Government Insurance and Out-of-Pocket Healthcare Expenditure in Ghana

Abstract The relationship between membership in Ghana’s national health insurance scheme and out-of-pocket healthcare expenditure (OOPHE) was quantified using data from two rounds of the Ghana Living Standards Survey for 2013 and 2017. Censored quantile regressions were evaluated with and without instrumental variables. The results show that going from having no insured household member to all insured predicted less OOPHE (by 19% at the median). We find statistically significant differences between the 2013 and 2017 estimates. Insurance reduced OOPHE in 2013 but had a statistically insignificant effect in 2017. The effect on spending on outpatient care was greater than that related to medicine and medical supplies. There was no statistically significant relationship with hospitalisation fees. Falling government health spending and growing reliance on private healthcare financing have been observed. The insurance scheme has become less generous, and it was therefore less effective in 2017.


Introduction
High out-of-pocket healthcare expenditures (OOPHEs) can push low-income households into extreme poverty, just as avoiding the necessary treatment could. Consequently, several developing countries have since 2004 introduced public health insurance schemes intended to reduce OOPHE and promote universal health coverage (Acharya et al., 2013). 1 Despite these efforts, weak administrative systems and implementation failures have led to low take up of health insurance in many developing counties.
In Ghana, participation in the national health insurance scheme (NHIS) has dropped from 41% in 2015 to 35% in 2017 (Nsiah-Boateng & Aikins, 2018). The proportion of Ghanaian households in which a member paid the largest share of their medical bills increased by 18 percentage points between 2012/2013 and 2016/2017, the NHIS notwithstanding (Table S.1 in the Supplementary Material). The increases in private financing might discourage formal healthcare use, especially by poor people. The Ghana Statistical Service (2014, 2019) reports a 21-percentage point decline in the proportion of ill or injured people who consulted a medical Furthermore, health insurance could be endogenous in the OOPHE equation because it is a choice variable. In practice, informal sector workers in Ghana have voluntary NHIS participation. A sure way to obtain reliable estimates is to apply an experimental design (ideally a randomised controlled trial, RCT) that randomised access to health insurance. However, neither this nor the previous studies used RCTs due to their high cost in time and money and the ethical and practical difficulties of ensuring similarity between households in the treatment and control groups. Aron-Dine, Einav, and Finkelstein (2013) argue that the RAND experiment is unlikely to be replicated on cost grounds alone. Consequently, researchers rely on quasi-experimental designs for identification.
We apply Chernozhukov's censored quantile instrumental variable (CQIV) estimator to incorporate endogenous regressors in the CQR (Chernozhukov et al., 2015). Two instrumental variables were considered: (a) The proportion of household members eligible for the NHIS premium exemption (exempt category IV); and (b) the proportion of NHIS subscribers in the community, exclusive of those in the observed household (community insurance IV). Also, we implement an IV estimator under plausible (not perfect) exogeneity (see Conley, Hansen, & Rossi, 2012). That method relaxes the exclusion restriction condition of the classical IV estimators. A Supplementary Material with additional analyses (individual-level analysis and effect on catastrophic OOPHE) is added.
The remainder of the paper is organised as follows. Section 2 presents the NHIS institutions and challenges. The data used are discussed in Section 3. The following section presents the econometrics framework and estimation techniques. The penultimate section, Section 5, reports and discusses the empirical results. And Section 6 concludes.

Institutions and challenges for the NHIS
In 2003, the Ghana government passed the National Health Insurance Act (Act 650), which instituted the National Health Insurance Authority (NHIA) to implement the NHIS and provide universal health insurance coverage (NHIA, 2013). By the end of 2006, the NHIS had been fully implemented in all the districts. The NHIA regulates the NHIS through its district offices. It provides accreditation to several health facility types: both public and private. More than half of the certified health facilities (54% in 2012) were state-owned (NHIA, 2012). 4 Enrolment in the NHIS is designed to be compulsory for all residents and at the individual level. Its subscribers are either in the premium-exempt or premium-paying groups. In 2012, the premium-paying group (comprising informal sector workers) accounted for about 36% of NHIS Subscribers (NHIA, 2012). Due to lack of data and income fluctuations, the premiumpaying subscribers pay a flat yearly premium ranging from GH¢15.00 in rural areas to GH¢22.00 in the cities (NHIA, 2013;Nsiah-Boateng & Aikins, 2018). This rate is regressive and overburdens low-income subscribers.
The NHIS is designed to make richer and healthier people pay more to cover the less healthy, poor, and vulnerable (NHIA, 2013). Achieving this has been easy with the formal sector workers since their contributions are deducted at source. However, the many informal sector workers make enforcing the mandatory participation rule difficult. Young children, pregnant women and the unhealthy are more likely to self-select into the scheme (Khalid & Serieux, 2018;Salari, Akweongo, Aikins, & Tediosi, 2019;van der Wielen, Falkingham, & Channon, 2018).
The NHIA reports that the scheme's primary challenge is its financial sustainability (NHIA, 2013), but patients also note the low quality of care received (Wang, Otoo, & Dsane-Selby, 2017). That often motivates subscribers to pay out-of-pocket for some covered services and medicines (Nketiah-Amponsah, Alhassan, Ampaw, & Abuosi, 2019). These unofficial payments also relate to the scheme's failure to reimburse healthcare providers, making them unwilling to continuously provide free services and drugs to subscribers (Aryeetey et al., 2016;Okoroh et al., 2020). The scheme's participation dropped by 3 percentage points between 2013 and 2017 from an already low base. Most people, especially in the rural areas, had dropped out from the NHIS by 2017, citing "no money" as their primary reason for not renewing cover. Previous studies have shown that the scheme is pro-rich and pro-urban in practice, contrary to its objectives.

Data
This section discusses our construction of the key variables: OOPHE and health insurance. We document the sampling procedure, key variables' description, in the Supplementary Material. The control variables are likewise defined and described in the Supplementary Material.
The GLSS asked respondents to indicate the amount of Ghanaian cedis (GH¢) they spent on medical goods and services in the two weeks preceding the survey. The study's OOPHE was per capita household medical expenditure (regionally deflated). That definition comprises outpatient payments (expenses on registration, consultation, diagnosis, and treatment), inpatient (hospitalisation) fees and payments for medicines and medical supplies (tablets, capsules, syrups, bandages, plaster, cotton wool and any item used for treatment). Transport costs to health facilities were not included because the NHIS does not cover such expenses. Adjusting the cedi values for inflation using the survey's price indices allowed comparing OOPHE values over time. Each survey provides rural and urban price indexes for each region.
The GLSSs also report data on individuals' health insurance participation. For the principal analysis, the proportion of household members with valid NHIS membership was taken as quantifying insurance, excluding people with just private health insurance. Only a negligible 208 out of the 46,764 sampled individuals ever registered with a health insurance scheme by 2017 (0.45%, unweighted) had just private health insurance. Hereafter, insurance and NHIS membership will be used interchangeably.

CQR
In identifying any significant relationship between insurance and OOPHE, a straightforward strategy would compare the OOPHEs of insured and uninsured people (as in Supplementary Table S.8). But omitted variable bias may cause identification problems. One technique for dealing with this problem is to control for observable factors that could drive NHIS participation and OOPHE. The CQR permits additional exogenous regressors.
The conditional quantile function of log(OOPHE þ 1), Q logðOOPHE i þ1ÞjZ ji , is represented by where OOPHE is the health expenditure data with many zeros; logðOOPHE i þ 1Þ is the log of OOPHE (we performed a monotonic transformation by adding one before taking the logs); Z 0 ji is a set of explanatory variables; 5 s represents the quantile points; h j ðsÞ is a vector of estimated quantile coefficients for the s th conditional quantile of log(OOPHE þ 1), and C i denotes the censoring point. The quantile coefficients on insurance will be biased if the factors that drive NHIS participation and OOPHE are unobserved (e.g. chronic or hereditary disease or other risk factors). The inability to control for the relevant unobserved confounders can create a spurious correlation between insurance and the error term in the OOPHE equation, leading to inconsistent estimates. Thus, we complement the CQR with the two instrumental variable estimators discussed below.
Insurance and out-of-pocket healthcare expenditure 401

CQIV
The CQIV estimator draws from Powell's (1986) CQR but extends Chernozhukov and Hong's (2002) algorithm using a control function approach to handle endogeneity. Although enrolment in the NHIS is individual, the main analysis was by household. Doing that permits using the CQIV estimator, which requires a continuous endogenous regressor (see Chernozhukov et al., 2015). A discrete endogenous regressor does not satisfy the assumptions necessary for the control function approach of this estimator (Kowalski, 2016). Angrist and Pischke (2008, pp. 142-144) discuss why estimating the first stage of a two-stage least squares estimation using the probit formulation is bad. They argue that only an OLS estimation guarantees first-stage errors that are uncorrelated with the covariates and fitted values.
Consider the following set of triangular quantile equations. (Refer to Chernozhukov et al., 2015;Kowalski, 2016, for further detail.) where all variables remain as defined; logOOPHE i is the logarithm of the observed/censored outcome for the ith household; logOOPHE Ã i is the logarithm of the latent outcome variable is the transformation function censoring logOOPHE Ã i from below at zero; HIPR i is the proportion of household members insured and Q HIPR represents the conditional quantile function of HIPR i given X ji and IV i ; X ji is a vector of exogenous covariates (with j indexing the various observed confounders); IV is the instrumental variable; i and e i are the independent error terms for log OOPHE Ã i and HIPR i respectively, so i is independent of HIPR i , X ji , IV i , e i , and C i : Following the two-stage least squares rationale, equation (4) is the first-stage quantile regression that estimates the predicted insurance values d HIPR i : Equation (3) is then estimated with d HIPR i replacing HIPR i to deal with endogeneity. The quantile coefficients of d HIPR i denote the insurance impacts for the specified quantiles (see Chernozhukov et al., 2015;Kowalski, 2016).
The estimations were performed in Stata 16. The household sampling weights provided by the surveys were applied to make the estimates nationally representative. Both the CQR and the CQIV were estimated using the command cqiv. The 95% confidence intervals were obtained via cluster-bootstrapping procedure with 200 replications. The bootstrap method is used to get standard errors for statistical inference when analytical expressions are problematic (Cameron & Trivedi, 2010, p. 415). It does this by resampling from the sample. The cluster-bootstrap method is suitable for clustered data, and standard errors are possibly correlated within clusters. Cameron and Miller (2015) vividly describe the cluster-bootstrapping procedure.

Instrumental variable.
A good IV must first be relevant, i.e. associated with insurance participation. The stronger the correlation between the instrument and endogenous regressor, the stronger the identification (Cameron & Trivedi, 2010, p. 175). Second, the IV must satisfy the exclusion restriction condition to be valid. That is, it ought to affect OOPHE only through insurance. While an IV's strength can be tested empirically, its validity is not verifiable. Conley et al. (2012) note that a strong IV can be used to offset the exclusion restriction's degree of violation. Finding an instrument that satisfies both conditions is challenging.
Consider first the exempt category IV. The NHIS provides a premium exemption for children, the elderly (70 plus), formal sector workers and pensioners, people with a mental disorder, pregnant women, and indigents. Supplementary Table S.5 shows the distribution of people eligible for the NHIS premium exemption. Theoretically, being entitled to the NHIS premium exemption is a good reason to join the scheme. The literature supports this belief. For example, van der Wielen et al. (2018) found that public sector workers and old adults are more likely than others to participate in Ghana's NHIS. Salari et al. (2019) also found that children, pregnant women, and people in paid employment had a higher probability of enrolling in the scheme.
Does this IV satisfy the exclusion restriction? The NHIS exogenously determines eligibility for the premium exemption. However, the factors that make people eligible for the premium exemption might affect their healthcare utilisation, and consequently, OOPHE. Children, the elderly, pregnant women, and the disabled usually have greater healthcare needs. This instrument is plausibly exogenous given income. 6 Income is the key determinant of OOPHE in many developing countries (see Ampaw, Nketiah-Amponsah, Agyire-Tettey, & Senadza, 2020, Ghana; Malik & Syed, 2012, Pakistan;Olasehinde & Olaniyan, 2017, Nigeria;Rous & Hotchkiss, 2003, Nepal). The poor put a higher premium on affordability than need when considering what to consume. Besides, older children who dominate the NHIS exempt category are the least prone to illness or injuries (Ghana Statistical Service, 2014). And controlling for age, disability, and pregnancy variables can remove any direct impact the instrument could have on OOPHE.
We now turn to the community insurance IV. Cheung and Padieu (2015) and Aryeetey et al. (2016) used a similar IV. Households in communities with higher NHIS participation rates are more likely to have positive health insurance perceptions. Thus, they should thus have a higher NHIS participation rate. Although intuitively sound in the Ghanaian context, given the communal culture, no empirical literature supports the above assertion to the best of our knowledge.
On the validity assumption, this instrument could correlate with OOPHE if healthcare infrastructure is not considered. Here is the logic. A community with fewer health facilities and personnel might charge higher fees for healthcare services, medicines, and medical supplies, leading to higher OOPHEs. And to cushion this effect, individuals in those communities may resort to enrolling in health insurance. Health infrastructure is unobserved in the urban data. The GLSS administered the community questionnaire in rural localities only. We include health facility availability, distance to health facility variables, and regional dummies in the rural analysis. But we include just the region dummies in the urban and national due to the data limitation. We have no secondary evidence of how well the region dummies control for unobserved health infrastructure, meaning the exclusion restriction could be violated.
Inference about the coefficients on HIPR i in Equation (3) will be exact if the exclusion restriction holds. That is, the instruments are uncorrelated with the unobservables that affect OOPHE. Our IVs may be only approximately exogenous. Therefore, we further explore the method proposed by Conley et al. (2012). That method is flexible on the exclusion restriction condition.

IV estimators under plausible exogeneity
Insurance and out-of-pocket healthcare expenditure 403 Again, all variables remain as defined. E½HIPR l 6 ¼ 0 and E IV 0 l ½ ¼ 0; b is our parameter of interest (the coefficient on the endogenous variable, insurance); c is the coefficient on the instrumental variable; ø denotes the coefficients on the control variables; P captures the correlation between the instrument and the endogenous regressor (first-stage coefficients); and l and e are error terms (unobserved).
The inclusion of IV ð Þc is the main distinction between Conley's approach and the classical IV estimators. This method draws from the 2SLS, which does not address the excess zero issues. (We did not exclude the zeros, nonetheless.) To satisfy the exclusion restriction, the 2SLS assumes that c is exactly zero. In contrast, Conley's method takes prior information suggesting that c is not exactly zero but close to zero.
The authors propose four methods for inference about the coefficient on the insurance variable (b) without assuming c ¼ 0 (see Conley et al., 2012; for details). The associated Stata code (plausexog, see Clarke & Matta, 2018) permits two methods: the union of confidence with c support assumptions and c local-to-zero approximation. Conley and colleagues report that estimates from the unions of prior-weighted confidence intervals and local-to-zero methods are close (see Conley et al., 2012).
We apply the union of confidence interval (UCI) approach, assuming that c falls into a small positive range (c 2 ½0, þ d). The fact that the plausible violations of the exclusion restriction are signed implies that we can think about the sign of the bias generated by incorrectly assuming that c ¼ 0: If P > 0 in (6), as we in fact find, then the estimate ofb in (5) is biased up if we incorrectly assume that c ¼ 0 when actually c > 0:  Table 1. The table reports the marginal effects of the CQR, which show the insurance effect on the conditional OOPHE distribution by quantile. The equations were estimated jointly for the two surveys, specifying OOPHE in both levels and log form. The log specification results will be discussed to emphasise the proportionate effects. (Please refer to the Supplementary Material for the absolute specification results.) Table 1 also reports the marginal effects of the Tobit formulation using a latent dependent variable. Although sensitive to extreme values, the Tobit result serves as a benchmark for the 50th quantile (median) estimates of the CQR. confidence intervals from a cluster bootstrapping procedure with 200 replications. Bold signifies statistically significant estimates (p < 0.05). Household sampling weights provided by the surveys were applied for nationally representative estimates. Stata command cqiv fits the CQR estimator. The control variables are head characteristics (a female dummy, a Christian dummy, educational attainment, and marital status). Also, household characteristics (household size, consumption expenditure per capita), proportion under 18 years, the proportion aged 60-69, the proportion aged 70 plus, the proportion that used medical services, proportion disabled, and proportion pregnant); geographical characteristics (an urban dummy and region fixed effects); and a wave dummy.

Results
The CQR shows that NHIS membership predicted less OOPHE. The estimates are negative and statistically significant between the 30th and 90th quantiles. Those quantiles' marginal effects are statistically significant because the lower and upper bounds of the 95% confidence interval do not include zero. The lower quantile estimates are zero due to the excessive zeros in the OOPHE data. The NHIS membership predicted 19% less OOPHE at the median. Also, the Tobit result suggests a 24% reduction.
Our finding is comparable to that of Powell-Jackson and his colleagues (Powell-Jackson, Hanson, Whitty, & Ansah, 2014). Using a randomised experiment between 2004 and 2005, they found that removing healthcare user fees reduced average OOPHE by 27% in Ghana. Also, a group led by Garcia-Mandic o reported (Garcia-Mandic o et al., 2021) a 20% reduction in average OOPHE, and Strupat and Klohn (2018) reported a 41% reduction. Both studies used data from the GLSS5 (2005GLSS5 ( /2006).
The CQR further shows that the crowding-out effects were smaller at the extreme tails of the OOPHE distribution than in the other quantiles. Being insured predicted 11% and 16% reduction in OOPHE at the 40th and 80th quantiles. The greatest impact was at the 60th quantile (23%). These results suggest that the NHIS had a more significant effect on people with moderate OOPHEs. These results are difficult to compare with those published previously because the earlier studies estimated the effects separately for inpatient and outpatient OOPHEs (see Barnes et al., 2017;Cheung et al., 2016;Zhang et al., 2017).

Heterogeneity in the effects.
Of course, health insurance may affect various groups with varying socioeconomic backgrounds or OOPHE needs differently. We test the statistical significance of the null hypothesis that the coefficient on health insurance is the same across waves, rural-urban localities, and income groups (see Supplementary Table S.11 for the results). We reject the null that the coefficients do not change across waves and fail to reject it in the other instances. The change in the insurance effect between the two waves was massive, and the effect was greater in 2013 than in 2017. We provide separate estimates for the 2013 and 2017 samples. The results are reported in Supplementary Table S.13.

Disaggregating by wave.
The results show that going from having no insured household member to all insured predicted less OOPHE in 2013 but showed no statistically significant relationship in 2017. The 2013 estimates are negative and statistically significant from the 40th quantile onwards. Insurance predicted 29% less OOPHE at the median in 2013. Like the joint estimates, the CQR further shows that the crowding-out effects in 2013 were smaller at the extreme tails of the OOPHE distribution than in the other quantiles. Being insured predicted a 22% reduction in OOPHE at the 40th and 90th quantiles. In between, the insurance effect was reasonably stable. The greatest impact was at the 60th quantile (34%).
What accounts for the different results across waves? We discuss two plausible explanations for why the NHIS appears to have been less effective in the latter year. The first is increased moral hazards, and the second is falling government health expenditures.
Ex-ante moral hazards are unlikely in Ghana's health system (Powell-Jackson et al., 2014;Salari et al., 2019). These are behavioural changes that predispose the insured to greater health risks, causing them to use more healthcare. The group led by Powell-Jackson found that free healthcare did not influence people's preventive measures or the incidence of self-reported illness. However, there are health-seeking behaviour changes after getting sick (ex-post moral hazard). Previous studies show a positive relationship between NHIS subscription and healthcare utilisation (Bagnoli, 2019;Blanchet, Fink, & Osei-Akoto, 2012;Brugiavini & Pace, 2016;Khalid & Serieux, 2018). Also, the NHIS may be associated with higher OOPHE in 2017 if it motivated private healthcare use.
Household head data were used to test to what extent NHIS membership had a stronger relationship with healthcare utilisation and private healthcare use in 2017 than in 2013. The joint Insurance and out-of-pocket healthcare expenditure 405 estimates reveal that the insured were more likely to use healthcare. The users of healthcare were less likely to visit private facilities. We further test the statistical significance of the null hypothesis that the coefficients did not change across waves (Supplementary Table S.12). We fail to reject the null in both cases. Therefore, we cannot attribute the results to increased (expost) moral hazards. Figures S. 6 and S.7 show that Ghana's government health expenditure has declined in absolute and relative terms since 2011. The government spent US$99 per capita in 2012, US$74 in 2013, US$55 in 2016 and US$49 in 2017 (all at purchasing power parity). Low government health spending leads to unofficial payments. Ensor and Thompson (2012, p. 154) define unofficial payments as "those given to providers outside official channels for services that should be covered by the (public) health care system." The falling government health spending and growing reliance on private healthcare financing suggest that the NHIS has become less generous, probably why it was less effective in 2017.

Disaggregating by OOPHE types.
Here, we disaggregate the total OOPHE into its components. We estimate the insurance association with outpatient OOPHE, inpatient OOPHE (hospitalisation fees) and OOPHE for medicine. In 2017 there was still no statistically significant relationship.
Applying Koenker and Bassett's (1978) quantile regression (QR) estimator to estimate any relationship between insurance and inpatient, outpatient, and medicine OOPHE highlighted three critical points for 2013. (1) There was significant evidence of crowding-out of outpatient spending only at the 90th quantile. (2) Insurance had no statistically significant relationship with inpatient spending. (3) There was a significant negative relationship with medicine expenditure beyond the 60th quantile. The reduction was more considerable for the outpatient OOPHE than for medicine in 2013.
Garcia-Mandic o's group has also reported (Garcia-Mandic o et al., 2021) that the introduction of the NHIS did not reduce average inpatient OOPHE in 2005/06. On the other hand, groups led by Barnes (Barnes et al., 2017) and Zhang (Zhang et al., 2017) report contrasting findings for India and China. Barnes and his colleagues report that insurance reduced inpatient OOPHE in India at higher quantiles, even as Zhang found the same at the lower and medium quantiles in China.
This study's outpatient OOPHE results for 2013 contradict Garcia-Mandic o, however. That group has reported that the introduction of the NHIS did not reduce average outpatient OOPHE. Zhang found that in China, insurance reduced outpatient OOPHE at the 20th quantile and increased it only at the 90th. This study has been the first to estimate the insurance effect on the distribution of medicine OOPHE. Garcia-Mandic o's reported average medicine OOPHE for 2005/2006 does, however, support these 2013 findings. The NHIS has not reduced inpatient OOPHE nationally, and its effect on outpatient and medicine OOPHE varies. Table 2 reports the CQIV and Tobit IV estimates for the pooled and disaggregated samples. The evidence corroborates the earlier results. Specifically, the Tobit IV estimates reveal a negative and statistically significant relationship between insurance and OOPHE (Panel A). The disaggregated analyses further suggest that the relationship was statistically significant in just 2013 (Panel C). The CQIV estimates are statistically significant at the 80th (35%) and 90th (41%) quantiles (Panel D). Although consistent with the 80th quantile estimate from the CQR (16%), the CQIV's 80th quantile estimate is higher.

CQIV
A description of the Tobit IV method is presented in the Supplementary Material. The same set of covariates used in the CQR was used in the disaggregated analyses. However, the pooled estimation achieves convergence only when we exclude the proportion under 18 years, the proportion aged 60-69, the proportion aged 70 plus, and region variables. The results in Figure 1 suggest a negative and statistically significant relationship between insurance and OOPHE even in the absence of a perfect IV (c 0). Therefore, if we relax the exclusion restriction and permit c > 0, the estimate of b becomes more negative. The interval estimates for b with support restriction of c 2 ½0, 2, for instance, is roughly [À51, À2]. Although the confidence set if the exclusion restriction is correct is approximately [À7, À2], the

Conclusions
Using pooled data from the 2013 and 2017 Ghana Living Standards Surveys, this paper documents a negative relationship between health insurance and OOPHE. We find statistically significant differences between the 2013 and 2017 estimates. The scheme appears to have had a greater impact in 2013 than in 2017. The separate analyses reveal that membership in the insurance scheme predicted less OOPHE in 2013 but had a statistically insignificant effect in 2017. We also find that insurance had a statistically insignificant relationship with hospitalisation fees even in 2013. Its effect on outpatient OOPHE was more significant than on medicine OOPHE. Insurance also reduced the risk of catastrophic health expenditure in 2013 but had a statistically insignificant effect in 2017. Some potential limitations of this study must be acknowledged. First, the OOPHE data were unverified and might suffer from erroneous recall, although the recall period was short and the Ghana Statistical Service applied several quality control measures during the surveys. Second, the CQR results for inpatient OOPHE could be dubious because of the small sample of those who spent out-of-pocket on hospitalisation in the two weeks before the survey. Third, our identification strategy might not entirely remove unobserved confounders. Hence, we are careful in giving causal interpretations to our findings. A cleaner identification strategy (randomised experiments) might be worth exploring to complement the current study. Results from randomised experiments are hardly generalisable, which is one of our paper's strengths since we use nationally representative samples.
The research supports the United Nation's SDG 3, Target 3.8, which aims to " … achieve universal health coverage, including financial risk protection, access to quality essential healthcare services and access to safe, effective, quality and affordable essential medicines and vaccines for all" (Loewe & Rippin, 2015). Ghana's NHIS is among the first national health insurance in sub-Saharan Africa and an exemplary model for many countries in the region. Therefore, the findings of this research might be relevant to other low-and middle-income countries. The Ghana government might consider reviewing the NHIS to fulfil its mandate. As a matter of urgency, it could contemplate enforcing the compulsory membership of informal sector workers. Achieving universal health coverage with the NHIS would be possible only if every political party subscribes to improving the scheme without seeking points for itself. Hence, there is a need for both main political parties to embrace the NHIS and commit to making it work. It is also high time the government revised its financial commitment to the health sector.
Future studies could extend these results beyond 2018. That was when the Insurance Authority implemented a mobile renewal policy. That policy, which digitised the NHIS membership renewal and some aspects of its management, has enhanced the scheme's accessibility and financial management (NHIA, 2019).