Back to Annual Meeting Page
133rd Annual Meeting & Exposition
December 10-14, 2005
William T. Robinson, PhD, HIV/AIDS Program, Louisiana Office of Public Health, 234 Loyola Ave, 5th floor, New Orleans, LA 70112, 504 568 5200, email@example.com and Leann Myers, PhD, Department of Biostatistics, Tulane University School of Public Health & Tropical Medicine, 1440 Canal St., 20th Floor, New Orleans, LA 70112.
There are many occasions in the analysis of public health data where the differences between group means are tested in the presence of a covariate. For example, the difference between men and women in ounces of alcohol consumed, controlling for age, may be of interest, but variability may be higher among males than females. Typically this type of analysis would call for the use of Analysis of Covariance (ANCOVA), which assumes that the variability within each group is equal across all groups. Unfortunately ANCOVA is not robust to violations of this assumption, particularly when group sample sizes are unequal. The current study demonstrates the extent to which violation of this assumption affects observed Type I error rates using traditional ANCOVA methods. Under conditions where smaller variances are paired with smaller group sizes the observed Type I error rates often drop dramatically, while under conditions where larger variances are paired with smaller group sizes the observed Type I error rates are highly inflated. Adaptations of the well-known Welch-Aspin (F*) technique for dealing with heterogeneity of variance in Analysis of Variance (ANOVA) as well as the computationally simpler Alexander's (A) statistic are presented and shown to be robust to violations of this type.
Keywords: Statistics, Simulation
Presenting author's disclosure statement:
I wish to disclose that I have NO financial interests or other relationship with the manufactures of commercial products, suppliers of commercial services or commercial supporters.
The 133rd Annual Meeting & Exposition (December 10-14, 2005) of APHA