One of Thase’s responses to questions regarding the effect size calculation was that MMRM is becoming “rapidly adopted as the method of choice for clinical trial data sets” – however, I have not seen it used to the extent that I’d call it the “method of choice.”
The main issue remains: the apparent magnitude of the effect of Seroquel on depressive symptoms increased by 50% when MMRM was used. That makes the treatment look more effective than when the more conventional method is used. Thase responded that “we did not report effect sizes using the [traditional] LOCF method largely because we didn't anticipate that someone would actually want to see them” – why wouldn’t people want to see them – given that the MMRM method inflated the apparent effect of the treatment by 50%, this is certainly news worth reporting in the article!
Mind you, the interpretation of effect sizes has always been based on traditional methods – we have no idea if the interpretation of effect sizes based on MMRM is accurate. I can buy into using MMRM as a method of testing statistical significance (was the treatment more effective than a placebo) but not into using MMRM to generate effect sizes (how much more effective was the treatment than a placebo). When a change of 50% can be seen in the magnitude of treatment depending on the type of analysis used, it is controversial and requires further explanation.
Thase has done a lot of good work and this post should not be taken as a personal attack on him. Indeed, I was glad to see him respond to Dawdy -- that is a good sign of collegiality and openness and he is commended for responding. I strongly disagree with the way data were presented and I don’t buy Thase’s argument that the traditional analysis was just not interesting enough to publish.