re: Re: st: Using AIPW for missing data purposes in RCTs?

["Ariel Linden, DrPH" <ariel.linden@gmail.com>]

The CONSORT statement* (which is an agreed upon set of guidelines for
reporting clinical trials) states the following:

"Similar recommendations apply to analyses in which adjustment was made for
baseline variables. If done, both unadjusted and adjusted analyses should be
reported. Authors should indicate whether adjusted analyses, including the
choice of variables to adjust for, were planned. Ideally, the trial protocol
should state whether adjustment is made for nominated baseline variables by
using analysis of covariance. Adjustment for variables because they differ
significantly at baseline is likely to bias the estimated treatment effect.
A survey found that unacknowledged discrepancies between protocols and
publications were found for all 25 trials reporting subgroup analyses and
for 23 of 28 trials reporting adjusted analyses."

I interpret this to mean that if you have conducted an RCT where there is
supposedly no imbalances on covariates (because the randomization worked),
you should not be using adjusted data for reporting the outcomes. In fact,
it explicitly states that even if there are significant imbalances on
baseline covariates, adjusting for them can further bias the treatment
effect estimates)... Finally, you should use adjustments if your original
proposal/plan called for such analyses, but then you need to state why you
did it both ways.

I submitted a paper on an RCT in which I provided adjusted outcomes (because
of some imbalanced covariates), and the reviewers told me to provide only
the unadjusted results.

As for the -dr-, it is not intended for longitudinal data (nor is aipw). For
that, I refer back to the reference I provided earlier...

Ariel

* http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2844943/


Date: Wed, 26 Jun 2013 17:45:55 -0400
From: Steve Samuels <sjsamuels@gmail.com>
Subject: Re: st: Using AIPW for missing data purposes in RCTs?

Ariel:

Adjustment for covariates in clinical trials can correct for chance
imbalances and also reduce the variability of estimated treatment
effects (Piantadosi, 2005, pp. 470-475). Adjutment can also be useful
for creating prognostic models and for subgroup analyses (interactions
with treatment), but the latter run into the multiple-comparison
problem.

Reference: Steven Piantadosi. 2005. Clinical Trials: 
A Methodologic Perspective. Hoboken, NJ: Wiley-Interscience.

IPTW estimators, originally designed for observational stuies
are  also applied to randomized studies:

http://www.ncbi.nlm.nih.gov/pubmed/22510495

http://biostatistics.oxfordjournals.org/content/early/2012/04/17/biostatisti
cs.kxs009.abstract

http://www4.stat.ncsu.edu/~davidian/synergytalk.pdf

Van Der Laan, Mark. 2011. Targeted Learning: Causal Inference for
Observational and Experimental Data. Springer.


Stephen:

A "doubly-robust" estimator similar, but not identical to, Stata's
- -teffects aipw- estimator is Mark Lunt's -dr- which can be obtained by:

. net from http://personalpages.manchester.ac.uk/staff/mark.lunt
. net install dr


Steve
sjsamuels@gmail.com



*
*   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/