How are the correlations calculated?Linear (Pearson) correlations were computed by assigning numeric values to variables as follows.
|Main answers||For each possible answer A to a given question Q, we assign the variable 'Q:A' either an undefined value or a defined a value between -2 and 2 as follows:
|Metasurvey accuracy||The variable 'metasurvey:accuracy' encodes the negation of respondents' normalized average error (?).|
|Chronological variables||Values are left unchanged.|
|Gender||'gender:female' = 2 if the respondent specified 'female' as gender, -2 if the respondent specified 'male' as gender, and undefined otherwise.|
|Tradition||For the purposes of calculating linear correlations, we have considered only the analytic-continental dimension. The variable 'tradition:analytic' = 2 when the respondent specified 'analytic' as tradition, -2 when the respondent specified 'continental' as tradition, and 0 otherwise.|
|All other answers||All other answers were treated as categorical variables. For example, an individual has 2 for 'AOS::Metaphysics' if he or she has specified Metaphysics as AOS, and -2 otherwise.|
Often, positively (or negatively) correlated variable pairs have corresponding negatively (or positively) correlated variable pairs. We have sought to display only the most natural correlation (usually the positive correlation).
Significance of correlationsWe have not included significance measures, partly because these are problematic when so many correlations are being calculated. But to give a rough idea, a one-off correlation coefficient of 0.1 (our cutoff for inclusion on the "most correlated answers" list) over a population of 931 subjects (the number of target faculty respondents, whose answers are used here) yields a p-value of 0.002.
Of course given that there are 5,675 pairs of answers to correlate, one would then expect that even in the absence of genuine correlations, there would be around 11 correlation coefficients that are over this cutoff by chance alone.