Multivariate fixed- and random-effects models for summarizing ordinal data in meta-analysis of diagnostic staging studies

For many diseases (e.g. rectal cancer and the Crohn disease), more than two stages exist and as treatment mostly depends on disease stages, correctly determining this by a diagnostic test is very important. To determine their role in clinical practice, the value of these tests should be carefully evaluated, and summarizing results in meta-analysis should also be done appropriately. A multinomial model for meta-analyzing data with more than two categories has previously been developed; these data were considered as nominal categories. However, there is an ordinal character within staging data. In this study we extended this multinomial model to three ordinal models (models for the logits of adjacent-categories, for continuation-ratio logits and for proportional odds logits) to summarize the ordinal character of staging data. Both fixed-and random-effects approaches were developed and compared. The principles of the multinomial model as well as three ordinal models are shown by fitting these models using the data on staging of rectal cancer by endoluminal ultrasonography and magnetic resonance imaging. The proportions of patients correctly staged, understaged, and overstaged per stage are obtained by these models. Because of the increased interest in meta-analyses for evidence-based guidelines, these models can be helpful in summarizing staging data. Copyright (C) 2010 John Wiley & Sons, Ltd.