This article examines characteristics of Japanese public opinion surveys during 1990-2009 and makes pertinent comparisons with the past. Public opinion surveys continued to increase with cities being the major contributor to that growth. Surveys conducted face-to-face and those that used drop self-administered questionnaires decreased, whereas mail surveys increased. Telephone surveys remained a relatively minor method, but increased substantially compared to before 1990. The use of Voter Registration Lists decreased; whereas there were increases in surveys that sampled either from the Basic Resident Registers or did not employ lists. For the 1990-09 period as a whole, the median response rate was 58.9% which suggests a steeper decline than the decreases observed in the past. Surveys sponsored by central government agencies had the highest rates and those by universities the lowest. Among studies that used probability sampling, the highest median response rates were for drop self-administered questionnaires followed by face-to-face and mail surveys.
The purpose of this study is to find less biased effect size index in one-way analysis of variance (ANOVA) by performing a thorough Monte Carlo study with 1,000,000 replications per condition. Our results show that contrary to common belief, epsilon squared is the least biased among the threemajorindices, while omega squared produces the least root mean squared errors, for all conditions. Although eta squared results in the least standard deviation, this does not necessarily make it a good estimator because a considerable amount of bias still occurs when the sample size is small.
Bayesian estimation of reliability by MCMC is proposed as an alternative to coefficient alpha, regarded as a classic method of estimating reliability. However, coefficient alpha is derived from a structural factor model, various kinds of which are used in modern data analysis, e.g., SEM, sometimes used to analyze reliability. SEM emphasizes covariances to analyze categorical items that are estimated by polychoric correlations. The methods proposed in this study directly apply the MCMC algorithm to individual examinees responses. Thus, covariances are not estimated but implicitly represented by factor models. For categorical items, models for continuous items are modified and applied based on an ordinal probit regression. The performances of the proposed models and algorithms are investigated by simulations and successful results are obtained.