Analysis No. 01 · Flagship

Was it the food, or the day?

I went looking for a food. I found a pattern in my days instead. Of everything I logged for nearly six months, the things that best predicted a bad reflux day weren't on my plate. They were how much I moved and how stressed I'd been.

By the author·Jul 2026·6 min read·179 days of data

01 — The question

Elimination diets never quite worked for me. I'd cut something, feel better for a week, then relapse for no reason I could name. So I stopped guessing and started measuring, including things that aren't food at all. The question I most wanted answered: were my symptoms really driven by what I ate, or by how I'd spent the day?

02 — The data

For nearly six months I logged every meal (ingredients, portion, time), every symptom (severity, type, duration), and daily context: sleep, stress, and how sedentary I was. Symptoms were logged when they happened, not on a fixed schedule, so a day with nothing recorded is an assumed clear day, not a confirmed one. Imperfect, but consistent. Consistency is the whole trick.

Fig. 1

Weekly share of days with a symptom, 26 weeks.

03 — The food came up empty

I sorted every ingredient into 18 reflux-relevant categories (tomato, spicy, fatty, fried, chocolate, and more), then tested each with Fisher's exact test, correcting for multiple comparisons across all 18. Tomato and oats had the lowest raw p-values, but once I corrected for testing 18 categories at once, nothing survived. The lowest adjusted p-value was 0.20.

Category	Relative risk	Meals	p-value
Tomato	1.72×	54	0.02
Fatty	0.94×	376	0.40
Chocolate	0.63×	59	0.26
Fried	0.74×	42	0.52
Oats	0.57×	119	0.02

Relative risk of a symptom within 4 hours of eating, vs. baseline.

Unexpected result

Fatty food, my largest category at 376 meals, sat just below baseline (0.94×). Tomato scored highest at 1.72×, but on 54 meals and without surviving correction, it's not something I would bet on. The triggers I'd spent years avoiding weren't showing up.

04 — The day did the predicting

When I grouped days by how sedentary and how stressed I'd been, a clean gradient appeared, running from 25% of my active, calm days having a symptom to 100% of my stressed, sedentary ones. A four-fold gap, and the clearest thing in the whole dataset.

Fig. 2

Share of days with a symptom, by type of day.

The finding

A logistic regression agreed it wasn't noise. Stress was the clearest predictor of a symptom day (OR 1.96, p = 0.030), with sedentary posture just behind (OR 1.64, p = 0.051). More sleep was protective (OR 0.67, p = 0.047). Workouts and eating late showed no independent effect.

05 — What I make of it

I stopped agonising over ingredients and started watching the shape of my day: a short walk after lunch instead of straight back to the desk, protecting my sleep, and treating a stressful workday as a genuine risk factor. That doesn't mean food never matters. It means that for me, with my fairly repetitive diet, the day was the louder signal.

Important caveat

This is n = 1. My diet barely varies, which may be exactly why food stayed quiet. Someone eating more variably might find the opposite. Stress and sedentary posture also both track "a bad desk day"; the model separates them statistically, but not perfectly at this sample size.

Key takeaway

The strongest lever I found was a kind of day to avoid, not a food to cut out.

Track your own data for a few weeks and I'll run this same analysis on it.

Start tracking yourself