Abstract The absence of an objective disease biomarker puts studies of Myalgic Encephalomyelitis/Chronic Fatigue Syndrome (ME/CFS) under the curse of imperfect diagnosis. This problem leads to frequent reports that fail to reproduce prior published studies. To address the impact of imperfect diagnosis on the robustness of studies' conclusions, we conducted a simulation study to quantify the statistical power to detect a disease association with a hypothetical binary factor in the presence of imperfect diagnosis. Using the classical case-control design, studies with sample sizes of less than 500 individuals per group could not reach the target power of at least 80% to detect realistic disease associations. We then recreated serological association studies in which the chance of imperfect diagnosis was combined with the probability of misclassifying a binary factor, as it happens in a typical serological association study. In this case, the target power of 80% could only be achieved for studies with more than 1000 individuals per group. Given the current sample sizes of ME/CFS studies, our results suggest that most studies are likely to be underpowered due to imperfect diagnosis alone. To increase reproducibility across studies, we provided some practical recommendations, such as the use of standard case definitions together with multi-centric study designs, and routine reporting of power calculations under a non-negligible chance of misdiagnosis. Our results can also inform the design of future studies under the assumption of misdiagnosis. https://www.medrxiv.org/content/10.1101/2022.06.08.22276100v1
Very useful observation, and good to have some statistical modeling. I've been thinking something similar for a while. My guess is that the misdiagnosis rates may be as high 25% even in studies conducted by people who are familiar with the topic. The remaining ME/CFS cases may actually be multiple diseases too. Under these circumstances you really need large studies to make progress, but the funding for that hasn't been there. ME/CFS may be a problem that cannot be solved without very large studies that are able to clearly distinguish between subgroups. We don't know for sure, but it's a possible reason for the lack of meaningful progress.
With the 80/20 assumptions there should still be an effect, though the sensitivity/specificity will obviously be unimpressive if there are important subgroups or misdiagnosis. In general, I am reminded of Lakens 'justify your alpha' approach - focusing on the smallest effect size of interest and then considering reasonable sample sizes and set the alpha based on a reasonable ratio of false positives vs false negatives. https://lakens.github.io/statistical_inferences/ The preprint discusses sample sizes of 1000+, but that is for serological association studies where the effect size might be quite small (genetic association studies have similar problems). Other studies don't require such large samples because the effect size of interest is much larger.
The paper states: "Using the classical case-control design, studies with sample sizes of less than 500 individuals per group could not reach the target power of at least 80% to detect realistic disease associations" But I suspect that power is still mostly determined by effect size so that the 500 individuals per group probably only applies for small to medium effects (odds ratio < 3) and that for large effects you would need far less participants even if there is substantial misdiagnosis.
This is shockingly bad. I am minded to write to the journal to demand an urgent correction. It should be fewer than 500 individuals, not less than!
