. We fit design-corrected logistic regression models to estimate associations between giving birth during the Ebola epidemic and FBD. While we had a priori bases to expect certain variables to be associated with FBD, we lacked a theoretical basis to expect particular variables to be associated with the EVD period and, therefore, to constitute potential confounders [64]. Therefore, we examined bivariate associations and constructed three multivariable models to assess for associations between the pre- and intra-EVD time period and FBD. The bivariate model assumes that Ebola is completely exogenous, so it presents only unadjusted relationships. In multivariable model 1, which serves as our primary model, we included only those predictors that were associated with both the EVD period and FBD during the pre-Ebola period at or below the p = 0.10 level [57]. Multivariable model 2 includes all variables associated with either Ebola or FBD at or below the p = 0.10 level. Multivariable model 3 includes all considered variables identified in the literature as potentially associated with FBD. Multivariable models 2 and 3 principally serve as robustness checks. We plotted levels of FBD in the post-Ebola period compared to preceding years adjusted for the controls in multivariable model 1 using predictive margins with covariates held at their observed levels. As a robustness check, we also graphically depicted FBD as a function of continuous calendar time using local polynomial regression (see S1 Appendix). To test whether observed associations between time period and FBD might be due to secular trends unrelated to Ebola, we conducted a stratified analysis that investigated FBD rates among (1) those who reported a belief that health facilities were a definite or uncertain risk for Ebola transmission and (2) those who believed that health facilities were not a risk for Ebola transmission. Because health facilities did not close in Rivercess County during the epidemic, we considered fear to be the greatest Ebola-related barrier to healthcare utilization in this area [65]. As such, we hypothesized that if Ebola were causally related to alterations in FBD, we would expect to identify a lower rate of FBD during the EVD period among those who perceived that health facilities posed a risk for transmission, and a lesser reduction among those who did not hold this belief. We assessed for differences in FBD rates in these stratified sub-analyses using survey-design-corrected logistic regression. All analyses accounted for our survey’s stratified design and incorporated clustering at the village and household levels. Taylor linearization was used to adjust standard errors for clustering in all parametric analyses; jackknifed errors were used with Somers’ D because it is not amenable to linearized errors. Because sampling probabilities differed by stratum, sampling weights were incorporated as the inverse probability of selection and corrected for non-response at the stratum level [42,66]. We incorporated finite population corrections at both sampling levels. Based on estimates of FBD rates and survey design effects from prior household-based surveys in rural Liberia [27,45], we estimated that approximately 870 births would be required to detect a 10% reduction in FBD with 80% power, which equated to approximately 4 y of birth data. We used Stata version 14 for all analyses. Replication datasets (S1 Dataset) and statistical code (S1 Code) are provided as online supplements. This study is reported per STROBE guidelines (S1 Checklist). Details of the analysis and any changes to the analysis plan are included in S1 Text. We conducted several sensitivity analyses to assess for robustness to potential biases. First, because we did not survey currently hospitalized women, the survey risked underestimating FBD among recent births. We addressed this by fitting additional models that excluded women who gave birth within 2 wk before survey administration began. Second, we may have excluded births in the pre-EVD period to women who turned 50 before our survey was administered and therefore did not meet inclusion criteria. We addressed this with models restricted to births to women aged 45 or less so that the entire cohort met survey inclusion criteria. Third, we fit models excluding births from women who had moved since their last birth to avoid misallocation of household demographic variables. Fourth, we ran an analysis that combined the preceding three sensitivity analyses. Fifth, because the exact date when the perceived threat of EVD became salient is unknown (and likely varied between people), we varied our a priori definition of the start of the EVD epidemic. The three alternate dates chosen were: May 29, 2014, when Liberia’s second wave began; July 15, 2014, by which point Ebola transmission was widespread in the country; and August 6, 2014, when Liberia declared a national emergency. Sixth, because inclusion of more years in the control period may increase susceptibility to bias from secular trends, we ran analyses restricted to only 1 and 2 y of control data. Seventh, because date-of-birth heaping was observed for the first day of each month, we randomly redistributed these birth dates across each month, which was expected to have a mild effect on both inclusion and allocation between the pre-EVD and EVD periods. Finally, because variance estimation is sometimes sensitive to the approach chosen [67,68], we present findings with standard errors calculated by the jackknife method instead of linearization.
