Impact Threshold of a Confounding Variable (ICTV) test and Konfound package for non-linear (logit) model with interactions

[Roger Clements]

Hello all,

I am trying to testing for omitted variable bias. Specifically, how strong a correlated omitted variable would have to be to overturn the main results using the user-written Impact Threshold of a Confounding Variable (ICTV) package Konfound. I have a non-linear (logit) model with an interaction term.

So the command in Stata for my logit model is:

logit treat firm_size##firm_performance, nolog

Then I run Konfound as:

konfound firm_size firm_performance, sig(.05) non_li(1)

A few questions:
1. Is it correct to use Konfound (ICTV) for a non-linear model? I see some papers originally saying no, but Konfound has an option to specify non-linear models in Stata (non_li(1)) so I am wondering whether this has been updated/changed.

2. Is my code correct for Konfound? I am getting results for just the first variable (firm_size) in the list after running Konfound so not sure how to fix that because the package won't allow interact terms (##).

3. Any other advice would be much appreciated.

I have exhausted the ssc help for Konfound and all web resources so your help is appreciated.

Thank you.

Roger

[Phil Bromiley]

You didn't get a quick answer. You'll increase your chances of a useful answer by following th FAQ on asking questions - provide Sttaa code in code delimiters, readable Stata output, and sample data using dataex.
With a user written routine like Konfound, getting help on this list serve depends on someone happening to use it. You may need to contact the author.

[Ran Xu]

Roger,
This is a tricky question because of the non-linear model and the interaction term. First of all, in such a non-linear case the “impact threshold for a confounding variable” is not the best option. Instead, look at the % bias to invalidate the inference, which does not assume linearity. This can be interpreted as the % of cases you would have to replace with null hypothesis cases to invalidate the inference. Next, to work in the interaction term it might be easiest just to make the term outside of the command by making a new variable that is the product of firm size and firm performance. You can then include that in the model. Then what is a null hypothesis case for an interaction term? It is one where the effect of firm size does not depend on firm performance. So a null hypothesis cases would assign the same effect to firm size regardless of the firm’s level of performance. Note you’re not really performing this replacement – it’s just a thought experiment.