Surveys are inherently difficult to present in a neutral fashion, especially when attempting to determine correlation. Take the following (simplified) survey for example:
I like Cheerios:
[Yes] [No] [Sometimes]
Rate your proficiency at math:
[Excellent] [Good] [Average] [Poor]
Now, let's say you found a statistically significant correlation between people who like Cheerios and people who are excellent at math. Congratulations! You just did not find a correlation related to math proficiency at all.
What you did just find is a correlation between people who selected the first option in your survey.
Now, randomizing your answers is a good start and will resolve the above issue. However, there are hundreds of other things which can affect your results and there is an entire survey industry formed around these problems. The immediate problems that spring to mind about the survey in TFA is:
-Respondents must have internet access
-Respondents must have signed up to Amazon's mechanical turk
-Respondents were paid for the survey
-Respondent proficiency at math/language was self-assessed
-Respondents must be able to comprehend English
Anyway, I could go on but my point here is this: despite the fact that a statistically-significant correlation that was found, that correlation may not stem from the questions themselves.