Re: st: FW: how to do binary mediation with survey data

[Steve Samuels <sjsamuels@gmail.com>]

As a followup to Ariel's post:  Both -medeff- and  -khb-  accept pweights and vce(cluster), so you can get a valid survey analysis with both and benefit from the superior accuracy that Ariel reports.  (You would be ignoring strata, which might inflate SEs a bit, but would not affect the estimates)

Steve

Craig, you are asked in the FAQ to provide a link to user-written
commands. -binary_mediation- is by Phil Ender and -findit- will provide
a link.

You can use -binary_mediation- with survey data if you -bootstrap- the
command, identifying strata and clusters (PSUs) in the -bootstrap-
option list.  The -help- for -binary_mediation- provides basic -bootstrap- code.
(If the survey data set provides bootstrap, brr, sdr, or jackknife
replicates, then you can -svyset- your data first and use the -svy-
prefix.) You can't analyze the weighted data with plain -bootstrap-, but
for causal modeling, ignoring the weights would be acceptable to many ,
especially if you include in the covariate list variables related to
weighting (household size comes to mind).

Steve


On Nov 15, 2012, at 9:33 AM, Ariel Linden, DrPH wrote:

Craig

I suggest you consider -medeff- or -khb- for categorical outcome variables
(both are user-written programs [findit medeff, findit khb].

I am not sure if either program works with -svy-, however both approaches
allow for pweights (though I am not sure how that will impact the results). 

In a paper that Kristian Karlson and I currently writing, we ran large scale
simulations to test the accuracy of several mediation modeling approaches.
Both of these programs (-medeff and -khb-) recover the true mediation
effects with the highest accuracy while binary_mediation provided rather
inaccurate results.

I hope this helps 

Ariel

Date: Wed, 14 Nov 2012 12:03:46 +0000
From: "Morgan, Craig" <craig.morgan@kcl.ac.uk>
Subject: st: FW: how to do binary mediation with survey data

Hi

I want to do a mediation analysis using survey data.  I've previously used
the binary_mediation' command, but this doesn't allow the prefix 'svy:'. 
Can anyone advise on how I can do a mediation analysis with prefix 'svy:'
when the outcome variable is dichotomous?

Thanks

Craig Morgan



*
*   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/



*
*   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/