Cox models on subgroups

[F. Danjou]

Dear everybody,

I need help regarding a statistical problem for which I couldn't find
any answer yet.

I realized 5 different Cox proportional hazards models for 5 outcomes
that represent 5 subgroups of my sample (discontinuation for functional
remission, discontinuation for other reasons, discontinuation for side
effects, etc...). I do not compare my Cox models between each others, my
interest is only to individuate predictors for each outcome, which I did
according to the evolution of likelihood ratio for each model. However
each different Cox model regards a different subgroup of my population
against the rest, so that I have the doubt that I might have to adjust
my models for something like "multiple testing".

Do I need to adapt in a certain way my different models? If yes what
should I adapt, as the only statistic I have for each model is a
likelihood ratio and I do not have any overall statistic (that actually
does not interest me as I do not want to compare models between them)?
And again, if yes, do I have to report in a way the eventual correction
on predictors p values within each model ?

Thank you very much in advance, best regards,

Fabien.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Hector Maletta]

To compare the subgroups, you may use the subgroup (reason for
discontinuing) as a variable defining strata. So you'll have four different
base hazard functions (one per stratum) and could compare the different
outcomes in a single analysis.
Another possibility is treating alternatively each outcome as THE outcome,
and the other outcomes as reasons for CENSORING a case if they happen
before. That is, if you are looking at discontinuation because of functional
remission, all other discontinuations before functional remission would be
regarded as censored cases, but information on those cases, up to the time
of their discontinuation, is used in the model. A further approach, not
feasible with simple Cox regression, is a model with competing risks.
Hector
-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of F.
Danjou
Sent: 20 March 2009 08:30
To: [hidden email]
Subject: Cox models on subgroups

Dear everybody,

I need help regarding a statistical problem for which I couldn't find
any answer yet.

I realized 5 different Cox proportional hazards models for 5 outcomes
that represent 5 subgroups of my sample (discontinuation for functional
remission, discontinuation for other reasons, discontinuation for side
effects, etc...). I do not compare my Cox models between each others, my
interest is only to individuate predictors for each outcome, which I did
according to the evolution of likelihood ratio for each model. However
each different Cox model regards a different subgroup of my population
against the rest, so that I have the doubt that I might have to adjust
my models for something like "multiple testing".

Do I need to adapt in a certain way my different models? If yes what
should I adapt, as the only statistic I have for each model is a
likelihood ratio and I do not have any overall statistic (that actually
does not interest me as I do not want to compare models between them)?
And again, if yes, do I have to report in a way the eventual correction
on predictors p values within each model ?

Thank you very much in advance, best regards,

Fabien.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[F. Danjou]

Dear Hector,

thank you very much for your help. The second option you describe is
actually the one I used to analyze data (others outcomes are censored
when I analyze one outcome), so that effectively as you made me realize
information I used in the model information on overall groups every time.

Thank you very much for the time you dedicated to this, best regrads,

Fabien.

++++++++++++

Hector Maletta wrote:

> To compare the subgroups, you may use the subgroup (reason for
> discontinuing) as a variable defining strata. So you'll have four different
> base hazard functions (one per stratum) and could compare the different
> outcomes in a single analysis.
> Another possibility is treating alternatively each outcome as THE outcome,
> and the other outcomes as reasons for CENSORING a case if they happen
> before. That is, if you are looking at discontinuation because of functional
> remission, all other discontinuations before functional remission would be
> regarded as censored cases, but information on those cases, up to the time
> of their discontinuation, is used in the model. A further approach, not
> feasible with simple Cox regression, is a model with competing risks.
> Hector
> -----Original Message-----
> From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of F.
> Danjou
> Sent: 20 March 2009 08:30
> To: [hidden email]
> Subject: Cox models on subgroups
>
> Dear everybody,
>
> I need help regarding a statistical problem for which I couldn't find
> any answer yet.
>
> I realized 5 different Cox proportional hazards models for 5 outcomes
> that represent 5 subgroups of my sample (discontinuation for functional
> remission, discontinuation for other reasons, discontinuation for side
> effects, etc...). I do not compare my Cox models between each others, my
> interest is only to individuate predictors for each outcome, which I did
> according to the evolution of likelihood ratio for each model. However
> each different Cox model regards a different subgroup of my population
> against the rest, so that I have the doubt that I might have to adjust
> my models for something like "multiple testing".
>
> Do I need to adapt in a certain way my different models? If yes what
> should I adapt, as the only statistic I have for each model is a
> likelihood ratio and I do not have any overall statistic (that actually
> does not interest me as I do not want to compare models between them)?
> And again, if yes, do I have to report in a way the eventual correction
> on predictors p values within each model ?
>
> Thank you very much in advance, best regards,
>
> Fabien.
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>
>
>
>

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD