Causal Inference: Methods for Estimating Causal Effects from Observational
Data
Introduction
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
Estimating causal effects from data is a key goal in many scientific fields like medicine,
economics and epidemiology. However, randomized controlled trials, considered the gold
standard method, are often infeasible or unethical. As a result, researchers must rely on
observational data which typically lack random treatment assignment. This poses challenges
as treatment groups may differ systematically, making it difficult to isolate the causal effect
of interest from other confounding factors.
Causal inference methods aim to overcome these issues and draw valid causal conclusions
from observational studies. This paper provides an overview of the core challenges of causal
inference and reviews techniques for estimating average treatment effects from observational
data. We start by formalizing the potential outcomes framework and defining treatment effect
estimands. Key methods including matching, stratification, regression adjustment,
instrumental variables and the front-door criterion are then described. Applications in a
variety of domains demonstrate the importance of causal inference methods. The paper
concludes with perspectives on current open problems and future developments.
Potential Outcomes Framework
The fundamental framework for causal inference is the potential outcomes model or Rubin
causal model. It focuses on the comparison of potential outcomes under treatment and control
conditions for each unit.
For unit i:
- Yi(1) is the potential outcome if treated
- Yi(0) is the potential outcome if control
- Yi is the observed outcome = Yi(1) if treated, Yi(0) if control
- Ti = 1 if treated, 0 if control
The average treatment effect (ATE) is defined as E[Yi(1)-Yi(0)], the expected difference
between potential outcomes. This causal estimand contrasts the average outcomes under two
hypothetical scenarios: if everyone was treated versus if no one was treated.
Key assumptions include stable unit treatment value assumption (SUTVA), no hidden
versions of treatment, and unconfoundedness - given a set of observed covariates, assignment
to treatment is independent of potential outcomes.
Matching Methods
Propensity score matching aims to balance observed covariates between treated and control
groups, reducing confounding bias. The propensity score e(x) is the probability of receiving
treatment given pre-treatment characteristics:
e(x) = P(Ti = 1|Xi = x)
Units are matched on e(x), and treatment effects are estimated within matched sets. Optimal
matching algorithms like nearest neighbor matching with caliper distance are commonly
used. Variants include covariate matching and full matching. Post matching standardization
can further balance distributions.
Stratification and Standardization
Stratification divides units into buckets defined by covariate values, estimating treatment
effects within strata and aggregating using weights. Standardization reweights the treatment
groups to balance covariate distributions between them, similar to matching. These methods
provide decompositions of overall effects but cannot fully address hidden biases.
Regression Adjustment
Regression models like linear or logistic regression can flexibly control for measured
confounders when estimating ATE. Treatment is included as a predictor, and coefficients on
other predictors imply confounding adjustment. This relies on correct model specification and
measurement of all confounders. Variants include difference-in-differences designs for
panel/time-series data.
Instrumental Variables
When unobserved confounding may still bias estimates despite controlling observed factors,
instrumental variable (IV) methods can provide alternative identification. An IV is a variable
correlated with treatment receipt but affecting outcomes only through its effect on treatment.
Two-stage least squares estimation of a structural equation model identifies local average
treatment effects under IV assumptions.
Front-door Criterion
This graphical criterion provides graphical conditions for identifying causal effects from non-
experimental data. It applies when there exists a set of variables Z such that all backdoor
paths from treatment T to outcome Y are blocked after conditioning on Z, and all front-door
paths from T to Y pass through Z. Identifying and adjusting on such a set Z allows
identifying the causal effect of T on Y.
Applications and Extensions
Causal inference techniques are applied across disciplines to estimate varied treatment
effects. For example:
- Estimating vaccine effectiveness from observational healthcare databases
- Evaluating education policies’ effects using national panel survey data
- Assessing effects of smoking/alcohol on health using Mendelian randomization
- Estimating returns to schooling using changes in compulsory attendance laws
Extensions include Bayesian hierarchical modeling, sensitivity analyses, and modern
machine learning methods for causal inference.
Challenges and Future Directions
Key open challenges in causal inference include modeling heterogeneous/localized effects,
estimating effects in the presence of many time-varying/intermediate confounders/mediators,
and transporting effects across varying populations/contexts. Recent research also focuses on
relaxinvg assumptions like endogeneity of treatment, developing robust estimation under
weak unconfoundedness, and improving methods for complex/high dimensional designs
involving networks/graphs. Combining novel experimental and observational data sources
holds promise for advancing the field. Overall, addressing methodology for causal questions
will continue driving progress across many applied domains.
Conclusion
This paper overviewed core causal inference methods that can help answer important
scientific and policy questions using observational data rather than experimental studies. By
clarifying counterfactual quantities of interest and developing techniques to account for
potential confounding, researchers can aim to estimate average and distributional treatment
effects with validity. Matching, regression adjustment, instrumental variables and other
methods provide a toolkit for addressing identification problems arising from non-random
treatment assignment. Continued growth of applied causal inference relies on further
strengthening methods and expanding their scope alongside new data opportunities.
