[R] psm/survreg coefficient values ?

John Logsdon j.logsdon at quantex-research.com
Tue Jun 19 11:17:48 CEST 2007


In survreg() the predictor is log(characteristic life) for Weibull (= 
exponential when scale=1) - ie the 63.2%ile.  For the others the predictor is 
log(median).

This causes problems when comparing predictions and a better way IMHO is to 
correct the Weibull prediction by a factor (log(2))^(1/scale).  This is only 
a simple multiple unless the shape parameter is also being modelled, when a 
completely different solution may arise.  Such heterogeneity modelling cannot 
of course be done within survreg().

On Monday 18 June 2007 22:56:54 Frank E Harrell Jr wrote:
> sj wrote:
> > I am using psm to model some parametric survival data, the data is for
> > length of stay in an emergency department. There are several ways a
> > patient's stay in the emergency department can end (discharge, admit,
> > etc..) so I am looking at modeling the effects of several covariates on
> > the various outcomes. Initially I am trying to fit a  survival model for
> > each type of outcome using the psm function in the design package,  i.e.,
> > all  patients who's visits  come to an end  due to  any event other than
> > the event of interest are considered to be censored.  Being new to the
> > psm and  survreg packages (and to parametric survival modeling) I am not
> > entirely sure how to interpret the coefficient values that psm returns. I
> > have included the following code to illustrate code similar to what I am
> > using on my data. I suppose that the coefficients are somehow rescaled ,
> > but I am not sure how to return them to the original scale and make sense
> > out of the coefficients, e.g., estimate the the effect of higher acuity
> > on time to event in minutes. Any explanation or direction on how to
> > interpret the  coefficient values would be greatly appreciated.
> >
> > this is from the documentation for survreg.object.
> > coefficientsthe coefficients of the linear.predictors, which multiply the
> > columns of the model matrix. It does not include the estimate of error
> > (sigma). The names of the coefficients are the names of the
> > single-degree-of-freedom effects (the columns of the model matrix). If
> > the model is over-determined there will be missing values in the
> > coefficients corresponding to non-estimable coefficients.
> >
> > code:
> > LOS <- sort(rweibull(1000,1.4,108))
> > AGE <- sort(rnorm(1000,41,12))
> > ACUITY <- sort(rep(1:5,200))
> > EVENT <-  sample(x=c(0,1),replace=TRUE,1000)
> > psm(Surv(LOS,EVENT)~AGE+as.factor(ACUITY),dist='weibull')
> >
> > output:
> >
> > psm(formula = Surv(LOS, CENS) ~ AGE + as.factor(ACUITY), dist =
> > "weibull")
> >
> >        Obs     Events Model L.R.       d.f.          P         R2
> >       1000        513    2387.62          5          0       0.91
> >
> >               Value          Std. Error      z         p
> > (Intercept)     1.1055    0.04425  24.98 8.92e-138
> > AGE             0.0772    0.00152  50.93  0.00e+00
> > ACUITY=2     0.0944    0.01357   6.96  3.39e-12
> > ACUITY=3     0.1752    0.02111   8.30  1.03e-16
> > ACUITY=4     0.1391    0.02722   5.11  3.18e-07
> > ACUITY=5    -0.0544    0.03789  -1.43  1.51e-01
> > Log(scale)    -2.7287    0.03780 -72.18  0.00e+00
> >
> > Scale= 0.0653
> >
> > best,
> >
> > Spencer
>
> I have a case study using psm (survreg wrapper) in my book.  Briefly,
> coefficients are on the log median survival time scale.
>
> Frank



-- 
Best wishes

John

John Logsdon                               "Try to make things as simple
Quantex Research Ltd, Manchester UK         as possible but not simpler"
j.logsdon at quantex-research.com              a.einstein at relativity.org
+44(0)161 445 4951/G:+44(0)7717758675       www.quantex-research.com



More information about the R-help mailing list