CausalAnalysis function results for Market access problem [Question]

[salman-moh]

Hi, without giving too much information we are using the DML-RF with this CausalAnalysis function to get out point estimates for a given treatment. The outcome is market share and the problem is some point estimates are turning out to be negative but out assumption states that given T=1 or T=0 the treatment can never cause to negative market share, so I'm wondering how can we impose such an assumption? Or is it wrong to impose such an assumption? (this is an assumption from the business team).

Many thanks in advance :)

[kbattocchi]

Without more information it's hard to give advice, but there is no way to constrain the output of the CausalAnalysis class to be positive. You could always post-process the results to truncate any negative effect to zero if that is a hard requirement.

However, in general I'd say that if you are receiving surprising results, then it's worth further investigating what might be going on. One thing to check is the size of the confidence intervals (e.g. maybe it would not be surprising to have a slightly negative point estimate at the center of an enormous confidence interval that does not rule out some positive values). If that's the case, you might choose only to report estimates for treatments that result in tight confidence intervals and omit the treatments that result in loose bounds. Alternatively, if there are still negative point estimates with tight bounds, it might be worth examining your data more closely, or revisiting the business assumption that treatments must always have positive effects.