ISS is the international Institute of Social Studies of Erasmus University Rotterdam

Transcription

1 ISS is the international Institute of Social Studies of Erasmus University Rotterdam

2 Analysing Risk and Adverse Selection Using Subjective Expectations of Health Expenditure: Evidence from Rural Ethiopia Zelalem Yilma, Owen O Donnell, Anagaw Mebratie, Getnet Alemu, Arjun S. Bedi Presentation at the University of East Anglia: 18/19 th September 2013

3 Outline of presentation Contribution of the paper Study setting and context Belief elicitation instrument Validity of the subjective expectations data Subjective distribution of health expenditure Modeling the health expenditure process Adverse selection in CBHI Conclusion

4 Why this paper? Extent and distribution of medical expenditure risk is central for the demand for voluntary health insurance and welfare case for social health insurance Little known about risk perceptions Little known about the incidence and magnitude of such risk in developing countries..why?

5 1. In cross-section surveys: Cross-section variance confounds uncertainty with predictable differences in health expenditure 2. Temporal variance (Newhouse et al 1989; van Vliet 1992; Feenberg and Skinner 1994; French and Jones 2004) few panel data sets with sufficient waves to identify stochastic properties Even if they are available: validity of rational expectations w.r.t health expenditure is absent even in high income countries Very little known how individuals forecast their health expenditure in any context (Breyer et al. 2012)

6 How are expectations of health expenditure formed? Vital question to understand: Perceptions of financial risk Degree to which the demand for insurance is related to risk To examine adverse selection directly Test of adverse selection using coverage and realized expenditure is plagued by the difficultly of distinguishing selection from moral hazard effects (Chiappori, 2000) Functioning of insurance markets?

7 Research questions? 1. Is it possible to collect data on subjective expectations that are valid in the sense of being logically consistent and plausibly informative of beliefs about future spending on health care. Consensus is that it is possible (Attanasio, 2009; Delavande et al, 2011). 2. What is the extent to which expectations appear to be based on past realized expenditures (level and change) among other characteristics? 3. Whether uptake of community based health insurance (CBHI) is related to expected health expenditure.

8 Relevance of the study Policy relevance: for CBHI scale up & for the operation of voluntary health insurance in low income settings Methodological: first study to elicit beliefs about future spending on health care in either high- or low-income settings Although experimented with other economic outcomes (McKenzie et al, 2007; Delavande, 2008; Attanasio, 2009; Delavande et al, 2011; Delavande and Kohler, 2012).

9 CBHI is vulnerable to adverse selection: voluntary and small size of pool (see Wang et al, 2006; Parmar et al, 2012) Highly persistent medical expenditure will jeopardize the sustainability of insurance/cbhi If individuals are aware of this, incorporate it in their expectations and use the information for CBHI take up can they? If insurer is not able to observe these major determinants of expectations of expenditure or is not allowed to adjust the premium

10 Study setting and context part of a larger project designed to evaluate CBHI pilot scheme launched in June-July Household survey in 4 regions of Ethiopia (Amhara, Oromiya, Tigray and SNNPR), 16 rural districts, 96 Kebeles, 1632 households, 9455 individuals...(march-april 2011) Follow-up survey (1599 households): after CBHI...(March-April 2012)

11 Data (2012 survey mainly) Belief elicitation instrument Imagine that no member of your household contracts a NEW serious illness or injury in the next 12 months. In such a case what would be the MINIMUM amount of money your household would have to pay for health care and medicines (including transport costs) over the next 12 months? Imagine that at least one member of your household contracts a NEW serious illness or injury that requires treatment in a hospital in the next 12 months. In such a case what would be the MAXIMUM amount of money your household would have to pay for health care and medicines (including transport costs) over the next 12 months?

12 Calculate three thresholds Belief elicitation instrument 0.25(max-min) Min (max-min) Amount A A+0.25(max-min) Amount B B+0.25(max-min) Amount C How likely is that the amount you would spend on health care (including transport) will be greater than threshold A, B and C in the coming 12 months?

13 Validity of responses # of respondents Total number of observations 1599 Non-response Non-response to minimum and/or maximum value, and so missing 11 all 3 probabilities Zero reported for both min. and max., and so missing on all 3 7 probabilities Valid min. and max. but non-response on at least 1 probability 1 Total # of non-response 19 Non-missing responses 1580 Enumerator calculation errors Errors in calculation of any/all thresholds 229 Errors resulting in wrong ordering of thresholds 38 Total usable observations 1542

14 Illogical responses Validity of responses # of respondents (%) Total usable observations 1542 Illogical responses One violation of monotonicity: P(X>A)<P(X>B) or P(X>B)<P(X>C) Two violations of monotonicity: P(X>A)<P(X>B) and P(X>B)<P(C>X) 66 (4.28) 65 (4.22) All reported probabilities zero: 46 (2.98) Observations with illogical responses 177 (11.48) Total observations with logical responses 1365 (88.52) Notes: A=min+k, B=A+k and C=B+k, where k=(max-min)/4. 38 observations for which calculation errors led to incorrect ordering of thresholds are not included (see Table 1), so A<B<C by construction.

15 Illogical response rate: comparable to Dominitz and Manski s (1997): 10 % (US) before prompting respondents to revise their answers. higher than Attanasio and Augsburg (2012): 4% (India) using a similar instrument No practice question in our study Test of difference in means of covariates in our model (logical vs illogical sample) shows illogical response sample is on average: Smaller, poorer, less healthy, non-cbhi member And is headed by more females, uneducated, older people

16 0 Percent Distribution of responses likelihood of spending atleast threshold A likelihood of spending atleast threshold B likelihood of spending atleast threshold C Note: Respondents are invited to indicate the likelihood on a scale of 0-10 which has been transformed to 0-100%

17 Validity of responses # of observations who reported [...] 0% 50% 100% # of observations Threshold A (0%) (16.8%) (4.4%) 1365 Threshold B (5.0%) (12.5%) (0.8%) 1365 Threshold C (13.9%) (3.9%) (0.7%) 1365 Note - Our survey asks likelihood on a scale of 0-10 but for presentation purposes, here, we use 0%, 50% and 100% which correspond to 0, 5 and 10 respectively.

18 Validity of responses bulk of the sample of poor, low educated and rural respondents (1,365 of 1,599, or 85 percent) were able to manage a rather abstract exercise of contemplating future scenarios with respect to health and associated medical expenses and reporting their likelihoods. the high response rate the vast majority of responses being logically consistent, the lack of bunching in the responses and the sensible shifting location of the probability distribution as the threshold is raised

19 Subjective distribution of health expenditure Mean probability of medical expenditure lying in each equally-spaced, household specific interval 0 Min - A A - B B - C C - Max

20 Subjective distribution of health expenditure Piecewise uniform distribution: used to compute the first and second moments of the distribution of forecast health expenditure in the next 12 months for each household Attanasio and Augsburg (2012) Realized expenditure in the last 12 months computed as: Simple extrapolation (outpatient times 6 plus inpatient expenditure) Cost function extrapolation (predict outpatient cost, multiply by 6 and add to inpatient expenditure)

21 Subjective distribution of health expenditure Obs Min Max Median Mean Std. dev. Forecast health expenditure next year (parameters of household-specific distributions) Mean Standard deviation Coeff. of variation Minimum Maximum Realized health expenditure last year Simple extrapolation Cost function extrapolation 1347 ~ Conditional on non-zero realized outpatient expenditure (simple extrapolation) Obs Min Max Median Mean Std. Dev Forecast health expenditure next year Mean Standard deviation Coeff. of variation Minimum Maximum Realized expenditure last year Notes: Simple and cost function extrapolations of realized expenditures refer to method of estimating annual spending on outpatient care from reported expenditure in past two months

22 Density Density Subjective distribution of health expenditure Fig.3a: Full sample (n=1365) Fig. 3b: Non-zero realized outpatient expenditure (n=373) Log expenditure Log expenditure Forecast mean log expenditure Realized log expenditure kernel = epanechnikov, bandwidth =

23 Distribution of health expenditure Substantial overestimation of risk exposure that is likely to arise from utilization of risk measures based on cross-sectional variance. Such measures confound risk with predictable heterogeneity across households.

24 Correlation coefficients Full sample Realized expenditure Realized expenditure Change ( ) 2011 (n=1365) (0.218) (0.000) 2012 (n=1365) (0.000) Expected expenditure 2013 (n=1365) (0.000) (0.000) (0.689) Change (n=1365) (0.551) (0.000) (0.000) (expected 2013 realized 2012) Conditional on realized outpatient health expenditure > 0 Realized expenditure 2011 (n=454) (0.085) (0.000) 2012 (n=373) (0.000) Expected expenditure (0.004) (0.003) (0.008) Change (0.087) (0.000) (0.000) (expected 2013 realized 2012) Notes: Realized expenditures calculated using the simple extrapolation of payments for outpatient care.

25 Modeling the health expenditure process lnoop lnoop X e t t 1 t t e lnoop t lnoop t 1 (lnoop t 1 lnoop t 2) X t t lnoop e t ( )lnoop 1 lnoop t t 2 X t t

26 Modeling the health expenditure process Potential source of bias in estimation of (1) measurement error in actual health expenditure time invariant unobservables that are necessarily correlated with lagged actual expenditure 2SLS using Kebele mean expenditure Arellano and Bover (1995) trick of instrumenting lagged expenditure with the change in expenditure Instrument actual outpatient expenditure with predicted values from health cost function Alternative definition of forecast expenditure also used

27 Modeling the health expenditure process VARIABLES (1) (2) (3) (4) (5) (6) (7) Mean E[lnOOP] Mean E[lnOOP] Mean E[lnOOP] Mean E[lnOOP] Mean E[lnOOP] Mean E[lnOOP] Mean E[lnOOP] Ln [outpatient 2012] * *** ** (0.0135) (0.0589) (0.0256) (0.0667) (0.0164) (0.0619) (0.0759) Ln [ inpatient 2012] *** *** *** *** *** (0.0177) (0.0615) (0.0223) (0.0196) (0.0278) (0.122) (0.0291) D[ln outpatient] *** *** (0.0122) (0.0282) (0.0232) D[ln inpatient] (0.0217) (0.0885) (0.0220) CBHI member * * (0.0567) (0.0549) (0.0571) (0.0558) (0.0579) (0.0584) (0.0610) Constant 4.517*** 4.536*** 4.538*** 4.528*** 4.623*** 4.616*** 4.697*** (0.522) (0.541) (0.508) (0.538) (0.508) (0.531) (0.556) Observations 1,345 1,345 1,345 1,345 1,345 1,345 1,330 R-squared

28 Modeling the health expenditure process VARIABLES (1) (2) (3) (4) (5) (6) (7) Ln [mid point] Ln [mid Ln [mid Ln [mid Ln [mid Ln [mid Ln [mid point] point] point] point] point] point] Ln [outpatient 2012] * *** ** (0.0143) (0.0598) (0.0270) (0.0637) (0.0171) (0.0632) (0.0741) Ln [ inpatient 2012] *** *** *** ** ** (0.0199) (0.0737) (0.0243) (0.0209) (0.0294) (0.140) (0.0300) D[ln outpatient] ** *** (0.0128) (0.0294) (0.0245) D[ln inpatient] (0.0226) (0.0953) (0.0228) CBHI member 0.108* 0.109* ** * (0.0618) (0.0599) (0.0617) (0.0614) (0.0626) (0.0657) (0.0658) Constant 5.144*** 5.082*** 5.178*** 5.155*** 5.229*** 5.082*** 4.925*** (0.653) (0.669) (0.621) (0.682) (0.640) (0.661) (0.622) Observations 1,345 1,345 1,345 1,345 1,345 1,345 1,330 R-squared Note: Notes under table 8 also apply here.

29 Further on adverse selection Mean Health expenditure before and after CBHI Members (CBHI district) Non-members (CBHI district) post-enrolment pre-enrolment Non members (control district)

30 Further on adverse selection Estimating if CBHI take up is related to pre-enrolment expenditure and health measures

31 Further on adverse selection There is a positive correlation between CBHI and past period outpatient and inpatient expenditure (albiet insignificant) Self assessed health status and paralysis predict the probability of taking up insurance.

32 Conclusions Cross-sectional data on subjective expectations and realizations of health expenditures offer a low-cost alternative to panel data for learning about the persistence of medical spending. Uncertainty is greatly overestimated by cross-section variability of health expenditure The health expenditure process is serially correlated but is far from persistent (we reject the null of unit-roots)

33 Conclusions There is an indication of adverse selection in CBHI take up But the transient relationship between expected and actual spending suggests that adverse selection is unlikely to threaten the financial sustainability of the CBHI scheme.