http://www.andrewheiss.com/blog/2020/02/25/closing-backdoors-dags/

[andrewheiss]

Ways to close backdoors in DAGs | Andrew Heiss
http://www.andrewheiss.com/blog/2020/02/25/closing-backdoors-dags/

[andrewheiss]

Originally posted by Steve Miller on 2020-02-25 12:29:17

So glad you wrote this as it's giving me some ideas for how to retool
my grad-level policy methods class in its next iteration.

[andrewheiss]

Originally posted by GrumbleSquasher Huang on 2020-02-26 03:14:43

Hi-- How different is this from just using the simplest approach (not
shown in the post) which gives the effect specified (10) with the
smallest standard errors (.26)?

>
> > m1 <- lm(final_grade ~ math_camp + undergrad_gpa + gre_quant,
> data = math_camp_guessed_need)
> > tidy(m1)
> # A tibble: 4 x 5
> term estimate std.error statistic p.value
> <chr> <dbl> <dbl> <dbl> <dbl>
> 1 (Intercept) 56.5 3.85 14.7 1.77e- 46
> 2 math_campTRUE 10.0 0.257 38.9 2.72e-247
> 3 undergrad_gpa 4.80 0.300 16.0 2.60e- 54
> 4 gre_quant 0.452 0.0291 15.5 2.47e- 51

[andrewheiss]

Originally posted on 2020-02-26 03:20:36

In this case, just controlling for the backdoors in a regression model
like that works just fine. That will only work, however, if all the
relationships between the confounders and treatment and outcome are
linear, which they are here because I made them that way. If there are
any non-linearities or anything that's better modeled
non-parametrically, using controls in regular OLS won't work (but IPW
and matching should work)

[andrewheiss]

Originally posted by GrumbleSquasher Huang on 2020-02-27 02:08:16

Interesting-- When using the propensity scores-- should you use the
squared or cubed terms as well to account for nonlinearities (if they
are present) or are you suggesting that those are not needed (and will
just work)? Thanks.

[andrewheiss]

Originally posted on 2020-02-27 04:14:46

That I'm not entirely sure of. I've seen people talk about using any
methods to get the most accurate propensity scores (even things like
random forests and neural networks), though based on this
tweet,
those techniques can overfit propensity score models. So maybe include
squared and cubed terms? And maybe not? :)

[andrewheiss]

Originally posted by huts on 2021-11-12 23:05:34

Great introduction to methods for causal inference (CI) with
observational data. I am still new to the field, but this article
provides a great resume to the techniques explained in Coursera's crash
course introduction to CI. Thank you!

If I am correct, the techniques you detail in this article (matching,
IPWT) are relevant when the confounders are observed & measured. I
thought one could always find a way out when confounders were measured
but you say that "it's possible that [these methods] [...] don't
work when there are non-linear relationships". Can you point to
references showing when things don't work anymore, even when measuring
all possible confounders?