Poverty and health care demand in Kenya

We can develop the theoretical framework shown in Figure 1 below based on [20].

The starting point in the figure is that individuals within a household, are assumed to derive utility from the consumption of non-health goods (such as clothing) and good health. To produce good health, however, the individuals have to consume health care and other health-enhancing goods. The poverty status of the household, however, limits the kind of health care, other health-enhancing goods, and the non-health goods that the household members can consume. The individuals are assumed to choose their consumption levels of health care, other health-enhancing goods, and non-health goods in such a way that their overall utility is maximized.

We present in Figure 2, shown below, a conceptual framework for analyzing the effect of poverty status on the health care demand by an individual. The figure is constructed based on [16],[19].

According to the figure, health care and other factors determine an individual’s health status. The individual, therefore, demands health care as an input into his/her health production process. The demand for health care is influenced by predisposing factors such as age, sex, marital status, education, religion, household size, education, etc; enabling factors such as the poverty status of the individual/household; illness level; and provider characteristics such as price of care and quality of care. The poverty status is, in turn, influenced by the predisposing factors, the demand for health care (mainly due to catastrophic health care expenditures) [23]-[25], and the illness level. The arrows in the figure do not necessarily imply causality but do indicate the direction of influence.

Whenever we estimate a model based on survey data we need to, in general, worry about three main issues that have the potential to bias our parameter estimates: sample selection bias, endogeneity of some of the covariates, and unobserved heterogeneity.

In our dataset, individuals were asked whether they were sick or injured in the four weeks period immediately preceding the survey. Those who answered in the affirmative were then asked whether or not they consulted a health provider. In this case, sample selection bias will arise if the unobservable factors influencing the decision to report whether or not one consulted a health provider are correlated with the unobservable factors influencing actually consulting a health provider [26]. This problem does not, however, exist in our case since 99.95% of those who reported being sick or injured in the four weeks prior to the survey also reported whether or not they had consulted a health provider. It is only 0.05% we are not sure of, which could actually be a case of data-entry errors.

An individual’s poverty status is potentially endogenous in our model due to the potential reverse causality between the poverty status and the health care provider choice [27],[28]. There is, therefore, the risk of inconsistency of our estimated coefficients and the further risk of inability to infer causality between the poverty status and health care provider choice [28]. We can, however, use the method of Two-Stage-Residual-Inclusion (2SRI) to consistently estimate our model [29]. This technique involves two steps. In the first step, we estimate a poverty status equation and obtain the generalized residuals using the methods discussed in [30]. In the second step, we estimate the health care provider choice equation where both the poverty status variable and its generalized residuals are included in the set of explanatory variables. Poverty status will be endogenous in the health care provider choice equation if the coefficient of the generalized residuals is statistically different from zero [31].

Unobserved heterogeneity will occur in our model if there is a non-linear interaction between unobservable factors and poverty status that cause the effect of poverty status on health care provider choice to differ amongst population subjects [32]. We control for unobserved heterogeneity using the control function approach [33]. This approach generally involves including in the health care provider choice equation interactions between the generalized residuals from the poverty status equation and the poverty status variable. A coefficient of the resulting interaction term that is statistically significantly different from zero is indicative of the presence of unobserved heterogeneity in the health care provider choice equation.

Since it is only the poverty status variable in our model that is assumed to be endogenous, identification of our model requires one exclusion restriction. The exclusion restriction, also referred to as the instrumental variable, should not be correlated with the stochastic error term in the health care provider choice equation (that is, it should be valid), should be correlated with the poverty status in the health care provider choice equation (that is, it should be relevant), and should be excluded from the health care provider choice equation [34].

We use the proportion of children that are severely underweight in each district as an instrumental variable for poverty status. This choice is motivated by two main reasons. First, we do not expect the proportion of children who are severely underweight in a district to affect the demand for health care at the individual/household level. Second, as shown in the literature (see, for example, [35]), there is a high correlation between childhood undernutrition and poverty. We, therefore, expect that households found in districts with higher proportions of severely underweight children to have a higher probability of being poor and vice-versa.

Based on the conceptual framework in Figure 2, we can formulate the following health care demand function for an individual in a particular household who reports having been sick or injured in the four weeks immediately preceding the survey

Specifically, the Health Care Provider (HCP) variable is defined as follows:

while the poverty status variable is defined as

As explained earlier, the first step of the 2SRI technique requires that we estimate the poverty status equation and obtain generalized residuals from it which are then used as an additional explanatory variable in the second step. The poverty status equation is given by

To control for potential endogeneity of poverty status in Equation (1), the equation is re-estimated with the generalized residuals included as an additional independent variable. Controlling for potential unobserved heterogeneity involves further including in the equation interactions between poverty status and the generalized residuals.

Since the poverty status variable is binary, we can either use the probit model or the logit model to estimate Equation (4) [36]. We choose the probit model since according to the literature, it does not matter in general whether we use the probit or logit model as the coefficients obtained using one model can be transformed into those obtained using the other model [36],[37].

The HCP variable, on the other hand, is polychotomous. Two main models for estimating such a variable include the multinomial logit model and the multinomial probit model [36],[38]. The multinomial logit model, however, has the disadvantage of imposing the assumption of the independence of irrelevant alternatives (IIA) on our choices, which may be untenable in practice [36]. To avoid the problems posed by the IIA assumption, therefore, we use the multinomial probit model to estimate the health care provider choice models [38].

We use data from the Kenya Integrated Household Budget Survey conducted in 2005/2006 [39]. The survey collected data on various socio-economic aspects from a national representative sample of over 10,000 households [39]. Data on poverty status of the households come from a report on poverty compiled using the survey data [40]. Data on the proportion of children who are severely underweight for each district is obtained from Table 6.1 of the survey report [39].

The data used in this study were not collected by the author directly but by the Kenya National Bureau of Statistics (KNBS), a public body mandated to collect data from the Kenyan population. Most of the ethical issues that would, therefore, arise from data collected by the researcher do not arise in this case. Chapter 2 of the report accompanying the data contains detailed information on the design of the survey and the actual process of data collection [39]. It can also be inferred from the survey questionnaire that participation in the survey was voluntary for the sampled households. The survey questionnaire that was used by the Kenya National Bureau of Statistics to collect the data used in this study has been provided as an Additional file 1.

Table 2 shows the variable definitions for the variables in our models.

The descriptive statistics are shown in Table 3.

From the table, we can observe that about 47.47% of the individuals in our sample come from poor households. We can further observe from the table that the number of members in the household ranges from 1 to 29. The table also indicates that about 42.61% of the individuals reporting being sick or injured had fever or malaria while the average age of infants in our sample is 6.5634 months. The table further shows that the average distance to the nearest health facility ranges from 2.6 kilometres to 58 kilometres.

Table 4 shows the type of health care provider visited by those who were sick or injured in the last four weeks preceding the survey.

The various health care providers have also been classified in the table as either none, non-modern or modern. We consider a health care provider to be modern if one can obtain formal health care goods and services from the provider [16]. One cannot obtain such goods and services from non-modern health care providers.

In the table, we can observe that the health care providers visited range from referral hospitals to not visiting any health care provider. Health care providers classified as modern include referral hospitals, district/provincial hospitals, public dispensaries, public health centers, private dispensaries/hospitals, private clinics, missionary hospitals/dispensaries, and pharmacies/chemists. The non-modern health care providers include traditional healers, kiosks, faith healers, herbalists, and facilities classified as “other”.

The table shows that 30% of individuals in our sample who reported being sick or injured in the four weeks preceding the survey did not consult any health care provider, 7.56% consulted non-modern health care providers, while 62.48% consulted modern health care providers. A further look at the table shows, for example, that 1.35% visited a referral hospital, 0.15% visited a traditional healer, while 13.92% visited a public dispensary.

We first estimate the poverty status models. We have decided to subdivide our sample into four categories: infants (aged under 1 year), children (aged 1 to 5 years), children (aged 6 to 14 years), and adults (aged 15 years and above). The reasons for the subdivision of the sample into subsamples based on age groups are twofold. First, for some of the age groups some of the variables are measured in different units. For example, age is measured in months for infants but in years for the other age groups. Second, some of the variables do not make sense for certain age groups. For example, marital status and level of education do not make sense for infants, children aged 1 to 5 years, and children aged 6 to 14 years. In such models, therefore, we include the education level of the household head and the marital status of the household head, instead.

For each age group, we estimate a probit model. We report average marginal effects. The average marginal effects are obtained as follows: we first compute the marginal effect for the respective covariate for each observation in the subsample, and then find the arithmetic mean of these marginal effects for all the observations [36]. The results are shown in Table 5 below.

The results in the table indicate that for all age-groups, the predictors of poverty include large household sizes and longer distances to the nearest health facility. In particular, an increase in the household size by one member increases the probability of being poor among infants by 0.03, among children aged 1 to 5 years by 0.046, among children aged 6 to 14 years by 0.042, and among adults by 0.044, holding other factors constant. Further, an increase in the average distance to the nearest health facility by 1 kilometre increases the probability of being poor by 0.007 among infants, by 0.005 among children aged 1 to 5 years, by 0.008 among children aged 6 to 14 years, and by 0.007 among adults, holding all the other factors constant.

Among the children, poverty is higher for those children who live in households where the household head has no formal education compared to those who live in households where the household head has formal education, holding other factors constant. In particular for children aged 1 to 5 years, poverty is higher among households where the household head has no formal education as compared to those where the household head has formal education by 0.126, holding other factors constant. Among the children aged 6 to 14 years, poverty is higher for those households where the household head has no formal education compared to those households where the household head has formal education by 0.13, holding other factors constant.

Among the adults, poverty is higher amongst those suffering from headaches (as compared to other illnesses), the older ones, and those without formal education. For example, the results show that compared to adults with formal education, those without formal education have a higher probability of being poor by 0.121, holding other factors constant.

The results also show that among adults, those who are married are less likely to be poor compared to those who are not married, holding other factors constant. In particular, being married as opposed to not being married reduces the probability of being poor by 0.032, holding other factors constant.

We also estimate health care provider choice models for all the age groups. For each age group we estimate three models: a basic model, a model that controls for endogeneity of poverty status, and a model that controls for unobserved heterogeneity. The estimation results show that even though poverty is endogenous in health care provider choice equations, there is no unobserved heterogeneity in our models. The estimation results (average marginal effects) are shown in the Appendix. As such, the appropriate models are those that control for the endogeneity of poverty status. We show in Tables 6, 7, 8 and 9 the estimation results for the models that control for the endogeneity of poverty status. Each of the tables shows the average marginal effects for the various types of health care providers for the different age groups. Table 6 shows the results for infants, Table 7 provides the results for children aged 1 to 5 years, Table 8 shows the results for children aged 6 to 14 years, while Table 9 shows the results for adults.

The results in Table 6 show that among infants, poverty increases the probability of not visiting any health care provider when ill but reduces the probability of visiting a modern health care provider, holding other factors constant. Specifically, the results show that among infants, poverty increases the probability of not visiting any health care provider by 0.375 but reduces the probability of visiting a modern health care provider by 0.449, holding other factors constant.

From Table 7, we can observe that among children aged 1 to 5 years when other factors are held constant, poverty increases the probability of visiting non-modern providers and that of not visiting any provider, but reduces the probability of visiting a modern health care provider. In particular, for this age group, poverty increases the probability of not visiting any provider by 0.401, increases the probability of visiting a non-modern health care provider by 0.184, but reduces the probability of visiting a modern health care provider by 0.585, holding other factors constant.

Table 8 also shows that holding other factors constant, for children aged 6 to 14 years, poverty increases the probability of not visiting any health care provider by 0.224, it increases the probability of visiting a non-modern health care provider by 0.222, but reduces the probability of visiting a modern health care provider by 0.447. The results in the table also show that living in the rural areas as opposed to living in urban areas increases the probability of visiting a non-modern health care provider by 0.070 but decreases the probability of visiting a modern health care provider by 0.080, holding all other factors constant. The table also shows that an increase in the average distance to the nearest health facility by one kilometre reduces the probability of visiting a non-modern health care provider by 0.005 but increases the probability of visiting a modern health care provider by 0.005, holding other factors constant.

Table 9 results show that among adults, poverty increases the probability of not visiting any provider by 0.271, it increases the probability of visiting a non-modern health care provider by 0.156, but it reduces the probability of visiting a modern health care provider by 0.427, holding all the other factors constant. The results in the table also show that when other factors are held constant, adults living in rural areas compared to those living in urban areas have a higher probability of visiting a non-modern health care provider by 0.047 but a lower probability of visiting a modern health care provider by 0.048. We can also see from the table that adult males are less likely to visit modern health care providers compared to adult females by 0.026, holding other factors constant. The results in the table further show that among adults, an increase in the average distance to the nearest health facility by one kilometre reduces the probability of visiting a non-modern health care provider by 0.003 but increases the probability of visiting a modern health care provider by 0.002, holding other factors constant.

The association of large household sizes with increased probability of being poor may be due to the fact that larger households have larger demands in terms of the amounts of resources needed to satisfy the household’s basic food and non-food needs. This association is confirmed by other studies in the literature [41]-[44].

The association of longer distances to health facilities with poverty may be because health facilities that are farther away from where people live are less accessible to the majority of the people and this makes it difficult for people to seek modern health care when ill [45], decreasing their chances of engaging in income-generating activities. The finding is supported by other studies from the literature [46]-[48].

Not having formal education may be a predictor of poverty due to the fact that education opens up a range of income-generating opportunities, such as employment in the formal sector, which may not be available to those without formal education. The literature actually shows that the higher one’s education level is, the higher the private returns to education [49]. The association of lack of formal education and poverty is supported by studies in the literature [41],[43],[44].

The negative correlation between being married and being poor could be because marriage may increase the resources available to the household by, for example, having a spouse who earns a higher income. Studies in the literature support this correlation [43],[44].

A possible explanation for the negative relationship between poverty and the probability of visiting a modern health care provider when ill is that there are both direct and indirect costs associated with consulting modern health care providers [20]. These costs could be too substantial for individuals from poor households to bear. The result that poverty negatively affects the demand for modern health care is supported by findings from the literature in three main ways. First, the literature shows that compared to the non-poor, the accessibility of health care services by the poor is low [22],[50],[51]. Second, according to the literature, the poor tend to use non-modern health care providers such as traditional healers [52] or not seek health care at all [53],[54]. Our results actually support this finding from the literature. Third, it is also shown in the literature that the amount of money spent by households on health care services is positively related to household income [55]. There are, however, other studies in the literature that show that income is generally an unimportant determinant of health care provider choice [56].

The lower likelihood of individuals aged between 6 and 14 years, and adults in rural areas compared to those in urban areas visiting a modern health care provider when sick or injured could be due to ease of accessing modern health care facilities in urban areas compared to rural areas [57],[58]. This finding is consistent with findings from the literature where, for example, it is reported that in the case of treatment of childhood malaria, caretakers of the children in the rural areas are more likely to resort to self-treatment while their urban counterparts are more likely to take the children for treatment in private or government health care facilities [58]. It is also reported in the literature that in the case of treatment of acute illnesses, more rural residents are more likely to visit faith healers than urban residents [59].

The result that male adults are less likely to seek health care from modern health care providers is, however, not supported by some studies in the literature which show that women are less likely to seek care compared to men [60] and that women and men are equally likely to use medical services [57].

The positive effect of distance to nearest health facility on the demand for modern health care is contrary to expectation. This is because, distance in our health care provider choice models is a proxy for price of modern health care. It actually proxies the indirect cost of modern health care. The positive effect on the probability of visiting a modern health care provider, however, provides evidence of bypassing of nearer facilities to seek health care from farther away facilities [57]. Patients bypass nearer facilities due to mainly price and quality concerns [61],[62].