Reader Comments

Post a new comment on this article

Different effects on reading scores in subgroups?

Posted by pspjxr on 14 Nov 2012 at 10:55 GMT

The results reported for reading scores show a statistically significant effect in favour of DHA for the subgroup with baseline reading score ≤20th centile. Within that subgroup the effect appears to be due mainly to the subgroup with baseline reading score ≤10th centile. In fact, calculations that can be made using the data provided in the text (and repeated in Figure 2 and Table 4) appear to show that for the subgroup comprising the ≤20% to >10th centile there was a negligible benefit of DHA, and that for the subgroup comprising ≤33% to >20th centile the effect was, if anything, in favour of placebo. (These calculations assume, among other things, equal or nearly equal numbers of active- and placebo-supplemented participants in these subgroups.)

It would be helpful if the authors would provide the results to illustrate this interaction; that is, the mean change scores for the children in the ≤33% to >20th centile following active and placebo supplementation (total n=138) and the same for the children in the ≤20% to >10th centile (total n=119) and test whether these differ from the scores of the children in the ≤10th centile subgroup. Recent discussions highlight the importance of testing for an interaction to demonstrate differences between subgroups, even where subgroup analyses are pre-planned (Brookes et al, 2004).

Reference: Brookes et al, 2004. Subgroup analyses in randomized trials: risks of subgroup-specific analyses: power and sample size for the interaction test. Journal of Clinical Epidemiology, 57, 229-236.

Comment posted by Peter J Rogers (University of Bristol, UK) and Katherine M Appleton (Bournemouth University, UK).

No competing interests declared.

RE: Different effects on reading scores in subgroups?

AJRichardson replied to pspjxr on 11 Feb 2013 at 11:51 GMT

Many thanks for your comment. We share your concern for the subgroup effects and the problems when testing their significance. Conducting a post-hoc interactiontest for the separate subgroups (33%-20%; <20%-10%, <10%) demonstrates that the treatment effect on reading seems concentrated on the arbitrarily defined <10% group of readers (Result for the interaction effect: Active*<10%group = 3.170***(SE: 1.115)). However, we stress the somewhat contrived nature of the 10% grouping here, as only the analysis of the <20% subgroup was pre-planned. We reported the results for the 10% group to give at least an indication that the treatment effect was potentially due to the poorest readers, in line with an “at-risk” principle. Essentially however, the result is for the <20% readers is the most rigorous, that is unbiased result, as it was pre-planned. A closer responder analysis might, for example, reveal that the treatment effect was due to the bottom 11% or 9% readers. Thus such a result would be open to a criticism of “cherry picking”. Instead we will be testing this question by investigating potential differential subgroup effects in a replication study currently under way.

As requested we provide below the mean reading change scores by separated subgroups.

Reading Change Scores:
Subgroup: Mean : S.E.: N:
Total Sample:
below 10% 1.91 .434 105
10%-20% 0.99 .336 119
20%-33% 1.17 .405 138
Treatment Group:
below 10% 3.06 .638 49
10%-20% 1.16 .489 61
20%-33% .671 .553 70
Control Group:
below 10% .911 .563 56
10%-20% .845 .464 58
20%-33% 1.69 .59 68
Total 362

Competing interests declared: Corresponding author of the article