Determinants of healthcare utilisation and out-of-pocket payments in the context of free public primary healthcare in Zambia

Background: Access to appropriate and affordable healthcare is needed to achieve better health outcomes in Africa. However, access to healthcare remains low, especially among the poor. In Zambia, poor access exists despite the policy by the government to remove user fees in all primary healthcare facilities in the public sector. The paper has two main objectives: (i) to examine the factors associated with healthcare choices among sick people, and (ii) to assess the determinants of the magnitude of out-of-pocket (OOP) payments related to a visit to a health provider. Methods: This paper employs a multilevel multinomial logistic regression to model the determinants of an individual’s choice of healthcare options following an illness. Further, the study analyses the drivers of the magnitude of OOP expenditure related to a visit to a health provider using a two-part generalised linear model. The analysis is based on a nationally representative healthcare utilisation and expenditure survey that was conducted in 2014. Results: Household per capita consumption expenditure is significantly associated with increased odds of seeking formal care (odds ratio [OR] = 1.12, P =.000). Living in a household in which the head has a higher level of education is associated with increased odds of seeking formal healthcare (OR = 1.54, P =.000) and (OR = 1.55, P =.01), for secondary and tertiary education, respectively. Rural residence is associated with reduced odds of seeking formal care (OR = 0.706, P =.002). The magnitude of OOP expenditure during a visit is significantly dependent on household economic well-being, distance from a health facility, among other factors. A 10% increase in per capita consumption expenditure was associated with a 0.2% increase in OOP health expenditure while every kilometre travelled was associated with a K0.51 increase in OOP health expenditure. Conclusion: Despite the removal of user fees on public primary healthcare in Zambia, access to healthcare is highly dependent on an individual’s socio-economic status, illness type and region of residence. These findings also suggest that the benefits of free public healthcare may not reach the poorest proportionately, which raise implications for increasing access in Zambia and other countries in sub-Saharan Africa.

The statistical analyses in this study are based on a cross-sectional dataset from the Zambian Household Health Expenditure and Utilisation Survey (ZHHEUS) conducted in 2014 by the Central Statistical Office. The ZHHEUS sampled a cross-section of households in all 10 provinces of Zambia using a sampling design, which was aimed at achieving national representativeness. The Central Statistical Office, with support from the Ministry of Health, Lusaka, Zambia and the University of Zambia, Lusaka, Zambia conducted the survey, yielding a total of about 12 000 households, including some 59 500 individuals, in all 10 provinces of Zambia. A two-stage stratified cluster sample design was used. In the first stage, standard enumeration areas were selected within each stratum using the probability-proportional-to-estimated-size procedure to select a total sample of 599 clusters or primary sampling units (psu) from each of Zambia’s 10 provinces of which 250 were from urban areas and the rest (ie, 349 psu’s) from rural areas. A full census (or listing) of all households in each psu was conducted prior to sampling of sample households. In the second stage, a fixed proportion of 20 households were selected from each psu using a systematic random sampling procedure. Thus, the sample size was powered to be representative at the cluster, provincial and national levels. The survey response rate was 99.4%. At each sampled household, all members were enumerated for all modules except for the maternal health section, which was restricted to female members aged between 12 and 49 years. The survey included modules on health status (self-rated health status and self-reported illness experience); illness experiences associated healthcare utilisation (visits, admission, type of providers sought, health expenditure); and a quality of care assessment. Specifically, individuals were asked if they had experienced any illness or injury in the 4 weeks preceding the survey, or if they had been admitted to a health facility in the 6 months preceding the survey. OOP health expenditure included charges for consultation, drugs, medical investigations, and other fees incurred at facilities, as well as transportation costs and other costs related to a visit to a health facility. The empirical model applied in this paper is based on the Grossman model of demand for health and healthcare, which describes how individuals make choices regarding healthcare utilization.21,22 In the Grossman theoretical framework, utilisation of healthcare is optimally chosen as an attempt to attain or maintain optimal health. When individuals fall sick, they demand healthcare in order to restore their health capital. An important contribution of the Grossman model is in providing a theoretical framework for testing the relationship between characteristics of an individual and his or her health behaviour. Since Grossman, empirical studies have examined the marginal effects of characteristics such as income, age, education, health insurance, health status, distance to a health provider, and so on, on health decisions and healthcare consumption.23-25 The Grossman model postulates that apart from expanding an individual’s ability to pay, higher wages lead to a substitution of medical consumption for time or resources invested in health promotion or prevention. In other words, a higher wage induces an individual to dedicate less time to health promotion or prevention and more time to earning a wage.21,22 In contexts where healthcare utilisation is dependent on OOP payments, income works through price to relax the consumer’s budget constraint. Hence, income is expected to increase the likelihood of seeking healthcare as well as the magnitude of health spending. In this survey, we used household consumption expenditure which is widely considered to be a more reliable measure of household wealth than self-reported income. It is less sensitive to short-term fluctuations. Consumption expenditure also captures the value of home production, which is important to appropriately measure wealth or economic capacity in many rural settings.26 Although Grossman had predicted a negative relationship between education and demand for healthcare on account that education increases an individual’s health prevention ability, through health knowledge, healthy lifestyle, processing health information, and so on, which should imply less need for medical care consumption, empirical studies have shown a positive relationship between education and healthcare utilisation.27,28 Empirical studies hypothesise that more years of schooling make individuals choose better healthcare options which include the ability to seek effective medical care following an illness experience. With regard to age, theory predicts that with increasing age, more healthcare is needed to offset the effect of increasing depreciation of health capital. However, studies have suggested a non-linear relationship as at some point in age, the marginal cost of investing in renewing health exceeds its marginal benefits, at which point this relationship becomes negative.29,30 Also, the literature shows the demand for healthcare to be higher among children under the age of five years and among the elderly.31 Empirical extensions of Grossman’s work have included other factors such as gender and region of residence. In this paper, the set of explanatory includes gender, age, household per capita consumption expenditure, highest level of education attained by the head of the household, employment status of the household head, residential location of the household, type of illness reported by an individual and the type of healthcare provider visited. Household consumption expenditure was used as a proxy for household income or wealth. Table 1 provides the full list of variables used and their definitions. Abbreviations: OOP, out-of-pocket; ZMK, Zambian Kwacha. Empirically, our approach leads us to estimate two models. First, we specify a multinomial logistic regression model of an individual’s decision regarding healthcare utilisation. In the second part of the analysis, we analyse determinants of the magnitude of OOP healthcare expenditure conditional on visiting a health provider using a two-part estimation procedure. The logistic regression to estimate the probability of an individual incurring a positive expenditure and the generalised linear model to analyse the determinants of the magnitude of OOP healthcare expenditure. Given that the response variable is at three mutually exclusive levels (sought formal care, performed self-medication, or did nothing), and our intention to model effects (on the response variable) that operate at the community level, we fit a multilevel multinomial logistic regression model. In a multinomial logistic model, the probability of an individual i living in primary sampling unit j, choosing care option p is given by πijp = Pr (care option = p). Thus, p = 1,..,q (q = 3). One of the response categories is taken as the reference category. In this case, the “did nothing” was the “reference” response category. We estimated a simultaneous set of q-1 logistic regressions for the other two “care option” categories, contrasting each category with the reference category. Thus, a separate intercept and slope parameter was estimated for each of the categories, as indicated by the p superscripts. The multilevel multinomial logistic regression model is specified using the following logit link: Probability,πijp, is a function of a vector of covariates denoted by X, and the specified community level (i.e. psu) random effects (ujp). The term represents random variation in the likelihood of doing nothing relative to formal care, or doing nothing compared to self-medication, at psu level. The parameter β(p) represents the fixed part of the model which is interpreted as the change in the odds of being in category p relative to the “reference” category associated with a 1‐unit increase in the explanatory variable denoted by X, if X is continuous. In the case of discrete explanatory variables, β(p) represents the change in the odds associated with being in one category (eg, living in a rural area) relative to being in the reference category (being in an urban area). The model assumes that uj ~ N(0,σu2). Further, the residual error term denoted εpij is random error at individual level which is assumed to have a logistic distribution with mean zero and variance π23 . In this hierarchical structure, we take account of variations in choice of care option that operate not just at the individual or household levels but also at the community level. Community level or neighbourhood in this case is defined by the survey clusters called primary sampling unit (psu). It is plausible that observed variations in healthcare choices might be partly explained by community level influences on health-related behaviour.32-34 In this approach, unobserved variations in healthcare choices are captured as random effects operating at the community level through the parameters uj.32 The model was estimated using the maximum likelihood method in the generalised linear latent mixed model (gllamm) framework. The gllamm procedure provides an estimation algorithm that is more robust than either ordinary least squares (OLS) or traditional maximum likelihood estimators.32 In this part of our analysis, we model the determinants of the magnitude of outpatient OOP healthcare expenditure. In the survey, only individuals who reported a visit to a health provider were asked about expenditure incurred during a visit. Those who did self-medication were not asked to state how much they may have spent. In estimating the health expenditure model, we considered a number of methodological challenges commonly reported in the literature.35,36 Specifically, the distribution of OOP expenditure shows a high density at zero and a right-skewed continuous distribution of positive amounts. These findings are because, as stated earlier, user fees at all public and mission primary level healthcare facilities have been abolished for primary health services. For example, if an individual did not incur any transportation (ie, if they may have walked to the facility) or other health visit related expenses, their total health expenditure for the visit would be zero. Thus, the zeros in the data are real zeros, and not due to censoring (censoring typically referring to reporting a zero OOP expenditure simply because an individual did not fall sick or did not seek care) but akin to a semi-continuous dependent variable problem.37 In such a case, applying the standard OLS estimation procedure would result in biased and inefficient estimates.35 Based on guidance from the literature,35,36,38,39 we applied a two-part estimation procedure. The first part of this model constructs a logistic regression to estimate the probability of an individual incurring a positive expenditure among, which is expressed as follows: where OOPi is the level of OOP expenditure on an outpatient visit by individual i, x is a vector of covariates (as defined earlier, plus distance travelled to a facility, in kilometres), β denotes coefficients of the corresponding estimates, εi is the stochastic error term. The second part of the model predicts the magnitude of OOP expenditure for a visit, conditional on expenditure being positive, using the set of explanatory variables identified above. In this part of the model, we estimate the expected OOP expenditure given the same set of explanatory variables as defined earlier (with the inclusion of distance travelled), denoted by x. Thus, E(OOPx) = xβ. In estimating the second part of the model, we considered the generalised linear model (glm) and OLS estimators, taking into account the trade-offs in bias and efficiency each estimator brings. Given a kurtosis coefficient on the log-scale residuals of the glm estimator of 3.18 and the overall superiority over OLS, we chose the glm estimator.40 Specification of the glm framework requires the analyst to choose a link function that models how the dependent variable is connected to the explanatory variables, and a distribution function that models the relationship between the mean and variance of the expectation of the dependent variable. We tested alternative link-distribution combinations appropriate for the type of data at hand. The best fitting model was the glm with a gamma distribution and a log link function. The gamma-log glm yielded the lowest value of the Akaike information criterion (AIC) (2.9) and the deviance statistics (7483.7) compared to other glm specifications. It is also theoretically a more appropriate fit for our data.41 To improve on model efficiency, robust standard errors were clustered at the primary sampling unit (psu) level. All computations were done using the tpm routine in Stata 13.

N/A

• masiye, felix
• kaonga, oliver