Better together, or the same years?
A director can look better with a partner because they worked together in the director's best years. This page lists the pairs with the biggest gap, checks each against the director's other films from the same years, and gives a verdict only where no single film changes it.
The finding
Of 184 director–collaborator pairs with enough films to compare, 30 show the director hitting at least 20 points more often with the partner than without. Checked against the director's other films from the same years, one holds up, one closes, seven are too close to call, and 21 cannot be tested.
How the pairs were chosen
A pair is a director and anyone credited on their films as an actor, writer, cinematographer, editor or composer. Only the director's films are compared: those with the partner credited against those without. To be listed, the director needs at least 8 films with the partner and 8 without, and to hit at least 20 points more often with. The films are the director's counted features, less documentaries, music films and anthologies, so the director's figures here can be lower than on their page. Years are release years.
The same-years check
Directors have good decades and bad ones, and a long partnership usually covers a good one. So the check compares the films with the partner against the director's films without them released in the same window: from the second shared film's year to the second-to-last's, so neither end of the partnership sets its own control. The window is shown on each row. With fewer than 5 films without the partner in it, the pair can't be tested.
Too close to call
A verdict is reported only if it survives every one of these changes, taken one at a time: dropping any single film from either side, and moving any film within 1 point of the 80/100 bar across it. If any change flips the verdict, the pair is too close to call. So is a pair where dropping one film would leave fewer than 8 films with the partner or fewer than 5 without them in the window, since there would be nothing left to compare. A verdict that holds describes what happened, not proof that the partnership caused it.
Related
How these numbers work: Methodology