Hi Rimal,

Survival analysis is a "time-to-event" analysis, so you are measuring both the mortality rate, but also the median time until death (or mean if you use a parametric model) . So you can achieve your desired outcomes using this approach. I am attaching a link to a paper I wrote on survival analysis a few years ago. There is a comprehensive description in the paper on this approach and what it all means...

I hope this helps

Ariel

Article Evaluating Disease Management Program Effectiveness: An Intr...

Dear Ariel:

Thanks very much for your help. I read your article. It is a very useful and comprehensive article.   

However, I want to predict the survival time for each individual considering the baseline covariates . Cox regression gives the predicted relative hazard for each individual,while did not predict the individuals' survival times. Is there any suggestion on how to predict the individuals' survival times in STATA.

That is different than what the model is intended for, however the paper in this link looks like it offers a possibility. Let me know how it goes.

http://www.ncbi.nlm.nih.gov/pubmed/24292992

Actually, you can get to the predicted median time for each individual by using a parametric survival model and then using predict . For example:

webuse kva, clear

streg load bearings, d(weibull)

predict medtime, median time // will give median time

predict meantime, mean time // will give mean time

Hi Rimal,

The right approach is to use parametric survival models. As dr. Linden says, the Weibull function is one option for that analysis.

You should looking for these books:

Applied Survival Analysis: Regression Modeling of Time to Event Data. David W. Hosmer, Jr., Stanley Lemeshow, Susanne May, John Wiley & Sons, 2011.

Modeling Survival Data: Extending the Cox Model. Terry M. Therneau, Patricia M. Grambsch. Springer Science & Business Media, 2000.

Best regards. Andrés.

Hi Rimal,

Really, as Andres suggests, if you want to predict patients' life expectancy considering other covariates, the right approach is to use parametric survival models. Using a Cox proportional hazard model, you can get only punctual estimations of survival function on observed death times.

The use of log- linear regression models may help you solve the problem of positively skewed data. Also may help you perform point estimations of survival function. But, if you are analyzing cancer survival data, the hazard function usually grows at first, reaching a maximum and then decreases, and I believe that, in these cases, the log- normal model will be more adequate that the Weibull model.

Dear Ariel, Andres, and Nieves.

Many thanks for your help .I followed your recommendations on using a parametric survival model to predict the survival times. I used the accelerated failure model and tested the Weibull, loglogistic and lognormal distribution using AIC criteria. 

I used the following commands :

xi: streg i.RX_cat _mmws , d(weibull) time

Predict Survival_time 

Where RX_cat is the treatment types, _mmws is the weights provided for each individual through propensity score method.

What I have noticed that the predicted survival times, adjusted for the weight derived from the propensity score method, were hugely different from actual observed survival times.

I tried to test different distributions,however, all ended up with similar results. i.e. patients who died from breast cancer( event of interest), or died from other causes, had higher predicted survival times. Any suggestion on what would the solution be? Also any suggestion on how to adjust for competing risk in an accelerated failure model.

Consider using stcrreg for competing-risks regression

I'm sure by now you've figured this out, but I thought I'd just comment as I find these analyses interesting...

Depending on your sample size, you could also consider using a skew-normal regression approach (see paper by Moser et al, 2014). 

For your weighting technique, I'm not 100% sure why you are weighting, however, I do know that stata has special commands for weights:

"fweights, iweights, and pweights may be specified using stset; see [ST] stset. However, weights may not be specified if you are using the bootstrap prefix with the streg command."

For competing risks, if you're interested in using a flexible parametric survival model, there is some detail on how to do this in the book 'flexible parametric survival analysis...' by Royston and Lambert.

estat ic can do the trick