In the literature, two main models explain households’ choice of cooking fuel: the energy ladder and fuel stacking theory. The energy ladder model postulates that as households’ income increases, they move away from more costly, more polluting, and less efficient technologies to more costly, less polluting, and more efficient technologies (Muller and Yan, 2018). The model has been used by scholars such as Hiemstra-van der Horst and Hovoka (2008) and Treiber (2012) to explain household fuel-switching patterns. The model has, however, been criticized by Hiemstra-van der Horst and Hovorka (2008) and Jebaraj and Iniyan (2006) for focusing only on income to the detriment of social and cultural factors on fuel choice, and the consumer rationality assumption of linear path movement from one fuel source to another.

Contradicting the energy ladder model, the fuel stacking theory argues that households adopt a multiple fuel use approach in which modern fuels are added to traditional fuels and not completely removed as households’ income increases (Martins, 2005; Treiber, 2012). In addition toincome, the theory perceived factors such as fuel accessibility and availability, household cooking practices, and health impact as the main drivers influencing households’ fuel-switching decisions previously neglected by the energy ladder model (Hosier and Dowd 1987). Multiple fuel use is practiced by many households in developing countries by climbing up and down the energy ladder instead of the traditional linear fuel switching (Leach, 1992; Martins, 2005).

Models on the impact of households’ cooking fuel choice on health and consequently healthcare expenditure have centered on the environmental health pathway developed by Smith and Pillarisetti (2017). The concept posits that indoor air pollution starts with sources of pollution, moves to environmental levels, then to human exposures and doses within the body, and finally to health impacts. It provides a better understanding of pollution risks associated with using household cooking fuel, especially biomass fuels.

Empirically, what determines household healthcare expenditure in the developing world has been well researched with many diverse results. Using the 2010 household income and expenditure survey and the ordinary least square (OLS) technique, Molla et al. (2017) investigated the determinants of out-of-pocket healthcare expenditure in Bangladesh. The authors concluded that urban households spend more on healthcare than rural dwellers and those factors such as income, ill-health, and household size were the main predictors of healthcare expenditure. Similar research carried out in China by You and Kobayashi (2011) using the 2004 China Health and Nutrition Survey data and the Heckman selection model examined the determinants of out-of-pocket health expenditure. They concluded that age, educations, and income of households were positive predictors of household health expenditure while urban households compared to rural spend more on health.

Malik and Syed (2012) explored factors influencing healthcare expenditure in Pakistan using the 2004/5 Pakistan Standard of Living Measurement Survey dataset and the OLS and established that non-food household expenditure and household characteristics were significant determinants of household healthcare expenditure. In general, factors such as the presence of illness, income, health insurance, and residence location, among others are prominently highlighted in many studies related to determinants of healthcare expenditure (Akanda et al., 2011; Molla et al., 2017).

Many studies have also examined the impact of household cooking fuel choices on health and the environment in the developing world. A recent study by Khan and Lohano (2018) found that children in households using modern fuels are less likely to have symptoms of respiratory infection than those in households using traditional fuels using the 2012/2013 Pakistan Demographic and Health Survey and the logistic model. A similar study by Acharya et al. (2015) found that the use of solid fuel in kitchens is a risk factor for acute respiratory tract infection among under-five children in Nepal using the 2011 Nepal Demographic and Health Survey.

Investigating the impact of residential fuel combustion on health expenditures in 44 Sub-Saharan African countries, Badamassi et al. (2017) employed the general method of moments (GMM) technique. Their results showed that residential fuel combustion was significantly correlated with higher health expenditures. The authors proposed health policies that boost households’ access and use of modern fuels and improved cookstoves use in SSA. Capuno et al. (2018) in their study revealed that the use of clean fuel can lower the incidence of severe coughing in young children by 2.4 percentage points in the Philippines using the propensity score matching method and the 2013 National Demographic and Health Survey.

Using baseline data from a randomized controlled trial of an improved household energy initiative and the logit model in Rwanda, Das et al. (2018) concluded that children are more likely to experience symptoms of respiratory infection, illness with cough, and difficulty in breathing from indoor pollution, while evidence from Khan et al. (2017) in Bangladesh revealed that indoor use of solid fuel increases the risk of acute respiratory infection, pregnancy complication, cesarean delivery, and low birth weight. Baumgartner et al. (2011) established that exposure to indoor biomass combustion is associated with an increase in blood pressure among rural Chinese women using a mixed-effects log-linear regression model. In a related study, Ofori et al. (2018) examined the link between household biomass fuel use and blood pressure among women in southern Nigeria using survey data from 389 women and established a significant link between household biomass fuel use and high blood pressure.

The work of Duflo et al. (2008) in India observed a significant association between the use of traditional fuels/stoves and respiratory illness and recommended an increase in ventilation and subsidizing cleaner fuel technologies, while Boy et al. (2002) examined the relationship between exposure to smoke during pregnancy and birth weight in Guatemala and established a positive relation. Other studies have also established specific health effects of biomass use including pneumonia in children (Bautista et al., 2009; Kurmi et al., 2010), tuberculosis (Lin et al., 2007), and age-related cataracts (Zodpey and Ughade, 1999). In gist, the literature points to the negative impact of household cooking fuel on health outcomes and by implication on household healthcare. However, very little research has been carried out on the dynamics of fuel choices and the impact on healthcare expenditure, justifying the need for this study.

Following Parker and Wong (1997), we assume that households derive utility ( U i ) primarily from members’ health and other consumptions and is expressed as: U i = u ( H i , C i ) , (1) i = 1, 2, 3,…..,n,where H i is a member’s health and C i is other consumption. To consume health, households produce it by combining inputs of production such as healthcare services ( H S ) and members’ time (   T ) as follows: H i = h ( H S ,   T ) . (2)

Substituting Eq. 2 into (1) gives (3) that shows household health utility function: U i = u [ h ( H S , T ) , C i ] . (3)

We assume that households face monetary constraints to produce health for consumption. Following Parker and Wong (1997), we assume that healthcare service utilization is a demand derived from the demand for health. Therefore, healthcare demand could be formulated through an aggregate utility function for all members of the household. The maximization problem for the household is therefore expressed as: max U i   [ h   ( H S , T ) , C i ]                           subject   to                           P H S H S + P C C + T ≤ Y , (4) where P H S and P C are the prices of healthcare services and other consumption, respectively, and Y is household income. Following Karimu (2015), the price of other consumption and household members’ time is normalized to one. The solution to the optimization problem gives the household demand for health ( Q ) as: Q i = q ( P H S , Y ) . (5)

Eq. 5 is then augmented with a vector of household main cooking and social and demographic variables ( Z ) to obtain the household’s derived demand for health as: Q i = q ( P H S , Y   ,     Z ,   ) . (6)

The Tobit model deals with censored regression, and therefore households with zero healthcare expenditure are censored and used for the analysis. The functional form for the Tobit model is given as: Y ∗ = β X i + ε i ,                               i = 1,2 , … , n , (7) where Y ∗ is the latent (unobservable) variable representing the dependent variable, X i is the vector of independent variables, β ' s are the parameters, and ε i is the normally distributed random error term. The observable variable Y i (healthcare expenditure) is defined as: Y i = { Y ∗                   i f   Y ∗ > 0 , 0                   i f   Y ∗   ≤ 0.

The conditional expectation of the dependent variable given the independent variables is presented in Eq. 8: E ( Y i   / X 1 i = x 1 , X 2 i = x 2 , … X n i = x n ) =   β 0 +   β 1 X 1 +   β 2 X 2 + …   +   β n X n . (8)

Eq. 8 can be written in a linear form as Eq. 9: L o g H E i =   β 0 +   β i X i +   ε i , (9) where L o g H E i represents the logarithm of households’ healthcare expenditure and β i is the set of coefficients to be estimated. Independent variables considered include household main cooking fuel use, household head’s characteristics, and household socioeconomic and demographic characteristics. Details of the description and summary statistics of the variables-mean, standard deviation, and minimum and maximum values are provided in Table 1.

The difference between the maximum and minimum values of the variables helps to determine the spread. The bigger the gap of a variable, the larger the standard deviation of the said variable.

The study uses two nationwide cross-sectional survey data—the Ghana Living Standards Survey VI and VII (GLSS VI and VII) conducted by the GSS in 2012/13 and 2016/2017, respectively. Out of the 16,000 and 14,000 households surveyed in 2016/17 and 2012/13, about 10,797 and 8,769 households representing 67.5 and 64.6 percent provided information on the three main cooking fuels, respectively. The analysis is therefore made using these households. The data contain detailed information on households’ demographic and socioeconomic characteristics as well as other key variables like migration, employment, remittances, housing, fuel use, and household expenditure, among others.

Two main estimation techniques have been used to study the determinants of healthcare expenditure: Tobit estimation techniques used by Mugisha et al. (2002) and the OLS used by Molla et al. (2017) and Malik and Syed (2012). In this study, we use the Tobit regression model and the OLS model for robustness check. The reason is that the OLS is usually used for data with less zero observations (Yeboah, 2018). However, in the GLSS data, there is some zero healthcare expenditure. Ignoring such zeros biases the results and destroys the linearity assumption making the use of OLS not the best (Jelani and Tan, 2012). The Tobit regression model, however, addresses such issues and the problem of heteroscedasticity. We also employed the logarithmic transformation of the household healthcare expenditure to reduce heteroscedasticity. This generated a problem for using the log of zero observations which does not exist, as some households had zero healthcare expenditure. To address this problem, a value of one is assigned in place of zero household healthcare expenditures so that after log transformation they could remain in the dataset following You and Kobayashi (2011). For household head’s years of education, one is added to each observation so that after log transformation zero years of education observations could remain in the dataset.

Price is considered a vital variable for estimating expenditure functions. However, Deaton (1987) and Deaton et al. (1989) argue that the price variable is mostly considered unobservable in most empirical studies on expenditure functions. The available data lacked price information and quantities of health services purchased and therefore considered unobserved. Parker and Wong (1997) argue that in such a case, one can approximate the variation in prices using health insurance coverage of households and regional locations. We, therefore, include health insurance coverage and region of residence of households in our model to serve as an indicator of regional variation in prices to reflect the realities in Ghana. Sawdust, animal waste, kerosene, electricity, and other fuels were not included in the analysis because they were used by only 1.5 and 1.6% of households in 2012/13 and 2016/17 for cooking, respectively. We use the maximum likelihood estimation and STATA 15.0 econometric software for the estimations.

Results of the Tobit estimation are presented in Table 2. In 2012/13, relative to households using wood as cooking fuel, households using charcoal and LPG are 54.40 and 115.09 percentage points less likely to spend on healthcare services, respectively, at a 1 percent significant level. However, in 2016/17, relative to households using wood as their main cooking fuel, those using charcoal and LPG are 28.15 and 103.25 percentage points less likely to spend on healthcare services at 10 and 1 percent significant levels, respectively. The reason could be that emissions from the combustion of various cooking fuels have varying effects on the health of users (Singh et al., 2016). Whiles modern fuels like LPG generate less harmful emissions, traditional fuels such as wood and charcoal generate more harmful emissions. (Badamassi et al., 2017). Hence, households using LPG and charcoal are less exposed to harmful emissions than those using wood. In Ghana, many households depend on wood as their main cooking fuel, especially in rural areas (Mensah and Adu, 2015). However, the use of wood as cooking fuel has been found to have adverse effects on human health (WHO, 2016; Badamassi et al., 2017) and hence driving up healthcare expenditure as the results indicate.

The differences in the marginal effects of households using LPG and charcoal are 12.65 and 26.25 percentage points, respectively, less likely to spend on healthcare over the 5 years using wood as the reference point. This suggests the impact has marginally reduced over the 5 years for households using LPG and charcoal. This could be attributed to the enhanced programs on the use of cleaner fuels such as the Rural LPG Promotion Program and the Cylinder Recirculation Model and crudely suggests that these government policies are having a positive impact.

With regards to household head characteristics, almost all the variables were statistically significant except the sex of the household head in both 2012/13 and 2016/17. In 2012/13, a percentage increase in the age of the household head results in a 0.21 percent increase in per capita household healthcare expenditure. This is consistent with the life cycle hypothesis that older household heads can afford more healthcare than the younger ones who face financial constraints and consistent with the study of Van Minh et al. (2013) and You and Kobayashi (2011). However, in 2016/17, a percentage increase in the age of the household head reduces per capita household healthcare expenditure by 0.46 percent and supports the work of Rous and Hotchkiss (2003). An explanation for such a change in the sign of the age of household heads over the 5 years could be the increase in the National Health Insurance Scheme (NHIS) enrollment from 65.31 percent in 2012/13 to 73.60 percent in 2016/17 (Aryeetey et al., 2016) with higher enrollment for older people due to the premium exemption policy under the scheme for the aged that has significantly helped them (Duku et al., 2015).

In 2012/13 and 2016/17, a percentage increase in households’ head years of education resulted in a 0.36 and 0.41 percentage decrease in per capita household healthcare expenditure, respectively, suggesting that household heads with higher education are less likely to spend on healthcare in both years. An explanation could be that household heads with higher education are more aware of the negative effects of health which pushes them to focus on preventive than curative than those with low education. The results support the study of Rous and Hotchkiss (2003) but contradict that of Malik and Syed (2012).

In 2012/13, relative to households headed by persons with reported illness, those without reported illness are 72.46 percentage points less likely to spend on healthcare services. However, in 2016/17, those without reported illness are 116.07 percentage points less likely to spend on healthcare than those with reported illness. This is consistent with studies by Brown et al. (2014) and Molla et al. (2017). The difference in the coefficients could be associated with the success of NHIS in improving the health status of households in Ghana over the 5 years (Aryeetey et al., 2016).

Household heads covered by health insurance has a negative and significant (at 10 percent) effect on per capita household healthcare expenditure. Specifically, in 2012/13, relative to household heads not covered by health insurance, those covered by health insurance are 12.10 percentage points less likely to spend on healthcare services. This finding confirms the financial protection role of health insurance coverage and is in line with Van Minh et al. (2013) but contradicts the results by You and Kobayashi (2011). However, in 2016/17, household heads covered by health insurance had a positive influence on per capita healthcare expenditure but was not statistically significant. The reason for the variation in the results between 2012/13 and 2016/17 may be attributed to the few challenges confronting the NHIS during the period that made many households resort to the cash and carry system (Alhassan et al., 2016).

Expectedly, in 2012/13 and 2016/17, a percentage increase in household expenditure resulted in households being 1.59 and 1.54 percentage points, respectively, more likely to spend on healthcare suggesting that healthcare is a normal good. This result is in line with studies by Brown et al. (2014), Molla et al. (2017), Parker and Wong (1997), and You and Kobayashi (2011). The changes in the coefficients from 1.59 percent in 2012/13 to 1.54 percent in 2016/17 could be attributed to the National Health Insurance Scheme (NHIS) covering part of the health cost of households (Aryeetey et al., 2016).

In 2012/13 and 2016/1, a percentage increase in household size resulted in the household being 0.84 and 0.48 percentage points less likely to spend on healthcare services, respectively, suggesting that larger households are generally associated with poorer households that do not have enough income to afford healthcare service (Brown et al., 2014) and hence less per capita household healthcare expenditure. This result confirms the study of Brown et al. (2014) and Van Minh et al. (2013) but contradicts that of Molla et al. (2017) and Rous and Hotchkiss (2003).

With regards to location, in 2012/13, urban households were 44.38 percentage points less likely to spend on healthcare than their rural counterparts. This is in line with studies by Van Minh et al. (2013) but contradicts that of Malik and Syed (2012) and Molla et al. (2017). Urban households may use more LPG as cooking fuel than their rural counterparts who may use charcoal or wood (Mensah and Adu, 2015) and therefore get sick and spend more. However, in 2016/17, a negative influence of urban households on per capita household healthcare expenditure was observed though not statistically significant.

In 2012/13, relative to households in the Western region, households in Ashanti, Greater Accra, Central, Volta, Northern and Brong Ahafo regions are 69.27, 27.23, 29.85, 59.98, 66.91, and 30.58 percentage points, respectively, more likely to spend on healthcare services. Households in the Eastern and Upper West regions are, however, 4.89 and 77.52 percentage points less likely to spend on healthcare services than those in the Western region. More need to be carried out to explain the less likelihood of households in the Eastern region spending more on healthcare than those in the Western region. However, the Upper West region could be attributed to the low level of income as the region is one of the poorest regions in Ghana (GSS, 2019). However, in 2016/17, relative to households in the Western region, those in Ashanti, Greater Accra, Central, Eastern, Volta, Northern, and Upper East and Upper West regions are 102.01, 109.93, 131.24, 211.67, 261.77, 197.47, 146.33, and 83.62 percentage points, respectively, more likely to spend on healthcare services. The increase in the magnitude of the coefficient in 2017/18 compared to 2012/13 in many of the regions suggests that households in those regions pay more than those in the Western region over the 5-year period and could be attributed to additional payments in the form of consultation and medicines by many households over the 5-year period (Okoroh et al., 2018).

Results of the OLS estimation presented in Supplementary Appendix S2 are quite similar to those of the Tobit estimations confirming the robustness of our estimation, though the magnitudes of the effects are a little higher for the Tobit than the OLS.