The effect of health insurance and health facility-upgrades on hospital deliveries in rural Nigeria: A controlled interrupted time-series study

Background:â € Access to quality obstetric care is considered essential to reducing maternal and new-born mortality. We evaluated the effect of the introduction of a multifaceted voluntary health insurance programme on hospital deliveries in rural Nigeria. Methods:â € We used an interrupted time-series design, including a control group. The intervention consisted of providing voluntary health insurance covering primary and secondary healthcare, including antenatal and obstetric care, combined with improving the quality of healthcare facilities. We compared changes in hospital deliveries from 1 May 2005 to 30 April 2013 between the programme area and control area in a difference-in-differences analysis with multiple time periods, adjusting for observed confounders. Data were collected through household surveys. Eligible households (n = 1500) were selected from a stratified probability sample of enumeration areas. All deliveries during the 4-year baseline period (n = 460) and 4-year follow-up period (n = 380) were included. Findings:â € Insurance coverage increased from 0% before the insurance was introduced to 70.2% in April 2013 in the programme area. In the control area insurance coverage remained 0% between May 2005 and April 2013. Although hospital deliveries followed a common stable trend over the 4 pre-programme years (P = 0.89), the increase in hospital deliveries during the 4-year follow-up period in the programme area was 29.3 percentage points (95% CI: 16.1 to 42.6; P < 0.001) greater than the change in the control area (intention-To-Treat impact), corresponding to a relative increase in hospital deliveries of 62%. Women who did not enroll in health insurance but who could make use of the upgraded care delivered significantly more often in a hospital during the follow-up period than women living in the control area (P = 0.04). Conclusions:â € Voluntary health insurance combined with quality healthcare services is highly effective in increasing hospital deliveries in rural Nigeria, by improving access to healthcare for insured and uninsured women in the programme area. Kwara State is part of the north central region of Nigeria with a total population of ∼2.5 million based on the 2006 National Population Census. The 2013 Nigerian demographic health survey reported that, in Kwara State, 76.7% of women delivered in primary health centres or hospitals (National Population Commission (NPC) [Nigeria] 2014). The KSHI programme began providing health insurance to households in the Asa Local Government Area in Kwara State, Nigeria (the programme area) in July 2009. In the 2 months before the insurance was introduced (May–June 2009), the programme facilitated quality improvements in the participating hospitals. The Ifelodun Local Government Area in Kwara State was chosen as the control area, as it was comparable to the programme area in terms of socio-demographic and socio-economic characteristics. The quality and services provided in the healthcare facilities in the two areas were also similar before the introduction of the programme (see the Supplementary Materials for a figure of the study area). Enrolment in the health insurance scheme was voluntary and on an individual basis. At the time of this study, the annual insurance premium was ∼2.4 USD per person per year, which corresponded to ∼0.5% of average yearly per capita consumption among the 1500 surveyed households in 2009. The insurance package provided coverage for consultations, diagnostic tests and medication for all diseases that could be managed at a primary care level, as well as limited coverage of secondary care services. Secondary care services provided included antenatal care, vaginal and caesarean delivery, neonatal care, immunizations, radiological and more complex laboratory diagnostic tests, hospital admissions for various diseases, minor and intermediate surgery and annual check-ups. Excluded from the programme were high technology investigations (computed tomography and magnetic resonance imaging), major surgeries and complex eye surgeries, family planning commodities, treatment for substance abuse/addiction, cancer treatment requiring chemotherapy and radiation therapy, provision of spectacles, contact lenses and hearing aids, dental care, intensive care treatment and dialyses (Hendriks et al. 2014). Quality and efficiency of healthcare were monitored through independent audits by an international quality improvement and assessment body called SafeCare, a partnership between the PharmAccess Foundation, the American Joint Commission International and the South-African Council for Health Services Accreditation of Southern Africa. Prior to enrolment in the KSHI programme, a baseline assessment of the clinic or hospital was conducted by SafeCare and a quality improvement plan was formulated. The provider specific improvement plans consisted of specific targets in 13 different domains, including management and leadership, human resource management, patients’ rights and access to care, management of information, risk management, primary healthcare services, inpatient care, operating theatre, laboratory, diagnostic imaging, medication management, facility management and support services. The improvement plans were implemented by the healthcare providers with technical and financial support from Hygeia Community healthcare. SafeCare monitored the progress on quality improvement through annual follow-up assessments with the SafeCare Quality Standards. Examples of quality improvement interventions included implementation of treatment guidelines and protocols for waste management and hospital infection control, training of staff in guideline-based care and adequate medical file keeping, hospital renovation, upgrading of laboratory equipment and training of laboratory staff in basic laboratory testing and assurance of continuous essential drug supplies (Hendriks et al. 2014) (see the Supplementary Materials for additional information on the KSHI programme). We applied a controlled interrupted time-series design to measure the impact of the KSHI programme 4 years after its introduction. We used stratified two-stage cluster sampling, with stratification by area of residence (programme or control) and distance to the nearest (potential) programme hospital (within 5 km or within 5–15 km) resulting in four subareas. Based on the 2006 National Population Census those four subareas were divided into 300 enumeration areas, of which a random sample of 100 enumeration areas was drawn. Subsequently a random sample of 1500 households [900 (60%) households in the programme area and 600 (40%) households in the control area] was drawn from those 100 enumeration areas, such that the resulting sample was representative of the Asa and Ifelodun areas. As a 50–60% uptake of insurance among households was expected, households in the programme area were over-sampled compared with the control area. The target sample size of 1500 households was based on sample size estimates required to study the effect of the programme on healthcare utilization and financial protection in the overall population. Therefore, no formal sample size calculations were performed using hospital deliveries changes as a main outcome measure. However, a fixed sample size of 1500 households would allow us to measure a minimum impact of a 21.2 percentage points increase in hospital deliveries, with a power of 80% using a two-tailed test and a 0.05 level of significance. Data were collected in three consecutive population-based household surveys that were simultaneously conducted in the programme area and control area. A baseline survey was carried out in May 2009, shortly before the introduction of the programme, and two follow-up surveys were carried out among the same households in June 2011 and 2013, respectively. Households were included in the surveys after written informed consent was obtained from adult household members. Consent was obtained from the head of household for those under 18. All respondents (including the respondents under 18) were explicitly asked to assent to respond to the pregnancy questionnaire within the household surveys. All deliveries during the 4-year baseline period (1 May 2005–30 April 2009) or 4-year follow-up period (1 May 2009–30 April 2013) from women aged 15–45 years at the time of delivery were eligible for this study (see the Supplementary Materials for more information on the survey questions, potential recall bias and data construction). The study protocol was approved by the Ethical Review Committee of the University of Ilorin Teaching Hospital in Nigeria (04/08/2008, UITH/CAT/189/11/782). Hospital delivery was defined as delivery in any hospital or clinic where skilled delivery care was provided and where caesarean sections were possible, as opposed to at home or in a primary healthcare centre, as reported by the women during the household survey. Primary healthcare centres in rural Kwara State did not provide skilled delivery services and were therefore not included in the definition of hospital delivery. We measured the intention-to-treat effect of the KSHI programme by using a difference-in-differences method. All women living in the programme area had access to improved quality maternal and child healthcare services in the upgraded programme hospitals, with or without being enrolled in the health insurance, although uninsured women had to pay for these services. Therefore, all women in the programme area were considered to be in the intervention group irrespective of whether they were actually insured. Such an intention-to-treat approach avoids the bias introduced by self-selection into (or out of) the health insurance and incorporates the independent effect of the quality improvements in the programme hospitals on uninsured women in the programme area. In difference-in-differences analysis the intention-to-treat effect (or impact) was estimated as the increase in percentage of hospital deliveries from the pooled 4-year baseline period to the pooled 4-year follow-up period in the programme area, controlled for the change in percentage of hospital deliveries in the control area (Lee and Kang 2006; Blundell and Dias 2009; Wooldridge 2010). The key identifying assumption behind the difference-in-differences method is that hospital deliveries in the programme and control area followed a common constant pre-intervention trend (the common trend assumption). Let yi be a dummy variable that equals 1 if woman i delivered in a hospital. The intention-to-treat programme effect on hospital deliveries was estimated by the following linear probability multivariable difference-in-differences model (Blundell and Dias 2009): where Programmei is the treatment indicator that equals 1 if woman i was living in the programme area, and 0 otherwise. The treatment indicator captured possible differences between the programme and control area prior to the introduction of the programme (measured by ϑ). Postt is a dummy variable equal to 1 for the pooled 4-year follow-up period, which captured aggregate factors that would cause changes in hospital deliveries in the absence of the programme (measured by δ). The interaction term Programmei · Postt equals 1 if woman i was living in the programme area during the pooled 4-year follow-up period, and 0 otherwise. The interaction term measured the intention-to-treat programme effect, which is identified by γ under the common trend assumption. The vector Xi captured the effects of the observed confounders on hospital delivery (measured by the β’s) and the error-term εi captured unobserved factors affecting hospital delivery. The common trend assumption was assessed in the controlled interrupted time series analysis, where we estimated a fully flexible difference-in-differences model. This model compared changes in hospital deliveries from the 4 pre-programme years to hospital deliveries in the 4 years after the introduction of the programme, allowing for fully flexible pre- and post-programme trends in the programme and control area. The following fully flexible multivariable difference-in-differences model was estimated by including the separate 4 pre- and 4 post-programme years (Mora and Reggio 2012): where Pret is a dummy variable equal to 1 for the tth pre-programme year and Postt is a dummy variable equal to 1 for the tth post-programme year. The interaction term Programmei · Pret equals 1 if woman i was living in the programme area during the tth pre-programme year and the interaction term Programi · Postt equals if woman i was living in the programme area in the tth post-programme year. The common trend assumption was assessed by testing the H0: ρ1 = ρ2 = ρ3=ρ4, which would suggest that indeed the difference-in-differences model in equation (1) is appropriate (Mora and Reggio 2014). In addition, we assessed whether the intention-to-treat effect was constant in the follow-up period by testing the H0: γ1 = γ2 = γ3=γ4, which would suggest that the programme reached its maximal impact in the first post-intervention year already and remained stable at its maximum in the years that followed (Mora and Reggio 2014). In pre-specified heterogeneity analysis we assessed whether the programme’s impact varied with distance to the nearest (potential) programme hospital (within 5 km vs within 5–15 km). Hereto, we augmented the difference-in-differences model in equation (1) by including interactions with distance to the nearest (potential) programme hospital: where Distancei is a dummy variable, which was equal to 1 for women living more than 5 km away of a (potential) programme hospital. The intention–to–treat programme effect is measured by γ1 for women living within 5 km of a programme hospital and by γ1 + γ2 for women living more than 5 km away, where we expect γ2 to have a negative sign. The 95% CI and P-value of the programme’s impact among women living more than 5 km away of a programme hospital were calculated by employing the delta method after estimation. In sensitivity analysis using the subsample of women who delivered in both periods, the multivariable difference-in-differences model was estimated with individual fixed-effects to control for the effect of unobserved time-constant confounders as well (Wooldridge 2010): where αi is the individual fixed-effect representing unobserved time-constant characteristics of the women and δt is the time fixed-effect representing the trend in the control group. The vector Xit captured the effects of the observed time-varying confounders on hospital delivery. Observed time-constant confounders were not included in this model. Since hospital delivery is a variable taking only values 0 and 1 one might assume that a non-linear difference-in-differences method would be preferable to linear difference-in-differences (linear probability model) but Blundell and Dias (2009) showed the opposite. They showed that difference-in-differences loses much of its simplicity even under a very simple non-linear specification and requires additional strong assumptions which are often not met. Moreover, Angrist and Pischke (2008) showed that in practice the results of the linear probability model are just as good as those of non-linear models. To verify this, in a sensitivity analysis the results of the linear difference-in-differences model were compared with the results of the logistic difference-in-differences model (noting that this logistic difference-in-differences model is not favourable). If the programme affected the odds of getting pregnant (or not getting pregnant) then this would bias the difference-in-differences results. Therefore we additionally estimated whether the probability of pregnancy during the follow-up period was significantly different between women of reproductive age with access to the programme and women in the control area without access to the programme, by using multivariable logistic regression analysis. Finally, deaths could be pregnancy related which can bias the programme’s impact, as well. Therefore we assessed whether the number of deaths among women of reproductive age were similar over time in both study areas. Confounders were selected based on Gabrysch and Campbell’s (2009) conceptual framework. This framework distinguishes four sets of variables related to hospital delivery, namely (i) perceived needs [age, parity, complications during (previous) delivery and desire to become pregnant], (ii) socio-demographic factors (religion and ethnicity), (iii) socio-economic factors (marital status, female head of household, educational level head of household, and household wealth) and (iv) physical accessibility [distance to nearest health facility and distance to nearest (programme) hospital]. In addition, a context specific factor was added, namely whether the delivery date coincided with one of two public sector health worker strikes in Kwara State (a 40-day strike in May–June 2011 and 3-week strike in December 2011–January 2012). Household wealth was estimated by an asset index derived using principal component analysis. In the multivariable models all a-priori selected confounders were included, irrespective of whether they were statistically significant. Age, parity, complications during (previous) delivery, marital status, female head of household, distance to nearest health facility, distance to nearest (potential) programme hospital, and the strike variable were included as time-varying variables and desire to become pregnant, religion, ethnicity, educational level of the head of household, and household wealth were included as time-constant variables, measured at baseline. Educational level of the head of household and household wealth were measured at follow-up as well, but included at their baseline value to avoid endogeneity problems. In a sensitivity analysis the strike variable was removed from the list of confounders to assess the estimation bias of not controlling for the effect of the strikes. Finally, we provided an estimate of the independent effect of the quality improvements in the programme hospitals on uninsured women in the programme area by estimating the increase in hospital deliveries from the baseline period to the follow-up period among uninsured women in the programme area who delivered in both periods, by using multivariable logistic regression analysis. In all analyses we corrected for clustering within enumeration area, household, and individual. Data were analysed using STATA, version 12.1 (StataCorp LP, TX, USA).