Thread by @isabelfulcher

Out today is our paper on estimating causal effects on a network! This was joint work with twitterless Eric Tchetgen Tchetgen and Ilya Shpitser.

This project occupied a significant portion of my PhD… and about 10% of my life… so enjoy!

tandfonline.com/doi/full/10.10…

1/n

We wanted a method to estimate causal effects in complex settings:

- The outcome of one person (1) is influenced by others’ outcomes (2) depends on others’ exposures
- Persons cannot be divided into distinct clusters (“network hairball”)
- The exposure is not randomized

2/n

Specifically, we were interested in the expected change in a person’s outcome corresponding to a hypothetical change in:

their exposure, while the exposures of other persons stay the same (direct)

other persons’ exposures, while their exposure stays the same (spillover)

3/n

The identification of these causal effects requires – to no one’s surprise – assumptions! We assumed network versions of the consistency, positivity, and unmeasured confoundedness standard in causal inference.

4/n

Another *key* assumption is that the network’s outcome, treatment, and covariate vectors arise from a chain graph model.

IMO, this is an intuitive way to capture conditional independences between individuals - especially in a social network setting.

5/n

For estimation, we utilized Besag’s auto-models for the outcome and covariates AND a glorious amount of Gibbs samplers (shoutout to @hms_rc and #Rcpp).

And, voila, auto-g-computation!

6/n

2017:

2018:

2019:

2020: HOW CAN THIS METHOD BE USED TO UNDERSTAND #COVID19 TRANSMISSION?!

Me: Perhaps we can use observational data to estimate the probability of COVID-19 infection for an unvaccinated individual under various network vaccination rates!

7/n

In practice, this requires that the network structure AND all relevant variables (confounders!) are *accurately* measured.

In addition, careful consideration of how network connections are defined is necessary. Does the chain graph model hold?

8/n

Thanks for tuning in! 👏

Please see the parallel work by @BetsyOgburn, @ildiazm & team (arxiv.org/pdf/1705.08527…) as well as additional thoughts on using chain graphs models with social network data (arxiv.org/abs/1812.04990).

9/n