Sugi 28 statistics and data analysis analysis

SUGI 28 Statistics and Data Analysis

Analysis
Univariate Analyses
Using PROC FREQ and PROC UNIVARIATE, an initial
univariate analysis of the demographic, exposure, and
deployment variables crossed with hospitalization
experience was carried out to determine possible
significant explanatory variables to be included in the
model runs. An exploratory model analysis was then
performed to explore the relations between the variables
while simultaneously adjusting for all other variables that
had influences on the outcome of interest. After
investigation of confounding, all variables with p-values of
0.15 or less were considered possible confounders and
were retained for the model analysis. Additionally, the
distributions of attrition were checked to see if attrition
rates differed for the categories of exposure over the study
period.
Multivariable Cox Modeling Approach
Dummy variables were created for reference cell coding of
the categorical variables. These were necessary for the
output of measures of association using the reference
category of choice. Starting with a saturated model, PROC
PHREG was run using a manual backward stepwise
model building approach. This created a final model with
statistically significant effects of explanatory variables on
survival times while controlling for possible confounding
of exposure effects.
SAS Programming
PROC PHREG data=analydat;
model inhosp*censor(0)=expose1-expose6
pwhsp status1 sex1 age1-age3 ms1
paygr1-paygr2 oc_cat1-oc_cat9 ccep
/ rl ties=efron ;
title1 'Cox regression with exposure
status in the model';
run;
The options used in this survival analysis procedure are
described below:
DATA=ANALYDAT names the input data set for the
survival analysis.
RL requests for each explanatory variable, the 95% (the
default alpha level because the ALPHA= option is not
invoked) confidence limits for the hazard ratios.
TIES=EFRON gives the researcher the approximations to
the EXACT method without using the tremendous CPU it
takes to run the EXACT method. Both the EFRON and
the BRESLOW methods do reasonably well at
approximating the EXACT when there are not a lot of ties.
If there are a lot of ties, then the BRESLOW
approximation of the EXACT will be very poor. If the
time scale is not continuous and is therefore discrete, the
option TIES=DISCRETE should be used.
Stratification By Exposure Status
These data were then stratified by exposure and the models
were run with the exposure flag covariate withdrawn from
the model. This allowed for inspection of confounding
between exposure status and covariates. Running these
separate models also allowed for the computation of
survival function estimates using the BASELINE function
in PROC PHREG. The survival curves (which are really
step functions, however there are such numerous events
that they appear continuous) were now available to
compute the cumulative distribution function for the
separate exposure categories.