Impact of fee subsidy policy on perinatal health in a low-resource setting: A quasi-experimental study

Background A national subsidy policy was introduced in 2007 in Burkina Faso to improve financial accessibility to facility-based delivery. In this article, we estimated the effects of reducing user fees on institutional delivery and neonatal mortality, immediately and three years after the introduction of the policy. Methods The study was based on a quasi-experimental design. We used data obtained from the 2010 Demographic and Health Survey, including survival information for 32,102 live-born infants born to 12,474 women. We used a multilevel Poisson regression model with robust variances to control for secular trends in outcomes between the period before the introduction of the policy (1 January, 2007) and the period after. In sensitivity analyses, we used two different models according to the different definitions of the period “before” and the period “after”. Results Immediately following its introduction, the subsidy policy was associated with increases in institutional deliveries by 4% (RR = 1.04, 95% CI: 0.98–1.10) in urban areas and by 12% (RR = 1.12, 95% CI: 1.04–1.20) in rural areas. The results showed similar patterns in sensitivity analyses. This effect was particularly marked among rural clusters with low institutional delivery rates at baseline (RR = 1.44, 95% CI: 1.33–1.55). It was persistent for 42 months after the introduction of the policy but these increases were not statistically significant. At 42 months, the delivery rates had increased by 26% in rural areas (RR = 1.26; 95% CI: 0.86–1.86) and 13% (RR = 1.13; 95% CI: 0.88–1.46) in urban areas. There was no evidence of a significant decrease in neonatal mortality rates. Conclusion The delivery subsidy implemented in Burkina Faso is associated with short-term increases in health facility deliveries. This policy has been particularly beneficial for rural households. Burkina Faso is a West African low-income country with a population of more than 19 million inhabitants in 2017. The institutional delivery rate increased from 38% in 2003 to 66% in 2010, and the neonatal mortality rate ranged from 31 per 1,000 in 2003 to 28 per 1,000 live births in 2010 [20]. We used data from the latest Demographic and Health Survey (DHS) conducted in Burkina Faso in 2010. The DHSs are regular cross-sectional household survey conducted in several low-income countries. The DHS methodology, questionnaires and reports are available online (http://dhsprogram.com/). The women’s questionnaire helps collect sociodemographic characteristics and the reproductive life history of all 15- to 49-year-old women from the surveyed households. The vital status of each live-born child prior to the survey is collected for each woman. The age at death (days, months or years) is recorded when a child dies. Using a quasi-experimental approach, we tested the impact of the delivery policy subsidy by comparing outcomes between the period before and the period after the introduction of the policy. The subsidy was implemented at the national level; thus, formation of a control group of women from subsidy-free areas was impossible. Of the 14,947 households selected, 96.5% (14,424) were surveyed. Of the 17,363 women aged 15 to 49 identified in these surveyed households, 17,087 responded to the women’s questionnaire, yielding a response rate of 98.4%. Most of the information is self-reported. The data are organized depending on the following ascending hierarchical structure: new-born, woman, household, household cluster and region. The national subsidy for deliveries and emergency obstetric and neonatal care represented the exposure and has been described elsewhere [12, 21]. Briefly, this national subsidy policy covered approximately 60 to 80% of direct medical expenses (depending on whether the delivery occurred at a hospital or a health centre). The remaining 20 to 40% of the expenses were borne by the parturient. This policy was adopted in October 2006 and was officially implemented on 1 January, 2007 [12]. However, depending on the health district, the policy was introduced between January and April 2007 [7, 12, 22]. Several studies [12, 23, 24] indicated that this national subsidy policy was not evenly established throughout the country. The DHS data do not allow a connection between household clusters and health districts. The subsidy exposure was set to 1 for exposed and 0 for non-exposed, indicating whether each live birth occurred after the introduction of the subsidy policy. Two binary variables were considered for each live birth included in the study (birth place and vital status at 28 days of life). The place of birth was given a value of 1 for a health facility delivery (public or private) and 0 for delivery outside a health facility. The information is available for live births that occurred during the 5 years prior the survey. Multiple births were considered as one delivery. The new-born’s vital status was coded as 1 for new-borns who died within the first 28 days of life and 0 for new-borns who remained alive [25]. The information for age at death is available for 98.2% of children who died; the remaining 1.8% was imputed. Three causes (infections, complications of preterm birth and intrapartum-related neonatal death) account for 80% of neonatal deaths in low-income countries, and access to assisted delivery can significantly reduce mortality related to these causes [25]. Therefore, the evaluation of this outcome does not require a latency period since a significant increase in health facility deliveries can result in an immediate decrease in neonatal mortality if obstetric and neonatal care are good quality. Only live births that occurred during the 10 years prior to the survey were included in the neonatal mortality analysis to consider the following: first, the cross-sectional nature of certain control variables, and second, whether this effect contributed to limiting a possible secular trend in variation that could be overshadowed by a long pre-intervention period. Based on the literature, [6–9, 19] the following four effect-modifier variables were considered: the parturient’s education and literacy, the area of residence (rural vs urban) and the household wealth. The woman’s education comprised three categories [7–9] (no education, primary and secondary or higher). Literate women (those who could partially or fully read a sentence) were differentiated from illiterate women. We assume that the categorization of this second variable should be more homogeneous, since women with similar education levels may have different reading skills. Household wealth is a wealth index ranking of all surveyed households by income quintiles. This index was provided by DHS and was built by principal components analysis [26] using the source of drinking water, type of sanitary system, housing characteristics and the household’s possession of some durable goods (i.e., television, telephone, and refrigerator). Several variables associated with health facility delivery and/or neonatal death reported in the literature [1, 27] were considered as control variables. These variables are the maternal age at the time of delivery (categorized by 5-year intervals), number of births per delivery (single vs multiple), birth order (first, 2 to 4th and 5th or more), interval since precedent birth (36 months or more, less than 36 months and first delivery), new-born gender (female vs male) and the woman’s occupation status during the survey period (working vs not working). The work category included paid or unpaid work. The woman’s marital status (98.5% married) was discarded because it was non-selective. Breastfeeding and birth weight variables were not included because they were only reported for births in the 5 years preceding the survey. In addition, birth weight accounted for more than one-third (34.6%) of the missing data. The distance of households to the nearest health centre was not reported. The unit of analysis is the new-born. A modified multilevel Poisson regression [28, 29] analysis was performed to obtain the rates directly. Three levels were considered: level 1 (new-born), level 2 (woman or household) and level 3 (household cluster). A segmented regression analysis including a continuous time variable (in months) counted from the start to the end of the observation period was performed to isolate the secular trend and make adjustments for autocorrelation of the observations [30, 31]. A variable (post-time) with value 0 before the subsidy and counting the number of months since the introduction of the subsidy was also included to assess the change in the slope induced by the subsidy [30, 31]. In addition to the random intercepts, variable coefficients allocated to the subsidy, time and post-time variables were considered depending on the clusters [32]. The choice of a model with random intercept and coefficients is also based on the results of a previous study that showed that the effects of this policy varied by district [33]. Education and literacy were separately introduced into the models due to their strong correlation. Based on these specifications, the following 3-level models were analysed: where i = ith new-born, j = jth woman or household, k = kth household cluster, Yijk = outcome status, Tijk = length of time since the beginning of the observation (January 2000 or June 2005 depending on the outcome), post-timeijk = length of time since the introduction of the subsidy, Iijk = subsidy status, Ruralk = cluster localization, SESjk = household wealth, Educjk = education or woman’s literacy, and Xijk = control variables. For this model, β1 estimates the secular trend, β2 represents the immediate effect of the subsidy, and β3 evaluates the change in the secular trend after the introduction of the subsidy. The sum of β1 and β3 represents the secular trend after the introduction of the subsidy. γ0jk and γ00k represent the level 2 and 3 random intercepts, and γ1k, γ2k and γ3k constitute the level 3 random effects related to β1, β2 and β3, respectively. The interactions between the subsidy and the area of residence, the household wealth and the woman’s education are respectively estimated by coefficients β7, β8 and β9. By considering 1 January 2007 as the subsidy start date, some non-exposed births will be classified as exposed. To assess the consequences of these misclassifications on the results, we conducted two sensitivity analyses: the first one considered 1 April 2007 as the subsidy introduction date. This strategy leads to the classification of certain exposed births as non-exposed births. The second one excluded births that occurred between 1 January and 31 March 2007, as suggested by some authors [34, 35], because whether these births were exposed or not could not be determined. The three-level models were tested against two-level models with the cluster at level 2. In the birth place analysis, the woman and household levels showed no variance, and the analysis was therefore limited to two levels. The interactions were evaluated after controlling for confounding factors. The three interactions were tested together and then separately for each dependent variable. The interaction between the area of residence and the subsidy for the birth place was the only significant interaction, and the analyses were conducted separately in subgroups for urban births and rural births. Over-dispersion was tested by introducing a random intercept at level 1 (new-born) [36]. Standard errors were adjusted for regional level clustering for all analyses. From the final models, a seasonal variable representing the calendar month of birth was introduced to control for potential seasonal fluctuations [31]. Only the tests of interactions, random effects and hierarchical structure used the likelihood ratio test with a p-value at the 0.05 threshold. The other comparisons (confounding and seasonality) focused on the subsidy-modifying coefficient of at least 10%. We checked the multicollinearity of independent variables by using the variance inflation factors. The likelihood ratio test was used to test the fit of nested models, and we analysed the residual statistics to assess the fit of the models. A model-based standardization [37, 38] was conducted to calculate the rate ratios and rate differences associated with the subsidy for different periods (0 to 42 months) at the population level. Each new-born’s probability of death or birth in a health centre was predicted for each period under two scenarios (with and without the subsidy) and by replacement of the time and post-time variables with their corresponding time values. These predictions incorporated the random-effects predicted values (obtained by empirical Bayes prediction) [36] for each woman, household and cluster [37]. Then, the probabilities were averaged for each scenario and period, and their ratios and differences were calculated. Calculations were also performed for the three and two-level models, and similar results were obtained. The results are reported here for the two-level models for parsimony. Confidence intervals were obtained using the delta method [38, 39]. The analyses were conducted with the Generalized Linear Latent and Mixed Models (GLLAMM) [36] program in Stata SE 14.2 (StataCorp LP).