From: Time-dependent ROC curve analysis in medical research: current methods and applications
Definition and marker time | Sensitivity and specificity | Estimation method and R software (when available) | Pros/Cons | |
---|---|---|---|---|
C/D t = 0 | \( \begin{array}{l} S{e}^C\left( c, t\right)= P\left({X}_i> c\Big|{T}_i\le t\right)\\ {} S{p}^D\left( c, t\right)= P\left({X}_i\le c\Big|{T}_i> t\right)\end{array} \) | CD1 survivalROC | Pro: Easy Cons: a) Produce non-monotone sensitivity and specificity b) Not robust to marker-dependent censoring | Pro: Clinically relevant since many clinical experiments aim to discriminate individuals with disease prior to specific time and healthy individual beyond that time Con: Use redundant information in separating cases and controls |
CD2 survivalROC | Pros: a) Produce monotone sensitivity and specificity b) Allow censoring to depend on marker Con: Does involve smoothing parameter | |||
CD3 Programme code | Pro: Does not involve any smoothing parameter Cons: a) Does involve recursive computation b) Produce non-monotone specificity | |||
CD4 survAUC | Pros: a) Produce monotone sensitivity and specificity if the score is produced from a survival model b) Allow censoring to depend on marker Con: Not invariant to an increasing transformation of the marker | |||
CD5 timeROC | Pros: a) Produce monotone sensitivity and specificity b) More robust than CD1 and CD3 Con: Does not robust to marker dependent censoring | |||
CD6 timeROC | Pros: a) Produce monotone sensitivity and specificity b) Robust to marker-dependent censoring c) More less biased than CD2 when censoring strongly depends on marker | |||
CD8, VL Cox survival | Pro: Straightforward to implement | |||
VL Aalen timereg | ||||
VL KM prodlim | ||||
C/D A longitudinal time point | \( \begin{array}{l} S{e}^C\left( c, t\right)= P\left({Y}_{i k}> c\Big|{T}_i-{s}_{i k}\le t\right)\\ {} S{p}^D\left( c,{t}^{*}\right)= P\left({Y}_{i k}\le c\Big|{T}_i-{s}_{i k}> t\right)\end{array} \) | AD4 (ECD2) | Pros: a) Produce monotone sensitivity and specificity b) Allow censoring to depend on marker Con: Does involve smoothing parameter | Pro: Use the most recent marker value prior to prediction time Con: Just use a marker value at a particular time instead of using all serial of marker value |
I/D t = 0 | \( \begin{array}{l} S{e}^I\left( c, t\right)= P\left({X}_i> c\Big|{T}_i= t\right)\\ {} S{p}^D\left( c, t\right)= P\left({X}_i\le c\Big|{T}_i> t\right)\end{array} \) | ID1 risksetROC | Pros: Produce consistent sensitivity and specificity if the control set is unbiased | Pro: Allow time-averaged summaries that directly relate to a familiar concordance measures such as Kendall’s tau or c-index Con: Require an exact time of interest which often just a few individual has an event at a particular point |
ID2 | Pro: Potentially more robust than ID1 Con: Computationally intensive | |||
ID3 Programme code | Pros: a) Easier especially when involve a large number of marker b) Understandable since it is a “regression-type” model | |||
I/S t = 0 | \( \begin{array}{l} S{e}^I\left( c, t\right)= P\left({Y}_i> c\Big|{T}_i= t\right)\\ {} S{p}^S\left( c,{t}^{*}\right)= P\left({Y}_i\le c\Big|{T}_i>{t}^{*}\right)\end{array} \) | None | Pro: Allow separation of long-term survivors from healthy individual within a fixed follow-up Con: Require an exact time of interest which often just a few individual has an event at a particular point | |
I/S A longitudinal time point | \( \begin{array}{l} S{e}^I\left( c, t\right)= P\left({Y}_{i k}> c\Big|{T}_i-{s}_{i k}= t\right)\\ {} S{p}^S\left( c,{t}^{*}\right)= P\left({Y}_{i k}\le c\Big|{T}_i-{s}_{i k}>{t}^{*}\right)\end{array} \) | IS1 | Pro: Provides unbiased estimates of model parameters of sensitivity and specificity Con: Computationally intensive since involve spline functions | Pro: Use the most recent marker value prior to prediction time Con: Just use a most recent of marker value instead of all marker values |
IS2 | Pro: Use the most recent marker value prior to prediction time Con: not a natural companion to hazard models | |||
Other All longitudinal time points | ROC(t, p) = S[a 0(T ik ) + a 1(T ik )S − 1(p)] | AD1 Programme code | Pro: Use all marker value along visit times in the estimation of ROC curve Con: Do not incorporate censored outcomes |