The Story You Tell Your Data
You started magnesium three weeks ago. Your sleep scores have been higher than they've been in months. You feel more rested. You tell a friend: "Magnesium fixed my sleep."
It might be right. But it might also be that:
- Daylight hours in your region increased by 45 minutes over the same period
- A high-stress work deadline passed at the end of week one
- You started going to bed 20 minutes earlier without consciously noticing
- You cut back on alcohol during the same stretch
- All of the above
This is the correlation-causation problem in personal health data, and it trips up even careful trackers. Understanding it does not require a statistics degree — it requires a few practical habits and a healthy skepticism toward your own conclusions.
Related: Want to put this into practice? Try our Experiment Builder to get started, and check out 30-Day Sleep Experiment: Optimize Bedtime With Data for more context.
Correlation Is Not Nothing
First, the important clarification: correlation is not useless. When your data shows that a supplement coincides with measurable improvement across multiple metrics, that is genuinely informative. It means the compound is a candidate for what helped. That is exactly what a self-experiment is designed to tell you.
The problem comes when you stop at correlation and declare it causation. "My data suggests a relationship between starting magnesium and better sleep" is a very different claim from "magnesium improved my sleep." The first is honest. The second requires more evidence.
Common Correlation Traps
Seasonal confounding. Sleep tends to improve naturally as days get longer and temperatures warm up in spring. HRV often tracks with training load, which frequently shifts with seasons. If your experiment runs from January through March, any improvement you see is partly competing with a known seasonal trend.
The "worst week ever" effect. People often start a new supplement during a particularly bad stretch — they are exhausted, HRV is low, sleep is terrible. This is a well-documented phenomenon in statistics called regression to the mean. If you started at your personal low point, you may have improved regardless of the supplement simply because your body was recovering.
Regression to the mean in practice: If you started a supplement during your worst week of sleep in three months, your sleep was statistically likely to improve in the following weeks no matter what you did. Always establish a baseline during a typical period — not your worst or best week.
Multiple simultaneous changes. You started magnesium and also bought blackout curtains and started a wind-down routine. All three changes happened in the same week. Your sleep improved. Which one did it?
Expectation effects. If you believe a supplement will help, you may report feeling better, sleep more consistently, and even unconsciously make other small changes that support the outcome you expect. This is the placebo mechanism, and it is real and powerful. Wearable data helps here — your Oura ring does not know you started magnesium.
The ABA Design: A Simple Test for Causation
The most practical way to move from "my data suggests a correlation" to "my data suggests a causal relationship" is the ABA protocol:
- A Phase (Baseline): Several weeks without the supplement, tracking your metrics.
- B Phase (Active): Add the supplement and track the same metrics.
- A Phase (Washout): Remove the supplement again and track whether your metrics return toward baseline.
If your metrics improve during B and return toward baseline when you remove the supplement in the second A phase, that is much stronger evidence of a real effect. The compound is behaving as if it is driving the outcome — not just coinciding with it.
ABA Protocol Template
A Phase 1 (Weeks 1–2): No supplement. Log daily metrics. B Phase (Weeks 3–6): Add supplement. Log same metrics. A Phase 2 (Weeks 7–8): Remove supplement. Log same metrics.
Ask: Do metrics improve in B? Do they return toward baseline in A Phase 2?
Confounds You Can Control For
You cannot eliminate every confounding variable. But you can minimize the most common ones:
Log confounds alongside your metrics. Add a daily field for alcohol consumption, exercise intensity, unusual stress events, travel, and sleep opportunity (time in bed). When you review your data, you can look for whether improvements correlate with the supplement — or with one of these other variables.
Use consistent measurement conditions. Weigh yourself at the same time, under the same conditions. Review wearable data at the same point in your morning. Subjective ratings taken at different times of day are hard to compare.
Run experiments during normal life. Avoid starting or ending a phase during a vacation, a high-stress work period, or a significant lifestyle change. The more stable your background conditions, the more signal your experiment produces.
When to Trust Your Data
Your data deserves more confidence when:
- The change appeared within a plausible timeframe (not immediately for slow-acting supplements, not weeks later for acute compounds)
- Multiple metrics moved in the same direction
- The improvement persisted across the full active phase rather than just the first few days
- An ABA design reproduced the effect
- Objective wearable data agrees with your subjective rating
Be the first to try Prova
We're building an app to track whether health data correlation actually works. Join the waitlist.
Holding Conclusions Lightly
The goal of self-experimentation is not certainty — it is better evidence than you had before. "My data suggests magnesium may be contributing to my improved sleep" is not a weak conclusion. It is an honest one. It gives you a reasonable basis for continuing the supplement while staying alert to other explanations.
The worst outcome is not finding out a supplement does not work for you. The worst outcome is confidently believing it works when it does not — and spending years, money, and health decisions on a correlation you never tested.