Error rates for multiple comparison methods: Some evidence concerning the frequency of erroneous conclusions.

Abstract

Seven methods currently used for paired comparisons among all group means following analysis of variance were tested for sensitivity to violation of stated requirements. The t-test methods and the Duncan multiple range test were found to produce far more Type I errors under null conditions than is generally acceptable. The methods of Scheffe' and Tukey produced the fewest Type I errors and are least susceptible to violation of requirements. Type II error rates for all methods were also generated. The appropriate choice of error rate and multiple comparison method is discussed.

Hippocampal formation lesions produce memory impairment in the rhesus monkey

“…These analyses were followed, when appropriate, by additional a posteriori comparisons by using tests of simple main effects (Kirk, 1982;Hayes, 1988). All such multiple comparisons were assessed using the Tukey B correction to provide ''experiment wise'' control of Type I error (Petrinovich and Hardyck, 1969).…”

K‐ABC sequential‐simultaneous processing and language‐impaired preschoolers

“…As shown in Table 1, univariate analyses of (ANOVAs) indicated significant differences between group means for all K-ABC scales and for the Mental Processing Composite. Post hoc comparisons showed that the CON and BP subjects obtained higher Mental Processing Composite and Simultaneous Processing scores than the two language-impaired groups (Tukey Multiple-Range Test, /?<.05; Petrinovich & Hardyck, 1969). On Sequential Processing, the CON subjects obtained higher scores than the two groups with language impairment, and the BP subjects obtained higher scores than the LI-BP group (Tukey Multiple-Range Test, p < .05).…”

Do recognition-free recall discrepancies detect retrieval deficits in closed-head injury? demonstrating the inaccuracies of a reviewer's critique

“…It is further known, over decades of computer simulation studies (beginning with the seminal work of Norton's 1952 doctoral dissertation--see also Boneau, 1960;Lindquist, 1956;Petrinovich & Hardyck, 1969;and Cicchetti, 1994), that parametric techniques are quite robust with respect to violations of assumptions underlying their usage. Specifically, there are three parameters that need to be considered in the comparison of any two groups: sample sizes, or Ns; the ratio of the sizes of the two variances; and the shape of the distribution for each of the samples being compared.…”