Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: svy variance estimation: delete 1 jackknife & svymlogit and syvclogit?

From   [email protected] (Jeff Pitblado, StataCorp LP)
To   [email protected]
Subject   Re: st: svy variance estimation: delete 1 jackknife & svymlogit and syvclogit?
Date   Mon, 18 Jul 2005 09:48:08 -0500

Alexander Cavallo <[email protected]> asks two general questions
about the survey data capabilities in Stata:

1.  What happens when you use -svy jackknife- with data from a design with 2
PSUs in each stratum?

2.  Why isn't there a -svyclogit- command?

Question 1:

> I am using complex survey data with stratified and clustered design.  I 
> have over 200 strata and 2 PSUs per strata (or are these clusters??).
> If I want to use the delete 1 jackknife for variance estimation, wouldn't 
> each replication involve a strata with a singleton PSU?

Yes.  However -svy jackknife- uses the "replicated" point estimates from the
delete-1 fits to compute it's own variance estimator for the model parameters.
These delete-1 fits do not employ the linearized variance estimator; in fact,
the variance from each of these delete-1 fits are entirely discarded.

> Can I do jackknife variance estimation in this case?

Yes, the above procedure works.

> Could I drop an entire strata and use the jackknife calculations for delete
> 1 strata instead?

Only if you change the way you -svyset- your data.  I'm not recommending this,
but you could treat your strata variable as a PSU variable in your -svyset-,
then -svy jackknife- will delete your strata instead of the PSUs.

Question 2:

> I am interested in estimating the CLOGIT model with complex survey data. I 
> am estimating a choice model with some variables that are attributes of 
> the choice X(j) and others that are characteristics of the individual 
> Z(i).
> The model is
> Pr(i choses j) = exp[beta*X(j) + gamma(j)*Z(i)] / sum k=1 to M { 
> exp[beta*X(k) + gamma(k)*Z(i)}
> The model can be estimated by converting the data to long format where 
> each observation becomes M observations with a single instance of the 
> indicator S=1 for the chosen category.  To estimate the gamma(j) 
> paramaters I need to create interactions of Z(i) with with dummy variables 
> for the choices.
> I am planning to use jackknife deletion of one PSU to estimate the 
> variance of the parameter vector.
> My questions are:
> Why is there no svyclogit - is there a theoretical reason that it is not
> implemented?

No.  Until recently there were some technical reasons why we hadn't
implemented -svy: clogit-.  I can make no guarantee, or even give a timeline;
but we are working to get the -svy- prefix to work with other estimation
commands, including -clogit-.

For the time being, you can use -svy jackknife- with -clogit- in the following

1.  Make sure to -svyset- using -iweights- instead of -pweights-, -clogit-
currently will not accept -pweights-.

2.  Prefix your call to -clogit- with -svy jackknife _b:-, this will bypass
-svy jackknife-'s check for supported estimation commands.  This allows users
to experiment with -svy-'s resampling methods on commands that are not
identified as supported.

3.  Use -svy- to replay the estimation results, -clogit- does not recoginize
survey estimation results and does not display a proper header.

[email protected]
*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index