统计问题第105问：荟萃分析（7）

Question

Researchers undertook a meta-analysis of the efficacy of tricyclic antidepressants in the treatment of migraine and tension-type headaches. Randomised placebo controlled trials were included if the active treatment group received a tricyclic antidepressant regularly at any dose as a single intervention for at least four weeks.

The primary outcome was burden of headache, typically measured by recording the frequency and intensity of headaches; larger values were indicative of greater burden. For each trial, the standardised mean difference in headache burden was derived because trials did not use the same method of measurement. The standardised mean difference was calculated as the burden in the active treatment group minus that in the placebo group. For each subgroup—tension-type headaches and migraine headaches—a subtotal estimate was derived, then an overall effect estimate was obtained for both subgroups combined.

The results of the meta-analysis were presented in a forest plot⇓.

• Forest plot of the effectiveness of tricyclic antidepressants (amitriptyline, clomipramine, amitriptylinoxide, opipramol, or doxepin) compared with placebo in reducing the burden of headache.

Which of the following statements, if any, are true?

·a) For both subgroups combined, the overall estimate of headache burden with tricyclic antidepressants was significantly different from that with placebo at the 5% level of significance.

·b) For both subgroups combined, statistical heterogeneity existed between the sample estimates of the population mean difference in the burden of headache.

·c) For the migraine headaches subgroup, statistical heterogeneity existed between the sample estimates of the population mean difference in the burden of headache.

提示：这是一道多选题。

Answer

Statements a and b are true, whereas c is false.

The overall estimate of the standardised mean difference in headache burden between patients on tricyclic antidepressants and those receiving placebo (tricyclic group minus placebo group) was −0.96 (95% confidence interval −1.39 to −0.53). The 95% confidence interval did not include zero—that is, no difference in the effects of tricyclics and placebo. Therefore, a statistically significant difference at the 5% level of significance existed between tricyclics and placebo in burden of headache (a is true), with a lower burden in patients using tricyclics. As described in last week’s question, because the total estimate was based on standardised mean differences, it may not be straightforward to interpret. It might have limited value as a measure of treatment effect and serve only as a qualitative measure of the strength of the evidence against the null hypothesis of no difference.

It was essential that the meta-analysis incorporated a statistical test of heterogeneity. The Higgins I² test statistic is used to ascertain the extent of heterogeneity between the sample estimates of the population mean difference in the burden of headache. The Higgins I² test statistic when both headache subgroups were combined was 90.1%. Because the value of the statistic was greater than 50%, statistical heterogeneity existed (b is true) and the variation between sample estimates was not simply a result of sampling error, as would be expected when taking samples from the same population. Since statistical heterogeneity existed, a random effects meta-analysis was undertaken to derive the overall estimate. If statistical homogeneity had been present, a fixed effects analysis would have been used. The random effects meta-analysis produced a wider confidence interval for the overall estimate than would have a fixed effects meta-analysis, resulting in an overall estimate with less precision.

It was essential to explore why heterogeneity existed between the sample estimates. Heterogeneity may have existed because the samples came from different populations or estimated different clinical outcomes. The trials were split into the subgroups of tension-type headaches and migraine headaches, and a subtotal estimate obtained for each. The Higgins I² test statistic was 95.4% for tension-type headaches and 25.0% for migraine headaches. Therefore, statistical heterogeneity existed between the sample estimates for tension-type headaches but not between those for migraine headaches (c is false). For both subgroups, the subtotal estimate for the difference between tricyclics and placebo in headache burden was statistically different at the 5% level of significance, with a lower burden in patients using tricyclics.

Statistical heterogeneity was found to exist between the sample estimates when the headache subgroups were combined. Therefore, the validity of combining the sample estimates was questionable. Subgroup analysis for migraine headaches showed homogeneity between the sample estimates, thus the subtotal estimate could be generalised, at least on the basis of the treatments, dosages, and populations studied. Heterogeneity existed between the sample estimates for the subgroup of tension-like headaches. Although random effects methodology dealt with the presence of heterogeneity in the calculation of the subtotal estimate, it is questionable whether these sample estimates should have been combined—it may not be obvious to what population, intervention, or outcome the results can be generalised.

所以答案是选择 a b

每天学习一点，你会更强大！‍

‍