Overlap assumption and Calibration test

[FedericaPetru]

Dear Team,

I am working with the grf package to study the effect of different shocks on consumption and I have two questions I wanted to ask you.

1. Regarding the Overlap assumption, I understand that to avoid a deterministic decision of the treatment status, the estimated propensity scores should be close to one or zero.
   My question is: is there a specific accepted range? I am using the dummy variable drought as treatment and the range of the W.hat goes from 0.250 to 0.285 as in picture "W.hat.png". I was wondering whether this is too small and does not exclude the overlap or if you see any issue I am not considering.

2. Regarding the Calibration Test, according to the rule: a coefficient of 1 for mean.forest.prediction suggests that the mean forest prediction is correct and that a coefficient of 1 for differential.forest.prediction additionally suggests that the heterogeneity estimates from the forest are well calibrated.
   My question is again related to the accepted range: how far from the ideal value of 1 is acceptable to still assert that the prediction is correct and that the forest are well calibrated. Is a result of 1.5 or 0.6 considered too far from the ideal benchmark?

Thank you in advance for your support.
Federica

[erikcs]

Hi @FedericaPetru, 1) the function average_treatment_effect(forest) will give you a warning if overlap seems to be an issue. 2) The output gives you a standard error you could use to construct a confidence interval.