Notice: On March 31, it was **announced** that Statalist is moving from an email list to a **forum**. The old list will shut down on April 23, and its replacement, **statalist.org** is already up and running.

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

From |
"Ariel Linden. DrPH" <ariel.linden@gmail.com> |

To |
<statalist@hsphsun2.harvard.edu> |

Subject |
re: re: Re: st: binary mediation command |

Date |
Sat, 25 Jun 2011 13:26:19 -0700 |

Kristian does an excellent job of adding some important points to the discussion of mediation analysis. A couple of follow-up items: VanderWeele's marginal structural modeling approach is not currently available in Stata (Steve Samuels and I have been noodling this around for about a year, and we will hopefully write such a program before we're put out to pasture). One other interesting program that is in Stata and worthy of note is -gformula-. There is an accompanying paper (found with the example data provided with the program: Daniel, R. M., De Stavola, B. L., and Cousens, S. N. 2010. gformula: Estimating causal effects in the presence of time-varying confounding or mediation using the g-computation formula. The Stata Journal (submitted)) This approach should address the very real issues that Kristian raised earlier as threats to validity of mediation analysis and can handle them in a somewhat different fashion than that of marginal structural models. Ariel Date: Fri, 24 Jun 2011 09:53:58 +0200 From: Kristian Karlson <kristian.karlson@gmail.com> Subject: re: re: Re: st: binary mediation command Let me add a bit to Ariel's suggestions. The method you choose will often depend on the question--and data--at hand. The Imai et al. approach is focused on the identification of causal indirect effects/mediation, not "descriptive" or "associational" indirects effects. The former is important if you have experimental and quasi-experimental data, while the latter often is used as a summary measure of concomitant variation in three or more variables. Sociologist S.L. Morgan has an excellent article forthcoming in Sociological Methods and Research, which discusses these issues (in a sociological debate on the origins of class inequalities in educational attainment). The methods by Erikson et al., Karlson et al., and Fairlie are "associational". They do not stipulate anything about causality (but see Morgan's article). The Imai et al. approach concerns "causal" indirect effects, and to identify these effects they assume "sequential ignorability", which means (1) that selection into treatment (independent variable) is random conditional on pre-treatment covariates, and (2) that selection into treatment is random conditional on treatment and pre-treatment covariates. While the first assumption is credible when treatment is a randomized experiment, the second assumption is restrictive and untestable (although Imai et al. suggest some simulation techniques to evaluate the robustness of results to this assumption). The weakness of this second assumption is best explained by Sobel, in his 2008 article, which I consider one of the most important contributions to this literature (http://jeb.sagepub.com/content/33/2/230.short). Sobel also develops the conditions under which we can identify causal indirect effects with randomized experiments (and these conditions are very restrictive). Other methods exist for the estimation of "causal" indirect effects: VanderWeele has some developments within the marginal structural models approach (http://journals.lww.com/epidem/Abstract/2009/01000/Marginal_Structural_Mode ls_for_the_Estimation_of.6.aspx). This approach assumes that section is random conditional on observables, which may be an overly restrictive assumption. I do not know whether VanderWeele's approach is implemented in Stata (but as far as I know there have been some writings on the implementation of marginal structural models in Stata). Best Kristian - ------------- From "Ariel Linden, DrPH" <ariel.linden@gmail.com> To <statalist@hsphsun2.harvard.edu> Subject re: Re: st: binary mediation command Date Fri, 17 Jun 2011 09:32:47 -0700 There are several other approaches to mediation analysis. One technique that I have recently become enamored with is a non-model based approach that can accommodate any combination of data types for the mediator and outcome. This eliminates the concerns that arise with forcing a non-linear variable into a linear approach (using the standard Baron-Kenny model). This program by Imai et al is in R, but they are about to come out with a Stata version (it's not as complete as the R version but handles most jobs you'd ever need). See: http://imai.princeton.edu/projects/mechanisms.html, in particular: Imai, Kosuke, Luke Keele and Dustin Tingley (2010) A General Approach to Causal Mediation Analysis, Psychological Methods 15(4) pp. 309-334. Imai, Kosuke, Luke Keele and Teppei Yamamoto (2010) Identification, Inference, and Sensitivity Analysis for Mediation Effects, Statistical Sciences, 25(1) pp. 51-71. I hope this helps Ariel * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/

- Prev by Date:
**Re: st: Collapse - How does it collapse** - Next by Date:
**st: Apostrophe in labels breaks graph** - Previous by thread:
**re: re: Re: st: binary mediation command** - Next by thread:
**st: how to calculate cumulative number of organizations for stcox?** - Index(es):