Date of Award

8-1995

Degree Name

Doctor of Philosophy

Department

Psychology

First Advisor

Dr. Bradley E. Huitema

Second Advisor

Dr. Alyce Dickinson

Third Advisor

Dr. Alan Poling

Fourth Advisor

Dr. Joseph McKean

Abstract

Recent work by Huitema and McKean (1991, 1994a, 1994b, 1994c, in press) has shown that the most frequently used statistical methods for performing conventional time-series analyses lead to gross distortions of results when these approaches are applied in the context of the typical behavioral research study. Most of these problems could be avoided if researchers were aware that the time-series methods recommended in many areas are not generally needed. The appropriate evidence regarding the need for complex time-series methods requires a meta-analysis of the autocorrelation present in behavioral studies. The project involved: (a) sampling several hundred research articles published in the Journal o f Applied Behavior Analysis (JABA) during 1990-1995, (b) extracting the quantitative data from the time-series data displays in 200 sample articles, (c) analyzing each data set using recently developed statistical methods, and (d) performing a meta-analysis of the individual analyses using a new methodology developed for time-series data.

Conventional lag-1 autocorrelation estimates rl (known to be negatively biased) were computed on (a) the raw data and (b) the residuals obtained from regressing the raw data on time in order to remove the linear trend. New unbiased autocorrelation estimators (rF1 and rHM) were applied to both the raw and detrended data. New tests of the homogeneity of the coefficients in the distributions of rHM and rF1 were computed, and confidence intervals on the means of the distributions were computed. Finally, new tests of the significance of the obtained rF1 and rHM values were computed.

Results indicate that recent conclusions regarding the presence of the autocorrelation "problem" in typical behavioral experiments should be questioned. When the unbiased estimator (rF1) is used, there appears to be a significant proportion of behavioral data sets exhibiting autocorrelation, generally in the positive direction. Most of the apparent positive autocorrelation can be explained by linear trend in the data. When the data are appropriately detrended the autocorrelation distribution is similar to that expected by chance when sampling from a population of independent (nonautocorrelated) errors. The conclusion that behavioral data are not highly autocorrelated is drawn. In most cases the explanation for "apparent" autocorrelation is not that the errors are autocorrelated. Rather, the statistical model used to explain the data has been misspecified. Artificial autocorrelation is to be expected if trend (linear or otherwise) and other deterministic components in the data have not been specified in the statistical model. Conventional general linear model solutions are satisfactory for many studies in applied behavior analysis.

Access Setting

Dissertation-Open Access

Share

COinS