Causal Inference for Everyday Life
Did this change actually cause the result? The logic behind causal inference isn't just for statisticians — it's the most useful thinking tool most people never learned.
The Most Important Question You're Probably Not Asking
You start taking magnesium before bed and sleep better for a week. You add morning sunlight and your mood improves. You cut alcohol and your HRV goes up.
The natural interpretation: these changes worked. And maybe they did. But "things happened after I made changes" is not quite the same as "my changes caused the improvements."
Causal inference is the branch of statistics concerned with this question. It sounds technical. The underlying logic isn't. And once you understand the basic ideas, you'll never look at a self-experiment — or a headline — quite the same way.
Why Correlation Isn't Worthless (But Isn't Enough)
The phrase "correlation is not causation" is repeated so often it has become almost meaningless. It's true, but it doesn't quite explain the problem.
A correlation tells you that two things move together. That's useful information. If HRV and sleep quality correlate in your data, that's worth knowing — it might tell you something about your nervous system, or suggest a variable worth investigating.
The problem is that correlations have three possible explanations:
- A causes B
- B causes A
- Something else causes both A and B
When you observe a pattern in your own data, all three explanations are always on the table. The correlation tells you that a pattern exists. It doesn't tell you which explanation is correct.
Causal inference is the set of methods for distinguishing between these explanations and arriving at one defensible conclusion.
The Fundamental Problem: You Can't Run Both Conditions Simultaneously
Here is the core challenge of causal inference, stated plainly: to know the causal effect of a treatment, you would need to compare the same person under two conditions at the same time. The person who took magnesium and the same person who didn't, on the same night, in the same circumstances.
That's impossible. The best you can do is approximate it.
The most powerful approximation is randomization. By randomly assigning which nights (or days, or weeks) you apply an intervention, you make the two groups — treated and untreated — similar in expectation. The things you can't control (mood, stress, ambient temperature, Tuesday versus Saturday) are distributed roughly equally across conditions, so they can't systematically bias the result.
This is why randomized experiments produce cleaner causal conclusions than observational studies. It's not because the researchers were more careful. It's because randomization — not effort or measurement quality — is what breaks the link between potential confounders and the treatment condition.
Confounding: The Main Reason Self-Experiments Fail
Confounding is when a third variable causes both the treatment and the outcome, creating a spurious association.
A classic example: people who carry lighters have higher lung cancer rates. Do lighters cause cancer? Of course not. Smoking causes both lighter-carrying and cancer. Smoking is the confounder.
In personal experiments, confounding is everywhere:
You start a morning walk routine in spring. Your mood improves. Was it the walking? The sunlight? The fact that you're outside before work? The seasonal change in daylight? The reduced phone time before starting your day? Any of these could be driving the result, and your "walking routine" is entangled with all of them.
You give up alcohol for a month. You sleep better and feel more energetic. Was it the sleep improvement from not drinking? The calorie reduction? The changed social patterns? The increased discipline and self-efficacy? The fact that you started other good habits at the same time?
Confounding doesn't mean you can't learn anything. It means you should be appropriately skeptical about causal claims from uncontrolled observations — including the ones you make about yourself.
The Three Tools That Help
You don't need a PhD to reason clearly about causation. Three ideas do most of the work:
1. Randomization When you randomly assign conditions across days or weeks, confounders are distributed evenly between conditions. A crossover experiment where you flip between A and B on a randomized schedule controls for everything that varies weekly, seasonally, or with your mood.
2. Isolation Change one thing at a time. This is the experimental design principle that prevents confounding by construction. If you're testing magnesium, don't also change your sleep schedule, your alcohol intake, or your exercise timing during the test. Isolation is your main defense against the "I don't know which change did it" problem.
3. Comparison groups Always ask: compared to what? If you only ever try the intervention, you have no baseline to compare against. The control condition — doing nothing different, or doing the alternative — is what makes a result interpretable. Without a comparison, "my sleep improved" just means things changed.
Regression to the Mean: The Subtler Problem
There is one more phenomenon worth understanding because it tricks even careful people: regression to the mean.
When something is bad enough that you decide to intervene, it is often at an extreme. Extreme values tend to be followed by less extreme ones, regardless of what you do. This is regression to the mean — a statistical property, not a real effect of your intervention.
You had three terrible nights of sleep. You started a new supplement. You slept better. Did the supplement work? Maybe. Or maybe you were at the worst end of natural variation and were going to improve regardless.
The fix is, again, randomization and enough data. If you only run the experiment when things are bad and measure "improvement," you will confuse regression to the mean with treatment effects. If you run the experiment across a representative range of conditions — good nights and bad, stressful weeks and calm ones — the regression effect averages out.
The Everyday Version
None of this requires a statistics course. The practical upshot is a set of habits for your own self-improvement reasoning:
When you notice an improvement, ask: what else changed? Could the improvement have happened without the intervention? Would it be happening anyway?
When you're planning to test something, ask: can I change exactly one variable? Can I randomize across days or weeks? What is the comparison condition?
When you read a headline, ask: how was the comparison made? Was it randomized? What could be confounding the result?
These questions don't guarantee causal certainty — that's not attainable outside of very large, carefully designed trials. But they raise your baseline level of skepticism to something appropriate for the actual quality of the evidence, which is usually much lower than the confidence with which it's presented.
Causal thinking isn't about being contrarian or dismissive of evidence. It's about being honest about what a given result can and cannot tell you — so you can make better decisions about what to test next.
Design a controlled experiment and start building your own causal evidence.