Dose-Response Testing¶

Why It Matters¶

Observing an effect at one treatment level doesn't prove causation - it could be a threshold effect, confound, or coincidence. Dose-response provides additional evidence: if increasing X increases Y proportionally, alternative explanations become implausible. It also reveals the shape of the relationship (linear, logarithmic, threshold, U-shaped), which is crucial for practical application.

The Key Move¶

Don't just test treatment versus control - test multiple levels of treatment. If testing exercise on mood, try 0, 20, 40, 60 minutes. Plot outcome against dose. Look for monotonic relationships (more X → more Y) or systematic patterns (diminishing returns, thresholds). Absence of dose-response undermines causal claims; presence strengthens them substantially.

Classic Application¶

Toxicology and pharmacology. "The dose makes the poison" - substances safe at low doses become harmful at high doses. Dose-response curves establish safe exposure limits and therapeutic windows. A substance that shows no dose-response (harm is unrelated to dose) is unlikely to be causally responsible for the observed effect.

Surprising Application¶

Debugging software or processes. When you change a parameter and see improvement, test multiple values rather than just "old" versus "new." If increasing the timeout from 1s to 2s helps, try 3s, 5s, 10s. If benefit plateaus at 3s, you've found the effective threshold. If there's no relationship between timeout and success rate, the parameter isn't the cause (something else changed coincidentally).

Failure Modes¶

Assuming linear dose-response when the relationship is non-linear (thresholds, U-shapes, plateaus). Testing doses too close together to detect differences. Confusing correlation of dose with outcome for dose-response (if people who self-select higher doses differ in other ways). Ignoring that multiple causal factors can make dose-response appear absent.

Go Deeper¶

Hill, A.B. (1965), "The environment and disease: Association or causation?" Proceedings of the Royal Society of Medicine 58(5); Howick, Glasziou, & Aronson (2009), "The evolution of evidence hierarchies," Journal of the Royal Society of Medicine 102(12).