Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

st: Question about xtlogit and temporal dummy variables


From   <[email protected]>
To   <[email protected]>
Subject   st: Question about xtlogit and temporal dummy variables
Date   Sun, 2 Mar 2014 11:11:13 +0000

(1) Note well the point by Steve Samuels about the interpretation of your parameter estimates. For a discrete time PH model, exponentiated coefficients on the predictors (other than those summarising duration dependence)  can be interpreted as hazard ratios. It is less meaningful to do so for the parameters summarising duration dependence. To interpret these, I recommend deriving predictions of the hazard rate for some fixed values of the predictors. See the Lessons at the URL below for worked examples.

(2) Fixed effects models are rarely used to model frailty in duration for precisely the reason you are discovering. FE models in effect with the observations for which there are events (changes from 0 to 1, so to say), and the other obs drop out -- which one doesn't want. The only way that you can make progress with FE models in this context is by having (typically lots of) repeated spells for each subject  (firms in your  case).   [Paul Allison has written on this: googling should find the relevant papers.]  Hence virtually all the models of duration with frailty use random effects specifications -- their benefits are gained at the cost of having to assume that the individual subject-level frailties and the predictors are uncorrelated.

Stephen
------------------
Stephen P. Jenkins <[email protected]> 
Survival Analysis Using Stata: http://www.iser.essex.ac.uk/survival-analysis




------------------------------

Date: Sat, 1 Mar 2014 11:21:07 -0800
From: Jacob Model <[email protected]>
Subject: Re: st: Question about xtlogit and temporal dummy variables

Thanks, Stephen. That was a really thorough (and helpful answer).

I think I'll go with xtcloglog. Given how close the cloglog and the
logit results were in Beck's paper, I'd be surprised if there was a
substantive difference between the specifications. In fact, when I ran
it the results were similar when running it with logit, cloglog or
xtlogit or xtcloglog...

As far as worrying about frailty... I have firm-year observations. The
other thing I tried to do to control for time-invariant unobserved
heterogeneity across was to run xtlogit (with the time dummies) with
fixed effects for firms. This specification has it's own problems
(e.g., some firms in my sample become excluded by construction because
they never experience an event).

- -Jacob

On Sat, Mar 1, 2014 at 7:54 AM,  <[email protected]> wrote:
> Jacob Model <[email protected]>:
>
> If you simply applied -xtlogit- to your TSCS (time-series cross-section) data set-up, you would be assuming that the (discrete) hazard were constant.
>
> The point of the Beck et al. paper is to show to quant political science analysts that TSCS data with a binary outcome variable (BTSCS) are of exactly the same structure that one would use to fit a discrete time hazard regression model.
>
> Hence they recommend creating a set of dummy/binary variables that correspond to the amount of time since the start of the spell (e.g. if modelling onset of peace in year T, then the binary variables indicate the years since war broke out). By using a set of variables in this way, the interval-censored "baseline hazard" is non-parametrically specified. Parametric specification of time-at-risk are also possible. In fact, note that to apply their proposed method 'out of the box' you would use -logit- not -xtlogit-.
>
> Note that if you really want a discrete time PH hazard model, then use -cloglog- rather than -xtlogit-.
>
> Note also the Beck et al.'s section 3.3 on "complications", especially on how to handle multiple spells and left-censored spells.  For the former aspect, you could control for correlations of unobserved factors across spells using -xtcloglog- (allows for normally distributed frailty) -- or -xtlogit- if you wish to persist with a logistic model. For country-year data, -xtset- the data: the iis variable is country and tis variable is year. Left-censoring is rather difficult to deal with, without advanced methods (and assumptions).
>
> I suggest that you consult some standard texts on discrete time survival analysis. There are some citations at the website below my signature
>
> Stephen
> ------------------
> Stephen P. Jenkins <[email protected]>
> Survival Analysis Using Stata: http://www.iser.essex.ac.uk/survival-analysis
>
> ------------------------------
>
> Date: Fri, 28 Feb 2014 14:40:33 -0800
> From: Jacob Model <[email protected]>
> Subject: st: Question about xtlogit and temporal dummy variables
>
> I had a question about Beck, Katz and Tucker (1998)'s recommendation
> for using temporal dummies to approximate a proportional hazard model.
> (http://www-personal.umich.edu/~franzese/BeckKatzTucker.TakingTimeSeriously.AJPS1998.pdf)
>
> I was trying to understand exactly if temporal dummies are necessary
> (or desirable) when using xtlogit to estimate a discrete time event
> history model.
>
> I wasn't sure (and couldn't really tell in reading the documentation)
> to what extent xtlogit already incorporates this in its estimates or
> does the user have to specify them? Or if specifying them might
> interact with any correction that xtlogit does already.
>
> Thanks for your input!
>
> Best,
> - -Jacob
>

Please access the attached hyperlink for an important electronic communications disclaimer: http://lse.ac.uk/emailDisclaimer

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index