📊 statistics

Identifying Unmeasured Confounders in Panel Causal Models: A Two-Stage LM-Wald Approach

This paper introduces the Two-Stage LM-Wald (2SLW) approach, a diagnostic tool grounded in latent variable modeling that extends Lagrange Multiplier and Wald tests to detect unmeasured confounders in panel causal models, thereby enhancing the robustness of causal inferences through simulation and empirical application.

Original authors: Bang Quan Zheng

Published 2026-02-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Bang Quan Zheng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: Did Event A cause Event B?

In social science, researchers often use "panel data" to solve this. This is like taking a series of photos of the same group of people over several years to see how their lives change. For example, they might track if a person's personality (like being friendly) changes their political views (like their stance on immigration) over time.

The problem is that these models often rely on a huge, unproven assumption: "We have measured everything that matters."

But what if there's a hidden variable? What if a secret third factor—like a person's upbringing or a hidden news source they never mentioned—is actually pulling the strings on both their personality and their politics? If you don't account for this "ghost in the machine," your detective work is flawed. You might think Personality caused Politics, when really, the "Ghost" caused both.

This paper introduces a new tool called 2SLW (Two-Stage LM-Wald) to catch these ghosts. Here is how it works, explained simply:

The Problem: The "Blind Spot"

Think of a standard panel model as a GPS navigation system. It calculates the best route from Point A to Point B based on the roads it knows about.

The Assumption: The GPS assumes there are no hidden shortcuts or secret tunnels (unmeasured confounders) that could mess up the route.
The Reality: Sometimes, a secret tunnel exists. If the GPS doesn't know about it, it might tell you the route is direct, when actually, you're being diverted by something invisible.

The Solution: The 2SLW Detective Kit

The authors created a two-step process to find these hidden tunnels. They call it 2SLW.

Step 1: The "Hunch" (The LM Test)

Imagine you are a mechanic listening to a car engine. You hear a weird rattle. You don't know exactly what's wrong yet, but you have a hunch that a specific bolt is loose.

In the paper, this is the Lagrange Multiplier (LM) test.
It scans the entire model and says, "Hey, the math doesn't add up here! There is a 'rattle' (statistical error) suggesting we are missing a connection."
It gives you a list of the top 25 suspects (potential missing connections) that might be causing the noise.

Step 2: The "Proof" (The Wald Test)

Now, you have a list of suspects. But you can't just fix everything; that would be like replacing every bolt in the car just because of one rattle. You need to be sure.

This is the Wald test.
You take the top suspects from Step 1 and actually test them. You ask: "If we add this missing connection, does it actually fix the problem, or was it just a false alarm?"
If the connection is real and important, the test says, "Yes, this is the culprit!" If it's just noise, it says, "Ignore this one."

Why This Matters: The "Ghost" Revealed

The paper tested this method using computer simulations (creating fake worlds where they knew the ghosts existed).

Result: The 2SLW tool successfully found the hidden ghosts 99% of the time. It told the researchers exactly where the model was lying to them.

They also tested it on real data: Personality vs. Immigration Attitudes.

The Old Way (Without 2SLW): The model suggested that changes in immigration attitudes caused changes in personality. This sounds weird! Usually, personality is the stable foundation, and attitudes change because of it. The model was confused because it missed a hidden factor.
The New Way (With 2SLW): The tool found a hidden "rattle" (a missing connection between different time points). Once they fixed it, the model made sense again: Personality changes actually did lead to changes in immigration attitudes. The "ghost" was removed, and the truth was revealed.

The Big Picture

Think of this paper as giving social scientists a metal detector for their data.

Before, researchers had to guess if their models were missing something. Now, they have a systematic way to scan their work, find the "hidden tunnels" (unmeasured confounders), and fix them before drawing conclusions.

In short:

The Problem: We often miss hidden factors that mess up our cause-and-effect conclusions.
The Tool: A two-step check (Hunch + Proof) that finds these hidden factors.
The Result: More trustworthy science. We stop blaming the wrong things and start understanding the real reasons behind human behavior.

The authors even made this tool easy to use in a free software program called R, so any researcher can start using this "metal detector" today.

1. Problem Statement

Panel data models (e.g., Cross-Lagged Panel Models [CLPM] and Random Intercept Cross-Lagged Panel Models [RI-CLPM]) are foundational for causal inference in social sciences. However, they rely on the critical assumption of sequential ignorability: that no unmeasured variables jointly influence predictors and outcomes over time.

The Gap: Violations of this assumption (unmeasured confounding) lead to biased estimates of autoregressive and cross-lagged effects, conflating true causal dynamics with spurious correlations.
Limitations of Current Methods:
- Standard lagged variables do not resolve endogeneity if unobserved confounders have dynamic persistence.
- RI-CLPMs separate stable between-person differences but fail to address time-varying unmeasured confounders.
- Existing sensitivity analysis techniques are often limited to single predictors, ignore latent residual correlations, or require subjective manual variable selection.
- Standard model fit statistics (e.g., $\chi^2$ , CFI) often fail to detect specific structural misspecifications related to unmeasured confounding.

2. Methodology: The Two-Stage LM-Wald (2SLW) Approach

The author proposes a diagnostic framework that extends Lagrange Multiplier (LM) and Wald tests to longitudinal causal models. The method does not recover the latent confounders themselves but tests whether the restrictions implied by sequential ignorability (e.g., zero constraints on unmodeled paths or residual covariances) are statistically compatible with the data.

The Two-Stage Procedure:

Stage 1: LM Test & Screening (Detection)
- Mechanism: A univariate LM test is applied to the constrained model to identify parameters where relaxing the constraint would significantly improve model fit.
- Metrics: The procedure ranks candidate parameters by their LM $\chi^2$ statistics and Expected Parameter Change (EPC) values.
- Filtering: The top 25 candidates are retained, excluding those that violate temporal ordering, cause non-identification, or have negligible EPCs.
- Goal: To flag potential "locally omitted" paths or residual covariances that suggest unmeasured confounding.
Stage 2: Wald Test (Validation)
- Mechanism: A forward stepwise Wald test is applied to the candidates identified in Stage 1. Parameters are added one by one to the model.
- Criteria: Only parameters with $p < 0.05$ are retained.
- Goal: To confirm that the detected deviations are substantively meaningful and not just statistical noise or overfitting. This distinguishes between superficial fit improvements and genuine structural violations.

Statistical Foundation

LM Test: Evaluates the score vector at the boundary of the parameter space (local necessity). It approximates the change in parameters if a constraint were released ( $\Delta \theta \approx I^{-1}s$ ).
Wald Test: Evaluates the magnitude and significance of parameters once unconstrained (global estimability).
Complementarity: The LM test identifies where the model might be misspecified; the Wald test confirms if the omission is significant enough to alter substantive conclusions.

3. Key Contributions

Novel Diagnostic Framework: Introduces the 2SLW procedure as a generalizable tool for detecting unmeasured confounding in longitudinal SEMs, specifically tailored for the RI-CLPM framework.
Theoretical Distinction: Reframes LM-Wald testing from a tool for "model fit improvement" (common in CFA) to a diagnostic for identifying assumption violations (sequential ignorability).
Implementation: Provides a scalable, automated workflow implemented in the R package lavaan, reducing the subjectivity of manual sensitivity analysis.
Generalizability: While demonstrated on RI-CLPMs, the framework is applicable to standard CLPMs and higher-order longitudinal designs.

4. Results

A. Monte Carlo Simulations

The study simulated data from three distinct population models (4-wave RI-CLPM) containing unmeasured confounders:

Model 1 (Correlation): Dynamic confounding between residuals of different waves/variables.
Model 2 (Direct Effect): Unmodeled direct paths bypassing intermediate processes.
Model 3 (Mediation): Unmeasured mediators distorting direct effects.

Findings:

Detection: The LM test successfully flagged the correct parameters with high $\chi^2$ and EPC values.
Validation: The Wald test effectively filtered out false positives, retaining only the parameters that matched the population truth (e.g., in Model 1, it identified the specific residual correlations; in Model 2, it identified the specific direct path).
Bias Correction: Incorporating the identified parameters into the model significantly reduced bias in autoregressive (AR) and cross-lagged (CL) coefficients, bringing estimates closer to population parameters.

B. Empirical Application

Data: LISS panel data (Netherlands, 2008, 2014, 2020) examining Agreeableness and Immigration Attitudes.
Baseline Model: Suggested that immigration attitudes predicted changes in agreeableness, while the reverse effect was weak/unstable—a theoretically counterintuitive finding (personality is usually an antecedent).
2SLW Application: The procedure identified a significant residual covariance between time points 1 and 3 for immigration attitudes ( $WFY_1 \sim \sim WFY_3$ ).
Outcome:
- Model Fit: Improved drastically ( $\chi^2$ dropped from 679.5 to 130.8; RMSEA from 0.059 to 0.011).
- Substantive Shift: After correcting for the residual structure, the causal interpretation aligned with theory: increases in agreeableness predicted less exclusionary immigration attitudes, while the reverse effect weakened.
- Conclusion: The baseline model's "wrong" conclusion was driven by unmodeled residual dependence, which 2SLW successfully detected and corrected.

5. Significance and Implications

Enhanced Credibility: The 2SLW approach strengthens the validity of causal claims in political science, psychology, and sociology by providing a rigorous check against hidden confounding.
Beyond Fit Indices: It addresses the limitation of relying solely on global fit indices, which can be misleading in complex panel models.
Diagnostic vs. Confirmatory: The author clarifies that 2SLW is a diagnostic tool. It signals where assumptions may fail but does not automatically produce bias-corrected estimates without theoretical justification. It prompts researchers to investigate specific structural misspecifications.
Future Directions: The paper suggests integrating 2SLW with formal sensitivity analysis to quantify how large unmeasured confounding must be to overturn conclusions, and calls for further research into small-sample performance.

In summary, the paper provides a robust, statistically grounded method for researchers to detect and address unmeasured confounding in panel data, moving beyond the assumption that temporal ordering alone guarantees causal validity.