# Randomisation > Assigning treatments or conditions by a known chance process so the assignment design supports comparisons. - HTML version: https://robbiepalmer.me/ideas/randomization - Source: https://www.nber.org/papers/t0333.pdf Randomisation assigns treatments or experimental conditions using a known chance process. The researcher specifies the possible assignments and their probabilities before observing outcomes. That assignment mechanism becomes part of the evidence used to compare the groups. In a simple two-group experiment, each unit may have the same known probability of receiving the treatment. Other designs randomise within blocks, assign whole clusters, vary the probabilities, or restrict the possible allocations to improve balance. The analysis must reflect the procedure that generated the assignment. ## The assignment mechanism supports the comparison Randomisation breaks deliberate and systematic links between treatment assignment and the units' pre-treatment characteristics. It does not guarantee identical groups in the realised experiment. Chance imbalances can still occur. Instead, the known assignment probabilities support unbiased estimators under the design and allow randomisation tests to compare the observed result with results under other assignments that could have occurred. [Duflo, Glennerster, and Kremer](https://www.nber.org/papers/t0333.pdf) describe practical choices about the unit of randomisation, stratification, sample size, data collection, non-compliance, and spillovers. [Kennes and colleagues](https://pmc.ncbi.nlm.nih.gov/articles/PMC8366748/) show that restricted procedures trade some randomness for balance and that the planned analysis must match that choice. Allocation concealment is part of the mechanism. If someone can predict or change the next assignment, the nominal probabilities no longer describe the process that produced the groups. Blinding addresses a different problem. It limits changes in behaviour, treatment, or measurement after assignment. ## Random allocation is not random sampling Random allocation chooses a condition for units already in a study. Random sampling chooses units from a population. Allocation helps estimate a causal effect for the study units. Sampling helps generalise an estimate to the population represented by the sampling frame. A trial can randomise treatment among a convenience sample. The treatment comparison may have strong internal validity while saying less about people who were not recruited. A survey can draw a probability sample without assigning any treatment. The word "random" does not make these two operations interchangeable. [Study design](/ideas/study-design) decides which operation the question needs. [Sampling bias](/ideas/sampling-bias) concerns systematic distortion in who enters or remains in the observed sample, not imbalance created by treatment assignment. ## Statistical randomisation is an operational claim The originating essay, ["Are All Experiments Actually Quasi-Experiments?"](/blog/2022-03-28-quasi-experiments), asks whether objective randomness exists in nature. That is a philosophical question about indeterminism. Statistical randomisation makes a narrower claim. It requires a specified assignment procedure with known probabilities that people cannot manipulate using information about the units or their potential outcomes. A reproducible pseudorandom generator can run such a procedure when researchers choose the seed by an unpredictable process, conceal it until assignment, and keep outcome information out of the algorithm. The validity of the design does not require proof that the universe is metaphysically indeterministic. Randomisation also has limits. It cannot correct biased recruitment, outcome measurement, non-adherence, interference between units, selective attrition, or selective reporting. Those problems need their own design and analysis choices. ## Questions it prompts * What are the possible assignments, and what probability does each one have? * Can anyone predict or alter an assignment before a unit enters the experiment? * Does the analysis use the assignment mechanism that the experiment actually followed? * Is blocking, clustering, or restricted allocation needed for the scientific question? * Which claims concern the assigned units, and which require a separate sampling argument? ## Sources * Duflo, Esther, Rachel Glennerster, and Michael Kremer. ["Using Randomization in Development Economics Research: A Toolkit"](https://www.nber.org/papers/t0333.pdf). NBER Technical Working Paper 333, 2006. * Kennes, Leonhard N., et al. ["A roadmap to using randomization in clinical trials"](https://pmc.ncbi.nlm.nih.gov/articles/PMC8366748/). *BMC Medical Research Methodology*, 2021\. * Hernán, Miguel A., and James M. Robins. [*Causal Inference: What If*](https://miguelhernan.org/whatifbook). Chapman & Hall/CRC, 2020, updated online edition. * Palmer, Robbie. ["Are All Experiments Actually Quasi-Experiments?"](/blog/2022-03-28-quasi-experiments). 2022. Originating essay. ## Related ideas - [Sampling Bias](https://robbiepalmer.me/ideas/sampling-bias.md): Systematic error caused when the process that produces a sample distorts an estimate for its target population. - [Study Design](https://robbiepalmer.me/ideas/study-design.md): The plan that connects a research question to the population, data, comparisons, and analysis needed to answer it. ## Where it appears - Blog post: [Are All Experiments Actually Quasi-Experiments?](https://robbiepalmer.me/blog/2022-03-28-quasi-experiments.md) - Blog post: [Effective and Ethical Data Science Study Design](https://robbiepalmer.me/blog/2023-02-24-effective-ethical-data-science-study-design.md) --- Markdown index of this site: https://robbiepalmer.me/llms.txt