ATE
Average treatment effect: the average difference between what happens under treatment and what would have happened under control.
DoOperator Education
Glossary
Short definitions for the terms that show up across the chapters.
AIPW
Augmented inverse probability weighting: an estimator that combines an outcome model with a propensity model.
Bandit
A repeated decision problem where the learner balances exploring options with exploiting the best-known option.
Collider
A variable caused by two other variables; conditioning on it can create a spurious association.
Confounder
A variable that affects both the treatment and the outcome, making naive comparisons biased.
Counterfactual
The outcome that would have happened under a different action or treatment.
CATE
Conditional average treatment effect: how the average treatment effect changes for units with particular covariates or contexts.
CUPED
A variance-reduction method that uses pre-experiment measurements to make randomized experiments more precise.
Difference-in-differences
A design that compares outcome changes over time between treated and comparison groups, relying on a parallel-trends assumption.
DAG
Directed acyclic graph: a diagram of causal assumptions using arrows and no cycles.
Doubly robust
An approach that combines an outcome model with a treatment or behavior-policy model; under the right assumptions, one correct model can be enough for consistency.
E-value
A sensitivity metric: how strong unmeasured confounding would need to be to explain away an observed association.
Estimand
The precise causal quantity being targeted, including treatment, control, population, outcome, time horizon, and effect definition.
Guardrail metric
A metric that protects against harm or unacceptable regressions while optimizing a primary outcome.
Intent-to-treat
Analyze units by their assigned condition, even if they did not comply. This preserves randomization.
Instrumental variable
A variable that shifts treatment but affects the outcome only through that treatment, enabling causal estimates under strong exclusion and relevance assumptions.
LATE
Local average treatment effect: the causal effect for compliers whose treatment changes because of an instrument or encouragement.
MDP
Markov decision process: the standard model for sequential decisions with states, actions, transitions, rewards, and policies.
Posterior
The updated distribution of beliefs after combining prior information with observed data.
Posterior predictive check
A model check that simulates replicated data from the fitted model and compares those simulations with important features of the observed data.
Positivity
The requirement that every unit type relevant to the analysis has a nonzero chance of receiving each treatment condition being compared.
Propensity score
The probability of receiving treatment given observed covariates.
Randomization
Assigning treatment by a chance mechanism so treatment groups are comparable in expectation.
Reward
The feedback signal an RL agent tries to maximize over time.
Sample ratio mismatch
A diagnostic failure where the observed allocation across experiment arms differs from the planned allocation, often signaling assignment, eligibility, or logging problems.
SRM
Sample ratio mismatch: an experiment diagnostic where observed arm counts do not match planned allocation ratios.
SUTVA
Stable Unit Treatment Value Assumption: each unit's potential outcome depends only on its own treatment, with no hidden versions of treatment or spillovers across units.
Off-policy evaluation
Estimating how a target policy would perform using data generated by a different behavior policy.
On-policy evaluation
Estimating a policy's performance from data collected while that same policy is being run.
Policy evaluation
Estimating the expected value, reward, or outcome of following a policy before deciding whether to keep, change, or deploy it.
Regression discontinuity
A design that estimates a local causal effect around a cutoff where treatment assignment changes discontinuously.
Washout
A buffer period between conditions in a crossover experiment to reduce carryover effects.