Methodological Study On Meta-Analysis Of Ordinal Data Based Generalized Odds Ratio

BackgroundThe concept of Evidence-based medicine(EBM)was formally proposed in 1992 and became the focus of research in the medical field.Evidence-based medical systematic reviews based on randomized controlled trials(RCT)is considered to be the highest level of evidence,and is more and more used to provide guidance for clinical practice evidence for health care decisions.Meta-analysis is a statistical method for quantitative system evaluation in evidence-based medicine research.It is one of the sources of evidence-based medicine and the best scientific evidence.The development of evidence-based medicine relies on meta-analysis methodology.We know that ordinal data is very common in the health care field,but there is currently no effective meta-analysis method.The commonly used method is to split the ordinal data into two-category data or continuous data,which will undoubtedly result in loss of information and may even lead to wrong conclusions.The Generalized Odds Ratios(GOR)between the two groups of ordinal data proposed by Agresti in 1980 can be regarded as a generalization of the common odds ratio(OR)dealing with ordinal data,which can be used to describe the relationship between two groups of ordinal data.Predecessors derived combined effect size for GOR based on the inverse variance method(IV method),and show good statistical performance.ObjectiveThis study aims to enrich the weighting method based on GOR.We proposed derive Mantel-Hasenszel Method for GOR to obtain combined effect size.Based on the weight of MH method,we construct another one combined effect size,which called sample size weighting method.In order to obtain application strategies in different situations,Monte Carlo simulation will be used to compares the above three methods.MethodOnly for parallel design odinal data,the combined effect size and its 95%confidence interval were constructed based on the inverse variance method,the MH method and sample size weighting method.Monte Carlo simulation was used to investigate the statistical performance of the point estimation and the interval estimation of combined effect size obtained by the three weighting methods.The point estimation indicators include average deviation,average relative deviation and mean square error.The interval estimation indicators include coverage probability of 95%Cl and non-coverage ratio on both sides.ResultIn the case of small sample size:when the ratio is relatively extreme,the average deviation of the three methods increases with the increase of the effect size,and the coverage probability of the confidence interval decreases with the increase of the number of studies;when the ratios are relatively balanced,the deviation of the sample size method has the smallest deviation;in the case of medium samples size:when the ratio is relatively extreme,the MH method has the smallest deviation;when the ratio is relatively balanced,the inverse variance method has the smallest deviation;in the case of large samples size,The results of the MH method and the inverse variance method are similar and perform well.ConclusionThis study uses mathematical derivation and simulation studies to compare different GOR meta-analysis,which MH method and simple size method proposed in this paper and the inverse variance method proposed by the predecessors.According to the simulation results,the following conclusions are drawn.In the case of small sample size,when the ratio is relatively extreme,the deviations of the three methods are large and increases with the increase of the effect size,when the ratio of each level is relatively balanced,the deviation of MH method is small and the coverage probability of the confidence interval can be controlled near the set level,the MH method is recommended.In the case of medium sample size,when the ratio is relatively extreme,the MH method is recommended,when the ratio is relatively balanced,the inverse variance method is recommended.In the case of large sample size,the MH method and the inverse variance method perform well and both can be applied.