Application of Excess Zero Methodology to Oral Health-Related Quality of Life: PEARL Network Findings

Publication Type
Conference Paper
Year of Publication
Matthews, A; Vena, D
Society for Clinical Trials 32nd Annual Meeting
Date Published
Vancouver, BC, Canada
Oral health research has recently focused on evaluating quality of life (QoL), with the most common tool being the Oral Health Impact Profile (OHIP). The score for the 14-item version of the OHIP survey is the total number of impacts reported (0-56), but is not particularly sensitive to small changes in QoL, which results is many study participants reporting zero impact. One dental practice- based research network, Practitioners Engaged in Applied Research and Learning (PEARL), utilized the OHIP-14 in four of its studies. All 4706 scores were highly overdispersed using multiple tests, including the Fisher overdispersion test. Over 65% of subjects reported 0 impacts, and the distribution of scores has been shown to differ significantly across protocols. Exploration of poisson, negative binomial, zero-inflated poisson (ZIP) and zero-inflated negative binomial (ZINB) models indicated the best fit of the data was the ZINB using AIC. However, in this model the dispersion is highly significant, indicating that there is residual overdispersion. For each of the models, protocol was highly significant (all p<0.001), as anticipated. An alternative approach (Lachenbruch, 2001) is to consider a two-part composite statistic: a test of the difference in the proportion of subjects reporting zero impacts across protocols, and a test of the difference in the magnitude of reported non- zero scores across protocols. For the first component, we utilize the 3df chi-square test of association from the corresponding 2x4 contingency table. The second component utilizes the 3df chi-square global test of association assuming a negative binomial model, which indicated no overdispersion (Pearson s chi-square/DF<2). As expected, the 6df composite chi-square test was highly significant (p<0.001). These results are particularly interesting in that once the zero impacts are excluded, the data fit a NB model indicating the overdispersion of the total score was likely due to excess zero counts and not heterogeneity.