You are here:
Evaluation of a weighted test in the analysis of ordinal gait scores in an additivity model for five OP pesticides.
Citation:
ROBINSON, A., C. GENNINGS, AND V. C. MOSER. Evaluation of a weighted test in the analysis of ordinal gait scores in an additivity model for five OP pesticides. Presented at Society of Toxicology, Seattle, WA, March 16 - 20, 2008.
Impact/Purpose:
A CRL (continuation-ratio logit) additivity model was developed and may be compared to mixture points of interest to test for interaction.
Description:
Appropriate statistical analyses are critical for evaluating interactions of mixtures with a common mode of action, as is often the case for cumulative risk assessments. Our objective is to develop analyses for use when a response variable is ordinal, and to test for interaction in a mixture of chemicals. The typical cumulative logit model for ordinal data assumes proportional odds; when that assumption is violated, a continuation-ratio logit (CRL) model may be used. A CRL additivity model was developed and may be compared to mixture points of interest to test for interaction. The study exposed mice to an environmentally-relevant mixture of five organophosphate pesticides with Gait Score ranked in four categories: (1) no, (2) mild, (3) moderate, (4) severe abnormalities. For these data the proportional odds assumption was violated and a CRL model was used. With the CRL model, the multinomial distribution is factored into independent conditional binomial analyses so that standard software for binary data can be used. Generally, as the ordinal categories progress, the sample size in each analysis decreases. In an unweighted analysis, each independent statistic has equal weight although the sample sizes may not be equal. A weighted statistic was developed with more weight given to analyses with higher sample sizes and less weight to analyses with fewer observations. Both the weighted and unweighted analyses of the data revealed synergy at both a low and moderate dose. A simulation study based on these data was performed to evaluate the power of the weighting scheme with varying degrees of interaction. The power of the test depends on how extreme the probability distribution is for additivity, which is dose-dependent. The weighted statistic had greater power partly due to the decrease in the weighted degrees-of-freedom for the test.