Coping with the economic burden of Diabetes, TB and co-prevalence: evidence from Bishkek, Kyrgyzstan

This study uses data collected from a cross-sectional survey administered in 2010 as part of a larger study combining quantitative patient information and qualitative caregiver information to understand and improve the management of Diabetes and TB in Bishkek, Kyrgyzstan. Patient data was collected using an adapted questionnaire originally used to collect data on TB care seeking in Cape Town, South Africa [37]. The questionnaire was translated from English to Russian and then back-translated to ensure accuracy and coherence. The english version of the questionnaire can be found in Additional file 1. Exit interviews were conducted with 309 adult patients, and administered by experienced local health systems researchers at four health facilities which were most likely to provide care for Diabetes and TB patients, the primary and secondary care facilities for Bishkek (City Endocrinology Dispensary and the City TB Hospital) and the tertiary facilities, which receive referred patients from all of Kyrgyzstan (the Endocrinology Department of the National Hospital and the National TB Centre). Diabetes patients are interviewed in dispensaries and hospitals to include both outpatient and inpatient care, while TB patients are interviewed in the two hospitals which are most frequented. Ethical clearance was obtained from University College London (Project 0025/001). This data collection was part of the project “Diabetics in Kyrgyzstan”, which was also approved locally by the Committee on bioethics under the Ministry of Health of Kyrgyzstan. All adult patients in the waiting room on a given day were invited to participate in the survey after receiving their treatment and giving written informed consent. Interviews were on consecutive days, starting from 3 September 2010 and continuing until the sample size was achieved on 1 November 2010. Each interview lasted approximately one hour. A simple quota was used, resulting in a final sample of 138 patients with Diabetes (either Type 1 or Type 2), 139 patients with TB, and 32 patients with both illnesses.

The survey collected data on health expenditure (such as direct medical and non-medical costs, formal and informal treatment costs), socio-economic status (such as age, gender, education level, employment status, and household size) and financial coping strategies of patients (income and savings, social welfare and donations, support from social networks, borrowing money or selling household assets). Education was measured in levels of completed schooling (primary, secondary, tertiary). Employment status was categorised as: being out of work (unemployed or not able to work due to health issues), working in an informal arrangement (subsistence farming, self-employment or housekeeping), being formally employed (public or private sector) or retired (in retirement or pensioner).

The patient survey included specific information about household income and health expenditure in the national currency, Kyrgyz Som (KGS). The exchange rate was 46.14 KGS to 1.00 USD in 2011 [38]. Health expenditure included three components: the cost of travel to the health facility, informal payments made during the hospital visit, and all formal payments incurred during the visit including medicines, diagnostic tests, doctors’ fees, etc. Data were collected as aggregates for these components for the most recent overnight admission, observation or emergency room visit and outpatient visit. Patients were also asked for the reason for visiting to select only those visits related to Diabetes and TB. These cost components were then multiplied by the number of reported visits in the last 90 days, to get an estimate of health expenditure over this time span, while reducing the risk of recall bias. Health facility visits are compared for both single diseases and the co-prevalence.

Capacity to pay (CTP) was approximated by effective income [2, 39]. Effective income represents household income after subsistence needs are met. In urban settings, such as Bishkek, this effective income is often used to calculate CTP [2, 40, 41]. To reflect economies of scale in households with more members, household income was transformed into equivalence income. This was done by dividing household income by household size to the power of 0.56. This equivalence scale had been empirically derived from a multi-country regression of household survey data from 59 countries [2, 40]. Minimum food expenditure was subtracted from equivalence income and the remainder used as a measure of CTP for health expenditure and a proxy for poverty risk [40]. The average minimum food expenditure was derived from a World Bank calculation based on the Kyrgyzstan Integrated Households survey 2010 [42].

A commonly used measure of catastrophic health expenditure is to estimate health expenditure proportional to CTP. When using the CTP approach, a threshold of 40 % is commonly used [2]. If health expenditure exceeds the threshold and more than 40 % of the household’s CTP is being spent on health, a catastrophic impact on the household may be expected [2]. The catastrophic payment method was chosen over the measurement of impoverishment. The impoverishment approach measures if patients drop below a poverty line after health spending. Measuring impoverishment is however dependent on an agreed poverty line. While a national poverty line is available in Kyrgyzstan, a World Bank report in 2011 [26] found that this national poverty line is not representative of the actual subsistence level.

The selection and definition of coping strategies, is adopted from the analytical framework developed by McIntyre et al. [7]. To better represent the post-communist context of Kyrgyzstan a coping strategy called ‘social welfare support and donations’ is added. Additionally, social networks, such as family and friends, are included in this analysis and represent informal social protection mechanisms, as described by Russell [1]. As such, the complete list of coping strategies included in this study is as follows: income or savings, social welfare support or donations from employers or agencies; support from social networks like friends and family; borrow money; and raise money by selling assets. Patients were asked if they used these strategies for coping with the health care spending experienced in their health care seeking in the last 90 days, but could also indicate that they had used other coping strategies not listed in the response options.

Comparing catastrophic health spending and coping strategies within the three patient groups without controlling for confounding by differences in the socio-economic status can result in overestimation of the effect of the disease or co-prevalence. Matching the patient groups according to their different socio-economic statuses helps identify the unbiased effect of the disease on financial coping strategies and the occurrence of catastrophic health spending. Nonparametric matching, such as CEM, allows reducing these imbalances. While the regression models still need to control for the socio-economic imbalances, the model dependence is much lower than without matching [43]. Multilevel matching is a very common technique for the analysis of observational data and has been employed in many recent studies analysing the economic burden of disease [44‐47]. Coarsened Exact Matching (CEM) was used here to account for the socio-economic differences in the patient cohorts.

CEM aims at finding exact matches, but interprets “exact” as identical within a coarsened range of similar covariates. CEM finds categories of similar covariates based on distributions or intuitive division and matches exactly within these categories [47, 48]. The detailed methodology behind CEM can be read elsewhere [49, 50]. CEM has advantages over other matching algorithms like propensity score matching and exact matching, because it allows for the matching of groups with unequal numbers of observations and allows a degree of imprecision in the matching [47, 51]. This allows matching even when observations are not precisely exact, as long as they can be coarsened into similar categories. While CEM was originally designed to match one treated group to a untreated control group, it allows matching of multi-level treatment [43, 49].

In this study, CEM was used to match observations in the three patient groups, Diabetes, TB and co-prevalence. CEM was used here to control confounding by differences in the socio-economic background of patients affected by the two diseases or co-affected. The matching variables were: 1) gender, 2) age in years, 3) education level, 4) employment status and 5) equivalence income to control for differences in the capacity to pay, adjusted for household size.

Data were analysed using, version R 3.0.2. In the descriptive statistics, all mean calculations include 95 % confidence intervals. For variables with skewed distributions, such as income and expenditure, we include their medians. Item non-response was checked and found to be highest for household income (5.5 %) and employment status (5.1 %). Household income and employment were analysed for non-response bias in relation to other socio-economic variables such as education, age, household size and gender. Non-response bias could be ruled out and list-wise deletion was used in the analyses.

Regression models were used to analyse the association between socio-economic background and coping strategies. For each coping strategy, a logit model was conducted using the GLM function with a binomial logit link in R. The socio economic variables used in these models (as independent variables) were: gender, age, education, employment and equivalence income in tertiles¹.

Odds ratios were calculated to compare the likelihood of spending money on health care, engaging in catastrophic health spending and coping strategies between the three patient groups. Three patient group variables were generated (TB vs Diabetes, Co-affected vs Diabetes, Co-affected vs TB) to allow pairwise subgroup comparison. The odds ratios were determined with regression models for binary categorical variables using generalised linear models with logistic distribution functions i.e. logistic regressions [52]. As pointed out earlier, matching only reduces the imbalance between the patient groups, but does not completely eliminate it. Thus the regression models should include all covariates of the original, unmatched model [43, 49]. The regression models thus control for the covariates: 1) gender, 2) age in years, 3) education level, 4) employment status and 5) equivalence income. All odds ratios were estimated for matched and unmatched observations to allow a comparison of the association of the patient groups alone and the imbalanced association of patient group plus socio-economic background. While the models of the unmatched regressions produce estimates of the combined association of disease and socio-economic status, the logistic regressions of the matched observations, shows associations of the disease alone. All regression models were tested for influential outliers using Bonferroni statistics and influence index plots from the R package CAR, version 2.0-25. From the same R package, variance inflation factors were used to test multi-collinearity in all regression models.

Table 1 presents summary statistics of the main variables included in the analysis. On average, Diabetes patients are 55 years old, while TB patients are 30 years old (p < 0.001). Co-affected patients are approximately the same age (54 years) as the Diabetes patients. While TB patients have an equal gender ratio, Diabetes patients are mostly female (70 %) and co-affected patients are mostly male (63 %). Almost all patients have completed secondary school. The only differences are in the number of patients with completed tertiary education; while 54 % of Diabetes patients and 61 % of co-affected patients have completed tertiary education, only 31 % of TB patients have the same level of education. Half of Diabetes patients are retired, while 61 % of TB patients are in informal or formal employment. While co-affected patients are in the age range of Diabetes patients, they are only half as likely to be retired, but more than twice as likely to earn money in informal arrangements. TB patients are younger and accordingly, more likely to be in paid work than Diabetes patients. Although co-affected patients are the same age as Diabetes patients, they do not use pensions as often. Instead of pensions, co-affected patients earn money in informal arrangements.

Equivalence income is highest in TB patients with 4704 KGS (106 USD), but this is not significantly higher (p = 0.183) than in Diabetes patients with 3683 KGS (83 USD). Co-affected patients have the lowest equivalence income with 2393 KGS (52 USD) and are significantly poorer than Diabetes (p = 0.008) and TB patients (p = 0.004).

Figure 1 illustrates the differences in number of visits in matched observations. The mean number of inpatient admissions is around 1 and mean visits to the emergency room are around 0 for all patient groups. There are, however, differences in the number of outpatient visits and visits to collect medication. Diabetes and co-affected patients have a mean of 3 and 4 visits to outpatient facilities respectively, while TB patients only visit the outpatient facility once a month. Diabetes and co-affected patients need on average 2 visits per month to collect medication, while TB patients have on average 0 visits to collect medication. Significance tests prove that the differences between TB and Diabetes patients, as well as the differences between TB and co-affected patients, are significant. There is no significant difference between Diabetes and co-affected patients. In outpatient and medication collection visits, it becomes apparent that the co-affected patients face the double burden of Diabetes care and TB care, with the number of visits being exactly the sum of the other single disease patient groups.

The analysis of health expenditure in Table 1 shows that health expenditure is highest for Diabetes patients and lowest for TB patients. The medians in Table 1, illustrate however, that means comparison might not be appropriate, since there is a substantial number of patients who did not face any health expenditure in the last 90 days, despite using health services. From Table 5, it can be seen that Diabetes patients are significantly more likely to spend money for health care than TB patients and this finding is consistent after matching the observations for socio-economic status. Co-affected patients are in the middle between the single disease groups with a higher likelihood to face OOP than TB patients, but a smaller likelihood than Diabetes patients. Informal payments in hospitals and travel cost to total health expenditure can further be distinguished from co-payments for medication due to underfunding in the SGBP. Informal payments for outpatient care were not collected, which clearly limits this study with respect to the analysis of informal payments. In hospital settings, this study shows that all patient groups have median spending for informal payments of zero. Only few participants, all among TB or co-affected patients, reported to have informal payments. Similarly, travel spending was reported zero by most of participants, but for some patients getting to the health facility can be costly. This is more often the case for Diabetes patients with a mean spending of 254 KGS. TB and co-affected patients have mean spending of 49 KGS and 43 KGS. The remaining health spending can be seen as co-payments within the SGBP for medicines, diagnostic tests, consultation and others.

After demonstrating the higher economic burden for Diabetes and co-affected patients, the focus of this section of analysis is on the financial strategies for coping with the economic burden of care. Table 2 shows the number of coping strategies a patient or her/his household uses and the percentages of patients using a specific strategy for their last care seeking event. 208 patients used income or savings, 173 use social welfare and donations, 109 use support from social networks, 36 had to sell household assets and 31 borrowed money. Among the 208 households using income or savings, 34 % use this as their only coping strategy. Among the 173 households using social networks, 23 % use this strategy as their only financing source. Table 3 shows that social network is used most of all strategies as an additional funding source. Social welfare or donations are rarely used as a single strategy. Borrowing money, selling assets, donations and social welfare are also much more likely to be used in combination with other strategies. Borrowing money is combined most often with at least two or three additional strategies.

Additional regression analyses show that the socio-economic background influences the decision which a patient makes regarding what coping strategy to use. Five multivariate regression models were conducted to investigate association with socio-economic background variables for each of the five coping strategies. Table 4 shows significant association between employment status, age and equivalence income. Knowing that the socio-economic background varies within the three patient groups and knowing that socio-economic background is also associated with coping, raise this question whether choice of coping strategy is associated with disease or with socio-economic context of the patient arises. The next section compares coping in patient groups with balanced and imbalanced socio-economics and thus allows differentiating the association between patient groups and coping strategy.

Using CEM, 175 of 310 observations could be matched, reducing the imbalance score of L1 by 88 %. Table 5 shows the odds ratios for matched and unmatched observations for three patient groups. While the models of the unmatched regressions produce estimates of the combined or imbalanced association of disease and socio-economic status, the logistic regressions of the matched observations, find the association with the disease alone. The odds ratio in the unmatched observations show four significant results: 1) TB patients are more likely to use income or savings than Diabetes patients, 2) Diabetes patients are more likely to use social welfare or donations than TB patients, 3) co-affected patients are less likely to use social welfare or donations than Diabetes patients and 4) co-affected patients are more likely to face catastrophic health spending than TB patients. When using the CEM approach to match the observations for socio-economic status, the logistic regression models show three significant results: 1) TB patients are less likely to engage in catastrophic health spending than Diabetes patients, 2) Diabetes patients are more likely to use social welfare or donations than TB patients and 3) co-affected patients are less likely to use social welfare or donations than Diabetes patients.