Statistical analysis for experimental models of ocular disease: Continuous response measures

Abstract

Experimental designs in ophthalmologic research frequently treat both eyes of a subject in the same fashion: e.g., therapy with a specific drug or control. In these two-eye designs, observations from the same subject are often positively correlated. Failure to account for this correlation is a serious error which overstates the precision of studies, resulting in falsely significant results. This paper reviews the statistical methods appropriate for studies where endpoints are quantitative. We present: (1) the use of analysis of variance (t-test when there are 2 treatment conditions) to estimate differences between all experimental treatments, (2) the use of contrasts to estimate differences between specific treatments, and (3) methods for analysis of data from multiple experiments. Because of the ubiquity of incorrect analysis of data from two-eye designs in the ophthalmologic research literature and the serious consequences of this error, we propose a limited statistical review of manuscripts to ascertain if the statistical analysis matched the experimental design.