Factors for healthcare utilization and effect of mutual health insurance on healthcare utilization in rural communities of South Achefer Woreda, North West, Ethiopia

We conducted a cross-sectional study from February to March, 2016 in rural community of South Achefer, Northwest Ethiopia. In 20 kebeles, a total of 36,204 households were identified; of which 12,612(35%) households were enrolled to mutual health insurance [16].

House to house, community-based survey was conducted in South Achefer district. The objectives of this study were to determine healthcare utilization among insured and uninsured households and to identify factors linked with healthcare utilization. The main methodological challenges could be selection bias and variable omission because of unobserved heterogeneity during matching. Heterogeneity is a variation across individual units of observations, and the variation can’t be observed. The term heterogeneity refers to differences across the units being studied. For example, in an insurance model, members with non-members are assumed to differ from one another is said to have heterogeneous agents. We might have an omitted variable (as is often with cases of causal inference and selection bias) that we say we have endogeneity caused by unobserved heterogeneity. The effect of unobserved heterogeneity is endogeneity that is correlation between explanatory variables and error term. Unobserved heterogeneity is one instance in where correlation between observables and unobservable may be expected. In addition to the observed variables under study, there are other relevant variables which are unobserved, but correlated with the observed variables. Mutual health insurance scheme was voluntary based which could lead people to self-select themselves in to a particular group (either being insured or uninsured) [7, 14, 15]. Therefore, we applied a technique called propensity score matching to account for potential differences in background characteristics between insured and uninsured so that difference in the outcome of interest could be attributed to the intervention itself.

The study was designed to have 80% statistical power in order to get adequate sample size for justifiable statistical significant given efficient resource utilization. The level of significance was taken 5%. The ratio of insured to uninsured was 1:1. Assuming the proportion of health service utilization was 35% for insured households and 20% for uninsured household [17]. The sample size was estimated to be 296 for insured and 296 for uninsured households. The non-response rate of 10% was taken. The design effect was determined to be two for having two stages in selecting study subjects, the resulting sample size was 652 (326 insured and 326 uninsured households). The sample size was calculated for different objectives and the largest sample estimate was taken.

In this study, for large clusters of population, multi-stage sampling procedure was appropriate. Multi-stage sampling could divide large clusters of population into smaller clusters in several stages in order to make primary data collection manageable and cost wise. Of 20 Kebeles, six Kebeles were randomly selected from the South Achefer district. A total of 652 households were identified from six Kebeles. Proportional allocation of households per kebele was employed. Sampled households (insured and uninsured) were selected systematically from each Kebele.

Ten trained diploma graduate data collectors conducted face to face interviews with household heads. The data collection tools were structured questionnaires that were prepared in English, translated into Amharic and back translated to check its consistency. Pre-test was conducted for reliability and ease of understanding. We measured all variables at the household level. Enrollment to mutual health insurance was at the household level. Health services utilization was the dependent variable. Independent variables were age, sex, education, family size, distance from health institutions, insurance membership, wealth index, first choice place for treatment, chronic illness and health seeking behavior.

Utilization of healthcare was defined as health facilities visits by members of households at least once in the last 6 months for health services (diagnostic or treatment). Perceived health status was respondent’s report about their health status that was assigned numerical values according to the following scale: very good = 5, good = 4 medium =3 poor =2 and very poor = 1. Perceived quality of care was the extent of respondent’s view on the quality of healthcare delivery; it was measured by Likert scale questions. Chronic illness is a disease condition that lasts more than three months. Household wealth index was measured by using principal component analysis and categorized as poor, medium and rich. Households indexed in the mustered book of mutual health insurances schemes were recruited as insured while households which were not indexed to mutual health insurance schemes were recruited as uninsured.

Data were entered into Epi-info and analysis was performed by using STATA. Frequencies and proportions were used to describe the respondents in relation to the studied variables. To identify factors associated with healthcare utilization, logistic regression models were used. A bi-variable logistic regression model was fitted for each explanatory variable to identify the existence of associations among candidate and outcome variables. Multivariable logistic regression analysis (backward stepwise) was used to identify independent predictors for healthcare utilizations. All variables whose crude odds ratios greater than 0.2 in the bi-variable logistic regression model were fitted in to multivariable logistic regression model. An odds ratio with 95% confidence interval was used to measure the strength of association and identify statistically significant results. Propensity score matching was used to explain possible differences in the baseline variables between enrolled and un-enrolled households. In this data set a good-quality matching was observed since mean absolute bias was 4% which was < 5%. Therefore, the variation in healthcare utilization was due to average impact of mutual health insurance itself.

Ethical approval from the University of Gondar Ethics committee was obtained. In addition, ethical endorsement from Amhara Regional Health Bureau Research Core Process Unit was approved. Written permission to conduct the study was obtained from each Kebele administration involved in the study. The data collectors took written informed consent from household heads.

A total of 594 households participated in the study with a response rate of 91.1%. The average age of the respondents was 47.5 years (±14.6SD). Table 1 presents socio-demographic characteristics of study participants, enabling and needs factors. Among study subjects, more female household heads (47%) were utilizing healthcare services than male household heads (36%). Among respondents who were not able to write and read, higher percentage (68%) were not utilizing health care service than respondents who were able to write and read (51%). A considerable percentage of household heads who traveled less than five kilometers to reach the nearest health institutions (43%) were utilizing health services than household heads who traveled five kilometers and above (37%). Higher proportion of respondents with chronic illness in the households (52%) were utilized the health services than households without chronically ill respondents (37.4%) (Table 1).

Respondents, whose educational level was primary and above, were more likely to utilize health service (AOR = 1.84; 95% CI: 1.14, 2.98) than respondents who did not write and read. Households with presence of a chronic patient were more likely to utilize health services (AOR = 1.86; 95% CI: 1.13, 3.06) than its counter parts. Households, whose wealth index was categorized as rich, were more likely to utilize health service (AOR = 2.1; 95%CI: 1.29, 3.43) than households, whose wealth index was categorized as poor. Respondents whose first choice was health institution for treatment at the point of illness were more likely to utilize health service (AOR = 6.33; 95% CI: 2.97, 13.51) than household heads whose first choice was traditional healer. Respondents with poor perceived health status were less likely to utilize health service (AOR = 0.33; 95% CI: 0.18, 0.61) than respondents with perceived health status of very good. Households who enrolled to mutual health insurance scheme were more likely to utilize health service (AOR = 2.16; 95% CI: 1.45, 3.23) compared to households who did not enrolled to mutual health insurance scheme (Table 2).

Before matching was computed, healthcare utilization is significantly higher in households which were enrolled to mutual health insurance (50.5%) than households which were not enrolled to mutual health insurance (29.3%) (t = 5.4, p-value< 0.001). To minimize self-selection bias of insurance uptake, nearest neighborhood matching was done between who were enrolled (treated) to mutual health insurance (n = 297) and who were not enrolled (control) to mutual health insurance (n = 218). The average treatment effect on the treated group was 0.252 (t = 4.94, 95% CI: 0.14–0.35). Being a member to mutual health insurance contributed approximately 25.2% point increase healthcare utilization. The average incremental effect on healthcare utilization was because of enrollment to mutual health insurance scheme (Table 3).

As we conducted cross-sectional study, factors do not establish temporal relationship. Selection bias due to matching may result in variable omission; unobserved heterogeneity could be exhibited. An important source of potential bias did exist as participation. In general, mutual health insurance scheme was voluntary which could lead people to self-select themselves in to a particular group (either being insured or uninsured). This results in endogeneity.