The Data Executes the Hypotheses, and Apologizes to No One

CONTENTS

A study of 137 mobile-phone consumers executed, one by one, three of the industry’s favorite instincts. Instinct one: longer battery life, higher satisfaction. The data says the 18-hour battery option had a statistically significant negative effect on satisfaction — beta of minus 0.6154. Instinct two: the more AI, the better. The data says users without ChatGPT were more satisfied, coefficient 1.00677. Instinct three: the stronger the brand, the higher the satisfaction. The data says Apple scored lowest in brand perception, 1.88 against OPPO’s 2.74, yet the correlation between brand choice and satisfaction was 0.064 with p = 0.457 — effectively none. In the interviews, 10 of 12 participants also called recent iPhones lacking in innovation. Three instincts, all reversed.

Four Hypotheses Dying Together Is Not Noise

Some say 137 respondents is a small sample, so the counter-intuitive results are noise and deserve no attention. Doubting a single result is fair. But more than one hypothesis was rejected here: battery life negative, no-AI positive, brand irrelevant — all leaning against instinct in the same direction. A die landing on six once can be luck. Hypotheses falling to the same side one after another is not luck; the map drawn by instinct was wrong to begin with. When the errors cover a whole map, the weather is not to blame.

Rejection Is the Purpose, Not the Accident

Others say data that fails to support a hypothesis means the study was botched, and should be rerun until the results come out. The mark of a botched study is never a rejected hypothesis; it is tampering with the data after rejection. Hypotheses dying first is the purpose of research, not an accident — the information that research pays for hides precisely in the rejected column. The more honestly the death is recorded, the more information it carries; retouching the corpse burns the most valuable data. Good research does not exist to prove the researcher clever. It exists to let the hypotheses die first.

Absurdity Is a Lead, Not Trash

Still others say the findings are obviously absurd — an 18-hour battery lowering satisfaction, users happier without AI — and should simply be deleted. The words obviously absurd are instinct speaking. The questionnaire may well be flawed: bands of 4 hours and 1,000 yuan are crude, and that deserves fixing. What deserves fixing is the design, not the erasure of the result. When data says something absurd, the first response is to examine how the question was asked, not to pretend nothing was said. Absurdity is a lead; those who discard the lead are the ones leaving trash behind.

When Prior Collides with Posterior, the Prior Loses

A final voice: brand perception lowest at 1.88 yet unrelated to satisfaction — too contrary to common sense for anyone to act on. Common sense is prior belief; data is posterior evidence. When the two collide, the prior is what should lose — otherwise why do research at all; transcribing common sense would be enough. The field concedes the point itself: the finding challenges, as the study puts it, conventional marketing assumptions about brand perception’s direct influence on satisfaction. Common sense being challenged is no disgrace; pretending not to hear is.

The Deliverable Is a List of Causes of Death

Boundaries first: the study’s crude bands and sampling deserve a separate piece, and are not litigated here; nor does this extend into a judgment on Apple’s stock — the 70-dollar discount that failed to stop a further 5 percent share drop in early 2025 is background, not the argument, and the subject here is method. When the study is done, the deliverable is not a stack of conclusions that stand; it is a list of causes of death — which hypothesis died at the hands of which number, and how. The first job of a hypothesis is to be executed. No apology required.

Fengyu WANG
Fengyu WANG

Markets, investing, engineering — one person, one underlying logic.