Illustration of the Stata stcox command used for logistic regression estimation.
Results are compared to output from the logit procedure.
Background:
In SAS you can use the ties=discrete option in the model statement for moderately sized logistic regression analyses or in general for reasonable sized data sets if time is truly discrete.
The method is illustrated on sas.com for a conditional logistic regression, adding a strata statement – which in Stata compares to a strata variable specified in the strata option added to the stcox command. Have a look at support.sas.com here.
“Extra memory is needed for certain TIES= options. Let be the maximum multiplicity of tied times. The TIES=DISCRETE option requires extra memory (in bytes) of 4k*(p^2+4p), where k is the maximum multiplicity of tied times and p is the number of predictors/explanatory variables Source of citation on sas.com.
The equivalent option in Stata’s stcox is exactp, which however is not compatible with the vce(cluster) option. Can we mend it with the _robust command?
Stata code follows
*Demonstration use http://www.ats.ucla.edu/stat/stata/dae/binary.dta, clear *Benchmark estimates with the logit command logit admit gre, or /* Result ————————————————————————— admit | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] ————-+————————————————————- gre | 1.003589 .0009895 3.63 0.000 1.001651 1.00553 ————————————————————————— */ logit admit gre gpa i.rank, or /*Result ————————————————————————— admit | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] ————-+————————————————————- gre | 1.002267 .0010965 2.07 0.038 1.00012 1.004418 gpa | 2.234545 .7414652 2.42 0.015 1.166122 4.281877 | rank | 2 | .5089309 .1610714 -2.13 0.033 .2736922 .9463578 3 | .2617923 .0903986 -3.88 0.000 .1330551 .5150889 4 | .2119375 .0885542 -3.71 0.000 .0934435 .4806919 ————————————————————————— */ logit admit gre, or vce(cluster rank) /*Result (Std. Err. adjusted for 4 clusters in rank) ————————————————————————— | Robust admit | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] ————-+————————————————————- gre | 1.003589 .0005782 6.22 0.000 1.002456 1.004723 ————————————————————————— /* *Now similar logistic regressions with stcox capture stset ,clear stset time, failure(admit==1) origin(time 0) stcox gre, exactp /*Result ————————————————————————— _t | Haz. Ratio Std. Err. z P>|z| [95% Conf. Interval] ————-+————————————————————- gre | 1.00358 .0009882 3.63 0.000 1.001644 1.005518 ————————————————————————— */ stcox gre gpa i.rank, exactp /*Result ————————————————————————— _t | Haz. Ratio Std. Err. z P>|z| [95% Conf. Interval] ————-+————————————————————- gre | 1.002261 .001095 2.07 0.039 1.000117 1.00441 gpa | 2.229813 .7389021 2.42 0.016 1.164668 4.269082 | rank | 2 | .5099792 .1611582 -2.13 0.033 .2745141 .9474148 3 | .2627833 .0906066 -3.88 0.000 .1336925 .5165218 4 | .2128309 .0888118 -3.71 0.000 .0939374 .482204 ————————————————————————— */ *Note it is unwise to rearrange data and make a ‘fix’ for clusteret data using the stset command – due to memory limitations. A prober solution would utilize postestimation – which would also make comparison with unadjusted s.e.’s easy.
Comentários