The cutting room floor
the lab.
Work that has not earned a place on the main page.
Comparison
Travel days against ordinary days on the measures that answer: 1.9x more films, ratings unchanged.
Films per viewing day
1.9x more. 21 watches over 10 flight days against 783 over 699 others.
Share of days viewing more than one film
6.3x likelier. 7 of 10 against 78 of 699.
Mean rating with 95% confidence intervals
Unchanged. The two means differ by −1.6 points over 21 watches, and the 95% CI on that gap runs −7.1 to +3.8, so it includes no change at all. Both medians are 70.
Each whisker is a 95% CI for that side's mean, not the spread of the ratings, so the dashed median tick can sit outside it. The axis does not start at zero, which is why these are intervals and not bars: a clipped bar exaggerates a gap this size and a zeroed one hides it. The interval shows how little 21 watches pin down.
Against it: Three panels at equal weight, and the null result takes the most ink: the rating gap needs two whiskers and a paragraph explaining what a whisker is, while the 6.3x above it gets a bar pair and one line. Ink here follows how hard a measure is to draw, not how much it says.
Callout
No axes. The trips named, the films listed with the figures inline.
I watch roughly twice as many films on a day I spend flying. 2.10 per flight day against 1.12 on an ordinary viewing day ( 1.9x) and 7 of the 10 flight days held more than one film against 11% of the rest.
What does not change is what I think of them. The flight mean is 71.9 against 73.6, a gap of −1.6 points on 21 watches, smaller than the standard error of 2.8 that comes with it. Both medians are 70. No finding here about flying making a film worse, only about how many fit in a day.
Against it: Prose is not a visual. The ordinary-day baseline can only be a number in a sentence, so a reader asking whether 2.10 is a lot has nothing but this panel's word for it.
Mean rating per release year
Mean of ratings per release year.
Against it: Fifteen years hold one film, so their bar is that film. Mean punishes years with more ratings in the middle.
The grad school years
I was in graduate school for my Economics MS degree from August 2023 to May 2025. Two rolling lines over the whole log, rating and volume, on one time axis with the span shaded on both. Neither line does anything meaningful at the shading, which is the finding.
Mean rating
Trailing 10 watches, plotted at the watch the window closes on.
Films per month
Trailing 12 months, a monthly rate. The table below is per week.
Two windows, because the two measures need different ones. Films per month is a rate over time, so it takes a 12 month window; the span is 22 months, and a twenty-four month window would be wider than the thing it has to resolve. The rating takes a 10 watch window instead: a fixed stretch of days holds four watches one month and twelve the next, so a time-windowed rating would swing on how much I watched rather than on how I rated it. Inside the span one window covers 0.3 to 2.2 months, well short of the span. That is light smoothing, not a trend.
The trend walks straight through it.
The rating line enters the span at 85.0 and leaves it at 81.0, wandering in between. Across the span it travels 4 points net, around the 30th percentile of the 69 stretches of the same length. That is a middling stretch, not a still one: no stillness finding here. The climb that got it to 85 finishes before the shading starts, running up from 67.7 in 2020.
The pace line is not flat inside the span. It falls to 5.8 films a month by Apr 2024, recovers to 9.2 by Apr 2025, and finishes higher than it began. Neither edge breaks. The fall from the early years ends before the span opens, and the steepest decline in the log comes after it closes: 1.0 watches a week in the twelve months after against 1.8 inside.
| Watches | Per week | Mean rating | |
|---|---|---|---|
| the early years | 480 | 2.6 | 71.0 |
| 12 months before | 85 | 1.6 | 76.6 |
| in school | 169 | 1.8 | 77.4 |
| 12 months after | 53 | 1.0 | 76.8 |
What this cannot do is show that nothing happened.
A chart cannot establish an absence. Against the year on each side, the rating moves 0.8 points, inside what 169 watches can resolve, and the viewing rate is indistinguishable. The percentile above is a rank among overlapping stretches rather than a test. It locates the span, which is not evidence about it.
Set the span against everything outside it and the rating looks 4.9 points higher. That is the wrong comparison: 480 of the 635 watches outside the span are early years, at 71.0, so it mostly measures the distance from those. Against the neighbors in the table above, the gap goes away.
This is when I was in school and this is where the lines went. It is not evidence that school did it.
Why it is here and not on the main page: the section spends more words refusing the causal reading than stating the finding. That is the right ratio and a bad fit for a page a stranger skims, because a null result needs its method shown before it means anything.