[R] svycoxph and test statistics

[Chirag Patel]

Thank you, Dr. Therneau... I got a similar answer from Dr. Lumley:

On Mar 24, 2012, at 4:05 PM, Thomas Lumley wrote:

> As far as I know there isn't any theoretical justification for the
> t-distribution but it empirically works better.
> 
> You can get tests with a t or F reference distribution easily with
> regTermTest.  You can also get likelihood ratio tests that  way, which
> appear to have slightly better small-sample performance than the
> standard Wald tests.
> 
>   - thomas




On Mar 25, 2012, at 5:17 PM, Terry Therneau wrote:

> On 03/24/2012 06:00 AM, r-help-request at r-project.org wrote:
>> I have been using the function 'svycoxph' in the Dr. Lumley's survey package (version 3.26) to compute coefficient estimates for Cox regression.
>> 
>> I have noticed the p-values output are based on normal distribution (like in coxph); however in svyglm (and in other software, such as Stata or SAS) the p-values are computed via the t distribution with degrees of freedom equal to the number of PSUs minus number of strata.
>> 
>> I am wondering why there is a difference here?
> I'm not aware of any theory papers that back up the use of a t-distribution.  This is a Cox model, and "do what my Gaussian package does" is not usually the best approach.  I'm far from an expert in survey work though, so I'll yeild to Thomas L for a definitive answer.
>  In the case of mixed effects models I see the exact same leaning towards (approximate) REML vs ML; this is an area that I do know deeply and and the "REML better than ML" arguments from linear mixed effects models to NOT transfer over.
> 
> Terry Therneau
> 
> 
>