Article type
Year
Abstract
Background:
While conducting a meta-analysis researchers are occasionally faced with the problem that the relevant endpoint is reported in some studies as a continuous endpoint (e.g. group specific mean on a rating scale) and in other studies as a dichotomous endpoint (e.g. group specific number of responders), which makes the synthesis of primary trials difficult. Hence, effect sizes derived from dichotomous endpoints (e.g. odds ratios) are estimated out of effect sizes derived from continuous endpoints (e.g. standardized mean differences). Yet the reliability of currently available methods is insufficiently evaluated.
Objectives:
The objective is to empirically evaluate the performance of five methods to estimate both odds ratios and the corresponding standard errors.
Methods: A database of randomized controlled trials searched and selected for earlier meta-analyses on the effectiveness of psycho therapeutic, pharmacological, or combined treatments for chronic depression was used. Only those trials of the database were considered that reported both continuous and dichotomous endpoints. For all included trials odds ratios and the corresponding standard errors were calculated and additionally estimated using one of the five methods by Hasselblad & Hedges (H&H), Cox & Snell (C&S), Furukawa (F), Suissa (S) and Kraemer & Kupfer (K&K). Observed and estimated values were compared on a trial level and on the level of meta-analyses using relative odds ratios and percentage deviations.
Results: A total of 26 trials were included. On a trial level four out of five methods performed well in estimating odds ratios out of continuous endpoints (H&H, C&S, F, S). Regarding standard errors three out of four methods considerably underestimated the standard errors (H&H, C&S, F), whereas one method (S) reliably estimated the standard errors. On the level of meta-analyses all overall effect sizes were comparable with overlapping confidence intervals. Again, the method by Suissa performed best.
Conclusions: Odds ratios of response and the corresponding standard errors can reliably be estimated out of continuous endpoints with the method by Suissa performing best.
While conducting a meta-analysis researchers are occasionally faced with the problem that the relevant endpoint is reported in some studies as a continuous endpoint (e.g. group specific mean on a rating scale) and in other studies as a dichotomous endpoint (e.g. group specific number of responders), which makes the synthesis of primary trials difficult. Hence, effect sizes derived from dichotomous endpoints (e.g. odds ratios) are estimated out of effect sizes derived from continuous endpoints (e.g. standardized mean differences). Yet the reliability of currently available methods is insufficiently evaluated.
Objectives:
The objective is to empirically evaluate the performance of five methods to estimate both odds ratios and the corresponding standard errors.
Methods: A database of randomized controlled trials searched and selected for earlier meta-analyses on the effectiveness of psycho therapeutic, pharmacological, or combined treatments for chronic depression was used. Only those trials of the database were considered that reported both continuous and dichotomous endpoints. For all included trials odds ratios and the corresponding standard errors were calculated and additionally estimated using one of the five methods by Hasselblad & Hedges (H&H), Cox & Snell (C&S), Furukawa (F), Suissa (S) and Kraemer & Kupfer (K&K). Observed and estimated values were compared on a trial level and on the level of meta-analyses using relative odds ratios and percentage deviations.
Results: A total of 26 trials were included. On a trial level four out of five methods performed well in estimating odds ratios out of continuous endpoints (H&H, C&S, F, S). Regarding standard errors three out of four methods considerably underestimated the standard errors (H&H, C&S, F), whereas one method (S) reliably estimated the standard errors. On the level of meta-analyses all overall effect sizes were comparable with overlapping confidence intervals. Again, the method by Suissa performed best.
Conclusions: Odds ratios of response and the corresponding standard errors can reliably be estimated out of continuous endpoints with the method by Suissa performing best.