Abstract:
Logistic regression model is used to explain the relationship between explanatory variables and categorical response variable in many research fields. The maximum likelihood (ML) was generally used to estimate parameters, but the ML shows very poor results in the case of separate data. Exact Logistic Regression (ELR) and Markov chain monte carlo (MCMC) exact inference are the logical alternative to the ML. This research offers a comparison of the property of three methods for estimation. The study found that; In case of one continuous explanatory variable, when the percentage of the overlapping is less than or equal to 4, the three estimation methods have low efficient in parameter estimation. When the percentage of overlapping is higher than 4, the ML and ELR are equally efficient but the MCMC has the lowest efficiency. In case of two discrete explanatory variables, when percentage of responses is less than 50, the ML and ELR poorly perform in parameter estimation. However, when the percentage of response is equal to 50 at sample size less than or equal to 28, the ELR is more effective than the ML, but at sample size greater than or equal to 48, both methods are similarly effective. In interval estimation of two cases the results showed that, when the probability estimation coverage does not differ from the confidence level, the least confidence interval width estimation is ML. The ELR and MCMC methods commondly provided the infinite width of the confidence interval on average.