Menu
Menu
Menu
Menu
Zambia has been using output-based approaches for over two decades to finance whole or part of the public health system. Between 1996 and 2006, performance-based contracting (PBC) was implemented countrywide with the Central Board of Health (CBoH) as the provider of health services. This study reviews the association between PBC and equity of access to maternal health services in Zambia between 1996 and 2006. A comprehensive document review was undertaken to evaluate the implementation process, followed by a trend analysis of health expenditure at district level, and a segmented regression analysis of data on antenatal care (ANC) and deliveries at health facilities that was obtained from five demographic and health survey datasets (1992, 1996, 2002, 2007 and 2014). The results show that PBC was anchored by high-level political support, an overarching policy and legal framework, and collective planning and implementation with all key stakeholders. Decentralization of health service provision was also an enabling factor. ANC coverage increased in both the lower and upper wealth quintiles during the PBC era, followed by a declining trend after the PBC era in both quintiles. Further, the percentage of women delivering at health facilities increased during the PBC era, particularly in rural areas and among the poor. The positive trend continued after the PBC era with similar patterns in both lower and upper wealth quintiles. Despite these gains, per capita health expenditure at district level declined during the PBC era, with the situation worsening after the PBC era. The study concludes that a nationwide PBC approach can contribute to improved equity of access to maternal health services and that PBC is a cost-efficient and sustainable policy reform. The study calls for policymakers to comprehensively evaluate the impact of health system reforms before terminating them.
The study reviews three aspects: implementation process, level of spending and coverage of maternal health services. Although the PBC era was 1996−2006, the period 1987−2014 was analysed in order to establish the trend before, during and after PBC. To facilitate the analysis, a mixed-methods approach—that combined qualitative and quantitative research techniques was used. This approach is commonly used and recommended by a number of scholars as it enables a researcher to gain access to a variety of insights, to challenge and verify perceptions, and to form a broad basis on which to make conclusions (Creswell et al., 2011). Thus, the qualitative component of the study involved a comprehensive document review to gather information on the implementation process. For the quantitative part, data on district health expenditure and maternal health were collected and analysed. A comprehensive review of policy documents, strategic plans, independent studies and research articles which were produced in the early 1990s, 1994−1999, 2000−2005 and 2007−2014 was undertaken. The documents were carefully selected to acquire information on the historical thinking and motivation to introduce PBC (Ministry of Health, 1991); reform agenda, key institutions and implementation arrangements, and processes (Kalumba et al., 1994; Kalumba, 1997); and empirical evidence on the successes and challenges over time (Ministry of Health et al., 1997; Lake and Musumali, 1999; Ministry of Health, 2000, 2004a,b; Bossert et al., 2003; Chansa, 2009). Furthermore, district action plans, budgets and expenditure reports; planning handbooks; service contracts; and minutes and action taken reports from the health sector committee meetings were also reviewed. These rich data allowed us to track and assess the quality of the contracting and implementation process during the period under review. We analysed district-level health expenditure data through the public health system over the period 1995–2014 aimed at establishing the level of expenditure before, during and after the intervention period. These data were obtained from the Ministry of Health and comprises operational (recurrent) expenditure1 from both the Zambian government and external development partners to the district ‘basket’ which was subjected to annual action planning and budgeting. To adjust for inflation, the gross domestic product deflator (base year = 2010) was used. This is because Zambia moved to the base year of 2010 in 2014 in line with the 2010 economic census (Central Statistical Office, 2014). Total district recurrent per capita health expenditure was calculated by dividing the total annual district-level recurrent expenditure by the total annual national population. The annual population figures were obtained from the Central Statistics Office covering the period 1995−2014.2 Raw data on maternal health was accessed from five demographic and health surveys (DHS) for the years 1992, 1996, 2002, 2007 and 2014. DHS is a cross-sectional household survey that is conducted every 4–5 years and provides nationally representative coverage and outcome data on demography and population health status. Detailed information on the DHS methodology is available on the DHS website at https://dhsprogram.com/data/available-datasets.cfm. Individual recode STATA datasets from the DHS were used to aggregate data on two maternal health variables of interest, namely: antenatal care (ANC) and deliveries at health facilities (Table 1). These two indicators were among the key priority health indicators which were being monitored by the government during the period under review aimed at improving maternal and child health. To incorporate quality, only data on pregnant women who had at least four ANC visits at the stipulated times during their pregnancies were included. The resulting pooled dataset covers the period 1996–2014 for ANC and 1987–2014 for delivery care. Treatment variables, definitions, number of observations and time period The main steps in preparing the dataset for preliminary analysis included identifying the variables of interest in the datasets from the five DHS, reshaping the dataset from wide format to panel format, categorizing covariates and outcomes of interest into yearly intervals and merging the datasets from the five DHS into one pooled dataset. Each outcome variable was expressed as a proportion (Table 1) and stratified by wealth quintiles (bottom two and upper two quintiles). For 1992, 1996 and 2002, an additional step involved merging the individual recode dataset with the wealth index dataset because the individual recodes datasets did not have the wealth variable. For the 2007 and 2013/14 DHS, wealth variables were available. A summary of the outcome variables, definitions, number of observations and time periods are shown in Table 1. Data were inspected for presence of wild points, linear trend and seasonal patterns by plotting the outcome variable against time. The preliminary model was fitted and checked for presence of autocorrelation and partial autocorrelation by using the Durbin Watson test. In addition, the augmented Dickey Fuller test for stationarity was also conducted on the outcome variable (Table 2). Results from the autocorrelation test for ANC (Figure 2a) showed a significant spike at lag 1 and decreases after a few lags, implying an autoregressive term in the data. The partial autoregressive function showed that the order of the autoregressive term was 1, and therefore, we modelled our series using autoregression (AR) (1). From the autocorrelation test for delivery care (Figure 2b), there were correlations at the first and second lags, followed by correlations that are not significant. This implies that the series had a moving average term. We observed a significant lag of 10 which implied that our series had order 2 AR process and moving average of order 10 [autoregressive moving average (2,10)]. Test for autocorrelation and partial autocorrelation. Source: Authors’ construction from DHS data. Test for stationarity Initial segmented regression analysis was using the Newey–West model. The analysis was done based on residence (urban/rural) and wealth (upper/lower). Results from the autocorrelation test for delivery care in urban Zambia showed no autocorrelation at lags 1 and 3. Similarly, the autocorrelation test for delivery care in rural Zambia showed that there was no autocorrelation at lag 1. When we disaggregated the data by wealth, the results showed autocorrelation at lags 4 and 10 for the upper wealth quintile. For the lower wealth quintile, there was no serial correlation even at lag 1. For ANC, the results show autocorrelation at lag 3 for both urban and rural Zambia, and serial correlation for both the lower and upper wealth quintiles at lag 3. To correct all remaining serial correlation, we implemented the generalized least squares model, based on Prais–Winsten procedure. Prais–Winsten segmented regression analysis was used to assess trends in ANC coverage and deliveries at health facilities before, during and after PBC. Segmented regression analysis was used because it is helpful in evaluating population-level effects in interrupted time series data, and it has the ability to control for secular trends and serial correlation (Wagner et al., 2002; Linden, 2015). The data that were used for the analysis was ordered as a time series and a number of observations were available in both the pre-intervention and post-intervention periods to make a valid analysis (Wagner et al., 2002; Campbell et al., 2009; Wagenaar et al., 2016). This was stratified by residence (rural and urban) and by wealth (upper two and lower two). For deliveries, observations were available in the pre-intervention, intervention and post-intervention periods; whereas for ANC the observations were only available during and after the intervention, because the wealth variable was not included in the 1992 round of DHS. Therefore, for deliveries, we compared the periods before, during and after PBC whereas for ANC we compared the periods during and after PBC. The procedure that was used to prepare the data for analysis also conforms to suggested guidelines for analysis of time series data (Lagarde, 2012). After the data were ready for analysis, we used STATA version 15.1 to run the analysis. The STATA command for segmented regression analysis assumes a linear relationship between time and the outcome within each segment, and fits a least-squares regression line to each segment of the independent variable and time (Wagner et al., 2002). For intervention status j and k, at time point t, the outcome is estimated with the following equation: Where Yt is the aggregated outcome variable (ANC or delivery care), Tt is time since the start of the study, Xt is a dummy variable representing the intervention (pre-intervention = 0, otherwise = 1), XtTt is an interaction term. The coefficients β0 is the starting level of the outcome, β1is the trend of the outcome before PBC, β2is the change in level of the outcome in the period immediately following PBC, compared with pre-intervention levels and β3 is the difference between pre- and post-intervention trends. For ANC, the equation was implemented with one interruption (after the end of PBC), whereas for delivery care we placed two interruptions (before and after PBC). We did the analysis by residence and wealth.
N/A
