Picked together

Lift is how often two things are taken together against how often two things that popular would coincide by chance. A lift near 1 just means two popular choices met.

Two directors in one closet

A pair of films is too thin to say anything about: the highest co-occurrence anywhere in this data is 5 visits, and a lift on four visits is a coin toss with a decimal point. Directors are the same question at a unit with support, and the pairs below have had the films the two made together removed from both sides of the count: 18 pairs disappeared entirely when that was done, which is what a co-credit looks like when it is mistaken for a taste.

The other end is harder to read and more interesting. Among pairs where both directors are reached by at least 20 visits, these are the ones that almost never share a closet. Every interval below spans 1.0, so this is a list of candidates rather than a finding: with these counts the data cannot tell a real avoidance from chance.

Closest taste

Jaccard overlap between what two guests chose. Guests filmed together are left out: 32 pairs shared one list and would otherwise score a perfect 1.00 against each other.

Do two actors choose alike

Slightly, and it takes a permutation test to see it. Two visits by guests in the same trade share 0.0082 of their choices against 0.0062 for two guests in different trades, comparing only visits filmed in the same year. That difference survives 2,000 reshuffles of the trade labels within each year at a p-value of 0.0025. Both numbers are under one percent, so the honest sentence is that same-trade pairs are more alike than chance, not that they are alike.

Solo visits only, since a trade is attributed only to a guest filmed alone, and only visits with at least 5 choices. Pairs are compared within a filming year and the 2000 draws reshuffle trade labels inside each year, because the trades arrived in the series at different times and a pooled comparison would also be a comparison of eras. Labels are permuted per visit and the whole statistic is recomputed each draw, which is what makes comparing 33,000 overlapping pairs legitimate. Both means are under one percent: this is a very small average difference, and within_group_jaccard is the more interesting number, since the trades differ from each other in how much they agree among themselves. Jaccard also falls as lists grow, so mean_items is published beside it. A year needs 30 pairs on each side to get its own row, and 11 years are too thin for one: the test itself uses every same-year pair regardless, since the permutation handles small strata correctly and a display floor should not change a p-value.

Got there first

Being first is partly being early: the 2010 and 2011 guests had an empty field. Read this as who the series followed, not as who has the best taste.

Obscure and obvious

Popularity is how many visits took the same thing, counted at the level of what was chosen. A box set is one choice, not its contents.

Came back

Overlap is the share of everything a guest has ever chosen that they chose on every visit. Counted at item level, so returning for the same box set counts as repeating one choice. The baseline block is what two unrelated visits share, which is the only thing that makes an overlap of zero readable: without it a reader cannot tell whether these ten people are unusual or whether nobody repeats anything.