Summary of Bayesian Estimation in the Rasch Model H. Swaminathan and J. Gifford Journal of Educational Statistics (1982)
Problem: Estimate “ability” of each of N standardized test takers, based on a performance on a set of n test items
Rasch model Model used in psychometrics relating performance on a series of test items to ability It is a logistic regression model with a single parameter describing each test item;
Estimating N ability parameters, assuming b j ’s known where r i = # of items i th examinee answers correctly Estimate by ML
Bayes set-up
Posterior calculation Need to wrt 2 and
Posterior (con’t) No known distribution…
Computation In 1983, this joint posterior was too complicated to compute and use Authors suggested using modes as estimators Find maxima using single-valued Newton- Raphson; i.e.,
Estimating N ability parameters, and n difficulty parameters Same idea as before, except add hierarchical and prior structure for b j ’s Same structure as for ability parameters: Can compute joint posterior
Specification of priors Authors want prior to be proper and to have variance defined > 4 Recommend 5 15 Set (?)
Simulation Studies 1&2 Artificial data was generated according to logistic model Ability and difficulty parameters generated as uniform Conducted factorial simulation experiments: (1) n x N; (2) n x N x ( b and ө ) Calculated Bayes and ML estimators
Conclusions MSE smaller for Bayes estimators Varying has little effect except in smallest cases
Example: NAEP Math 8 th grade n=25, N = ? = 10 = 5,8,15,25 Conclusions Estimators similar except at extremes of ability/difficulty Bayes allows estimation of ability for perfect score