📊 statistics

Statistical Models for the Inference of Within-person Relations: A Random Intercept Cross-Lagged Panel Model and Its Interpretation

This paper positions the Random Intercept Cross-Lagged Panel Model (RI-CLPM) as a valuable tool for inferring within-person relations by clarifying its conceptual and mathematical relationships with other statistical models, highlighting its assumption of uncorrelated stable traits and within-person variability, and addressing its practical interpretation and application challenges.

Original authors: Satoshi Usami

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Satoshi Usami

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

The Big Picture: Who Are We Talking To?

Imagine you are a detective trying to solve a mystery: Does sleeping better make you happier, or does being happier make you sleep better?

For decades, psychologists have used a tool called the CLPM (Cross-Lagged Panel Model) to answer this. Think of the CLPM as a "Group Average Detective." It looks at a crowd of people and asks, "On average, do people who sleep more tend to be happier later?"

The problem? The CLPM is bad at telling the difference between who you are (your personality) and what you are doing right now (your daily mood).

This paper, written by Satoshi Usami, argues that we need a better detective. He introduces a new tool called the RI-CLPM (Random Intercept Cross-Lagged Panel Model). This tool is designed to solve the mystery of within-person relations: "When I sleep better than usual, do I become happier?"

The Problem: The "Group Average" Trap

To understand why the old tool (CLPM) is flawed, let's use an analogy.

Imagine a race with two runners:

Runner A is naturally very fast (a "stable trait").
Runner B is naturally slow.

If you just look at the data, Runner A is always ahead of Runner B. The old tool (CLPM) might conclude: "Being ahead causes you to run faster!" But that's wrong. Runner A is fast because of their natural talent, not because they are ahead.

In psychology, this is the Between-Person problem.

Between-Person: "People who generally sleep well are generally happier." (This is about comparing Person A to Person B).
Within-Person: "When I sleep better than my usual self, do I feel happier?" (This is about Person A changing over time).

The old tool mixes these two up. It can't tell if a change in happiness is because of a change in sleep, or just because the person is naturally a "happy sleeper."

The Solution: The "Personal Baseline" Detective (RI-CLPM)

The RI-CLPM is like a detective who gives every suspect a personal baseline.

Instead of comparing you to everyone else, the RI-CLPM asks: "How are you doing today compared to your own average?"

The Stable Trait (The "Anchor"): The model first identifies your "anchor." This is your stable personality. Are you naturally an anxious person? A calm person? A heavy sleeper? The model locks this down and says, "Okay, we know who you are. Let's ignore that for a moment."
The Deviation (The "Wobble"): Once the anchor is set, the model looks at the "wobble." Did you sleep 2 hours more than your usual amount today? Did you feel 2 points happier than your usual mood?
The Connection: The model then checks: "When the wobble in sleep goes up, does the wobble in happiness go up?"

The Analogy:
Imagine a boat on the ocean.

The CLPM looks at the whole ocean and says, "Boats in calm waters are higher than boats in rough waters." (It compares different boats).
The RI-CLPM looks at one specific boat. It ignores the tide (the ocean level) and asks, "When this specific boat bobs up a little higher than its normal floating point, does it move faster?"

By isolating the "bobbing" (within-person change) from the "tide" (between-person differences), the RI-CLPM gives a much clearer answer about cause and effect.

The "Other Suspects": Why Not Use Other Tools?

The author mentions that there are other detectives (statistical models) out there, like the LCM-SR, LCS, and GCLM. He compares them to detectives who are trying to do too many jobs at once.

The Over-Adjustment Trap: Some models try to map out the entire journey of a person's life (their growth curve) while also trying to figure out daily mood swings. The author argues this is like trying to watch a movie and count the frames at the same time. You end up throwing away the "baby with the bathwater"—you accidentally remove the very changes you are trying to study.
The "Accumulating" Factor: Some models assume that a person's past influences their future in a way that piles up over time (like a snowball rolling down a hill). The author suggests that while this is flexible, it's hard to know if that's actually how real life works.

The "Econometric" Cousin: The DPM

The paper also introduces a model from the world of economics called the Dynamic Panel Model (DPM).

Think of the RI-CLPM as a Sedentary Detective: It assumes your "anchor" (personality) stays perfectly still and doesn't interact with your daily changes.
Think of the DPM as a Dynamic Detective: It assumes your "anchor" might have been building up over time before the study even started.

The author points out that these two detectives are actually cousins. If you make certain assumptions, they give the same answer. But if the world is messy and complex, they might disagree.

The Bottom Line: What Should You Do?

The author concludes with a very practical piece of advice: Don't pick just one tool and stick with it.

Since we can't know the "true" way the world works (the data-generating process), the best strategy is Sensitivity Analysis.

Run the RI-CLPM (The "Anchor" model).
Run the DPM (The "Dynamic" model).
Compare the results.

If both models tell you the same story (e.g., "Sleep causes happiness"), then you can be very confident. If they tell different stories, you know the answer is complicated and depends on how you look at the data.

Summary in One Sentence

To truly understand how you change over time, you must stop comparing yourself to others and start comparing yourself to your own past, using a statistical model that separates your permanent personality from your daily fluctuations.

1. Problem Statement

The paper addresses a critical methodological issue in longitudinal research, particularly within psychology: the distinction between between-person relations (differences between individuals) and within-person relations (processes of change within an individual over time).

The Limitation of the CLPM: The Cross-Lagged Panel Model (CLPM) has been the standard for inferring reciprocal causal relations (e.g., does $X$ cause $Y$ or vice versa?). However, the CLPM fails to adequately control for stable individual differences (unobserved heterogeneity). Consequently, the cross-lagged coefficients in a CLPM represent a mixture of between-person and within-person effects, often leading to biased estimates that do not reflect true within-person causal processes.
The Rise of RI-CLPM: Hamaker et al. (2015) proposed the Random Intercept Cross-Lagged Panel Model (RI-CLPM) to solve this by separating stable trait factors (between-person variance) from within-person deviations. While the RI-CLPM has gained rapid popularity, the paper argues that:
1. Its mathematical relationship with other models (like the Dynamic Panel Model) is not clearly understood.
2. There is a lack of clarity regarding when the RI-CLPM is the appropriate choice versus other models (e.g., LCM-SR, LCS, GCLM, DPM).
3. The interpretation of the "stable trait factor" in the RI-CLPM from a causal inference perspective requires deeper scrutiny.

2. Methodology and Framework

The author employs a unified theoretical framework to compare various Structural Equation Modeling (SEM) approaches used for reciprocal relations. The methodology involves:

Mathematical Formulation: The paper derives and compares the algebraic representations of several models:
- CLPM: Standard cross-lagged regression without separating stable traits.
- RI-CLPM: Decomposes observed scores into a stable trait factor ( $I$ ) and within-person deviations ( $*$ ). Crucially, it assumes the stable trait factor is uncorrelated with within-person variability.
- LCM-SR (Latent Curve Model with Structured Residuals): Separates growth trajectories (intercept/slope) from residuals.
- LCS (Latent Change Score) & ALT (Autoregressive Latent Trajectory): Use "accumulating factors" where the effect of the common factor propagates through time via autoregressive paths.
- GCLM (General Cross-Lagged Panel Model): Includes stable traits and moving-average (MA) terms.
- DPM (Dynamic Panel Model): An econometric model using "unit effects" (accumulating factors) often estimated via GMM, but here analyzed via ML-SEM for comparison.
Unified Framework: The author categorizes these models based on three dimensions:
1. Type of Common Factor: Is it a stable trait factor (constant direct effect) or an accumulating factor (effect propagates over time)?
2. Mean Structure: Does the model explicitly model trajectories (growth factors) or just group means?
3. Measurement Error: Does the model account for measurement error (e.g., STARTS model)?
Empirical Application: The models are applied to longitudinal data from the Minnesota Adolescent Community Cohort (MACC, $N=4,671, T=6$ ) examining the relationship between movie smoking exposure and actual smoking intensity.
Simulation/Comparison: The paper analyzes estimation results under time-invariant and time-varying parameter assumptions to demonstrate how model choice affects coefficient estimates, standard errors, and model fit.

3. Key Contributions

A. Clarification of Model Relationships

The paper provides a rigorous mathematical mapping showing that:

RI-CLPM vs. DPM: Under specific constraints (time-invariant coefficients and specific initial condition assumptions), the RI-CLPM and the Dynamic Panel Model (DPM) can be mathematically equivalent.
The "Predetermined" RI-CLPM: If the RI-CLPM allows correlations between the stable trait factor and initial observations, it becomes equivalent to the DPM. However, this changes the interpretation of the trait factor from a "stable trait" to an "accumulating factor."
Accumulating vs. Stable Factors: The author distinguishes between stable trait factors (RI-CLPM, STARTS), which exert a constant direct effect, and accumulating factors (LCS, ALT, DPM, GCLM), whose effects accumulate over time through autoregressive paths.

B. Critique of Alternative Models

The paper identifies specific inferential pitfalls in models other than the RI-CLPM:

LCM-SR: Risks over-adjustment. By controlling for slope factors (which often represent the very within-person change of interest), it may distort the reciprocal relations ("throwing the baby out with the bathwater").
LCS/ALT/GCLM: These use accumulating factors. While flexible, they may not adequately represent time-invariant confounders and can lead to scale invariance issues (estimates change if variables are linearly transformed) unless strict constraints are imposed. The inclusion of Moving Average (MA) terms in GCLM complicates the interpretation of cross-lagged coefficients.

C. Causal Interpretation of the Stable Trait Factor

A major theoretical contribution is the re-evaluation of the stable trait factor in the RI-CLPM:

It is not a confounder in the traditional sense (it does not cause the within-person variability).
Instead, it functions as a centering operation (similar to person-mean centering in hierarchical linear models).
The RI-CLPM estimates within-person relations by orthogonalizing the variance into between-person (stable trait) and within-person components.
The paper argues that the RI-CLPM is valid only if the data-generating process assumes that unobserved heterogeneity is a stable trait uncorrelated with within-person change. If the heterogeneity is an "accumulating factor" (e.g., a time-invariant confounder that influences change over time), the RI-CLPM may be biased, and the DPM might be more appropriate.

D. Practical Guidelines

Time Points: The RI-CLPM requires at least 3 time points ( $T \geq 3$ ) for identification, whereas the CLPM requires only 2.
Improper Solutions: Models with measurement errors (like STARTS) or complex structures (LCS, GCLM) are more prone to improper solutions (e.g., negative variances) than the RI-CLPM.
Model Selection Strategy: Since the true data-generating process is unknown, the author recommends a sensitivity analysis approach: estimate multiple candidate models (RI-CLPM, DPM, Predetermined RI-CLPM) and compare fit indices, information criteria, and residual correlations to assess robustness.

4. Results from Empirical Example

Using the MACC smoking data:

Divergent Results: The CLPM and RI-CLPM produced significantly different cross-lagged coefficients (direction, magnitude, and significance).
RI-CLPM Fit: The RI-CLPM generally showed better model fit and smaller autoregressive coefficients with larger standard errors compared to the CLPM, reflecting the control of stable individual differences.
DPM vs. RI-CLPM: When assuming time-invariant coefficients, the Predetermined RI-CLPM and DPM yielded identical estimates. However, under time-varying assumptions, they diverged, with the DPM showing different significance patterns for cross-lagged effects.
Variance Decomposition: The study highlighted that variables like BMI and sleep duration had high between-person variance (high $R^2$ for stable traits), while depression showed higher within-person variability, suggesting the importance of model selection based on variable characteristics.

5. Significance and Future Directions

Paradigm Shift: The paper solidifies the shift in psychology from the CLPM to models that explicitly separate within- and between-person variance, validating the RI-CLPM as a primary tool while cautioning against its blind application.
Interdisciplinary Bridge: It bridges the gap between psychometrics (SEM) and econometrics (Dynamic Panel Models), clarifying that they are often mathematically related but differ in assumptions about initial conditions and factor types.
Causal Inference Rigor: It emphasizes that "within-person" does not automatically equal "causal." The validity of the inference depends on whether the model correctly captures the data-generating process (e.g., the nature of unobserved confounders).
Future Research: The author calls for:
1. More research on non-linear causal inference methods (e.g., Marginal Structural Models) applied to within-person data.
2. Better integration of substantive researchers (psychologists, epidemiologists) with methodologists to ensure models match theoretical hypotheses.
3. Continued development of methods to handle time-varying confounders that are not captured by simple stable traits.

In conclusion, Usami positions the RI-CLPM as a robust, interpretable, and generally superior method for inferring within-person reciprocal relations compared to the CLPM, provided the assumption of uncorrelated stable traits holds. However, he warns that researchers must carefully consider the nature of their data-generating process and consider alternative models like the DPM or conduct sensitivity analyses to ensure the validity of their causal inferences.