Biomedical subjects
Publications and source records attributed to C F Sheu.
Psychologists often use special computer programs to perform meta-analysis. Until recently, this had been necessary because standard statistical packages did not provide procedures for such analysis. This paper introduces linear mixed models as a framework for meta-analysis in psychological research, using a popular general purpose statistical package, SAS. The approach is illustrated with three examples, using SAS PROC MIXED.
This paper considers a regression approach to estimating signal detection parameters for rating data. The methodology is based on the statistical modeling of ordinal data and requires only standard statistical software such as SAS (SAS/STAT User's Guide, 1999) for computation. The approach is more efficient than the current practice of extracting the parameter estimates with the use of specialized software and analyzing the estimates with the use of a standard statistical package. It greatly facilitates exploration of the effects of covariates on model parameters. The method is illustrated using a published data set from a single factor multiple-alternative perceptual task, and data from a more complex factorial design examining recognition memory rating data.
The purpose of this paper is to describe and illustrate a regression approach to the analysis of correlated binary outcomes (Liang & Zeger, 1986). Ignoring the correlations between repeated observations can lead to invalid inferences. This approach extends logistic regression to account for repeated observations in each of a series of individuals. In this paper, I present a nontechnical introduction to the generalized estimating equations (GEE) approach. A fictitious example is used to demonstrate that GEE regression correctly adjusts for the correlations between repeated binary observations. The approach is illustrated with an analysis of safer sex practices among high-risk teenagers.
The Fourier transform method in conjunction with frequency domain smoothing techniques has been suggested as a powerful tool for examining components in a serial, additive reaction time model (P. L. Smith, 1990). Robustness and sensitivity to violations of the assumptions of serial model of this method are evaluated. When an incorrect distribution was used in recovering an unobserved component, results gave no information to show that an incorrect distribution was used, and the results were just as interpretable as those obtained using the correct distribution. These results demonstrate that the assumptions underlying the method cannot be assessed by the result of deconvolution, and the method cannot show that the purported component is actually from the serial combination.
We analyze how subjects make causal judgements based on contingency information in two paradigms. In the discrete paradigm, subjects are given specific information about the frequency a, with which a purported cause occurs with the effect; the frequency b, with which it occurs without the effect; the frequency c, with which the effect occurs when the cause is absent; and the frequency d, with which both cause and effect are absent. Subjects respond to P1 = a/(a + b) and P2 = c/(c + d). Some subjects' ratings are just a function of P1 while others are a function of delta P = P1-P2. Subjects' post-experiment reports are accurate reflections of which model they use. Combining these two types of subjects results in data well fit by the weighted delta P model (Allan, 1993). In the continuous paradigm, subjects control the purported causes (by clicking a mouse) and observe whether an effect occurs. Because cause and effects occur continuously in time, it is not possible to explicitly pair causes and effects. Rather, subjects report that they are responding to the rate at which the effects occur when they click versus when they do not click. Their ratings are a function of rates and not probabilities. In general, we argue that subjects' causal ratings are judgments of the magnitude of perceptually salient variables in the experiment.
Global memory models are evaluated by using data from recognition memory experiments. For recognition, each of the models gives a value of familiarity as the output from matching a test item against memory. The experiments provide ROC (receiver operating characteristic) curves that give information about the standard deviations of familiarity values for old and new test items in the models. The experimental results are consistent with normal distributions of familiarity (a prediction of the models). However, the results also show that the new-item familiarity standard deviation is about 0.8 that of the old-item familiarity standard deviation and independent of the strength of the old items (under the assumption of normality). The models are inconsistent with these results because they predict either nearly equal old and new standard deviations or increasing values of old standard deviation with strength. Thus, the data provide the basis for revision of current models or development of new models.