Bayesian Analysis of Health Care Provider Choice in Rural India
- Presenter:
Mon June 5, 2006 9:30-10:45 Room Alumni Lounge
This paper uses a Bayesian mixed multinomial logit (MMNL) framework to analyze the determinants of health care provider choice in rural India. Classical framework often becomes infeasible when random coefficients in MMNL model are allowed to be correlated and/or when distributions other than normal are assigned to the random coefficients. Allowing correlation between the random coefficients increases the number of parameters to be estimated, which often becomes difficult using classical methods. Normal distribution for the random coefficients is appropriate when there is some theoretical basis for the coefficient to be any real number. However, on many occasions, a random coefficient is expected to be either positive or negative. For example, price coefficient in the utility specification is expected to be always negative as people dislike higher price. In such cases, distributions with positive support (e.g., log-normal distributions) are more appropriate. However, when the distributions are assumed to be non-normal, the resulting likelihood surface often becomes very irregular, leading to failure in finding the maximum. Bayesian framework provides a natural way to deal with these complexities. In this paper, we assume both normal and log-normal distributions for the price coefficients. We use a Gibbs sampler with a layer of Metropolis-Hastings algorithm embedded within it to make random draws from the resulting joint posterior distribution. Our findings show that models with normal coefficients fit the data better. Price and distance are found to be important determinants of health care provider choice in rural India. Moreover, when individual-specific characteristics are interacted with provider attributes, model fit improves, indicating that these characteristics also influence health care provider choice decisions.
