How a Test Decides ยท five essays
At a Glance
A test starts from a guess it can count, and asks how rare the evidence would be if the guess were true. What it can answer is narrower than what we ask.
- 1Significant but WrongWhat does a tiny number assume before it convicts?
- 2What It Would Take, In CupsWhat would it take to believe a claim?
- 3One Napkin, Two FiresHow many voices are enough to overturn a story?
- 4The Seam Does Not Show AnymoreWhere do Fisher's test and Neyman and Pearson's meet?
- 5A Weed Called CrazyWhat can a test answer?
| Idea | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|
| The starting guess | |||||
| Shared, not separate | |||||
| The line comes first | |||||
| Not proven is not disproven | |||||
| Narrower than the question |
The starting guess
Every test begins from a default it can count: nothing is going on.
- Part 1Elffers had to assume Lucia was like any other nurse: innocent until proven guilty.
- Part 2The starting assumption is that she is guessing: the one story that can be counted.
- Part 3Miles's story stands until evidence overturns it; the burden fell on Andi.
- Part 4Both camps start from the same guess: the driver is keeping pace with traffic.
- Part 5Both models began from one guess: the three treatments stop bacteria equally well.
Shared, not separate
Chances multiply only when nothing is shared.
- Part 1A ward is a system, not a deck: its shifts share patients, seasons and handovers.
- Part 2The cups are shuffled, so a cup's place is independent of how it was poured.
- Part 3Four friends paid by one man were one decision spoken four times.
- Part 5Three wells on one plate share one pour, so the plates became blocks.
The line comes first
The threshold is set before the evidence comes in.
- Part 1A threshold guards the innocent only when the evidence was gathered blind.
- Part 2Fisher's usual line was 5 percent; eight right, at 1 in 70, clears it.
- Part 3Before any friend speaks, the court decides how many would be enough.
- Part 4Neyman and Pearson fix the cutoff beforehand; every driver over it gets the same ticket.
- Part 5Move the line from 1 in 100 to 1 in 20, and the 1.8 millimeter gap would count.
Not proven is not disproven
A test that fails to reject has not shown the guess is true.
- Part 2Six of eight would have left the test silent, and a silent test proves nothing.
- Part 3Failing to reject only means the evidence was not strong enough yet.
- Part 4The gun can show a driver was not keeping pace, and never prove one was.
- Part 5Not proven different is not proven alike, the way not guilty is not innocent.
Narrower than the question
What a test answers is smaller than what people ask of it.
- Part 11 in 342 million was the chance of her shifts if the ward were a deck, not of her innocence.
- Part 2A p-value is not the chance that the guessing story is true.
- Part 4"Significantly" says a line was crossed, with no size attached.
- Part 5A ring in a dish is not a healed wound; only a study of people can take that.