Finding causality using text data

Question

I have read several recent papers on new machine learning methods to find causality using textual unstructured data. Here are some famous examples, and I am still not sure if we can hold the SUTVA assumption in text data. For example, if we are interested in studying the impact of exposure to Russian propaganda accounts (i.e. our treatment variable) on voters' choice and preferences in the 2016 election, using Reddit data, is it sufficient to say that we have the following treatment and control groups:

1-treatment group refers to those in political Pro-Trump sub-reddits that were exposed to an account identified as a Russian propaganda account, and 
2-the control group refers to those in political pro-Clinton sub-reddits that were NOT exposed to Russian misinformation accounts.

Specifically, I am confused because even with the research design above, I am not sure if we can assume that individuals in the two groups don't influence each other (e.g. by participating in both Trump and Clinton sub-reddits). If this is indeed the case, then this makes it difficult to attribute online behavior and expressed political preferences to mere exposure to Russian accounts, correct?

A follow up question to the great points raised below:

How would we explain a null change in the behavior of a Clinton supporter in a pro-Trump subreddit (i.e. treated individual but with no statistically significant effect due to treatment), would that weaken the argument that exposure to Russian misinformation increases Trump support? 

Answer 1

First of all, considering only treated individuals that are also in pro-Trump subreddits and only untreated people that are also in pro-Clinton subreddits will not give you a good estimation of the impact of Russian propaganda accounts. You want to have treated and untreated in both subreddits. Otherwise, you don't know whether the voters' choice is really influenced by Russian or whether it is just the political orientation (expressed by what subreddit is read). See e.g. matching.

The SUTVA is independent of the above problem. It just says that

• the outcome of the treatment of an individual is independent of the treatment other individuals receive, and
• the outcome of the treatment must be independent of how the treatment was assigned.

E.g., as you already correctly pointed out, if exposure of an individual's friends to Russian propaganda will influence that individual's outcome of the treatment, the SUTVA is violated.

In summary, first, you have to make sure that the data has been collected in compliance with SUTVA. Then, to compute the average treatment effect, you have to use an appropriate method like matching.