Inquiry: re konfound for interaction terms

[David A]

Hello, 

First off, thanks so much to the KonfFound-It! team for all the work you and your colleagues have put into a major public good for the social sciences. I really appreciate it.

I'm currently working on a Revise & Resubmit for a journal article in which my co-author and I have been asked to provide a sensitivity analysis for our main findings, which revolve around an interaction term from an OLS regression (between a dichotomous variable and a continuous variable). I'm a STATA user.

I see from the website Q&A that the suggested approach to this is RIR, but that the moderation effects section which would speak to models with interaction terms is under development (for what it's worth, I've given it a try with the current command and gotten an error message "interactions not allowed"). Would you happen to have any suggestions?

Many thanks for your guidance and best regards,  

David Attewell (Aarhus University and University of Zurich)

[Frank, Kenneth]

Thanks David glad you have found the konfound-it helpful.

 

We have worked out the technical details for Robustness of Inference  to Replacement (RIR) for interactions but not yet published the result.  We have a draft of a paper.  But the calculations have been embedded into the konfound packages.

 

We leverage a general approach in which one replaces data with hypothetical cases for which the relationship between X and Y is zero, conditional on covariates in the model.  For an interaction effect, conceptualize replacing observed data with hypothetical data for which the correlation between the interaction term and the outcome is zero, conditional on the main effects already in the model.

 

Currently, these have to be run using the pkonfound command in Stata, manually entering the estimated effect, standard error, etc., into a the pkonfound command or app.

https://konfound-project.shinyapps.io/konfound-it/

screen shot below.

 

 

 

 

The corresponding stata code would be

 

ssc install konfound

pkonfound 0.125 0.05 6265 7 0.217 0.991 0.251, eff_thr(0) fr2max(0.61) indx(COP)

return list

 

The relevant output given is as below.  The last line is the one you want, in this case RIR=1351.  As a % it equals 1351/6174=.22.  To nullify the inference, 22% of the data would have to be replaced with data for which the interaction was zero, conditional on the main effects in the model.

This function also calculates conditional RIR that invalidates the statistical inference.

 

If the replacement data points have a fixed value, then RIR = 2000.604.

If the replacement data points follow a null distribution, then RIR = 1096.205.

If the replacement data points satisfy rxy|Z = 0, then RIR = 1351.435

 

 

 

 

 

Check out the R Shiny app for sensitivity analysis or the web site: https://konfound-it.org/ or Podcast

 

Our approach to school governance: https://www.balancingvoices.org/

 

NSF Training SIARM for STEM: https://voices.uchicago.edu/nsf-siarm/

 

Study of Ambitious Math Instruction: https://www.studyofelemmath.org/

 

 

Ken Frank (zoom https://msu.zoom.us/j/783760435)

MSU Foundation Professor of Sociometrics

Measurement and Quantitative Methods

Counseling, Educational Psychology and Special Education

Room 462 Erickson Hall

620 Farm Lane

Michigan State University

East Lansing, MI 48824-1034 phone: 517-355-9567 fax: 517-353-6393 kenf...@msu.edu

https://sites.google.com/msu.edu/kenfrank

 

pronouns he/him/his

--
You received this message because you are subscribed to the Google Groups "KonFound-it!" group.
To unsubscribe from this group and stop receiving emails from it, send an email to konfound-it...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/konfound-it/fdad7fe0-7dd4-48c2-bfe8-a18612f6d37cn%40googlegroups.com.