From: Analysis of randomised trials with long-term follow-up
Problem | Does the problem affect acute, intermittent or sustained interventions? | What causal effects can be estimated? | Methods | Methods suitable for cross-sectional or longitudinal analyses? |
---|---|---|---|---|
Non-compliance | most commonly affects sustained or intermittent interventions, but can also affect acute interventions | the complier average causal effect (i.e., the average effect in people who would adhere to whichever treatment they were assigned to) | instrumental variable regression, propensity score methods | when there is a summary measure of compliance across the entire treatment period, both methods can be applied to cross-sectional or longitudinal analyses |
Treatment switching | affects intermittent or sustained interventions | the average effect in people who do not switch treatments (or the average effect of some other defined treatment pattern) | marginal structural models, g-estimation | can be applied to cross-sectional analyses, but are most valuable when analyses are longitudinal |
Loss to follow-up | this is not a treatment-level issue: loss-to-follow-up can occur with any type of intervention | the average treatment effect | mixed longitudinal models, multiple imputation | mixed longitudinal models are applicable to longitudinal analyses, multiple imputation can be applied to cross-sectional or longitudinal analyses |
Truncation by death and other events | this is not a treatment-level issue: truncation can occur with any type of intervention | the survivor average causal effect (i.e., the average effect in people who would have survived no matter which treatment they were allocated to) | sensitivity methods, propensity score-based methods | methods are most well-developed for cross-sectional analyses, extensions of propensity score-based methods are available for longitudinal analyses |