Got this very nice response(thanks again!)from Margaret May-one of the
authors of the Lancet paper:
We used the ksm stata command to do kernel density smoothing of the
Nelson Aalen cumulative function (bandwidth =.40). The differentiation
was done point by point by calculating gradients. You may need to
smooth again to get a reasonable curve. The confidence intervals were
done using bias corrected bootstrap replicates.
We used stpm for the flexible parametric models which was written by
Patrick Royston - you can download this from stata website. stpm gives
the AIC as part of the estimates list. In general you can work out the
AIC from the log-likelihood and the number of parameters, n (AIC =
-2loglikelihood + 2n.)
We hope to publish a methodological paper on this work soon.
Best wishes
Margaret May
> Subject: st: kernel density smoothing Nelson-Aalenand AIC after streg?
>
> As new user of Stata, I'm currently enjoying reading the recent book "An
> introduction to survival analysis using Stata". In this weeks Lancet there
> is remarkable and interesting paper by Egger et al(Prognosis of
> HIV-1-infected patients starting highly active antiretroviral therapy) in
> which a prognostic survival model was made using Stata by several of the
> methods mentioned in the book.
>
> Although my questions are somehow specific to understanding this paper, I
> guess that they may have interest to other users:
>
> 1. Kernel density smoothing of Nelson-Aalen?: According to the paper: "We
> modelled the instantaneous rate (hazard) of progression to AIDS or death
> according to baseline CD4 cell count, using kernel density smoothing of
the
> Nelson-Aalen cumulative hazard function which was differentiated to give
the
> estimated hazard function"- However, I haven't been able to find any
> references in Stata on how to do this- any suggestions?
>
> 2. Calculation of AIC after streg?: "Candidate prognostic models were
> parametric survival models based on the Weibull, loglogistic, and
lognormal
> distributions"...". We .. additionally considered flexible parametric
> models, based on the same three distributions, which use cubic splines to
> model the baseline hazard allowing for the possibility that the hazard
might
> decrease and then subsequently increase....,We fitted models with one to
> five knots and compared models using Akaike's information criterion (AIC)
> which penalises more complicated models. We found, for each survival
> distribution, that two knots gave the lowest AIC. We fitted models with
one
> to five knots and compared models using Akaike's information criterion
> which penalises more complicated models." Is there any ado's which are
able
> to compare AIC's including AIC after flexible parametric models? I've
found
> fitstat, but haven't been able to find any ado's which are able to
calculate
> AIC after streg..
>
> thanks
>
> Jannik Helweg-Larsen, MD
> Copenhagen HIV programme
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
