stcox
Explains how long people last, using several columns at once.
What we have
- hazard ratios, with Breslow ties by default and
efronon request - factor notation, and a match with R's
coxph(checked)
Coming later
- weights, which are refused
- stratified and parametric models
- a test of the proportional-hazards assumption
How you write it
stcox covariates [, breslow | efron]
Square brackets mean optional. A vertical bar means choose one.
What it does
Cox proportional hazards regression. Where sts compares whole groups, this adjusts for several things together: age, department and grade at the same time.
Results come as hazard ratios. A ratio of 1.3 means that at any moment those people are 1.3 times as likely to have the event as the comparison group; a ratio below 1 means they last longer. The model never has to say what the underlying risk over time looks like, which is why it is so widely used.
Ties — two people leaving on the same day — are handled Stata's way (Breslow) by default; add efron for the more accurate treatment when ties are common.
Try it
Every example below was run after sysuse vira.sample03, clear.
. stset tenure_months, failure(left)
Survival-time data settings
Failure event: left != 0 & left < .
Observed time interval: (0, tenure_months]
Exit on or before: failure
--------------------------------------------------------------------------
4,000 total observations
0 exclusions
--------------------------------------------------------------------------
4,000 observations remaining, representing
1,652 failures in single-record/single-failure data
87,299 total analysis time at risk and under observation
At risk from t = 0
Last observed exit t = 60
The same thing in pandas and in R
pandas
import numpy as np import pandas as pd time = pd.to_numeric(df["tenure_months"], errors="coerce") status = pd.to_numeric(df["left"], errors="coerce") valid = time.notna() & (time > 0) df["_st"] = valid.astype(int) df["_d"] = ((status.notna() & (status != 0)) & valid).astype(int) df["_t"] = time.where(valid) df["_t0"] = np.where(valid, 0.0, np.nan)
R
library(survival) surv <- Surv(tenure_months, left) # This R has NOT been run against a real R yet: read it as documentation.
. stcox age i.level
Cox regression with Breslow method for ties
No. of subjects = 4,000 Number of obs = 4,000
No. of failures = 1,652 LR chi2(5) = 67.10
Time at risk = 87,299 Prob > chi2 = 0.0000
Log likelihood = -12565.2971
------------------------------------------------------------------------------
_t | Haz. ratio Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
age | 1.002756 .00319522 0.86 0.388 .9965129 1.009038
level=2 | .8415329 .05045858 -2.88 0.004 .748226 .9464756
level=3 | .6638695 .05131911 -5.30 0.000 .570535 .7724728
level=4 | .4582762 .06198606 -5.77 0.000 .351556 .597393
level=5 | .5444055 .1171195 -2.83 0.005 .3571093 .8299346
------------------------------------------------------------------------------
The same thing in pandas and in R
pandas
import patsy
from statsmodels.duration.hazard_regression import PHReg
usable = df[df["_st"] == 1]
X = patsy.dmatrices('_t ~ age + C(level)', usable, return_type="dataframe")[1].drop(columns=["Intercept"], errors="ignore")
t, d = usable["_t"], usable["_d"]
model = PHReg(t, X, status=d, ties="breslow").fit()
R
library(survival)
model <- coxph(Surv(tenure_months, left) ~ age + factor(level),
data = df, ties = "breslow")
summary(model)
# This R was run in a real R (4.6.1) and gave the same numbers.
The full built-in help (help stcox)
stset timevar, failure(eventvar[==#])
sts list | graph | test [, by(var) at(times)]
stcox covariates [, breslow | efron]
Survival analysis: how long until something happens, when for some subjects
it has not happened yet. Someone still employed when the study ends has told
you they lasted at least that long -- dropping them, or counting them as
stayers, both give the wrong answer. That is what "censored" means.
. stset months, failure(left) declares the time and the event
. sts list, by(dept) at(12 24 36) Kaplan-Meier survivor function
. sts graph, by(dept) the same as a chart
. sts test dept log-rank test across departments
. stcox age i.dept hazard ratios
stset makes _st (usable), _d (event), _t (time). failure(left) counts
any non-zero value; failure(status==2) only that value. A time of
zero, negative or missing is excluded, and it says how many
sts list survivor function with Greenwood standard errors and a log-log
95% interval. With at(), one row per interval: how many were at
risk at its start, and how many failed and were censored in it
sts test the log-rank test, with observed and expected events per group
stcox Breslow ties by default, as in Stata; add efron for Efron
Not supported, and refused rather than ignored: delayed entry, several records
per subject, time-varying covariates, stratified and parametric models,
weights, Nelson-Aalen, and tests other than log-rank.
Every number matches R's survival package (survfit, survdiff, coxph); none has
been compared with a real Stata run.