Researches are undertaken ever-increasingly to find innovative treatments, and thus published papers are accumulating at an enormous rate. Meta-analysis enable previously published study results to be reviewed and summarized, and the meta-analysis methodologies differ depending on continuous or discrete outcomes. Ordinal outcomes are often found in medicine a few examples are the number of injurious fall, the degree of dementia or degree of pain. The summary statistics of these ordinal data are Odds Ratio(OR), Relative Risk(RR) and Risk Difference(RD). In this paper the generalized odds ratio, generalized relative risk and generalized risk difference (Edwardes and Baltzan, 2000)are proposed as summary statistics of the 2×R contingency table data and meta-analysis method based on generalized association estimates is proposed.
Recently, Whitehead et al.(2001) suggested logistic regression models for meta-analysis of 2×R contingency table data and these models require th proportionality of odds ratio. In this paper we compare meta-analysis methods based on the generalized estimates with logistic regression models, using the data of Whitehead et al.(2001, 1994) and also the fall data which were collected in order to evaluate the efficacy of exercise programs. The proportional odds assumption does not hold for some study data in the exercise intervention programs and thus the generalized estimates would be more suitable. Simulations were carried out based on the above 2×R contingency table data to evaluate association measures, tests of treatment efficacy and homogeneity. Even with these limited simulations it was shown that much used RD with arbitrarily dichotomizing ordinal data are not a recommended strategy and that logistic regression models present sometimes wrong estimates when the assumption of proportional odds ratio does not hold
|