Health insurance and hospitalisation duration: empirical evidence from Ghana’s national health insurance scheme

Abstract

The study aims to explore the causal effect of Ghana’s National Health Insurance Scheme (NHIS) on hospitalisation duration. The analysis was based on the Ghana Socioeconomic Panel Survey datasets, comprising the second wave (2014/2015) and the third wave (2018/2019). The study employed the endogenous switching regression model for count data (ESRC) to control selection bias and unobserved heterogeneity. The ESRC estimates reveal that NHIS membership significantly reduces the length of stay in the hospital, indicating that health insurance has a negative association with hospitalisation. On average, insured people spend nearly five fewer days in the hospital than their uninsured counterparts. The findings further revealed common and heterogeneous determinants of hospitalisation care for both insured and uninsured individuals. Age, household expenditure, and self-assessed health were the main predictors of hospitalisation duration for insured and uninsured persons. Heterogeneously, gender, education, and physical inactivity are significant determinants of NHIS members’ hospitalisation care, while chronic illness affects the length of stay in the hospital of the uninsured. The paper concludes with a discussion of the policy options for increasing NHIS enrolment to reduce the length of stay in the hospital and improve individuals’ well-being.

Keywords:

• Endogenous switching regression model
• health insurance
• hospitalisation
• Ghana

Reviewing Editor:

• Joe Thomas, Institute of Health and Management, Australia

Subjects:

• Economics and Development
• Health & Development
• Economics

1. Introduction

Universal health coverage (UHC) asserts that everyone should have access to the care they need, regardless of their ability to pay for it. Sustainable Development Goal (SDG) 3.8 aims at attaining UHC, which includes financial risk protection, access to high-quality essential healthcare services, and access to safe, effective, high-quality, and affordable necessary medications and vaccinations (United Nations, 2017). To achieve UHC, people must have both financial and healthcare access. Financial hardship is one of the most significant effects of paying for healthcare out-of-pocket (OOP) at the time of usage. When people make official and informal OOP payments for healthcare services at a point of use that exceeds their ability to pay, they face financial hardship (Moreno-Serra et al., 2013). Inpatient care or hospitalisation is more likely to create financial hardship due to out-of-pocket healthcare expenses than outpatient care. For the uninsured, hospitalisation care can be expensive due to the many different aspects of care that add to the costs. These costs may include hospital beds, meals, treatment, health checks, blood tests, etc. Depending on where one lives, a hospital bed for the night can cost anything from £510 in Italy to £2,193 in Canada (AXA, 2022).

Financial hardship can limit healthcare access, wreak havoc on one’s health, worsen poverty, and accentuate health and socioeconomic disparities. Unfortunately, poorer households, as well as those who must pay for long-term treatment, are highly vulnerable. Countries worldwide are continuously making changes to strengthen their health systems and give financial protection to their populations to assuage the effects of financial difficulty in accessing healthcare. In developing countries, health insurance is becoming more and more popular as a way to protect households from the impoverishment that comes with paying out-of-pocket costs. The World Health Organisation acknowledges health insurance as a promising mode of achieving UHC (World Health Organisation, 2010).

The Government of Ghana implemented a national health insurance scheme (NHIS) in 2003 to finance healthcare in the country. The scheme covers about 95% of all disease burdens in Ghana. The foremost purpose of Ghana’s NHIS is to increase access to essential healthcare services, replace out-of-pocket expenditures, and provide financial protection against excessive healthcare costs. Outpatient, hospitalisation, maternity, eye, and dental healthcare services are all included in the list of services covered by the scheme (Jehu-Appiah et al., 2012). The scheme covers persons residing in Ghana. The purpose of this criterion is to guarantee that the services provided by the scheme are available only to persons in Ghana. Members of the annual premium-paying and exempt groups make up the two categories of NHIS subscribers. Those who register as annual premium-paying members pay annual premiums and processing fees for renewal. The exemption group that does not pay premiums and processing fees includes pregnant women, indigents, persons with mental disorders, active SSNIT contributors, SSNIT pensioners, persons over seventy years old (the elderly), and other categories prescribed by the Minister responsible for Social Welfare.

A thorough examination of the earlier research related to hospitalisation care showed that different authors used either yes or no binary analysis for hospitalisation in the last 12 months (Han-Kim & Lee, 2016; Van Der Wielen et al., 2018) or length of stay [i.e., the number of days spent in the hospital] (Hullegie & Klein, 2010; Tian et al., 2012) to operationalise hospitalisation. Studies that analysed hospitalisation as integer-valued count data primarily adopted the Poisson regression (Geitona et al., 2007), negative binomial model (Tian et al., 2012), and multivariate regression analysis (Dutta & Husain, 2013) but failed to address the issue of endogeneity of the health insurance variable. Few studies have acknowledged health insurance as endogenous when hospitalisation is measured as a count data. For instance, Riphahn et al. (2003) used the bivariate panel count data estimation technique to study demand for healthcare. Previous studies did not control for the event that individuals with health insurance may differ from the uninsured in observable and unobservable characteristics. This study aims to investigate the causal impact of health insurance on hospitalisation. This study contributes to the literature on health insurance and hospitalisation by employing the endogenous switching regression model for count data (henceforth ESRC model) to control selection bias and unobserved heterogeneity. Another benefit of using the ESRC model for this study is that it enables us to evaluate the variables that affect a person’s likelihood of enrolling in health insurance and the variables that influence their length of stay in the hospital.

The remainder of the paper is organised as follows: The next section presents the methodology applied in the study. Section three presents and discusses the results of the empirical analysis, and the final section concludes the paper.

2. Methodology

2.1. Data

Data was sourced from the Ghana Socioeconomic Panel Survey (GSPS) datasets. The demographics, assets, health (including insurance status, anthropometry, daily life activities, health in the past two weeks, health in the past 12 months, etc.), consumption module, and housing characteristics of households are among the themes explored by this national survey. Because the first wave had difficulties producing the household expenditure variable, which serves as a proxy measure of income, the study used the second wave (2014/2015) and the third wave (2018/2019). Table 1 provides the descriptive statistics and operational measurement of the variables used in the analysis.

Table 1. Variables, measurement and descriptive statistics.

Variable	Measurement	Obs	Mean	SD
Dependent variable
Hospitalisation	Continuous: number of hospitalisation days	862	6.4612	14.509
Independent variables
NHIS enrolment	Dummy: 1 = currently insured; 0 = otherwise	19,247	0.615	0.487
Gender	Dummy: 1 = male; 0 = otherwise	19,247	0.479	0.500
Age	Continuous: positive whole numbers in years	19,247	35.721	20.108
Age squared	Continuous: positive whole numbers of square of age	19,247	1680.313	1723.330
Education	Ordinal: measured on a five-point scale ranging from 0 for no formal education to 4 as the highest educational level attained	19,247	0.492	0.834
Locality (urban)	Dummy: 1 = resides in urban area; 0 = otherwise	19,247	0.3412	0.474
Fever	Dummy: 1 = if suffered Fever; 0 = otherwise	19,247	0.060	0.237
Chronic illness	Dummy: 1 = if exposed to chronic illness (sores, irritations and/or numbness); 0 = otherwise	19,247	0.325	0.468
Physical inactivity	Dummy: 1 = difficulty in participating in physical activities/roles; 0 = otherwise	19,247	0.139	0.346
Self-assessed health	Ordinal: unhealthy =1, somewhat unhealthy =2, somewhat healthy =3 and Very healthy =4	19,247	3.678	0.6108
Average household expenditure	Continuous: positive numbers in Ghana cedis (i.e. monthly household expenditure divided by household size)	19,247	149.908	178.856
Instrumental variable
Formal-sector work	Dummy: 1 = if household head is a formal-sector worker; 0 = otherwise	19,247	0.107	0.309

SD: standard deviation; Obs.: observation.

2.2. Econometric model

A self-selection problem is relevant in evaluating the impact of treatment when the decision to receive treatment is based on individual-specific heterogeneity, which cannot be observed by empirical researchers. For instance, insured people may differ from the uninsured in sociodemographics, risk choices, health behaviours, tendency to use healthcare, and baseline health state (Sengupta & Rooj, 2019; Waters, 1999). According to Hasebe (2020), an unaddressed selection of unobservables could make determining the causal inference of treatment effects more difficult. Count nonnegative data and integers are frequently used in empirical studies of health economics and other applied microeconomics. The current study used hospitalisation, measured as the number of hospitalisation days, as our dependent variable. Hence, our outcome variable is count data. The selection problem is a typical feature of empirical studies of applied microeconomics.

Propensity score matching (PSM), the inverse-probability-weighted regression adjustment (IPWRA) estimator, Poisson regression with endogenous treatment effects (PRETE), and endogenous switching regression for count-dependent variables have all been used by researchers to estimate the impact of a binary treatment variable on a count variable (Danso-Abbeam et al., 2021; Hasebe, 2020; Khachatryan et al., 2019; Zheng & Ma, 2023). PSM and IPWRA are nonparametric methods that solely account for selection bias resulting from observed factors. On the other hand, the PRETE model can reduce selection bias resulting from both observed and unobserved factors but only estimates one selection equation and one outcome equation. This is where the ESRC model has an advantage: it can simultaneously estimate one selection equation and two outcome equations (Regime 1 and Regime 0), mitigating both observed and unobserved selection biases.

This study follows Hasebe (2020), who developed a new STATA user-written command, ‘escount’, which executes an endogenous-switching regression model with count-data outcomes suggested by Terza (1998) using the maximum likelihood estimation method. Here, potential outcomes (i.e. hospitalisation), which are count data, vary across two different treatment statuses, and the treatment status (i.e. NHIS membership) is endogenous. Even after controlling for observable factors, the potential outcomes are not independent of the treatment. One uses the assumption of lognormal latent heterogeneity to model the relationship between the selection process and the outcome. The lognormal latent heterogeneity offers a flexible specification, as explained by Greene (2009).

In this method, we have a binary treatment (selection) equation and a count discrete outcome equation. Beginning with the binary treatment equation assume ${\text{NHIS}}_i^{*}$ is a latent variable that captures the expected benefits of membership and non-membership in NHIS. The latent variable for i^th individual can be specified as follows: (1) $${\text{NHIS}}_i^{\text{*}}={z^{\prime }}_i{\gamma }_i+v_i\text{\hspace{0.33em}with\hspace{0.33em}}\hspace{0.17em}{d}_i=\left\{\begin{array}{ll}1\hfill & \mathit{\text{if\hspace{0.33em}NHIS}}{\hspace{0.17em}}_i^{\text{*}}>0\hfill \\ 0\hfill & \mathit{\text{otherwise}}\hfill \end{array}\right.$$(1) where $z_i$ is the set of explanatory variables hypothesised to influence NHIS membership and $v_i$ is the error term, and the i^th individual membership in NHIS ($d_i$=1) should they experience a positive benefit from it (${\text{NHIS}}_i^{*}$> 0).

The outcome equation describing the hospitalisation duration of membership and non-membership of NHIS can be expressed as follows: (2a) $$\mathit{\text{Regime}}\hspace{0.17em}1\left(d_i=1\right):y_{1i}={x^{\prime }}_i{}^{\prime }{\beta }_{1i}+{\epsilon }_{1i}\hspace{0.17em}$$(2a) (2b) $$\mathit{\text{Regime}}\hspace{0.17em}0\left(d_i=0\right):y_{0i}={x^{\prime }}_i{}^{\prime }{\beta }_{0i}+{\epsilon }_{0i}$$(2b) $y_{1i}$ and $y_{0i}$ refer to hospitalisation duration for membership and non-membership of NHIS, respectively. ${\gamma }_i$ and ${\beta }_{\mathit{\text{ji}}}$ (j = 1, 0) are parameters to be estimated. ${\epsilon }_{\mathit{\text{ji}}}$ (j = 1, 0) are error terms. $x_i$ is a vector of independent variables that are expected to influence hospitalisation duration.

The lambda parameters from the selection equation, which fulfil non-linearity identification, are included in these independent variables. They represent the endogenous selection mechanisms that capture the likelihood of being in a specific regime. Our second identification strategy involves employing exclusion restrictions. The variable, formal sector work, is used as an exclusion restriction serving as an instrument. Formal sector work had been used as an instrument in earlier research conducted in the Ghanaian context (Sekyi, Adom, et al., 2023; Sekyi, Nhamo, et al., 2023). According to the authors, all employees in Ghana’s formal sector are by law to contribute to the Social Security and National Insurance Trust (SSNIT), making all workers in the formal sector SSNIT contributors. As per the statute that established Ghana’s NHIS, workers are to pay 2.5% of their SSNIT contributions to the NHIS to fund it. The law also mandates SSNIT contributors to enrol in the NHIS without premium payment. The authors found formal sector work as a good instrument for NHIS enrolment. The variable, formal sector work, is expected to influence enrolment but does not directly affect the length of hospital stay. We examined the instrument’s validity using a Pearson correlation coefficient analysis. The findings indicate that our instrument is valid because it has a significant and positive correlation with NHIS enrolment but not with hospitalisation days (refer to Table A1 in the Appendix). Added to Pearson correlation analysis, we also estimated a logit model for NHIS membership and a negative binomial regression model for hospitalisation days with the inclusion of the instrument (i.e. formal sector work). The results show that the instrument is statistically significant in influencing NHIS membership but not statistically significant in predicting hospitalisation days (refer to Table A2 in the Appendix). The results further reinforce the validity of our instrument.

We take into account a model with parametric distributional assumptions developed form Terza (1998). To model the dependency concerning $y_{\mathit{\text{ji}}}$ and $d_i$ for $j=0;1,$ we presume lognormal latent heterogeneity. Suppose that $y_{\mathit{\text{ji}}}$ follows either a Poisson or a negative binomial distribution depending on $x_i$ and ${\epsilon }_{\mathit{\text{ji}}}$, which is latent heterogeneity. The selection process counts on observable characteristics $z_i$ and an unobservable term $v_i:d_i=1(z_i^{,}\mathrm{\gamma }+v_i>0)$, where 1(.) is an indicator function. $y_{j }$ and $d$ depend on one another via the correlation between ${\epsilon }_j$ and $v$. We consider $v,{\epsilon }_0$ and ${\epsilon }_1$ to be jointly normal, $$\left(\begin{array}{c}\upsilon \\ {\epsilon }_0\\ {\epsilon }_1\end{array}\right)\sim N\left(\left[\begin{array}{c}0\\ 0\\ 0\end{array}\right],\left[\begin{array}{ccc}1& {\rho }_0{\sigma }_0& {\rho }_1{\sigma }_1\\ {\rho }_0{\sigma }_0& {\sigma }_0^2& {\rho }_{01}{\sigma }_0{\sigma }_1\\ {\rho }_1{\sigma }_1& {\rho }_{01}{\sigma }_0{\sigma }_1& {\sigma }_1^2\end{array}\right]\right)$$ where ${\sigma }_j$ is a standard deviation of ${\epsilon }_j$ and ${\rho }_j$ is the coefficient of correlation between $v$ and ${\epsilon }_j$ for $j=0;1$. The parameter ${\rho }_{01}$ represents the correlation coefficient between ${\epsilon }_0$ and ${\epsilon }_1$, however, it is not identified in the model.

Since $y_{1i}$ and $y_{0i}$ referring to hospitalisation duration are measured as count variables, the Poisson regression models should be used to estimate Equations (2a)(2a) $$\mathit{\text{Regime}}\hspace{0.17em}1\left(d_i=1\right):y_{1i}={x^{\prime }}_i{}^{\prime }{\beta }_{1i}+{\epsilon }_{1i}\hspace{0.17em}$$(2a) and (2b)(2b) $$\mathit{\text{Regime}}\hspace{0.17em}0\left(d_i=0\right):y_{0i}={x^{\prime }}_i{}^{\prime }{\beta }_{0i}+{\epsilon }_{0i}$$(2b) . Letting $\varnothing (.)$ and $\Phi (.)$ to denote the probability density function and cumulative distribution function of a standard normal distribution, respectively, and assuming normality, the joint probability $f_j(y_i,d_i|x_i,z_i)$ is expressed as follows (Hasebe, 2020): $$\begin{array}{l}{f}_j\left(y_i,d_i{x}_i,z_i\right)={{\displaystyle \int }}_{-\infty }^{\infty }{f}_j\left(y_i{x}_i,{\epsilon }_j\right)\Phi \\ \hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\left\{\frac{\left(2d_i-1\right)\left({z^{\prime }}_i\hspace{0.17em}\gamma +{\rho }_j{\sigma }_j^{-1}{\epsilon }_j\right)}{\sqrt{1-{\rho }_j^2}}\right\}\varphi \left({\epsilon }_j\right)d{\epsilon }_j\end{array}$$

Where $f_j\left(y_i|x_i,{\epsilon }_j\right)$ is $$f_j\left(y_i{x}_i,z_i\right)=\frac{\text{exp}{\left({x^{\prime }}_i{\beta }_j+{\epsilon }_j\right)}^{\mathit{\text{yi}}}\hspace{0.17em}\text{exp}\hspace{0.17em}\left\{-\text{exp}\left({x^{\prime }}_i{\beta }_j+{\epsilon }_j\right)\right\}}{y_i!}$$

Upon a suggestion from an anonymous reviewer, we refer readers to Hasebe’s (2020) work on the mathematical specifications for a negative binomial distribution with the over-dispersion for a Poisson distribution and the log-likelihood function.

In contrast to the endogenous dummy variable model which confines ${\sigma }_0={\sigma }_1$ and ${\rho }_0={\rho }_1$, the switching model permits more flexible unobservable heterogeneity behaviours. The endogenous treatment dummy variable typically has more accurate estimates than the switching model since it is more parsimonious. The endogenous treatment dummy variable model is biased and misspecified if the restrictions on $\rho$ and $\sigma$ are violated. The switching model is superior in these circumstances. In reality, applied researchers frequently lack advanced knowledge of the behaviours of unobservable heterogeneity. One benefit of the switching model is that after fitting it, one can quickly evaluate the restrictions using Wald and likelihood-ratio tests because it nests the endogenous dummy variable model. In such a situation, the command ‘escount’ offers applied researchers a model-selection tool added to model choice. Additionally, the ‘escount’ command supports negative binomial specification.

Assessing the heterogeneity of treatment effects is worthwhile and imperative in conducting policy analysis. Following executing the command ‘escount’, the command ‘teescount’ estimates various treatment effects. We estimated three treatment effects, the average treatment effect (ATE), the average treatment effect on the treated (ATT), and the average treatment effect on the untreated (ATU). For the mathematical specifications of ATE, ATT, and ATU refer to Hasebe’s (2020) work.

3. Results and discussion

3.1. The determinants of NHIS enrolment and intensity of healthcare utilisation

The estimates from the endogenous switching regression model for the count data outcome are presented in Table 2. An important finding is the sign and significance of the correlation coefficients that are the lambda parameters, which measure the relationship between the NHIS enrolment and hospitalisation models. The results reveal statistically significant coefficients for both NHIS members and non-members, demonstrating the presence of self-selection. The negative correlation between NHIS enrolment and insured hospitalisation care suggests that insured persons have higher hospitalisation care usage than if they did not enrol in NHIS. This finding suggests that separating the two models is most likely to produce bias and inconsistent estimates. Therefore, the ESRC model will assist in reducing this bias and yield consistent results. Furthermore, some covariates were shown to affect the outcome variables of insured and uninsured persons, indicating the presence of heterogeneity in the sample. The Wald test shows that our model fits the data reasonably well, with a significance level of 1%.

Table 2. Results of the endogenous switching regression model for count data outcome.

		Hospitalisation days
Variable	NHIS enrolment	Uninsured	Insured
Gender	−0.313***	0.131	0.415***
	(0.107)	(0.146)	(0.053)
Age	0.010***	−0.049***	0.029***
	(0.003)	(0.007)	(0.007)
Age squared	−	0.001***	−0.0004***
	−	(0.0001)	(0.0001)
Education	−0.039	−0.105	−0.107***
	(0.074)	(0.091)	(0.022)
Chronic illness	−0.003	0.327***	−0.009
	(0.101)	(0.085)	(0.059)
Physical inactivity	0.148	−0.312	0.530***
	(0.138)	(0.207)	(0.085)
Average household expenditure	−0.000	−0.001*	0.001***
	(0.000)	(0.000)	(0.000)
Self-assessed health	−	−0.571***	−0.317***
	−	(0.141)	(0.041)
Fever	−	0.100	−0.109
	−	(0.171)	(0.085)
Locality (urban)	−	0.165	0.043
	−	(0.134)	(0.058)
Formal-sector work	0.302**	−	−
	(0.150)	−	−
Constant	0.290**	4.635***	1.845***
	(0.147)	(0.610)	(0.180)
Lambda	−	0.612***	−0.543***
	−	(0.078)	(0.082)
Observations	827
Wald ${\chi }^2$	1,136.81***
Log pseudolikelihood	−2,643.051

Robust standard errors in parentheses; *p < .1, **p < .05, ***p < .01.

The first part of the ESRC model estimates the selection equation (i.e. NHIS enrolment). Here, three variables, namely gender, age, and formal-sector work, were statistically significant in influencing individuals’ NHIS enrolment decisions. The exclusion restriction variable (i.e. formal-sector work) is statistically significant in the selection equation model. The positive coefficient suggests that those employed in the formal sector have a higher likelihood of enrolling in the scheme compared to workers in the nonformal sector. This result corroborates previous studies in Ghana (Sekyi, 2022; Sekyi, Nhamo, et al., 2023). Gender (male) has a negative influence on enrolment. According to this finding, males are less likely than females to enrol in the NHIS. Results from previous studies lend credence to this conclusion (Sekyi et al., 2015; Wiredu et al., 2021). Age is a positive predictor of NHIS enrolment. This result implies that older people are more likely to enrol. Due to their higher risk of sickness, as they age, older people may invest more in their health, including purchasing health insurance, as a safeguard against unforeseen financial worries associated with illness. This result corroborates similar studies (Aregbeshola & Khan, 2018; Salari et al., 2019; Yadav & Mohanty, 2021).

The second-stage estimation technique of the ESRC model involves estimating the outcome equation (i.e. hospitalisation days). This result shows the factors influencing the duration of hospitalisation for NHIS members and non-members. Age, household expenditure, and self-assessed health were joint drivers of hospitalisation care for insured and uninsured persons. Our results support a non-linear (quadratic) relationship between age and hospitalisation duration. Age squared is positively associated with hospitalisation duration for the uninsured but negatively related to the length of hospital stays for the insured. The result suggests that age influences hospitalisation care negatively up to the age of 28¹, after which it positively affects the length of hospitalisation for those without insurance. However, in the case of insured persons, the number of days spent in the hospital increases with advancing age. But after nearly 38 years, the use of hospitalisation care diminishes. The result for the uninsured appears counterintuitive since the finding suggests that the number of days uninsured people spend in hospitals decreases with age. This finding defies the usual thinking that suggests older patients stay longer in hospitals due to their advanced age, which complicates healing and increases their risk of developing chronic illnesses. Might the early discharge of uninsured individuals be due to their inability to pay for hospitalisation care?

Household expenditure, often used as a proxy for income, positively predicts the duration of hospitalisation. This finding is consistent with a study conducted by Dutta and Husain (2013), who found a positive influence of household expenditure on hospitalisation care among insured persons in urban India. Self-assessed health is a negative determinant of hospitalisation care. This result implies that the likelihood of intensifying hospitalisation usage decreases with self-reported good health. In other words, the duration of hospitalisation is associated with poorly rated health status. Our result confirms findings in the literature, which show self-assessed health as a significant determinant of hospitalisation (Cislaghi & Cislaghi, 2019; Isaac et al., 2015; Tamayo-Fonseca et al., 2015; Tian et al., 2012).

Differences in the determinants of hospitalisation care were observed for insured and uninsured individuals. Gender, physical inactivity, and education significantly determine NHIS members’ hospitalisation duration, as chronic illness influences uninsured hospitalisation care. The gender variable is a positive determinant of hospitalisation for insured individuals. This finding means that the likelihood of hospitalisation care usage increases intensively for insured males relative to their female counterparts. The risk-loving nature of men and their traditional roles may explain this finding. Men are relatively more adventurous. Also, traditional roles of men, such as providing for the family, make them engage in risky activities that expose them to health hazards, including injuries resulting in higher hospitalisation. This result reinforces the findings of Dutta and Husain (2013) and Han-Kim and Lee (2016) but contradicts the findings of Cameron et al. (1988), Saeed et al. (2015), and Li et al. (2016).

Education is a negative determinant of hospitalisation for insured persons. This result implies that highly educated individuals spend fewer days in the hospital. Education enlightens individuals’ perceptive powers and increases the need to seek medical attention promptly, reducing the severity of illness and its attendant problems. Some previous researchers explained that less educated people are less conscious of the necessity for hospitalisation and have a lower financial capacity to cover hospitalisation costs (Dutta & Husain, 2013; Grossman & Kaestner, 1997). This finding is consistent with results from previous studies (Dutta & Husain, 2013; Han-Kim & Lee, 2016).

People who have difficulty participating in physical activities may have underlying health conditions that put them at risk of being admitted to the hospital. The variable that captures people’s difficulties in participating in physical activities is a positive driver of hospitalisation care for insured persons. This finding reveals that those having difficulty engaging in physical activities spend more time in the hospital when admitted. This finding reinforces a study by Cameron et al. (1988).

The fact that the chronic illness coefficient is positive means that the uninsured with chronic illnesses spend more days in the hospital than those without such conditions. This finding reinforces an earlier study by Hernandez et al. (2009), who discovered that each chronic condition was linked to a 30% increase in hospital admissions for the preceding year.

3.2. Effect of NHIS membership on the intensity of healthcare utilisation

After estimating the parameters of the ESRC model with the escount command, the teescount command was used to estimate various treatment effects. Table 3 presents these results. This command provides an avenue to estimate the causal influence of NHIS membership on the duration of hospitalisation. The overall impact of NHIS membership on hospitalisation duration is shown by the average treatment effect (ATE) and is statistically significant. This result implies that insured people spend almost five fewer days in the hospital than their uninsured counterparts. The impact of health insurance on NHIS participants alone, as evaluated by the ATT, indicates an eight-day reduction in hospitalisation days. These findings suggest that NHIS membership reduces the number of days spent in the hospital. We identify three possible pathways through which NHIS membership can reduce hospitalisation duration. First, having health insurance makes possible timely access to healthcare services, such as general care, specialist consultations, and preventative care. Early diagnosis and treatment can decrease the severity of health problems and the necessity for extended hospital stays. Second, having health insurance can help prevent or detect health problems before they become serious. Health insurance can lower the risk of acute health problems, which may necessitate hospitalisation, by encouraging preventative care and disease management. Lastly, prescription drugs serve as a benefit of health insurance, lower the cost and increase patient access to necessary prescriptions. Following a doctor’s prescription can help control symptoms, manage chronic conditions, and prevent disease from getting worse. It can also shorten hospital stays for exacerbations of chronic illnesses and even help to avoid the need for hospitalisation altogether. The result is similar to the findings of Hullegie and Klein (2010) and Geitona et al. (2007). For instance, Hullegie and Klein (2010) found that private insurance had a marginally negative significant impact on the number of hospital nights spent. Our findings, however, contradict those of Tian et al. (2012), Liu and Zhao (2014), Van Der Wielen et al. (2018), and Ta et al. (2020). The difference between our findings and those of Van Der Wielen et al. (2018) could be explained by the fact that our study included respondents of all ages, whereas theirs focused on rural older persons. Moreover, Tian et al. (2012) discovered that private health insurance positively influences inpatient length of stay, but our study revealed a negative impact of NHIS membership on hospitalisation.

Table 3. Estimates of the treatment effects of NHIS enrolment on hospitalisation days.

Effect	Estimate	Standard error	z
ATE	−5.129***	1.583	−3.240
ATT	−8.321***	1.602	−5.194
ATU	5.934***	1.006	5.899

ATE: Average Treatment Effect; ATT is the average effect of the treatment on the treated and ATU is the average effect of the treatment on the untreated; ***p < .01.

4. Conclusion

This study aimed to explore the causal influence of health insurance on hospitalisation duration. The study makes use of nationally representative data from the Ghana Socioeconomic Panel Survey to offer estimates at the national level. The study employed the ESRC model for count data to control selection bias and unobserved heterogeneity. The main finding of this study reveals a negative relationship between NHIS membership and hospitalisation, demonstrating that health insurance reduces hospitalisation duration. Specifically, on average, insured persons spend almost five fewer days in the hospital than their uninsured counterparts. The results show differences and similarities in the determinants of hospitalisation care for insured and uninsured persons. Hospitalisation duration for insured and uninsured people was jointly influenced by age, household expenditure, and self-assessed health. Gender, education, and physical inactivity significantly drive NHIS members’ hospitalisation care, as chronic illness influences the length of stay in the hospital of the uninsured.

The heterogeneous determinants of hospitalisation duration for insured and uninsured individuals have some policy implications. Given that chronic illness negatively impacts the length of stays in the hospital of the uninsured, the study suggests that efforts be made to automatically enrol people with chronic conditions into the national health insurance scheme, just as is currently done for pregnant women and the aged (70 years and above). Such support will help the effective management of chronic conditions via access to medicines, periodic monitoring, disease management programmes, and lifestyle interventions, thereby lessening the probability of acute exacerbations that will need hospitalisation. Additionally, acknowledging that education influences how long insured patients stay in hospitals, the study recommends policy reforms and funding for education at the local and national levels by government agencies, educational institutions, and civil society organisations to uphold education as a fundamental human right, encourage social inclusion and equity, and further sustainable development goals about educational equity, quality, and access. Funding for education and policy reforms may contribute to a decrease in hospitalisation care. This is so because increased education is frequently linked to improved health knowledge, awareness of preventative care, and comprehension of health risks. Highly educated people might know more about how to care for chronic illnesses, identify symptoms, and get medical help quickly. This knowledge could help with early diagnosis and treatment, which could lessen the severity of health conditions and the need for extended hospital stays.

Given the significant contributions of individuals’ enrolment in the NHIS to hospitalisation care, policymakers in Ghana should encourage more households to enrol on the NHIS. One of the better approaches for encouraging household participation in the NHIS might be raising awareness among other non-participant individuals. Policies aimed at improving health status will help reduce the length of stay in the hospital and improve people’s well-being. All studies, like ours, have their limitations. This study’s use of self-reported data on patients’ hospitalisation duration is recognised as a weakness. Since human memory is prone to error and can become distorted as time goes on, asking respondents to recall their hospitalisation information in the past 12 months may result in inaccurate responses because of forgetfulness, misremembering of events, and conflation of multiple situations.

Acknowledgements

Data was sourced from the Ghana Socioeconomic Panel Survey (GSPS). This dataset was collected through the shared effort of the University of Ghana’s Institute of Statistical, Social, and Economic Research and Yale University’s Economic Growth Centre. The authors are grateful to both the Economic Growth Centre at Yale University and the Institute of Statistical, Social, and Economic Research (ISSER) at the University of Ghana for using the Ghana Socioeconomic Panel Survey dataset.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are openly available in Harvard Dataverse at https://doi.org/10.7910/DVN/E5QP0F,referencenumberV1,UNF:6:JLtXxepgNXfzyX0ThGLDiw==[fileUNF]

Additional information

Notes on contributors

Samuel Sekyi

Samuel Sekyi graduated from the University of South Africa with a PhD in Economics. He is an Associate Professor of Applied Economics at the Department of Economics, Simon Diedong Dombo University of Business and Integrated Development Studies, Ghana. His main areas of expertise and interest are microeconometrics, health economics, development economics, and agricultural economics.

James Dickson Fiagborlo

James Dickson Fiagborlo is a Lecturer at the Department of Multidisciplinary Studies, Ho Technical University, Ghana. He obtained his PhD in Economics from the University of Cape Coast, Ghana. His research interests span areas such as applied econometrics, transportation economics, agriculture technologies, education studies, and health economics.

Gloria Essilfie

Gloria Essilfie is a Lecturer at the Department of Applied Economics, University of Cape Coast. Gloria is a graduate of the University of Cape Coast, where she obtained her BA, MPhil, and PhD in Economics. Gloria Essilfie’s academic background covers the fields of applied microeconomics and development economics.

Notes

1 Computed from the estimations in Table 2 as coefficient of age divided by -2(age squared).

Appendix A

Table A1. Pearson correlation coefficient analysis for testing the instrument’s validity.

Instrumental variable	Dependent variable	Correlation coefficient	p-value
Formal-sector work	NHIS enrolment	0.1280	.0000
Formal-sector work	Hospitalisation days	0.0063	.8540

Table A2. Logit estimation for NHIS membership and negative binomial regression estimation for hospitalisation days.

	NHIS membership	Hospitalisation days
Variable	Coefficient	Coefficient
Formal-sector work	0.816***	0.095
	(0.059)	(0.116)
Gender	−0.580***	0.472***
	(0.031)	(0.074)
Age	−0.000	0.009
	(0.001)	(0.007)
Age squared	−	−0.000*
	−	(0.000)
Education	0.248***	0.034
	(0.022)	(0.047)
Chronic illness	−0.079**	0.105
	(0.034)	(0.074)
Physical inactivity	0.590***	0.204**
	(0.048)	(0.098)
Average household expenditure	0.000***	0.001**
	(0.000)	(0.000)
Self-assessed health	−	−0.412***
	−	(0.053)
Fever	−	−0.099
	−	(0.110)
Locality (urban)	−	−0.029
	−	(0.073)
Constant	0.445***	2.620***
	(0.037)	(0.266)
Lnalpha	−	−0.180***
	−	(0.052)
Alpha	−	0.835***
	−	(0.043)
Observations	19,247	862
LR χ^2	996.37***	163.35***
Log likelihood	−12,332.381	−2,452.232
Pseudo R^2	0.039	0.0322
Model classification	63.07%	−

Standard errors in parentheses; *p < .1, **p < .05, ***p < .01.

Likelihood-ratio test of alpha = 0: chibar2(01) = 5605.11 Prob> =chibar2 = 0.000.