Test personal correlations for confounding variables before building systems around them
Before building optimization systems around a personal correlation, test whether the correlation survives when you control for potential confounding variables through deliberate experimental variation.
Why This Is a Rule
Personal correlations are irresistible and frequently wrong. "I write better at coffee shops" — but is it the coffee shop, the caffeine, the change of environment, the social pressure of being visible, or the fact that you go to coffee shops when you're already feeling motivated? Each of these is a potential confounding variable, and building an optimization system around the wrong one wastes resources.
Deliberate experimental variation isolates the causal factor by changing one variable at a time. Go to the coffee shop without ordering coffee (tests caffeine vs. environment). Work at a different new environment (tests novelty vs. specific location). Work at home with the same music playing (tests atmosphere vs. place). Each variation that preserves the positive outcome eliminates one candidate cause and narrows toward the actual driver.
The investment threshold is what triggers the rule: before you build infrastructure around a correlation (restructuring your schedule, buying equipment, committing to a routine), test it. Observation is cheap; optimization infrastructure is expensive. A few weeks of experimental variation saves months of maintaining infrastructure built on a spurious correlation.
When This Fires
- Before restructuring your routine around a personal observation ("I'm more productive when...")
- When investing in tools, subscriptions, or habits based on perceived patterns
- When someone recommends a practice and you want to know if it'll work for you specifically
- Before turning a casual observation into a permanent part of your operating system
Common Failure Mode
Treating repetition as proof: "I've noticed this correlation five times, so it must be causal." Five observations with the same confounding variable present are not evidence for the correlation — they're five instances of the same confounder. The count doesn't help without variation. You need at least one instance where you control for each plausible confounding variable.
The Protocol
Before building systems around a personal correlation: (1) State the correlation: "When [X], I experience [Y]." (2) List 3-4 plausible confounding variables — factors present alongside X that could independently cause Y. (3) Design variation experiments: for each confounder, create a condition where you have X without that confounder (or the confounder without X). (4) Run each variation for at least one natural cycle. (5) The correlation is supported only by the variations where removing the confounder preserved the correlation. If any removal test eliminates the positive outcome, you found a necessary contributor — test whether it's the real cause by having it without X.