Subset analyses

There is inevitably some subjectivity in setting the question to be posed in a systematic review (see Section 11.5.1) and there is likely to be a trade-off between pooling only very similar trials, and achieving high statistical power. One way of dealing with this dilemma is to group trials by important characteristics in subset analyses. By grouping similar trials together we are attempting to minimize the clinical and statistical variability within the subsets of trials. This approach allows us to look for consistency and inconsistency between subsets of trials. We are hoping to partition any overall heterogeneity in such a way that it is explained mostly by differences between subsets of trials rather than by variation within subsets (see also Section 11.6.2). Individual meta-analyses are done and pooled results calculated, for each subset of trials, and if appropriate, an overall pooled result calculated for all trials. For example, in an IPD meta-analysis of chemotherapy in non-small cell lung cancer [51], it was decided a priori to split the trials according to the type of chemotherapy that was used. This was because it was thought that older chemotherapies might be less effective than newer regimens. In fact the results indicated

Hazard Ratio

Long-term alkylating agents

MRC LU02 VASAG EORTC 08741 VASOG 5 WP-L7351 Total 1670 events / 2145 patients

Other drugs

OLCSG 1 (a) OLCSG 1(b) SGACLC ACTLC1 WJG 2 (2&3)

Total 338 events / 918 patients

Cisplatin-based

LCSG 801 OLCSG 1(c) FLCSG 1

SGACLC ACTLC2 IPC Chiba WJG 2 (1&3) LCSG 853 JLCSSG

Total 614 events / 1394 patients

Surgery+CT better

Test for interaction chi-square= 10.97, p =

Surgery better 0.004

Fig. 11.11 Surgery+chemotherapy versus surgery alone in non-small cell lung cancer-survival analysis. Modified and reproduced with permission from [50].

that rather than just the quantitative differences that were anticipated, there was evidence of a qualitative difference (difference in direction of effect) between the types of chemotherapy. Long-term administration of oral alkylating agents appeared to be detrimental, whereas other more modern types of cisplatin-based chemotherapy were suggestive of a benefit. (Fig. 11.11).

The top subset of trials using long-term alkylating agents lie to the right-hand side of the line and the pooled result for this subset of trials indicated by the hatched diamond immediately below indicates a significant detriment of the intervention. In contrast, the bottom subset of trials using cisplatin-based chemotherapy lie mostly to the left hand side of the line and the pooled estimate denoted by the hatched diamond immediately below this group indicates a result in favour of chemotherapy, although this is not conventionally statistically significant. The results for these subsets of trials are clearly different, shown formally by the chi-square test for interaction (see Section 9.4.8), which is conventionally significant. In this case it is not sensible to combine the results of all trials, statistically it would be improper and clinically it would be meaningless, so there is no solid black diamond at the foot of the plot. In other circumstances, where there is no difference between subsets of trials it maybe appropriate to show an overall pooled result.

With this type of analysis, because trials are simply grouped by trial level characteristics, it is possible to subset trials irrespective of whether summary or individual patient data are used. Planned subset analyses should, of course, be pre-specified in the systematic review protocol.

5 Easy Ways To Stop Smoking

5 Easy Ways To Stop Smoking

Your first day without cigarettes can be difficult, but having a plan will make it easier! Learn what steps to take on the day you quit smoking.

Get My Free Ebook


Post a comment