Sublingual feverfew/ginger (LipiGesic M) reanalysis of data.

Sublingual Feverfew/Ginger (LipiGesic M) Reanalysis of Data In July of 2011, a pilot double-blind, placebocontrolled multicenter study of a homeopathic combination of sublingual feverfew and ginger (LipiGesic M) in the treatment of acute migraine was published in Headache. Enrolled subjects met the International Headache Society criteria for migraine, with 2-6 attacks per month, and were randomized to treat all migraine attacks over a 1-month time period with either sublingual feverfew/ginger (LipiGesic M) or placebo 3:1. Subjects were asked to treat migraine early, and were encouraged but not required to treat when the headache was mild. Data were analyzed from 59 patients treating 221 migraine attacks. Subjects could take a rescue medication after 2 hours. The study was intended to parallel clinical practice. The statistical methods employed for the analysis included descriptive statistics to establish mean and standard deviation for each primary and secondary end point. A 2-way analysis of variance was conducted to measure the significance of response comparing active and placebo groups. We concluded that sublingual feverfew/ginger (LipiGesic M) was statistically superior to placebo in relief of headache, and pain-free response (prespecified primary end point) and associated migraine symptoms (secondary end points).The product was well tolerated. In the article, it was suggested that this product, as an early intervention, might be a unique and valuable addition to the treatment armamentarium of many migraine patients. Subsequent to publication, a colleague contacted us to discuss study design and statistical analyses performed in this study. He pointed out that while similarly designed studies have been used previously in migraine trials, there is concern for potential statistical errors based on the fact that multiple attacks per subject were included in the data analysis without analysis for within-subject repeated measures. This, in turn, may have artificially distorted the efficacy of the active compound. In an effort to clarify this concern, we agreed to have the study data reanalyzed using statistical modeling to account for repeated measures. Daniel Serrano, PhD, provided the reanalysis of the results. Dr. Serrano’s comments and reanalysis are as follows. “In the multiple-attack data, attacks per subject ranged from 1 to 6. Ignoring the multiple observations per subject distorts estimates of degrees of freedom by not differentiating between number of subjects and number of observations, commonly causing liberal statistical tests and inflated rates of efficacy detection. In the context of repeated measures, many degree of freedom estimators have been developed, some giving fractional estimates (Satterthwaite, and Kenward and Roger), which, while at first odd-seeming, provide the best estimates in certain conditions. Regardless, the simplest degree of freedom estimates in repeated measures are calculated as a function of the total independent pieces of information in an analysis minus the number of estimated parameters (k). For example, with 10 subjects each having 5 repeated measures, degrees of freedom could be calculated as 50-k, ignoring the within-subject dependence of repeated measures, or as 10-k, attending to the repeated measures data. “In the case of the LipiGesic M trial data, with N = 59 independent cases in the trial, a basic model including only the intercept and treatment effect should have 57 degrees of freedom.Were the repeated measures nature of the LipiGesic M trial data ignored, the degrees of freedom would be calculated using the total number of available data points (n = 221), resulting in 221-2 = 219 degrees of freedom for the same basic model – the issue of ignoring repeated measures within subject is, thus, rendered quite astonishing. In addition, standard errors that are denominated as a function of the independent pieces of information contributing to a given test will be excessively shrunken when the repeated measures nature of the data is ignored. This increases the odds of false positive findings for efficacy. In combination, the effects of liberal degrees of freedom and standard error estimates on statistical inference when repeated measures are ignored elucidate why it is essential to use models that can accommodate repeated measures data when seeking efficacy detection in multiple-attack trials. “A simple procedure for doing so is to use the generalized linear mixed model (GLMM) or generalized estimating equations (GEE), which are an extension of the generalized linear model. Herein I focus on the GLMM because it tends to provide better estimates of variance, is more robust to missing data, and can accommodate a wider