[CMay]

I'd strongly question having confidence in the wide applicability of the results in respect to class or wealth, much less loneliness which one would guess skews higher than a random sample due to their data source being Mechanical Turk.

Both this paper and another paper it links to make it a point to state possible concern with getting their data this way. Aside from that, there are several important factors you would want to account for and the more you want to split out the data to isolate influences the larger the sample size you need. Their total sample size was only ~950, making each nested subset sample size progressively smaller than that.