Table 1 Three-step approach to sensitivity analyses for data missing at random versus missing not at random in randomised controlled trials

Step 1: Missing data patterns and mechanisms

- Determine the percentage of missing data at each time point

- Examine missing data patterns (e.g. plotting the outcome for each missing data pattern)

- Predict participation in follow-ups using baseline characteristics (e.g. logistic regression)

- Use other available information (e.g. process data)

Evidence to support assumptions about the process that lead to missing data (MAR versus MNAR)

Step 2: MAR models

- Determine the best-fitting shape of growth in preliminary models

- Calculate unadjusted growth model regressing the latent growth factors on the participants’ group assignment

- Add covariates to the model

Evidence about intervention efficacy under the assumption that data are missing at random

Step 3: MNAR models

- Generate missing data indicators

- Predict missing data indicators by outcome at time point t and the previous time point t-1 (Diggle-Kenward model)

- Predict missing data indicators by latent growth factors (Wu-Carroll model)

- Calculate the growth model for different subgroups that share the same missing data pattern (pattern mixture model)

- Compare the results to determine the sensitivity of the conclusions for different assumptions regarding the missing data

Evidence about intervention efficacy under the assumption that data are missing not at random