TSM Models and Data Dredging

questions concerning analysis/theory using program MARK

TSM Models and Data Dredging

Postby constant survivor » Wed Oct 28, 2020 10:48 am

Hello everyone,
this time I have a general question about model-building and the topic of dredging.
I know that dredging is evil in general...

But assume following situation:

1.) Build TSM model phi(t/t) p(.)
2.) You will get some "survival estimates" for TSM 1
3.) Take these estimates and establish for example 3 groups of similar 'survival' in TSM 1
4.) Now you could reduce the parameter number drastically and (of course) deviance by pooling the TSM1 intervals according to the estimates of step 2.)

Obviously, this is dredging. But could you 'apologize' for that, as the dredging refers to TSM1 which is obviously not the studies target? Because the target estimates are these for TSM 2+ .

So the question is, if this procedure would be acceptable if in turn the estimates for TSM 2+ become more accurate (if they do at all) ??

Thanks for your thoughts about that.
Hannes
constant survivor
 
Posts: 36
Joined: Wed Dec 11, 2019 12:20 pm

Re: TSM Models and Data Dredging

Postby Jochen » Fri Oct 30, 2020 7:02 am

Hi,
knowing a bit of your models I guess that your Phi(t/t) model has too many non-estimable parameters - otherwise there should be no need to reduce their number. BTW I'd expect similar issues with p(t).

Your idea to reduce parameter numbers includes two highly arbitrary decisions. First, what is 'similar' survival to be grouped without any other reaon? And what is 'more accurate' for you? I suppose the estimates have narrower CIs, but even then they could be biased. You probably don't have other studies of the same species to compare.
Cheers
Jochen
Jochen
 
Posts: 26
Joined: Fri Nov 17, 2006 12:04 pm
Location: Germany

Re: TSM Models and Data Dredging

Postby constant survivor » Fri Oct 30, 2020 1:30 pm

Hi Jochen,
thanks for reply!

First, what is 'similar' survival to be grouped without any other reaon?


Actually, I simply ordered the estimates of TSM1 in ascending order and pooled them into either 3 or 5 groups. Therefore I built a model with time dependence in TSM1 e.g. phi(t/.)p(t). I tried too find 'best' model under those with time dependence in TSM1. Most of them had no non-estimable parameters.

And what is 'more accurate' for you? I suppose the estimates have narrower CIs, but even then they could be biased


Yes, this was the main goal. A narrower CI and smaller SE for the estimate(s) of phi for TSM 2+. But actually I found, that the effect is rather small or even absent compared to models without the grouping in TSM1. The main effect is on AIC and Deviance. But for this, it does not really make sense to do it I guess.

So I already think I am going to discard this procedure of grouping and simply leave it with constant, time-dependent and/or sex-dependence in TSM 1.

Best
Hannes
constant survivor
 
Posts: 36
Joined: Wed Dec 11, 2019 12:20 pm


Return to analysis help

Who is online

Users browsing this forum: Google [Bot] and 17 guests

cron