They got trolled. The story in Unherd and the National Review is based on the original version of this study which tracked, among other things, vaccine hesitancy by educational level.
This result got all the attention
Contrary to expectations, vaccine hesitancy is climbing sharply at the highest educational levels. No other study I am aware of has looked specifically at people with PhD’s and professional degrees, although every other study has shown vaccine hesitancy dropping with education to at least the Master’s level (92.4% of those with more than 4 years of college education report being vaccinated).
So what happened here?
The survey was a Facebook survey where users were asked to report their educational level. The fraction of self-reported PhD’s was small, only about 3% of the entire sample:
This is still roughly twice what you would expect from a random sample.
That alone isn’t too bad (Facebook use is slightly higher among the more educated) but …
Imagine you really hate vaccines. You want your opinion to be known. You are then asked to pick an educational level with no one checking …
Which one are you going to pick, if you are not replying honestly?
Probably the one with the highest perceived credibility. Because the PhD and professional categories are so small, it only takes a small percentage of anti-vax people to answer wrongly to completely skew those categories. A movement of only ~ 2% to the PhD category will skew the result upward by roughly 20%.
Is there any evidence this actually occurred?
The overrepresentation is one clue, but not the strongest one. The biggest clue is looking at what other responses the PhD holders gave.
From the latest version of the article:
“To be included in the analysis sample, participants had to complete the questions on vaccine uptake and intent, and report a gender other than “prefer to self-describe.”. This exclusion was made after discovering that the majority of fill-in responses for self-described gender were political/discriminatory statements or otherwise questionable answers (e.g. Apache Helicopter or Unicorn), and that as a group, those who selected self-described gender (<1% of the sample) had a high frequency of uncommon responses (e.g., Hispanic ethnicity [41.4%], the oldest age group [23.2% ≥75 years] and highest education level [28.1% Doctorate]), suggesting the survey was not completed in good faith.”
Once you remove elderly Hispanic self-identified Attack Helicopters the effect largely disappears
The moral of the story is that self-reported data should always be viewed with suspicion, even when people aren’t actively trying to screw with the survey. Small groups are especially prone to the effects of miscategorization and should be looked at with extra care.