Subgroup analyses

When assessing the role of any new treatment, an important additional question is whether the treatment is equally effective in well-defined subgroups of patients. For example, is the treatment more or less effective in males or females, or in old or young patients? As discussed in Chapter 9, the results of subgroup analyses can be very misleading because of the multiplicity of testing and there is a high probability that any observed difference is due solely to chance (see Section

In individual trials it is unusual to have sufficient numbers and statistical power to permit reliable subgroup analyses. However, provided that such data have been collected uniformly across studies, a meta-analysis of all trials may achieve sufficient power in each subgroup to permit a more reliable exploration of whether the effect is, in fact, larger (or smaller) for any particular type of patient. Although still potentially misleading, subgroup analysis within the context of a large meta-analysis may be the only reliable way of performing such exploratory investigations. Not only do the greater numbers give increased statistical power, but we can also look for consistency across trials. In all types of systematic review any subgroup analyses should be pre-specified and stated clearly in the review protocol.

In contrast to what we have termed subset analysis, which groups whole trials, subgroup analysis groups individuals according to characteristics such as age, sex or tumour stage. In a meta-analysis setting, subgroup analyses are carried out for each trial and then the trial level summary statistics for each subgroup are combined to give an overall pooled effect. For example, the effect of treatment compared to control is calculated for men, and the effect of treatment compared to control is calculated for women, within each trial. The individual trial results for men can then be combined to give a pooled estimate of treatment effect for men and the same done for women.

In conventional meta-analyses using aggregate data from publications, it is usually extremely difficult to extract sufficient compatible data to undertake meaningful subgroup analyses. It is unlikely that separate results would be presented for each of the characteristics of interest (e.g. presenting separate results for men and for women, for stage II stage III and stage IV tumours, etc.) for every trial. Meta-regression has been proposed as a method of identifying significant relationships between the treatment effect and covariates of interest, based on the type of information that maybe available in a trial publication. However (in the absence of IPD) the unit of regression is restricted to the trial and there are many problems with the approach (see Section 11.6.3). Consequently, it is unlikely that practically useful analysis of subgroup data will be possible in the context of a systematic review based on just summary data extracted from publications.

Subgroup analyses can be done using tabulated data supplied by trialists, for example, if separate data tables for men, for women and for each specific age group are provided for each trial. It is worth noting that this could be extremely time consuming for trialists, especially if the necessary cross-tabulations were not done for their own analyses. It is conceivable for a small trial that a trialist may have to generate as many data tables as there are patients. To look at categories of patients defined by more than one baseline characteristic, for example, post-menopausal women with oestrogen-receptor-positive and node-negative status, is likely to prove impractical for both trialist and reviewer.

In contrast, IPD permits straightforward categorization of individuals for subgroup analysis defined by single or multiple factors. For example, because age is collected for each individual patient, it is straightforward to categorize patients into the same age bands in all trials (say younger than sixty-five and sixty-five years and older) and then to look at whether the effect of treatment differs between the older and younger age group. It is important to note that these subgroup analyses, based on IPD, remain stratified by trial.

When interpreting the results of subgroup analysis plots, we should look at the overall pattern of results and note firstly to what extent the confidence intervals for each of the subgroups overlap. Where there are large overlaps there is generally no indication of a difference in the effect of treatment in the subgroups. This is tested formally by the test for trend or interaction (see Section 9.4.8). It is important not to use subgroup analysis to focus on individual subgroups of patients where the result for that group reaches significance. For example, in the hypothetical example below (Fig. 11.12) it would be incorrect and unwise to conclude that the investigational treatment worked in those with brown eyes, but not in those with blue or green eyes. There are simply larger number of individuals with brown eyes and there is no indication that the treatment is any more or any less effective according to eye colour.

Hazard Ratio

Blue eyes

Chemotherapy better Chemotherapy worse

Fig. 11.12 Hypothetical example of a forest plot for subgroup analysis.

0 0

Post a comment