Skip to main content

Table 4 Approximation of placebo-arm findings in RCTs as a result of “natural” remission of CM

From: Fluctuations in episodic and chronic migraine status over the course of 1 year: implications for diagnosis, treatment and clinical trial design

 

EM

CM

Negative Binomial Mixed-Model for Change From Baseline to Follow-Upe

1-Way ANOVA for Headache Frequency Change Score From Baseline to Follow-Up

Time Contrast

 

Baseline

Follow-Upb

Change Score

 

Baseline

Follow-Upb

Change Score

Baseline Headache Status by Time Interaction Random-Effect Model-Based

Baseline Headache Status ANOVA Effect

 

Sample sizea

μ1

μ2 d

μΔ

n

μ1

μ2 d

μΔ

Rawf

RR

P value

Mean Differencec

P value

Wave 1 vs. Wave 2

7759

3.98

4.14

0.16

829

19.77

14.61

−5.16

0.71

0.67 (0.64 to 0.71)

<0.0001

−5.32 (−5.69 to −4.95)

<0.0001

Wave 1 vs. Wave 5

4407

4.05

4.09

0.05

502

19.80

13.69

−6.12

0.68

0.64 (0.59 to 0.69)

<0.0001

−6.17 (−6.69 to −5.64)

<0.0001

  1. ANOVA analysis of variance, CM chronic migraine, EM episodic migraine, HA headache, RCT randomized controlled trial, RR rate ratio, SAS statistical analysis system
  2. aSample sizes were based on complete case data, and are therefore lower for the second time contrast than the first because less subjects chose to respond to the survey by wave 5 than wave 2. The combined sample sizes are less than the wave 2 or wave 5 sample sizes respectively (i.e., 4407 + 502 = 4909, which is <5915) because the complete case data restriction also required the same subjects contribute data on the past month HA frequency variable at both the baseline and follow-up assessments
  3. bBaseline Headache status by time interaction estimation: (CM_ μ2/ CM_ μ1)/(EM_ μ2/ EM_ μ1)
  4. cBaseline Headache status ANOVA mean difference: CM_ μΔ – EM_ μΔ
  5. dRandom-effect model-based estimate parameterized with a random subject-specific intercept, fixed intercept, main effect for headache status (EM vs. CM, EM reference), main effect for time contrast (baseline vs. follow-up, baseline reference), and the interaction between headache status and time
  6. eRandom-effect negative binomial models were fit in SAS’ GLIMMIX procedure. The model estimation was identical to the primary models presented in this manuscript and described in the methods section with adaptive Gauss-Hermite quadrature based on 13 quadrature points and 9 pseudo-likelihood initial iterations for start values, with no generalized linear model–based iterations
  7. fRaw means, raw RRs, and random-effect model-based rate ratios indicate that HA frequency for CM declines over time while HA frequency for EM remains stable. Moreover, CM declines more as the amount of time between assessments increases (2 waves vs. 5). Specifically, with only three months between assessment, HA frequency declines 33% more for CM than EM (100*[1–0.67]), while HA frequency declines 36% more for CM than EM (100*[1–0.64]) when the time between assessments is 12 months