Compared below is the ITT vs. Weighted Per-protocol API Comparisons.
This analysis was a weighted and replicated GEE with treatment weights, stage 1 and stage 2 weights. So a heavy drinker who participated in both stages had three weights multipled (treatment, stage 1 and stage 2). A light drinker who participated at stage 1 had two weights multiplied (treatment and stage 1). Then I ran the exact ITT analysis (weighted and replicated GEE) with these new weights.
Light drinker weights: \(\frac{I(C_1=1)}{Pr(C_{1}|A_1,\textbf{X})*.5}\)
Heavy drinker weights: \(\frac {I(C_1=1,C_2=1)}{Pr(C_2=1|\textbf{X},C_1=1)Pr(C_1=1|A_1,\textbf{X})*.25}\)
| Early/Coach | Early/Email | Late/Coach | Late/Email | P-value | Early/Coach | Early/Email | Late/Coach | Late/Email | P-value | |
|---|---|---|---|---|---|---|---|---|---|---|
| Binge drinking | ||||||||||
| ITT | 2.05 (0.28) | 1.65 (0.25) | 1.57 (0.23) | 1.59 (0.24) | 0.422 | 3.19 (0.47) | 2.58 (0.36) | 2.01 (0.30) | 2.21 (0.30) | 0.148 |
| Weighted Per-protocol | 2.10 (0.46) | 1.67 (0.41) | 1.40 (0.40) | 1.23 (0.33) | 0.446 | 2.96 (0.67) | 2.50 (0.52) | 1.84 (0.51) | 2.14 (0.47) | 0.595 |
| Alcohol consequences | ||||||||||
| ITT | 1.61 (0.15) | 1.64 (0.16) | 1.68 (0.16) | 1.63 (0.17) | 0.987 | 0.83 (0.12) | 0.95 (0.14) | 0.99 (0.15) | 0.79 (0.14) | 0.530 |
| Weighted Per-protocol | 1.61 (0.26) | 1.46 (0.19) | 1.66 (0.28) | 1.62 (0.33) | 0.927 | 0.74 (0.18) | 0.81 (0.19) | 0.94 (0.29) | 0.87 (0.24) | 0.946 |
| Early (SE) | Late (SE) | Ratio of ratios (95% CI) | P-value | Early (SE) | Late (SE) | Ratio of ratios (95% CI) | P-value | |
|---|---|---|---|---|---|---|---|---|
| Binge drinking | ||||||||
| ITT | 1.84 (0.21) | 1.58 (0.19) | 1.17 (0.84, 1.61) | 0.347 | 2.87 (0.35) | 2.10 (0.24) | 1.36 (0.99, 1.88) | 0.059 |
| Weighted Per-protocol | 1.87 (0.33) | 1.32 (0.27) | 1.42 (0.83, 2.43) | 0.198 | 2.72 (0.47) | 1.99 (0.36) | 1.37 (0.84, 2.23) | 0.209 |
| Alcohol consequences | ||||||||
| ITT | 1.62 (0.13) | 1.65 (0.13) | 0.98 (0.79, 1.23) | 0.876 | 0.89 (0.10) | 0.89 (0.12) | 1.00 (0.70, 1.43) | 1.000 |
| Weighted Per-protocol | 1.53 (0.17) | 1.64 (0.23) | 0.94 (0.66, 1.33) | 0.709 | 0.77 (0.14) | 0.90 (0.20) | 0.86 (0.49, 1.51) | 0.597 |
I did another analysis for stage 1 which was not a replicated GEE. For the students randomized to an API, I ran a geeglm model with log link and a exhangeable correlation structure with only stage 1 weights, since technically I don’t need treatment weights to compare the timing interventions. I was looking at significannce of the time1:A1 and time2:A1 interactions. I did this for binge drinking frequency and alcohol consequences as the outcome.
This analysis used the exact same model (weighted and replicated GEE with the same weights) as the analysis comparing the four adaptive interventions.
I also looked at the subset of heavy drinkers and did a geeglm with a log link and exchangeable correlation structure with stage 2 weights. I was looking at significannce of the time1:A2 and time2:A2 interactions. I did this for binge drinking frequency and alcohol consequences as the outcome.
I did not do a replicated GEE for this analysis. I took all the subjects, gave a weight of 1 to the control subjects and used the following weights for students in the API. - Light drinker weights: \(\frac{I(C_1=1)}{Pr(C_{1}|A_1,\textbf{X})}\) - Heavy drinker weights: \(\frac {I(C_1=1,C_2=1)}{Pr(C_2=1|\textbf{X},C_1=1)Pr(C_1=1|A_1,\textbf{X})}\)
I decided to model the probabilities of compliance at Stage 1 and Stage 2 separately because there could be time-varying confounding. Then I ran a gee with exchangeable correlation and a log link. I looked at the API*time interactions.