Skip to content

Randomization

Why It Matters

If you let subjects choose their condition, or assign based on any systematic rule, you introduce selection bias. People who choose treatment differ from those who don't in countless ways. Maybe healthier people seek treatment, or sicker ones do - either way, you can't distinguish treatment effects from pre-existing differences. Randomization breaks this: any baseline difference between groups is due to chance, quantifiable by statistics.

The Key Move

When you want to test a causal claim and have control over assignment, randomize. Use a random number generator, coin flip, or random sequence - not alternation, not judgment, not "quasi-random" rules. When you can't randomize (observational studies), explicitly acknowledge that causal inference is weaker and identify the likely confounds you can't rule out.

Classic Application

Agricultural experiments. Fisher invented randomization for crop trials: assign fertilizer treatments randomly to field plots. This ensures that soil quality, drainage, sunlight - variables you might not even think to measure - are balanced across conditions in expectation. Without randomization, any observed yield difference could be due to systematically assigning fertilizer to better plots.

Surprising Application

Personal experiments with productivity techniques. If you only try new techniques when you're already motivated, you'll overestimate their effects (selection bias). Better: randomize days - coin flip each morning determines if you use the new technique or your default approach. This controls for motivation fluctuations, external events, and other unmeasured confounds.

Failure Modes

Pseudo-randomization (alternating assignments, birth dates, convenience) that creates predictable patterns. Randomizing too small a sample where chance imbalances are likely. Randomizing but then selectively excluding subjects (defeats the purpose). Believing randomization controls for everything - it controls for confounds at baseline, not differential dropout or measurement bias during the experiment.

Go Deeper

Fisher, R.A. (1935), The Design of Experiments, Chapter 3; Imbens & Rubin (2015), Causal Inference for Statistics, Social, and Biomedical Sciences, Chapter 2.