1
$\begingroup$

I currently have a procedure doing propensity score matching, and I use the fitted propensity scores (obtained via a glm call) and match on those. It turns out that I have about 60-70% more fitted values that are less than 0.5 than those greater than 0.5.

Hence, my fitted scores are skewed left. I am wondering if this violates any rules or what the implications might be from this and if it is something that can be assumed away? Thanks.

$\endgroup$
  • 2
    $\begingroup$ There is no reason for fitted propensity scores to have mean or median equal to 0.50. No assumption is violated. At first instance, "clip them" slightly (i.e. drop instance outside $[0.025, 0.0975]$) and use the PS as planned (through IPTW, inclusion as additional covariate, etc.). $\endgroup$ – usεr11852 Nov 27 '18 at 23:59
1
$\begingroup$

There is no reason for fitted propensity scores to have mean or median equal to 0.50 as there is not strict assumption that dictates such ranges.

At first instance, I think it will be more relevant to "clip" your existing PS slightly (i.e. drop instance outside [0.025,0.0975]) and use the PS as planned (through IPTW, inclusion as additional covariate, etc.). Lee et al. 2011. Weight Trimming and Propensity Score Weighting provides more detailed advice on this. Lee and his co-authors found that clipping is not always beneficial for certain kinds of estimators.

Particular to the issue you describe I would also note the following:

  1. if the range of the estimated PS is too narrow this might suggest that a potentially confounder is unaccounted for. Simply put, we do not have ignorability. The potential outcomes $Y$ are not independent of the treatment variable $A$ even if we condition on some other variables $X$.
  2. we should always aim to have common support between treatment and control groups; if we do not, there is a possibility that the differences we are observing in PS ranges is due to some form of selection bias.
  3. we might simply overfit our treatment assignment data $A$. This would have the direct consequence that our PS estimates are artificially over-determined. We should focus into getting a well-calibrated and properly estimated method to estimate the PS to begin with.
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.