BSTA 6652 Survival Analysis Parametric Methods

/* Parametric models: AFT modeling */

/* Data described in Chapter 3 of P. Allison, "Survival Analysis

Using the SAS System." */

options ls =79;

data recidall;

input week arrest fin age race wexp mar paro prio educ emp1-emp52;

cards;

20 … 1

;

data recid;

set recidall;

drop emp1-emp52;

run;

/* Checking for the distribution of Y using strata fin */

proclifetestdata=recid outsurv=out;

time week*arrest(0);

strata fin;

run;

data b;

set out;

s=survival;

logits=log((1-s)/s); /* for log-logistic model */

logneglog=log(-log(s)); /* for weibull model */

lnorm=probit(1-s); /* for lognormal model; probit: inverse function of cdf of N(0,1) */

lweek=log(week);

run;

procgplotdata=b;

symbol1value=circle i=join;

plot logits*lweek=fin logneglog*lweek=fin lnorm*lweek=fin;

run;

/* Initial AFT model selection */

procliferegdata=recid;

class educ;

model week*arrest(0)=fin age race wexp mar paro prio educ

/ dist=gamma; /* generalized gamma distribution */

run;

procliferegdata=recid;

class educ;

model week*arrest(0)=fin age race wexp mar paro prio educ

/ dist=lnormal; /* log-normal */

run;

procliferegdata=recid;

class educ;

model week*arrest(0)=fin age race wexp mar paro prio educ

/ dist=llogistic; /* log-logistic */

run;

procliferegdata=recid;

class educ;

model week*arrest(0)=fin age race wexp mar paro

prio educ/dist=weibull; /* weibull */

run;

/* backward model selection */

/* Weibull is selected by minimal AIC and L-R tests */

/* drop paro by Wald test p-value */

procliferegdata=recid;

class educ;

model week*arrest(0)=fin age race wexp mar

prio educ/dist=weibull;

PROBPLOT;/* check for goodness of fit of the initial model */

run;

/* drop wexp */

procliferegdata=recid;

class educ;

model week*arrest(0)=fin age race mar

prio educ/dist=weibull; run;

/* drop educ */

procliferegdata=recid;

model week*arrest(0)=fin age race mar

prio/dist=weibull; run;

/* drop race */

procliferegdata=recid;

model week*arrest(0)=fin age mar

prio/dist=weibull; run;

/* drop mar */

/* the final model */

procliferegdata=recid;

model week*arrest(0)=fin age

prio/dist=weibull; run;

/* Final model with covariance matrix, median survival times and residuals */

procliferegdata=recid;

model week*arrest(0)=fin age prio

/dist=weibull covb;

outputout=a cdf=f xbeta=xb p=median STD=se;

probplot; run;

procprintdata=a; run;

/* residual analysis */

data res;

set a;

e=-log(1-f); run;

proclifetestdata=res plots=(ls) notable graphics;

time e*arrest(0);

symbol1v=none; run;

SAS Outputs:

Name of Distribution Gamma

Fit Statistics
-2 Log Likelihood / 633.200
AIC (smaller is better) / 661.200

Name of Distribution Lognormal

Fit Statistics
-2 Log Likelihood / 641.576
AIC (smaller is better) / 667.576

Name of Distribution Weibull

Fit Statistics
-2 Log Likelihood / 633.370
AIC (smaller is better) / 659.370

Name of Distribution LLogistic

Fit Statistics
-2 Log Likelihood / 634.109
AIC (smaller is better) / 660.109

1 | Page

BSTA 6652 Survival Analysis Parametric Methods

Weibull is selected;

check its probability plot:

All points are within the confidence band, so it fits fine. Initial model is Weibull AFT full model.

1 | Page

BSTA 6652 Survival Analysis Parametric Methods

Backward model selection:

Initial model:

Type III Analysis of Effects

Wald

Effect DF Chi-Square Pr > ChiSq

fin 1 4.2905 0.0383

age 1 5.1329 0.0235

race 1 1.3376 0.2475

wexp 1 0.3172 0.5733

mar 1 1.2359 0.2663

paro 1 0.2400 0.6242

prio 1 7.2105 0.0072

educ 4 4.5800 0.3332

Final model: It must be confirmed with L-R tests.

Name of Distribution Weibull

Fit Statistics
-2 Log Likelihood / 643.002
AIC (smaller is better) / 653.002

Type III Analysis of Effects

Wald

Effect DF Chi-Square Pr > ChiSq

fin 1 3.3064 0.0690

age 1 9.6566 0.0019

prio 1 12.0751 0.0005

Analysis of Parameter Estimates

Standard 95% Confidence Chi-

Parameter DF Estimate Error Limits Square Pr > ChiSq

Intercept 1 3.7738 0.3581 3.0720 4.4755 111.08 <.0001

fin 1 0.2495 0.1372 -0.0194 0.5184 3.31 0.0690

age 1 0.0478 0.0154 0.0176 0.0779 9.66 0.0019

prio 1 -0.0698 0.0201 -0.1092 -0.0304 12.08 0.0005

Scale 1 0.7141 0.0637 0.5995 0.8506

Weibull Shape 1 1.4004 0.1250 1.1756 1.6681

Estimated Covariance Matrix
Intercept / fin / age / prio / Scale
Intercept / 0.128202 / -0.004620 / -0.005069 / -0.001772 / -0.000258
fin / -0.004620 / 0.018828 / -0.000046101 / -0.000219 / 0.001239
age / -0.005069 / -0.000046101 / 0.000236 / -0.000000992 / 0.000247
prio / -0.001772 / -0.000219 / -0.000000992 / 0.000403 / -0.000316
Scale / -0.000258 / 0.001239 / 0.000247 / -0.000316 / 0.004062

Residual analysis for final model:

Probability plot for final model:

1 | Page