cureregr, cure model regression (split-population model) or, parametric-cure mo > del (PCM)
cureregr [varlist] [if exp] [in range] , [distribution_option({it:weibull | lognormal | logistic | gamma | exponential}) class_option(mixture | non-mixture) link_option(logistic | lml, log-minus-log | linear ) scale_option(varlist) shape_option(varlist) noconstant scale_noc shape_noc nodescriptionmodel noshow left]
Description
cureregr fits a parametric cure model in either the non-mixture or mixture class. important -- cureregr requires that the data be stset prior to use. cureregr will correctly estimate multiple records per subject, to include time-varying-covariates. Please note that for multiple records per subject the records must be in contiguous intervals defined in stset, [_t0, _t]; in other words, there may be no gaps (see: stdes).
Options
distribution_option(weibull | lognormal | logistic | gamma | exponential) define Kernel: weibull: exponential in tt lognormal: norm(ln(tt)) logistic: tt/(1-tt) gamma: gammap(shape, scale, time) exponential: gammap(shape==1, scale, time) class_option(mixture | non-mixture) define parametric-cure model class: mixture: pi+(1-pi)*(1-Kernel) non-mixture: pi^Kernel link_option(logistic | lml, log-minus-log | linear) define cure_fraction (pi) link funtion logistic: exp(xb_pi)/(1+exp(xb_pi)) lml: exp(-exp(xb_pi)) linear: exp(xb_pi) where xb_pi is the linear model for the cure_fraction, pi scale_option(varlist) varlist for scale model shape_option(varlist) varlist for shape model noconstant specifies that no constant term for cure-fraction model is to be fit scale_noc specifies that no constant term for scale model is to be fit shape_noc specifies that no constant term for shape model is to be fit -- left specifies that there are left censored subjects. the failures defined by the last value specified by the failure() argument of stset are taken to be left censored. this creates (or overwrites) a variable called _leftd and is indicated in e(depvar). -- nodescriptionmodel do not show cureregr model definition information noshow do not show st setting information
please note these definitions tt = (scale*time)^shape function of time, t-tilde scale=exp(xb_scale) for the scale model shape=exp(xb_shape) for the shape model
Example
. cureregr g1_ageyr, dist(weibull) class(n) link(logi) noconstant
. xi: cureregr g0_ageyr i.g0_wbc i.g0_plt, sc(i.g0_wbc i.g0_plt) sh(i.g0_plt) dist(lognormal) class(non) link(logi) scale_noc shape_noc
. xi: cureregr (i.study_r) i.g_liver i.race2 if analset1 , dist(weibull) link(lml) sc((i.study_r) ) sh((i.study_r) ) class(nonmixture) technique(nr)
predicting
predict [if exp] [in range] , [at(numlist) | all | survival | ses | ucs | lcs | hazard | fd] [gen(varlist)] [level(number)] [fmtprint(%8.6f)]
After model estimation the survival, standard error, and confidence interval can be displayed with predict, at() specified time points. The survival estimates are displayed for all covariate patterns and if or in may be used to limit the number of covariate patterns displayed; An example can be found in the ancillary file, cureregr_predict_eg.pdf. The fmt() option allows any acceptable numeric format specification for the displayed numeric results. Please note also that the confidence limits are based on a ln(-ln(S)) transform.
Six quantities may be estimated and placed into a new variable: survival (S), and its standard error (seS), lower confidence limit (lciS), and upper confidence limit (uciS). In addition, the hazard (haz), or the fail density (fd) may be estimated. One of the key words (s, se, l, u, h, fd, or all) can be specified. With the option gen(), the estimate requested by the keyword will be placed in the new variable, while the keyword all requires six new variable names in order: survival, ses, lcs, ucs, hazard, and fd. Where the default names are: S, seS, lciS, uciS, haz, and fd, respectively.
. predict if study_r==1 & g_liver==1 & race2==2, at(30(30)365 2922) . predict , all gen(s_v2 se_v2 lci90_v2 uci90_v2 haz_v2 fd_v2) level(90) . predict , hazard gen(myhaz) . predict , lcs gen(lci95_v1) level(95) . predict , ucs gen(uci95_v1) level(95)
Author Allen Buxton, CureSearch, Children's Oncology Group, Arcadia, CA - acknowledgements - Richard Sposto, Mark Krailo, & Todd Alonzo, Keck School of Medicine, University of Southern California, USA John Thompson & Clarie Weston, Department of Health Sciences, University of Leicester, UK
References Sposto R. Cure model analysis in cancer: An application to data from the Children's Cancer Group. Statistics in Medicine, 21: 293-312, 2002 Stephen Jenkins, spsurv (spsurv.ado for Stata, 23may2001), University of Essex, UK Maller RA. & Zhou, X. Survival Analysis with Long Term Survivors, John Wiley, 1997. Schmidt P. & Witte A. Predicting criminal recidivism using 'split-population' survival time models, Journal of Econometrics, 40: 141-159, 1989
Files main: cureregr.ado, & ml evaluators: PCMmmkkll.ado predict: _cureregr.ado & _cureregr_gen.ado - ancillary - pcm_key.pdf, a description of the models cureregr will fit. cureregr_predict_eg.pdf, example of predict. example_stata.dta, dataset in the above example.
Also see Manual: [R] stset