Survival models
Effects on the probability of the event by a chosen time
After streg, the marginal effects are changes in the predicted median survival time by default (the mean after mestreg). Predicted survival times extrapolate beyond the observed follow-up when spells are censored, and Jones and Metzger (2019) and Metzger and Jones (2022) recommend interpreting duration models through the probability of the event by a chosen time instead. With time(#), the marginal effects of stcox and streg models are changes in the probability that the event has happened by time #, one minus the survival probability. A Cox model (stcox) needs time().
The examples use Stata’s cancer data: 48 patients in a drug trial, followed for up to 39 months. died records a death, and drug is the placebo or one of two drugs, Other and NA.
sysuse cancer, clear(Patient survival in drug trial)
stset studytime, failure(died)Survival-time data settings
Failure event: died!=0 & died<.
Observed time interval: (0, studytime]
Exit on or before: failure
--------------------------------------------------------------------------
48 total observations
0 exclusions
--------------------------------------------------------------------------
48 observations remaining, representing
31 failures in single-record/single-failure data
744 total analysis time at risk and under observation
At risk from t = 0
Earliest observed entry t = 0
Last observed exit t = 39
A Cox model and a Weibull model
The effects of the drugs and of age on the probability of death by month 10, in a Cox model and a Weibull model:
quietly stcox age i.drug, vce(robust)
estimates store cox1
quietly streg age i.drug, distribution(weibull) vce(robust)
estimates store wei1
mecompare drug age, models(cox1 wei1) time(10)Predicting: Pr(event by t=10)
Marginal effects and cross-model differences (N_cox1=48) (N_wei1=48)
| ME # Estimate Robust SE P>|z|
---------------------------------+---------------------------------------
drug |
Other - Placebo |
cox1 | 1 -0.419 0.089 0.000
wei1 | 2 -0.439 0.093 0.000
Difference | 3 0.020 0.032 0.528
NA - Placebo |
cox1 | 4 -0.538 0.081 0.000
wei1 | 5 -0.531 0.085 0.000
Difference | 6 -0.007 0.050 0.894
---------------------------------+---------------------------------------
age + 1 (centered) |
cox1 | 7 0.021 0.004 0.000
wei1 | 8 0.023 0.005 0.000
Difference | 9 -0.002 0.002 0.234
NOTE: the stcox and streg rows are changes in the probability that the event has happened by time 10
> . See mecompare##survival.
Compared with the placebo, the Other drug lowers the probability of death by month 10 by 0.42 in the Cox model and by 0.44 in the Weibull model; the two models do not differ significantly (Difference 0.020, p = .528).
A Cox model and a logit
Other models in models() keep their default prediction. A logit of whether a patient died during the study can be compared with the Cox model, here at month 20:
quietly logit died age i.drug, vce(robust)
estimates store lg1
mecompare drug, models(lg1 cox1) time(20)Predicting: Pr(died) (lg1), Pr(event by t=20) (cox1)
Marginal effects and cross-model differences (N_lg1=48) (N_cox1=48)
| ME # Estimate Robust SE P>|z|
---------------------------------+---------------------------------------
drug |
Other - Placebo |
lg1 | 1 -0.551 0.139 0.000
cox1 | 2 -0.484 0.096 0.000
Difference | 3 -0.066 0.108 0.537
NA - Placebo |
lg1 | 4 -0.484 0.136 0.000
cox1 | 5 -0.719 0.092 0.000
Difference | 6 0.235 0.127 0.065
NOTE: the stcox rows are changes in the probability that the event has happened by time 20. See meco
> mpare##survival.
The logit’s rows are changes in the probability of dying during the study, and the Cox model’s in the probability of dying by month 20. The logit counts every patient alive at their last follow-up as a survivor, however short their follow-up; the Cox model takes each patient’s follow-up time into account.
Choosing the time
The time is in the units of the analysis time set by stset, here months. With a Cox model, a time later than the last analysis time, or earlier than the first event, is refused.
References
Jones, Benjamin T., and Shawna K. Metzger. 2019. “Different Words, Same Song: Advice for Substantively Interpreting Duration Models.” PS: Political Science & Politics 52(4): 691–695. https://doi.org/10.1017/S104909651900060X
Metzger, Shawna K., and Benjamin T. Jones. 2022. “Getting Time Right: Using Cox Models and Probabilities to Interpret Binary Panel Data.” Political Analysis 30(2): 151–166. https://doi.org/10.1017/pan.2021.14