Modern Causal Inference in Statistical Research

The modern mathematical formalization of causation originated in Jerzy Neyman's 1923 agricultural experiments and was generalized to observational studies by Donald Rubin in 1974 (the Neyman-Rubin Causal Model).

Consider a binary treatment indicator $D_i \in \{0, 1\}$, where $D_i = 1$ denotes exposure to the experimental treatment and $D_i = 0$ denotes exposure to the control condition. For each individual unit $i$ in a population, we define two Potential Outcomes:

• $Y_i(1)$: The outcome value that unit $i$ would realize if exposed to treatment ($D_i = 1$).
• $Y_i(0)$: The outcome value that unit $i$ would realize if exposed to control ($D_i = 0$).

The individual causal effect for unit $i$ is defined as the contrast:

This definition reveals the Fundamental Problem of Causal Inference: For any individual unit $i$, we observe only the outcome corresponding to the treatment actually received:

If unit $i$ receives treatment ($D_i = 1$), we observe $Y_i(1)$, while the counterfactual potential outcome $Y_i(0)$ remains unobservable. If unit $i$ receives control ($D_i = 0$), we observe $Y_i(0)$, while $Y_i(1)$ is counterfactual. Causal inference is fundamentally a missing data problem: for every individual, exactly one potential outcome is missing.

To identify causal effects, researchers shift from individual effects $\tau_i$ to population average effects:

• Average Treatment Effect (ATE):

• Average Treatment Effect on the Treated (ATT):

The Stable Unit Treatment Value Assumption (SUTVA)

The mathematical validity of the potential outcomes framework requires SUTVA:

1. No Interference (Spillover): The potential outcomes of unit $i$ are unaffected by the treatment status of any other unit $j \ne i$:

(Violated in vaccination programs with herd immunity, or social network interventions).

1. No Hidden Variations of Treatment: The treatment administered to unit $i$ is identical in dosage and delivery mechanism to the treatment received by any other treated unit (consistency: if $D_i = d$, then $Y_i = Y_i(d)$).

Decomposing Naive Observational Association

Consider the naive observational difference between the sample means of treated and untreated cohorts:

Using the potential outcomes identity, expand $\Delta_{\text{naive}}$:

Add and subtract the counterfactual quantity $\mathbb{E}[Y(0) \mid D = 1]$:

Naive observational comparisons confound the true causal treatment effect on the treated ($\tau_{\text{ATT}}$) with Selection Bias: the baseline disparity in potential control outcomes between individuals who choose treatment versus those who do not. In a completely randomized trial, $D \perp\!\!\perp (Y(1), Y(0))$, so baseline selection bias vanishes ($\mathbb{E}[Y(0) \mid D = 1] = \mathbb{E}[Y(0) \mid D = 0]$), guaranteeing $\Delta_{\text{naive}} = \tau_{\text{ATE}}$. In observational studies, selection bias must be neutralized through structural causal modeling.

Judea Pearl formalized causal relationships using non-parametric Directed Acyclic Graphs (DAGs). A causal DAG $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ consists of a set of vertices $\mathcal{V}$ representing random variables and a set of directed arrows $\mathcal{E}$ representing direct causal influences. The graph is acyclic: no directed path can loop back to its starting node.

A causal DAG encodes the complete qualitative causal structure of a data-generating system. All statistical associations between two variables $X$ and $Y$ flow along paths connecting them. Paths consist of combinations of three fundamental junction typologies:

1. The Chain (Mediation):

Variable $X$ causally influences mediator $M$, which causally influences $Y$.

• Unconditioned: The path is open; $X$ and $Y$ are statistically dependent ($X \not\perp\!\!\perp Y$).
• Conditioned on $M$: Conditioning on the intermediate mediator blocks the path; $X$ and $Y$ become conditionally independent ($X \perp\!\!\perp Y \mid M$).

1. The Fork (Common Cause / Confounder):

Variable $Z$ is a mutual common cause of both $X$ and $Y$.

• Unconditioned: The path is open; $Z$ transmits a spurious non-causal association between $X$ and $Y$ ($X \not\perp\!\!\perp Y$).
• Conditioned on $Z$: Conditioning on the common cause blocks the spurious path; $X$ and $Y$ become conditionally independent ($X \perp\!\!\perp Y \mid Z$).

1. The Inverted Fork / Collider (Mutual Effect):

Two independent causes $X$ and $Y$ both influence a mutual outcome (collider) $C$.

• Unconditioned: The collider naturally blocks the path! Without conditioning, no association flows between $X$ and $Y$; they are completely independent ($X \perp\!\!\perp Y$).
• Conditioned on $C$ (or any descendant of $C$): Conditioning on a collider opens the path! Conditioning induces a spurious, non-causal statistical correlation between previously independent causes ($X \not\perp\!\!\perp Y \mid C$). This phenomenon is known as Berkson's Paradox or Collider Stratification Bias (M-bias).

The $d$-Separation Criterion

Judea Pearl established the $d$-separation (directional separation) theorem, providing a purely algorithmic graphical rule for determining whether two variables $X$ and $Y$ are conditionally independent given an conditioning set $\mathbf{Z}$:

Definition: A path $p$ between $X$ and $Y$ is blocked (d-separated) by a set of conditioning nodes $\mathbf{Z}$ if and only if:

1. Path $p$ contains a chain ($A \to B \to C$) or a fork ($A \leftarrow B \to C$) such that the middle node $B$ is in $\mathbf{Z}$; OR
2. Path $p$ contains a collider ($A \to C \leftarrow B$) such that neither the collider $C$ nor any descendant of $C$ is in $\mathbf{Z}$.

If all paths connecting $X$ and $Y$ are blocked by $\mathbf{Z}$, then $X$ and $Y$ are $d$-separated by $\mathbf{Z}$, which implies that $X$ and $Y$ are conditionally independent given $\mathbf{Z}$ across all probability distributions compatible with the graph:

Pearl derived non-parametric identification theorems that dictate exactly which variables must be adjusted for—and which variables must never be adjusted for—to estimate the true causal effect of an intervention $\mathrm{do}(X = x)$ on outcome $Y$:

The Backdoor Criterion

A set of variables $\mathbf{Z}$ satisfies the Backdoor Criterion relative to an ordered pair of variables $(X, Y)$ in a DAG $\mathcal{G}$ if:

1. No node in $\mathbf{Z}$ is a causal descendant of $X$; AND
2. The set $\mathbf{Z}$ blocks (d-separates) every backdoor path from $X$ to $Y$ (where a backdoor path is any path connecting $X$ and $Y$ that begins with an arrow pointing into $X$: $X \leftarrow \cdots Y$).

Theorem (Backdoor Adjustment Formula): If a set of covariates $\mathbf{Z}$ satisfies the Backdoor Criterion, the causal effect of $X$ on $Y$ is non-parametrically identified by the adjustment formula:

The corresponding average causal effect is estimated by adjusting for $\mathbf{Z}$ via multivariable regression, propensity score matching, or inverse probability weighting.

The Frontdoor Criterion

When confounders $U$ between $X$ and $Y$ are completely unobserved (making backdoor adjustment impossible), Pearl discovered that causal effects can still be identified if an observed intermediate mechanism $M$ satisfies the Frontdoor Criterion:

1. $M$ intercepts all directed causal paths from $X$ to $Y$ ($X \to M \to Y$).
2. There are no unblocked backdoor paths from $X$ to $M$.
3. All backdoor paths from $M$ to $Y$ are blocked by $X$.

Under the Frontdoor Criterion, the causal effect is identified without measuring $U$:

When unmeasured confounding between treatment $X$ and outcome $Y$ exists and cannot be blocked, the Instrumental Variables (IV) approach achieves causal identification by exploiting an exogenous source of variation $Z$.

A valid instrumental variable $Z$ must satisfy three structural core conditions:

1. Relevance Condition (Instrument Strength): $Z$ has a non-zero causal effect on treatment $X$:

1. Exogeneity / Independence Condition: $Z$ is as good as randomly assigned; it is statistically independent of all unobserved confounders $U$:

1. Exclusion Restriction: $Z$ affects the outcome $Y$ only through its effect on treatment $X$. There is no direct causal arrow from $Z$ to $Y$, and no unblocked backdoor path connecting $Z$ and $Y$:

Under linear homogeneous models:

The Instrumental Variable estimator (Wald Estimator for binary $Z$) is:

Local Average Treatment Effects (LATE)

In their seminal 1994 paper, Joshua Angrist and Guido Imbens (2021 Nobel Prize in Economic Sciences) analyzed IV under heterogeneous treatment effects and non-compliance. They partitioned the population into four latent compliance strata:

1. Compliers: Units who take treatment if assigned, and do not if not assigned ($X(1) = 1, X(0) = 0$).
2. Always-Takers: Units who always take treatment regardless of instrument status ($X(1) = 1, X(0) = 1$).
3. Never-Takers: Units who never take treatment regardless of instrument status ($X(1) = 0, X(0) = 0$).
4. Defiers: Units who do the exact opposite of their assignment ($X(1) = 0, X(0) = 1$).

Under the Monotonicity Assumption (no defiers: $X_i(1) \ge X_i(0)$ for all $i$), Angrist and Imbens proved that the IV estimator identifies the Local Average Treatment Effect (LATE):

The causal effect is identified specifically for the subpopulation of Compliers—individuals whose treatment uptake was actively induced by the instrument.

Weak Instrument Problem and Stock-Yogo Diagnostic

If the instrument is weakly correlated with treatment ($\mathrm{Cov}(Z, X) \approx 0$), the denominator of $\beta_{\text{IV}}$ approaches zero. James Stock and Motohiro Yogo established that weak instruments cause severe finite-sample bias toward the confounded OLS estimate and produce invalid standard errors. Researchers must report the First-Stage $F$-statistic testing $H_0: \gamma = 0$ in $X = \pi_0 + \gamma Z + \nu$. As a rule of thumb, $F > 10$ (or Stock-Yogo critical values $F > 16.38$) is required to rule out weak instrument bias.

The Regression Discontinuity Design (RDD), introduced by Donald Thistlethwaite and Donald Campbell (1960), is widely considered the most credible quasi-experimental architecture because it exploits local administrative rules to mimic randomized trials.

In RDD, treatment assignment $D_i$ is determined strictly by whether an observed continuous running variable (forcing variable) $X_i$ crosses an exogenous cutoff threshold $c$:

• Sharp RDD: Treatment is deterministic:

(e.g., students scoring $\ge 80.0$ receive a merit scholarship; students scoring @@MATH_TOKEN_1@@lt; 80.0$ do not).

• Fuzzy RDD: The probability of treatment jumps discontinuously at the cutoff:

Identification Assumption: Continuity of Potential Outcomes

The foundational identification assumption of RDD is that in the absence of treatment, the conditional expectations of potential outcomes $\mathbb{E}[Y(0) \mid X = x]$ and $\mathbb{E}[Y(1) \mid X = x]$ are smooth, continuous functions of the running variable $X$ across the threshold $c$. Units scoring just above the threshold ($c + \epsilon$) are virtually identical in all baseline traits, unobserved ability, and potential outcomes to units scoring just below ($c - \epsilon$). Under continuity, the causal treatment effect at the cutoff is non-parametrically identified:

Estimation is executed using Local Polynomial Regression within an optimal bandwidth $h$ around $c$ (using Imbens-Kalyanaraman or Calonico-Cattaneo-Titiunik optimal bandwidth selectors). The McCrary Density Test verifies that individuals cannot precisely manipulate or sort themselves across the cutoff: a discontinuity in the marginal density of $X$ at $c$ indicates strategic gaming, invalidating RDD.

Difference-in-Differences identifies causal effects in longitudinal or panel settings where a treatment group is exposed to an intervention at a specific time point, while a comparison group remains unexposed.

Let $D_i \in \{0, 1\}$ indicate treatment group membership, and let $T_t \in \{0, 1\}$ indicate post-intervention time periods. The canonical Two-Way Fixed Effects (TWFE) regression model is:

The Difference-in-Differences estimator computes the double difference across groups and time:

The Parallel Trends Assumption

The fundamental identification assumption of DiD is Parallel Trends: in the counterfactual absence of treatment, the average outcome of the treatment group would have followed the identical temporal trajectory as the control group:

Parallel trends cannot be verified directly in the post-treatment era because counterfactual $Y_1(0) \mid D = 1$ is unobserved. Researchers assess plausibility using Event-Study Specifications across multiple pre-treatment periods:

Estimating $\beta_k \approx 0$ for all pre-treatment periods $k < -1$ provides empirical support for parallel trends.

Because observational causal inference relies on the untestable assumption of no unmeasured confounding (unconfoundedness), rigorous research methodology requires quantitative Sensitivity Analysis: assessing how strong an unmeasured confounder would need to be to overturn the study's conclusions.

1. Rosenbaum's Bounding Method ($\Gamma$):

Paul Rosenbaum (2002) formulated sensitivity bounds for matched observational studies. Let $\Gamma \ge 1$ denote the odds ratio of receiving treatment between two individuals matched on all observed covariates $\mathbf{X}$, caused by an unobserved confounder $U$:

When $\Gamma = 1.0$, there is zero unmeasured confounding. Rosenbaum bounds compute the threshold $\Gamma^*$ at which the $p$-value for the treatment effect crosses the non-significance boundary $\alpha = 0.05$. If a study achieves $\Gamma^* = 3.5$, an unobserved confounder would need to multiply the odds of treatment by $350\%$ to explain away the observed effect, demonstrating exceptional robustness.

1. The E-Value Metric (Ding & VanderWeele, 2016):

Tyler VanderWeele and Peng Ding introduced the E-value, defining the minimum strength of association that an unmeasured confounder $U$ must have with both the treatment $D$ and the outcome $Y$, conditional on observed covariates, to fully explain away the observed treatment-outcome risk ratio $\mathrm{RR}$:

For an observed risk ratio of $\mathrm{RR} = 3.0$:

To nullify the finding, an unmeasured confounder would have to be associated with both the treatment and the outcome by a risk ratio of at least $5.45$-fold each, beyond all measured controls.

In classical Difference-in-Differences, all treated units receive intervention at the identical calendar time. In real-world policy research, however, interventions adopt a Staggered Rollout architecture: different states, municipalities, or hospitals enact policies at different times.

Historically, researchers estimated staggered DiD using Two-Way Fixed Effects (TWFE) regression:

In a groundbreaking 2021 econometrics paper, Andrew Goodman-Bacon proved the Goodman-Bacon Decomposition Theorem: Under staggered timing, the TWFE coefficient $\hat{\beta}_{\text{TWFE}}$ is a weighted average of all possible $2 \times 2$ DiD comparisons across timing cohorts:

Critically, Goodman-Bacon proved that the comparison pairs include:

1. Early-treated units versus never-treated controls (Clean comparison).
2. Late-treated units versus never-treated controls (Clean comparison).
3. Late-treated units versus already-treated early units, where the early-treated units serve as the control group ("Forbidden Comparison")!

If the causal treatment effect evolves, grows, or decays over time (dynamic treatment heterogeneity), using already-treated units as controls subtracts the ongoing treatment trajectory of the early cohort, producing negative weights! In fact, $\hat{\beta}_{\text{TWFE}}$ can be negative even when every individual unit experiences a positive causal effect! Modern causal econometrics resolves this bias using robust staggered estimators: Callaway and Sant'Anna (2021) and Sun and Abraham (2021), which estimate group-time Average Treatment Effects $\mathrm{ATT}(g, t)$ using strictly clean, untreated comparison cohorts.

Modern causal inference provides the mathematical language to answer foundational scientific inquiries:

• Formalize causal questions using Potential Outcomes ($Y(1), Y(0)$), recognizing that individual causal effects are missing data problems.
• Construct causal Directed Acyclic Graphs (DAGs) and apply $d$-separation to differentiate mediators, confounders, and colliders.
• Apply the Backdoor Criterion to identify minimal sufficient adjustment sets, strictly avoiding collider stratification bias (M-bias).
• In the presence of unobserved confounding, deploy quasi-experimental identification strategies: Instrumental Variables (2SLS/LATE), Regression Discontinuity Designs (RDD), and Difference-in-Differences (DiD).