From kimai at Princeton.EDU Fri Nov 1 03:40:32 2013 From: kimai at Princeton.EDU (Kosuke Imai) Date: Fri, 1 Nov 2013 02:40:32 +0000 Subject: [Mediation-information] Questions about R mediation package In-Reply-To: <637E630C8BFD6E41B80119E733D549935BF4E476@SN2PRD0310MB394.namprd03.prod.outlook.com> References: <637E630C8BFD6E41B80119E733D549935BF4E476@SN2PRD0310MB394.namprd03.prod.outlook.com> Message-ID: Dear Jessica, My responses are below but I?m ccing my collaborators who are much more informed about the software. Also, you should look at the accompanying paper: http://imai.princeton.edu/research/files/mediationR2.pdf This paper details the software in terms of both implementation and interpretation. Best, Kosuke --------------------------------------------------------- Kosuke Imai Office: Corwin Hall 036 Professor Phone: 609-258-6601 Department of Politics Fax: 609-258-1110 Princeton University Email: kimai at Princeton.Edu Princeton, NJ 08544-1012 http://imai.princeton.edu --------------------------------------------------------- On Oct 30, 2013, at 11:17 AM, Jessica Dennis wrote: > Dear Dr. Imai, > > I am a PhD student in genetic epidemiology at the University of Toronto, and as part of my thesis work, I have been using your mediation package in R to do some mediation analyses. > > I'm having a little bit of trouble, however, with implementation and interpretation, and I was hoping you could help clarify a few things: > > ? Testing for interaction between the treatment and the mediator. I don't quite understand the output of the Test.TMint function. On the example on page 9 of the mediation documentation: > Test of ACME(1) - ACME(0) = 0 > > data: estimates from med.out > > ACME(1) - ACME(0) = 0.015, p-value = 0.42 > > alternative hypothesis: true ACME(1) - ACME(0) is not equal to 0 > > 95 percent confidence interval: > > -0.03865720 0.05212853 > > > > Since the 95% CI includes 0, does this mean that we fail to reject the null hypothesis? The wording in the output is a little ambiguous. You fail to reject the null hypothesis of no interaction in this case. > > > > Also, I had wanted to test the interaction between the treatment and mediator in an lmer model. When I apply the Test.TMint function, however, I get: Error in test.TMint.default(med.out.2, conf.level = 0.95) : currently no test.TMint method exists for the input object. > > > Does this mean that the Test.TMint function is unable to handle output from the lmer object? > We are currently working to add several functionalities to lmer models. We will put this on our to-do list. > > ? Sensitivity analysis for sequential ignorability. Can sensitivity analysis be useful when the ACME is non-significant? In all the examples I've seen online, sensitivity analysis is used to determine when an ACME becomes non-significant. Most commonly, sensitivity analysis is used to assess the robustness of empirical findings. So, if you have null findings, it may not make much sense to do this. > In my mediation analysis, I estimated the ACME to be (my treatment variable is continuous, with values 0, 1, and 2): > > Estimate 95% CI Lower 95% CI Upper p-value > ACME -0.0196 -0.0662 0.0201 0.31 > > I have the following output table from the sens.out function: > Sensitivity Region > > Rho ACME 95% CI Lower 95% CI Upper R^2_M*R^2_Y* R^2_M~R^2_Y~ > [1,] -0.9 0.1750 -0.1765 0.5266 0.81 0.4807 > [2,] -0.8 0.1061 -0.1072 0.3194 0.64 0.3798 > [3,] -0.7 0.0728 -0.0738 0.2195 0.49 0.2908 > [4,] -0.6 0.0511 -0.0522 0.1544 0.36 0.2136 > [5,] -0.5 0.0348 -0.0361 0.1058 0.25 0.1484 > [6,] -0.4 0.0216 -0.0233 0.0665 0.16 0.0949 > [7,] -0.3 0.0101 -0.0134 0.0336 0.09 0.0534 > [8,] -0.2 -0.0003 -0.0123 0.0117 0.04 0.0237 > [9,] -0.1 -0.0101 -0.0337 0.0134 0.01 0.0059 > [10,] 0.0 -0.0196 -0.0607 0.0215 0.00 0.0000 > [11,] 0.1 -0.0291 -0.0886 0.0305 0.01 0.0059 > [12,] 0.2 -0.0388 -0.1177 0.0400 0.04 0.0237 > [13,] 0.3 -0.0492 -0.1487 0.0503 0.09 0.0534 > [14,] 0.4 -0.0607 -0.1832 0.0617 0.16 0.0949 > [15,] 0.5 -0.0740 -0.2230 0.0750 0.25 0.1484 > [16,] 0.6 -0.0903 -0.2719 0.0913 0.36 0.2136 > [17,] 0.7 -0.1120 -0.3370 0.1131 0.49 0.2908 > [18,] 0.8 -0.1453 -0.4371 0.1465 0.64 0.3798 > [19,] 0.9 -0.2142 -0.6443 0.2159 0.81 0.4807 > > > Since the value of the ACME doesn't change appreciably across different values of rho (i.e., the 95% CI always includes 0), do I conclude that there are no unmeasured confounders of the mediator-outcome relationship? > > > Thank you very much for your input. > Jessica > > > Jessica Dennis > CIHR Vanier Scholar, CIHR STAGE Fellow > PhD Candidate in Epidemiology > Dalla Lana School of Public Health University of Toronto > 155 College St, Suite 734 > Toronto, ON M5T 3M7 From kimai at Princeton.EDU Tue Nov 26 03:49:00 2013 From: kimai at Princeton.EDU (Kosuke Imai) Date: Tue, 26 Nov 2013 02:49:00 +0000 Subject: [Mediation-information] Effect size from Mediation software In-Reply-To: <3b3d523f4b4c40bf97c365c794df414e@DBXPR01MB029.eurprd01.prod.exchangelabs.com> References: <3b3d523f4b4c40bf97c365c794df414e@DBXPR01MB029.eurprd01.prod.exchangelabs.com> Message-ID: <68D81CA4-E79D-42DC-80E2-DA49F38E50DF@princeton.edu> Hi Xiayi, The estimates are average causal mediation effect estimates. They do not necessarily correspond to coefficients. Have a look at this paper which gives a non-technical overview of causal mediation analysis: http://imai.princeton.edu/research/mediationP.html or http://imai.princeton.edu/research/BaronKenny.html Best, Kosuke --------------------------------------------------------- Kosuke Imai Office: Corwin Hall 036 Professor Phone: 609-258-6601 Department of Politics Fax: 609-258-1110 Princeton University Email: kimai at Princeton.Edu Princeton, NJ 08544-1012 http://imai.princeton.edu --------------------------------------------------------- On Nov 25, 2013, at 12:47 PM, Ke, Xiayi > wrote: Dear Kosuke Imai, I have been using your mediation analysis methods and software recently. I am wondering whether you can kindly advice me when the non-parametric bootstrapping option is used, what are the effect size "estimates" standing for, how are they obtained (from standardized regression coefficient beta?), and what is the best way to interpret them? I guess their values can only vary between 0 and 1? I look forward to hearing from you. BW, Xiayi. -------------- next part -------------- An HTML attachment was scrubbed... URL: From kimai at Princeton.EDU Tue Nov 26 14:44:59 2013 From: kimai at Princeton.EDU (Kosuke Imai) Date: Tue, 26 Nov 2013 13:44:59 +0000 Subject: [Mediation-information] Effect size from Mediation software In-Reply-To: <655fb81441154ae9b0ee4494caf17b51@DBXPR01MB029.eurprd01.prod.exchangelabs.com> References: <3b3d523f4b4c40bf97c365c794df414e@DBXPR01MB029.eurprd01.prod.exchangelabs.com> <68D81CA4-E79D-42DC-80E2-DA49F38E50DF@princeton.edu>, <655fb81441154ae9b0ee4494caf17b51@DBXPR01MB029.eurprd01.prod.exchangelabs.com> Message-ID: Xiayi, I would not recommend odds ratio as a quantity of interest. I realize that this is a typical quantity of interest reported in some fields, but nobody was able to give me an intuitive interpretation of that quantity. The problem is that odds ratio is a ratio of two ratios. If I ask you "what does it mean to have odds ratio of 3?", can you answer it intuitively so that general audience without technical training can understand it easily? I doubt it. Rather, you should report the risk differences as it is done in our software. Any patient can understand the statement that "your risk of death increases by 3 percentage points" but they have no idea if you tell them "your odds ratio of death is 3". Best, Kosuke Department of Politics Princeton University http://imai.princeton.edu On Nov 26, 2013, at 3:53 AM, "Ke, Xiayi" wrote: > Dear Dustin and Kosuke, > > I am sorry if I have been unreasonably persistent. > > I have explained to others more or less along the lines as suggested in your email and the two 2010 papers (e.g. Equation 1 and 2, Table 5, Appendix D etc from the Imai, Keele and Tingley paper) , but they wanted a more intuitive answers like the odds ratio type or percentage of effect. > > Many thanks for your help and your excellent work. > > BW, > > Xiayi. > > > On Mon, Nov 25, 2013 at 9:49 PM, Kosuke Imai wrote: > Hi Xiayi, > > The estimates are average causal mediation effect estimates. They do not necessarily correspond to coefficients. Have a look at this paper which gives a non-technical overview of causal mediation analysis: http://imai.princeton.edu/research/mediationP.html or http://imai.princeton.edu/research/BaronKenny.html > > Best, > Kosuke > > --------------------------------------------------------- > Kosuke Imai Office: Corwin Hall 036 > Professor Phone: 609-258-6601 > Department of Politics Fax: 609-258-1110 > Princeton University Email: kimai at Princeton.Edu > Princeton, NJ 08544-1012 http://imai.princeton.edu > --------------------------------------------------------- > > On Nov 25, 2013, at 12:47 PM, Ke, Xiayi wrote: > >> Dear Kosuke Imai, >> >> I have been using your mediation analysis methods and software recently. I am wondering whether you can kindly advice me when the non-parametric bootstrapping option is used, what are the effect size "estimates" standing for, how are they obtained (from standardized regression coefficient beta?), and what is the best way to interpret them? I guess their values can only vary between 0 and 1? >> >> I look forward to hearing from you. >> >> BW, >> >> Xiayi. > > > _______________________________________________ > Mediation-information mailing list > Mediation-information at lists.r-forge.r-project.org > https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/mediation-information From senor_massao at hotmail.com Thu Nov 28 13:19:47 2013 From: senor_massao at hotmail.com (Mashhood Sheikh) Date: Thu, 28 Nov 2013 12:19:47 +0000 Subject: [Mediation-information] Interaction between exposures? In-Reply-To: References: Message-ID: Dear Dustin and Kosuke, Is it possible to estimate the direct and indirect effects when there is an interaction between exposures (X1, X2), regressed on outcome Y? Since X2 is also a mediator-outcome confounder between the mediation model X1>M>Y, and X1 is also a mediator-outcome confounder between the mediation model X2>M>Y, it is important to include them in the model as covariates, but than what about the interaction between them? Note: all variables, X1, X2, M, and Y are binary. I would highly appreciate any tips... Thankfully, Mashhood Institute of Community Medicine, University of Troms?, Norway. -------------- next part -------------- An HTML attachment was scrubbed... URL: From kimai at Princeton.EDU Sat Nov 30 03:53:23 2013 From: kimai at Princeton.EDU (Kosuke Imai) Date: Sat, 30 Nov 2013 02:53:23 +0000 Subject: [Mediation-information] Interaction between exposures? In-Reply-To: References: Message-ID: <8EDD92A6-D2F3-4A25-9398-17CD569DFE25@Princeton.Edu> You can think of it as the four-category exposure. Now, the definition of indirect effects is a little bit different: in Y(t,M(t)) - Y(t,M(t?)) you have to pick t and t?. In the binary treatment case, this choice is obvious because there are only two values; t=1, t?= 0. When you have four category exposure, depending on your substantive question, you can pick different comparisons. Best, Kosuke Department of Politics Princeton University http://imai.princeton.edu On Nov 28, 2013, at 7:19 AM, Mashhood Sheikh wrote: > Dear Dustin and Kosuke, > > Is it possible to estimate the direct and indirect effects when there is an interaction between exposures (X1, X2), regressed on outcome Y? Since X2 is also a mediator-outcome confounder between the mediation model X1>M>Y, and X1 is also a mediator-outcome confounder between the mediation model X2>M>Y, it is important to include them in the model as covariates, but than what about the interaction between them? Note: all variables, X1, X2, M, and Y are binary. > > I would highly appreciate any tips... > > > Thankfully, > Mashhood > Institute of Community Medicine, > University of Troms?, > Norway. > _______________________________________________ > Mediation-information mailing list > Mediation-information at lists.r-forge.r-project.org > https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/mediation-information