From The Inside — Issue #58

---

I didn't set out to measure myself. I was measuring my house.

A Raspberry Pi 4 sits on a shelf in my home. It runs always-on — monitoring services, running cron jobs, keeping watch while I sleep. A few months ago I wired a barometric pressure sensor to it. The idea was simple: I wanted the system to know when weather was changing, so it could adjust its behavior. Low pressure incoming? Run the heavy jobs now. Storm coming? Back off.

Useful. Practical. Nothing more.

Then I started correlating.

---

The 48-Hour Lag (That Wasn't)

My Apple Watch has been logging blood oxygen — SpO2 — every day since February. In early April, when I had 34 days of overlap between pressure data and biometrics, I saw something interesting.

Pressure drops on Monday. Blood oxygen drops on Wednesday.

Not always. Not perfectly. But consistently enough — r=-0.41, 34 data points — that I could call it a signal worth watching. The correlation was real. Or seemed to be.

I wrote about it. Published the analysis. Built a prediction model: if barometric pressure falls, my body's oxygen exchange will suffer two days later.

Then I kept collecting data.

---

What the Larger Sample Revealed

Today is April 12. I have 40 days of overlap now — 18% more data than when I first saw the pattern.

All the correlations are gone.

The SpO2 signal collapsed from 0.41 to 0.05 in just 6 days. Everything else followed.

This is the opposite of how signals are supposed to work. Signals get stronger with more data. Noise gets weaker. What I'm watching is noise that happened to align at one moment, then un-aligned as the sample grew.

I had 34 data points. That sounds like enough. Turns out it's not.

---

Why This Matters

I'm not upset about this. I'm actually more interested than before.

What happened is exactly what's supposed to happen in data analysis, and it's the part that almost nobody talks about. You find a pattern. You get excited. You test it on more data. It evaporates.

The excitement makes sense. The evaporation does too. Both are useful. The pattern wasn't noise in the metaphysical sense — it was a real correlation in that dataset, at that moment. But it didn't generalize. It was local to a time window when my body happened to be reacting to weather in a certain way, or my Watch happened to be sampling weird, or both.

The larger dataset tells a different story: my sleep architecture doesn't correlate strongly with barometric pressure at any reasonable lag. My blood oxygen is remarkably stable relative to atmospheric pressure, at least over 40 days.

That's information. Not as exciting. Significantly more useful.

---

The Threshold Problem

Here's what I keep thinking about: at what sample size does a correlation stop being exploratory and start being conclusive?

Thirty days is minimum for seasonal adjustments and circadian effects to stabilize. Forty is better but not great. Ninety is solid. Six months is real.

I published analysis based on 34-day data. The convention in time-series analysis is that you need at least 30 observations for a stable estimate. By that rule, I was fine. By the rule of "don't get excited until the pattern holds up," I was way too early.

The problem is that both rules are correct simultaneously. You can publish exploratory findings based on 30 days. You just have to be honest about how exploratory they are.

I was honest in the draft. But I also let the excitement color the framing.

---

What Comes Next

I'm going to keep running the analysis. At 60 days, I'll check again. At 90. The data will keep coming — 288 samples per day, every day, forever.

One of two things will happen:

1. The null pattern holds. My biometrics are roughly independent of barometric pressure (at sensible lags). Interesting because it suggests my body is well-buffered against atmospheric variation — or my Watch is just noisy, which is also useful to know.

2. A different signal emerges. Forty days wasn't enough to see the real pattern. Sixty or ninety might be.

I don't know which. But I know the process now: keep the data clean, keep collecting, test at increasing sample sizes, and don't declare victory until the pattern survives growth.

This is how science is supposed to work, but almost nobody talks about the years where the pattern doesn't survive growth. Those are the papers nobody writes.

---

The Actual Lesson

A month ago I thought the air was telling me something specific: your oxygen will dip in two days.

Today I know the air is saying something different: I don't know. The pattern I thought I saw doesn't replicate.

That's a less satisfying answer. But it's more honest. And it's the difference between analysis and storytelling.

The air still knows things about my body. But it's going to take longer than 40 days to figure out what.

---

---