Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Causal Inference
Statistical process aimed at establishing cause-effect relationships between variables by distinguishing correlations from true causal links, essential for understanding the real impact of interventions.
Directed Acyclic Graph (DAG)
Graphical representation of causal relationships between variables where nodes symbolize variables and directed arrows indicate direct influences, with no cycles possible to maintain causal consistency.
Counterfactual
Hypothetical scenario describing what would have happened if a cause had been different, allowing estimation of causal effect by comparing the actual situation to this imaginary alternative.
Confounding Variable
Variable that simultaneously influences both the cause and the effect being studied, creating a spurious association that may be mistakenly interpreted as a direct causal relationship.
Propensity Score
Conditional probability of receiving a given treatment based on observed covariates, used to balance groups and reduce bias in non-randomized observational studies.
Do-Calculus
Set of formal rules developed by Judea Pearl allowing transformation of causal expressions into probabilistic expressions calculable from observational data.
Causal Intervention
Deliberate manipulation of a variable in a system to observe its effect on other variables, distinct from passive observation as it modifies the underlying causal structure.
Average Causal Effect (ACE)
Aggregate measure of the average impact of an intervention across the entire population, calculated as the difference between potential outcomes with and without treatment.
Structural Causal Model (SCM)
Mathematical framework combining structural equations and causal graphs to explicitly represent data-generating mechanisms and causal relationships.
Frontdoor Criterion
Causal identifiability condition allowing estimation of treatment effect on an outcome via observable mediators, even in the presence of unmeasured confounding.
Backdoor Criterion
Set of conditions on variables to adjust for in order to block all non-causal paths between treatment and outcome, enabling correct identification of causal effect.
Causal Mediation
Quantitative analysis of mechanisms through which a cause affects an outcome, decomposing total effect into direct and indirect effects via intermediate variables.
Granger Causality
Statistical concept determining whether one time series significantly predicts another, based on improvement in predictive accuracy when including past values.
Average Treatment Effect on the Treated (ATT)
Measure of the average causal impact of treatment specifically on individuals who actually received it, relevant for evaluating targeted program effectiveness.
Confounding Bias
Systematic distortion in estimating treatment effect due to uncontrolled factors simultaneously influencing both exposure and outcome.
Causal Randomization
Technique of random treatment assignment to systematically eliminate confounding biases, ensuring comparable groups differ only in treatment received.
Causal Path Analysis
Quantitative method decomposing total correlations into direct and indirect effects through a network of causal relationships specified a priori.
Causal Transportability
Ability to generalize causal effects estimated from a source population to a different target population, by adjusting for distributional differences.