Table 5.
IPCW modelling considerations.
| Model feature | Consideration required |
|---|---|
| Selection of TVCs | Consider how best to determine which TVCs are important in predicting treatment change and outcome: consult clinical opinion; may be necessary to apply selection procedure (if numerous TVCs) |
| Functional form of covariates | Check optimal functional form using lowess curve of martingale residuals (from Cox model) |
| Extreme covariate values | Truncate at the 99th (or 95th) centile to avoid extreme weights (which in turn distort treatment effect estimate) due to influential outlying values of important predictors of treatment change/outcome |
| Time intervals (for discretised TVCs) | Strike the balance between greater accuracy (increases as interval length decreases) and computational intensity (increases with interval length) |
| Model type (Cox or PLR) | PLR is useful if using lagged variables or if TVCs change frequently (and therefore are too complicated to be analysed without discretising) |
| Cox modelling avoids the need to consider splines to mirror underlying risk function in PLR model | |
| Splines (for PLR only) | Create and use treatment-specific spline variables for WD model, but use overall splines for the WO model |
| Consider shape of underlying risk, in order to identify times where risk changes and thus inform positioning of knots | |
| CI estimation | Estimate CIs using bootstrapping to overcome correlation due to within-patient time-varying weights (in Cox model) and the reduction in SEs due to the weight estimation procedure |
| EPV ratio | Consider ratio of number of variables in model to number of treatment change events, in particular when considering the number of knots to use for spline variables (spline with knots requires variables) |
| Model assumptions | Consider plausibility of NUC assumption (whether all confounding variables have been accounted for) by seeking clinical expert opinion |
| Examine model weights for evidence of violation of positivity assumption: extreme weights may indicate unreliable pool of patients who do (or do not) change treatment at a particular time (within a given subgroup of patients defined by cross-classification of model covariates) |
TVC: time-varying covariate; PLR: pooled logistic regression; NUC: unmeasured confounders; WO: weighted outcome; WD: weight determining; IPCW: inverse probability of censoring weighting.