12

I understand the mechanics of calculating the weights using the propensity scores $p(x_i)$: \begin{align} w_{i, j={\rm treat}} &= \frac{1}{p(x_i)} \\[5pt] w_{i, j={\rm control}} &= \frac{1}{1-p(x_i)} \end{align} and then applying the weights in a regression analysis, and that the weights serve to "control for" or disassociate the effects of covariates in the treatment and control group populations with the outcome variable.

However on a gut level I don't understand how the weights achieve this, and why the equations are constructed as they are.

gung - Reinstate Monica
  • 132,789
  • 81
  • 357
  • 650
RobertF
  • 4,380
  • 6
  • 29
  • 46

1 Answers1

10

The propensity score $p(x_i)$ calculated is the probability of subject $i$ to receive a treatment given the information in $X$. The IPTW procedure tries to make counter-factual inference more prominent using the propensity scores. Having a high-probability to receive treatment and then to actually receive treatment is expected, no counterfactual information there. Having a low-probability to receive treatment and actually receiving treatment is unusual and therefore more informative of how treatment would affect subjects with low probability of receiving it; ie. characteristics mostly associated with control subjects. Therefore the weighting for treatment subject is $\text{w}_{i,j=\text{treat}} = \frac{1}{p(x_i)}$ adding more weight to unlikely/highly-informative treatment subjects. Following the same idea, if a control subject has a large probability of receiving treatment it is an informative indicator of how subjects in the treatment would be behave if they were in the control group. In this case the weighting for control subjects is $\text{w}_{i,j=\text{control}} = \frac{1}{1-p(x_i)}$ adding more weight to unlikely/highly-informative control subjects. Indeed, the equations at first instance can appear somewhat arbitrary but I think that they are easily explained under a counter-factual rationale. Ultimately all matching/PSM/weighting routines try to sketch out a quasi-experimental framework in our observational data; a new ideal experiment.

In case you have not come across them I strongly suggest you read Stuart (2010): Matching Methods for Causal Inference: A Review and a Look Forward and Thoemmes and Kim (2011): A Systematic Review of Propensity Score Methods in the Social Sciences; both are nicely written and serve as good entries papers on the matter. Also check this excellent 2015 lecture on Why Propensity Scores Should Not Be Used for Matching by King. They really helped me build my intuition on the subject.

usεr11852
  • 33,608
  • 2
  • 75
  • 117
  • Thank you, great answer! Of course, the reasoning behind the weight formulas is obvious in hindsight. I've looked at the 2015 King article. Very informative, although if I achieve excellent balance with propensity score matching w/out trimming, then why not use propensity scores? – RobertF May 09 '17 at 17:50
  • 1
    I am glad you find it helpful. Regarding King (2015): If we achieve excellent balance through PSM we should use PSM. The issue is that PSM commonly does *not* achieve excellent balance as we would have in a fully blocked randomized experimental design because it was not designed to do so. – usεr11852 May 09 '17 at 19:14
  • Brilliant reply, @usεr11852 – Nicg Oct 04 '19 at 14:44
  • Thank you. That's nice of you to say. – usεr11852 Oct 05 '19 at 22:38
  • One more thing I'll add to this answer - after multiplying each subject by an inverse propensity score weight, the total # of subjects within each variable category (age, sex, etc.) in the treatment and control groups are identical, permitting an unbiased estimate of the average treatment effect. – RobertF Feb 06 '21 at 17:10