From: Modelling of longitudinal data to predict cardiovascular disease risk: a methodological review
Method N(%)[refs] | Longitudinal outcome type | Disease outcome type | How the longitudinal data were used in the analysis N (%) [refs] | Reason for the use of method | Assumptions | Pros | Cons |
---|---|---|---|---|---|---|---|
Single-stage approaches (n = 40) | |||||||
Cox model, N = 25 (62.5) [18, 19, 21, 25, 28, 29, 32, 34,35,36, 38, 39, 41,42,43, 45, 47, 49,50,51, 53,54,55,56,57] | Continuous, Categorical | Time to event | Baseline only, N = 7 (17.5) [18, 21, 24, 43, 50, 53, 54] | To clinically relevant time point to be used for prediction | PH | Simple method | Dependence between measurement times is ignored |
Continuous | Time to event | To incorporate change over time | PH; Change is linear | Incorporates more than one time point | Only looks at pairs of time points | ||
Continuous | Time to event | To incorporate constant change in the survival model | PH; Change is linear | Incorporates more than one time point | Only looks at pairs of time points | ||
Continuous | Time to event | Average (categorized before use),a N = 1 (2.5) [36] | To incorporate the average change over time | PH; Constant between time points; Change is linear | Incorporates the average impact over time | Interpretation unclear | |
Continuous, Categorical | Time to event | Time-dependent covariate, N = 6 (15.0) [39, 42, 45, 47, 49, 51, 55] | To incorporate change in exposure variable over time | PH; Change is constant between two consecutive time points; Longitudinal data are measured without error | Incorporates time-varying measures over the follow-up period | Computationally slower as compared to time-fixed covariates; Computationally infeasible if the longitudinal outcome is measured at different time points for different individuals; Interpretation is difficult; Can lead to great overfitting of the data; must be used with caution | |
Continuous | Time to event | Summary measures(Standard deviation, number of drops between observations), N = 1 (2.5) [19] | To incorporate variability summaries of the longitudinal data | PH | Incorporates variability of measures into the model | Summary measures fairly specific to dataset | |
Continuous, Categorical | Time to event | Change in category between first and last time-point categorized, N = 2 (5.0) [34, 41] Change in continuous variable between time points categorized with manually defined cut-offs, N = 1 (2.5) [56] | To summarise trajectories in an interpretable way | PH | Results interpretable | Groups manually selected based on data which could lead to bias | |
Hierarchical Cox model to adjust for multiple studies, N = 1 (2.5) [20] | Continuous | Time to event | Continuous measurements categorized. Multiple time points also categorised as consistent/non-consistent, N = 1 (2.5) [20] | To summarise trajectories in an interpretable way adjusting for combining multiple studies | PH | Results interpretable; Adjusts for use of multiple studies | Groups manually selected based on data which could lead to bias |
Continuous | Binary | Baseline only, N = 1 (2.5) [31] | Allows clinically relevant time point to be used for prediction | Not applicable | Simple method | Dependence between measurement times is ignored | |
Categorical | Binary | Separate time points, N = 1 (2.5) [30] | To include all predictive values in model | Not applicable | Simple method | Caution needed for multicollinearity | |
Continuous | Binary | Summaries of repeated measures • Standard deviation • Mean • Mean change from baseline • Average daily risk rangeb • Range N = 1 (2.5) [48] | Includes different measures of variation | Not applicable | Simple method | Interpretation of different summary measures non-trivial | |
Continuous | Binary | Non-linear relationships considered through piecewise models or splines, N = 1 (2.5) [17] | To attempt to include a variety of shapes of relationships in the model using data from all time points | Not applicable | Includes all measured values of longitudinal variable with various relationships with risk | Splines harder to interpret; Produces population averages not individual predictions | |
Continuous | Binary | Multiple time points, N = 1 (2.5) [27] | To include values and change at all time points | Not applicable | Includes all measured values of longitudinal variable | Produces population averages not individual predictions | |
Continuous | Rates | Multiple time points, N = 1 (2.5) [37] Multiple time points categorized as stable, increasing (in the second or third time point), decreasing, unstable, N = 1 (2.5) [22] | To include all time points in predicting rates | Not applicable | Includes all measured values of longitudinal variable | Produces population averages not individual predictions | |
Continuous | Rates | To enable modelling of baseline rate | Not applicable | Enables modelling of baseline rate in a parametric manner | Dependence between measurement times is ignored | ||
Continuous, categorical | Continuous | To predict changes over time | Random effects are independent of covariates | Includes all measured values of longitudinal variable | None | ||
Fixed effects linear regression, N = 1 (2.5) [52] | Continuous, categorical | Continuous | The variable is transformed by subtracting patient-level mean to remove between patient variation. N = 1 (2.5) [52] | To predict changes over time | Not applicable | Includes all measured values of longitudinal variable; Relaxes assumption of independence of random effects from covariates; Computationally very easy to fit compared with mixed effects models | Lower statistical efficiency than mixed effects models |