[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
re: RE: st: propensity score analysis time-varying treatment
"Ariel Linden, DrPH" <firstname.lastname@example.org>
re: RE: st: propensity score analysis time-varying treatment
Wed, 14 Nov 2012 11:49:50 -0500
I just read through the Leon et al. article* (by the way, you gave an
incomplete reference. See below for the entire reference). The method is not
really longitudinal per-se (in the classical sense), but comparing
non-treated "periods" to treated "periods". I don't find this approach to be
as intuitive as the weighting approaches I gave you references for below. At
the very least, I think you should read through the following article:
Hernán MA, Brumback B, Robins JM. Estimating the causal effect of zidovudine
on CD4 count with a marginal structural model for repeated measures. Stats
in Med 2002;21:1689-1709.
All that said, Leon et al., describe how the propensity score was estimated
in section 2.2.1.
"Because subjects could have multiple exposure intervals, in which during
each interval they either received treatment or did not,
a longitudinal implementation of the propensity score that has been shown to
reduce selection bias  was applied here:
Propensity scores were calculated using parameter estimates from
mixed-effects logistic regression analyses, which included a random subject
effect to account for the multiple observations within each subject."
Next, they created quintiles of the propensity score and then tabulated the
number of periods exposed and unexposed for each quintile (table 1). Table 2
then shows the outcomes for each of these exposed vs unexposed categories.
In regards to your initial questions:
First, -predict- already gives you the predicted value (probability of
treatment) you need for analysis. There is no need to run -generate-.
Next, while you could run standardized differences here (and for that I
would suggest a user written program -pbalchk- [findit pbalchk], I am not
convinced of its utility here. After all, these are the same individuals in
both groups, so what would you expect to see different?
As for you last question about average treatment effects for the untreated
(or controls), again it points to a weighted model as an alternative to the
approach in the Leon et al. paper. Using weighting, you could estimate the
ATE (also known as IPTW), the ATT or the ATC with the following respective
gen iptw = cond(treatvar, 1/propvar, 1/(1-propvar))
gen att = cond(treatvar, 1, propvar/(1- propvar))
gen atc = cond(treatvar, (1-propvar)/propvar, 1)
I hope this helps
* Leon AC, Demirtas H, Li C, Hedeker D. Two propensity score-based
strategies for a three-decade observational study: investigating
psychotropic medications and suicide risk. Statist. Med. 2012, 31 3255?3260
Date: Tue, 13 Nov 2012 19:31:32 -0800
From: CJ Wilson <email@example.com>
Subject: RE: st: propensity score analysis time-varying treatment
Hi Ariel, thank you very much for your response. I am basing the approach on
the following references:
Leon AC, Demirtas H, Li C, Hedeker D. Two propensity score-based strategies
for a three-decade observational study: investigating psychotropic
medications and suicide risk. Stat Med. 012 Aug ..
Also from the book chapter: Analysis of Observational Health Care Data Using
SAS, "A Two-Stage Longitudinal Propensity Adjustment for Analysis of
Observational Data" Andrew C. Leon, Donald Hedeker, Chunshan Li.
> From: firstname.lastname@example.org
> To: statalist@hsphsun..harvard.edu
> Subject: re: st: propensity score analysis time-varying treatment
> Date: Sun, 1 Nov 012 0::2::7 -500<
> I am not sure why you have decided to use stratification when your data is
> longitudinal and treatment is time-varying? What would your rationale be
> that? A more sensible approach would be to use a marginal structural model
> or g-computation. I suggest you read the references below.
> Robins JM, Hernán MA, Brumback B. Marginal structural models and causal
> inference in epidemiology. Epidemiol 000;;1::50??0..
> Robins JM. Marginal structural models. In: 997 Proceedings of the Section
> on Bayesian Statistical Science. Alexandria, VA: American Statistical
> Association, 998::??0..
> Robins, J. M., and Hernan, M. A. 009.. Longitudinal Data Analysis, chap.
> Estimation of the causal effects of time-varying exposures, 53--99.. New
> York: Chapman and Hall / CRC Press.
> Fewell Z, et al. Controlling for time-dependent confounding using marginal
> structural models. The Stata Journal (004)) ,, Number ,, pp. 02??20<
> Daniel RM, et al. gformula: Estimating causal effects in the presence of
> time-varying confounding or mediation using the g-computation formula. The
> Stata Journal Volume 1 Number :: pp. 79--17<
> Date: Sat, 0 Nov 012 2::3::3 -800<
> From: CJ Wilson <cwil11111@@hotmail.com>
> Subject: st: propensity score analysis time-varying treatment
> Hi all, I posted the below message awhile ago but didn't get any
> Any help at all would be much appreciated. thanks!
> Dear Statalist,
> I?m trying to determine the correct Stata code to conduct a propensity
> stratification analysis in a longitudinal dataset. My variable for the
> treatment is time-varying, so some patients receive treatment at certain
> time points but not others. I haven?t been successful at finding any
> examples of Stata code online that are for longitudinal propensity scores.
> Here is my proposed approach:
> .. Estimate a random-intercept logistic regression model for the
> of treatment using xtlogit
> .. Calculate the propensity score as follows:
> predict prop_score
> gen prop_score = exp(prop_score)/(++exp(prop_score))
> xtile ps_quintiles = prop_score,, nq())
> tabulate ps_quintiles, ge(q)
> .. Determine whether there is a treatment by propensity interaction
> Here are a couple areas where I'm struggling:
> .. I would like to determine the standardized difference in means to
> whether balance is achieved. Does anyone have any sample Stata code
> to this?
> .. Does anyone have sample code to determine the average treatment effect
> among the untreated?
> Any advice would be much appreciated.
> Thank you!!
* For searches and help try: