EPPP: PART 1, KNOWLEDGE • DOMAIN 7: RESEARCH METHODS AND STATISTICS

Advanced Analysis — Interpret meta-analysis, factor analysis, and causal modeling findings

Mastering the quantitative techniques that synthesize evidence, uncover latent constructs, and test causal hypotheses in behavioral health research.

Historical Context & Motivation

Throughout the twentieth century, behavioral health researchers confronted a persistent challenge: individual studies often produced conflicting results, psychological constructs proved difficult to measure directly, and claims about causation remained contentious in non-experimental settings. Three families of statistical techniques emerged to address these problems. Meta-analysis provided a systematic way to aggregate findings across studies and estimate the true magnitude of an effect. Factor analysis offered a method for identifying the latent dimensions underlying batteries of observed measures—revealing, for example, that dozens of personality items could be distilled into a smaller number of coherent traits. Causal modeling (including path analysis and structural equation modeling) allowed investigators to specify and test theoretical networks of direct and indirect effects, moving beyond simple correlational statements toward rigorous causal inference.

1904
Spearman's Two-Factor Theory
Charles Spearman introduced factor analysis to psychology, proposing a general intelligence factor (g) derived from correlations among mental tests. This laid the mathematical foundation for all subsequent factor-analytic methods.
1918–1934
Path Analysis Emerges
Geneticist Sewall Wright developed path analysis to decompose correlations into direct and indirect causal effects, initially applied to animal breeding. Social scientists later adapted the technique for observational research on human behavior.
1976
Glass Coins 'Meta-Analysis'
Gene V. Glass formally named and systematized meta-analysis in his presidential address to the American Educational Research Association, applying it to synthesize psychotherapy outcome studies and demonstrating that therapy was, on average, effective.
1980s
Structural Equation Modeling Matures
Software packages such as LISREL and EQS made structural equation modeling (SEM) accessible to behavioral researchers. SEM unified confirmatory factor analysis and path analysis within a single framework, enabling simultaneous testing of measurement and structural models.
2000s–Present
Modern Extensions
Advances in computing power and Bayesian methods have expanded all three techniques. Network meta-analysis compares multiple treatments, bifactor models refine latent structure, and directed acyclic graphs (DAGs) formalize causal assumptions in observational health research.

Each of these techniques addresses a fundamental gap in the behavioral health evidence base. How large is an effect really, when dozens of studies disagree? What latent constructs account for the covariation among observed symptoms? And can we test whether a theorized chain of causation—stress → coping failure → depression—is consistent with the data? The sections that follow equip you to interpret findings from all three analytic families with the precision expected on the EPPP.

Core Principles & Definitions

Before diving into formulas and output tables, it is essential to grasp the conceptual architecture that underpins each technique. Although meta-analysis, factor analysis, and causal modeling serve different purposes, they share a common logic: each transforms observed data into estimates of quantities that cannot be measured directly—aggregate effect sizes, latent factors, or causal path coefficients. Understanding five foundational principles will anchor your interpretation of results across all three methods.

1

Effect Size as a Common Metric

Meta-analysis converts each study's result into a standardized effect size (e.g., Cohen's d, Pearson's r, odds ratio) so that studies using different scales can be compared and combined on a single continuum.
2

Latent vs. Observed Variables

Factor analysis distinguishes between observed variables (item scores, test responses) and latent variables (unobservable constructs such as anxiety or neuroticism). Factors are inferred from patterns of covariation among the observed indicators.
3

Heterogeneity & Moderators

A crucial meta-analytic concept is heterogeneity—the degree to which effect sizes vary across studies beyond what sampling error alone would predict. When heterogeneity is substantial, moderator analyses (e.g., by population, treatment dose) explain why effects differ.
4

Model Fit & Parsimony

Both confirmatory factor analysis (CFA) and structural equation modeling evaluate how well a hypothesized model reproduces the observed covariance matrix. Fit indices (CFI, RMSEA, SRMR) quantify this match, balancing accuracy against model complexity.
5

Causal Inference Requires Theory

Causal modeling does not prove causation from data alone. Arrows in a path diagram represent theoretically specified directional relationships. The model tests whether data are consistent with the proposed causal structure, but alternative equivalent models may fit equally well.
KEY TAKEAWAY
Think of these three techniques as different lenses on the same research landscape. Meta-analysis is like a wide-angle lens that panoramically integrates many studies into one big picture. Factor analysis is a microscope that reveals hidden structure inside a single data set. Causal modeling is a blueprint that maps the theorized wiring diagram among variables and checks whether reality matches the plan. Each lens answers a question the other two cannot.

Visual Explanation — How the Three Methods Relate

The following diagram illustrates the conceptual flow of each analytic family—from its input data, through its core operation, to its primary output. Notice how meta-analysis begins with multiple study-level results, factor analysis begins with a correlation matrix among observed items, and causal modeling begins with a theorized path diagram plus raw or summarized data. Each technique ultimately produces a different kind of interpretive product: an aggregate effect estimate, a set of latent dimensions, or a system of causal coefficients.

Each analytic family transforms a distinct type of input into a specific interpretive output. Note that structural equation modeling (SEM) (right) integrates confirmatory factor analysis and path analysis within a single unified framework, making it particularly powerful for behavioral health research that involves latent constructs and theorized causal pathways.

In the diagram, dashed lines converging on the SEM box illustrate that structural equation modeling is not a separate technique so much as a superordinate framework. When you encounter SEM in the literature, recognize that its measurement model is essentially a confirmatory factor analysis, while its structural model is essentially a path analysis. The EPPP expects you to understand this integration and to interpret SEM output accordingly.

Mathematical Framework

Meta-Analysis: Weighted Mean Effect Size

The core computation in meta-analysis is a weighted average of study-level effect sizes, where each study's weight reflects the precision of its estimate (typically the inverse of its variance). Larger, more precise studies exert greater influence on the pooled estimate. Two weighting models are commonly used: the fixed-effect model (assumes one true effect) and the random-effects model (assumes effects vary across studies due to real differences in populations, treatments, or contexts).

WEIGHTED MEAN EFFECT SIZE
ē = Σ(wᵢ × dᵢ) / Σ(wᵢ)
Where dᵢ = effect size from study i, and wᵢ = 1/Vᵢ (inverse variance weight). In a random-effects model, Vᵢ is replaced by (Vᵢ + τ²), where τ² is the between-study variance.
HETEROGENEITY STATISTIC (Q)
Q = Σ wᵢ(dᵢ − ē)²
Q follows a χ² distribution with k − 1 degrees of freedom (where k = number of studies). A significant Q indicates that effect sizes vary more than expected from sampling error alone. = (Q − df)/Q × 100% expresses heterogeneity as a percentage; values of 25%, 50%, and 75% suggest low, moderate, and high heterogeneity.

Factor Analysis: The Factor Model

Factor analysis posits that each observed variable is a linear combination of a smaller number of latent factors plus unique variance (error). The central equation expresses an observed score as the sum of contributions from each factor, weighted by the variable's loading on that factor.

COMMON FACTOR MODEL
Xᵢ = λᵢ₁F₁ + λᵢ₂F₂ + … + λᵢₘFₘ + eᵢ
Where Xᵢ is the observed score on variable i, λᵢⱼ is the factor loading of variable i on factor j, Fⱼ is the latent factor score, and eᵢ is residual (unique) variance. A loading of ≥ |.40| is conventionally considered meaningful.

Causal Modeling: Path Coefficients

INDIRECT EFFECT
Indirect effect = a × b
In mediation (X → M → Y), a is the path from X to mediator M, and b is the path from M to Y controlling for X. The total effect = direct effect (c') + indirect effect (a × b). The Sobel test or bootstrapped confidence intervals evaluate whether the indirect effect is statistically significant.
💡 EPPP Tip
The EPPP frequently tests your ability to distinguish between a mediator (explains the mechanism through which X affects Y) and a moderator (changes the strength or direction of the X–Y relationship). In path diagrams, mediators sit on a causal chain between X and Y, while moderators interact with X (often depicted as an interaction term entering the equation predicting Y).

Detailed Breakdown — EFA vs. CFA and Fit Indices

A critical distinction for the EPPP is the difference between exploratory factor analysis (EFA) and confirmatory factor analysis (CFA). EFA is used when the researcher has no strong prior theory about factor structure; the algorithm freely assigns loadings and the investigator interprets the resulting pattern. CFA, by contrast, is a hypothesis-testing procedure in which the researcher specifies which items load on which factors in advance and then evaluates how well that a priori model fits the data. Because CFA embeds within the SEM framework, it produces formal fit indices.

Left panel: In EFA, all items load freely on all factors (opacity indicates strength). Right panel: In CFA, the researcher specifies which items load on which factors and constrains cross-loadings to zero. CFA produces formal fit indices (CFI, RMSEA, SRMR) that evaluate how well the a priori model reproduces the observed correlations.
Common Fit Indices Used in CFA and SEM
Fit IndexGood Fit CutoffWhat It Measures
χ² (Chi-square)Non-significant p (ideally)Exact fit between model-implied and observed covariance matrices. Sensitive to sample size—often significant even with minor misfit in large samples.
CFI≥ .95Comparative Fit Index: compares the hypothesized model to a baseline (null) model. Ranges 0–1; higher is better.
RMSEA≤ .06Root Mean Square Error of Approximation: penalizes model complexity. Lower is better; values > .10 indicate poor fit.
SRMR≤ .08Standardized Root Mean Square Residual: average discrepancy between observed and predicted correlations.
TLI (NNFI)≥ .95Tucker-Lewis Index: similar to CFI but adjusts for parsimony. Can exceed 1.0 in over-fitting models.
🔄 Rotation in EFA
EFA solutions are rotated to achieve a simpler, more interpretable factor structure. Orthogonal rotation (e.g., Varimax) constrains factors to be uncorrelated—useful when theory predicts independence. Oblique rotation (e.g., Promax, Direct Oblimin) allows factors to correlate, which is generally more realistic in behavioral health because psychological constructs tend to overlap. The EPPP may ask which rotation method is appropriate when factors are expected to be correlated.

Worked Example — Interpreting a Meta-Analysis of CBT for Depression

Suppose a meta-analysis of 20 randomized controlled trials compares cognitive-behavioral therapy (CBT) to wait-list control for major depressive disorder. The forest plot and summary statistics are provided. Walk through interpretation step by step.

Interpreting a Meta-Analytic Forest Plot
1
Step 1 — Identify the Pooled Effect SizeThe summary diamond at the bottom of the forest plot shows a pooled Cohen's d = 0.82 with a 95% CI of [0.64, 1.00]. Because the confidence interval does not cross zero, the pooled effect is statistically significant. Using Cohen's benchmarks, this is a large effect (d ≥ 0.80), indicating that the average CBT participant improved by 0.82 standard deviations more than the average control participant.
Pooled d = 0.82, 95% CI [0.64, 1.00] — large, significant effect favoring CBT.
2
Step 2 — Evaluate HeterogeneityThe output reports Q(19) = 62.4, p < .001, and I² = 69.6%. The significant Q test tells us effect sizes vary more than sampling error alone would predict. An I² of roughly 70% means that about 70% of the total variance in effect sizes reflects real differences across studies, not random fluctuation. This is high heterogeneity, warranting moderator analysis.
I² = 69.6% — high heterogeneity; moderator analysis needed.
3
Step 3 — Interpret Moderator AnalysisA subgroup analysis reveals that individual CBT (k = 12, d = 0.95) produces a larger effect than group CBT (k = 8, d = 0.58), Q-between = 8.9, p = .003. Treatment format moderates the overall effect. The residual I² within each subgroup drops to 38%, suggesting that format explains a substantial portion of the original heterogeneity.
Individual CBT (d = 0.95) > Group CBT (d = 0.58); format is a significant moderator.
4
Step 4 — Assess Publication BiasThe funnel plot displays study effect sizes against their standard errors. An asymmetric funnel (more large effects among small studies) suggests publication bias. Egger's regression test yields an intercept of 2.1, p = .04, indicating significant asymmetry. A trim-and-fill analysis estimates the adjusted pooled d = 0.71, still clinically meaningful but somewhat attenuated.
Evidence of publication bias; trim-and-fill adjusted d = 0.71 (still a medium-to-large effect).

Strengths, Limitations, and Comparisons

Each technique carries distinctive strengths and vulnerabilities. Recognizing these is essential for critically evaluating published research and for selecting the appropriate method when designing your own studies. The table below organizes the key considerations for each analytic family.

Strengths and Limitations of Advanced Analytic Techniques
TechniqueStrengthsLimitations
Meta-AnalysisIncreases statistical power by aggregating samples; provides precise effect-size estimates; identifies moderators of effects; reduces reliance on single studies; promotes evidence-based practice.Vulnerable to publication bias (file-drawer problem); quality depends on quality of included studies ('garbage in, garbage out'); coding decisions introduce subjectivity; combining heterogeneous studies may obscure meaningful differences.
Factor Analysis (EFA)Reveals hidden structure in complex data sets; reduces dimensionality; aids scale development and construct validation; guides theory building when structure is unknown.Multiple criteria for number of factors (eigenvalue > 1, scree plot, parallel analysis) can yield different solutions; factor interpretation is subjective; sample size requirements are substantial (typically N ≥ 300 or 10:1 subjects-to-items ratio).
Factor Analysis (CFA)Tests a priori theory about factor structure; provides formal fit indices; allows comparison of competing models; integrates into SEM framework.Requires strong theory before data collection; χ² is sensitive to sample size; modification indices can lead to data-driven (atheoretical) model changes; requires large samples (N ≥ 200).
Causal Modeling (Path/SEM)Tests complex multivariate theories simultaneously; decomposes total effects into direct and indirect components; accounts for measurement error (in SEM); can model latent variables and structural paths together.Causation is assumed, not proven—cross-sectional data cannot establish temporal precedence; equivalent models may fit equally well; requires large samples and multivariate normality; misspecification can produce biased estimates.
KEY TAKEAWAY
No single technique is a silver bullet. Think of advanced statistical methods like diagnostic tools in a clinic: a meta-analysis is like reviewing a patient's complete medical history to see the overall trend, a factor analysis is like running a panel of lab tests to identify underlying conditions, and causal modeling is like mapping a clinical formulation that traces how stressors, vulnerabilities, and symptoms connect. Each tool is powerful, but each can mislead if applied without clinical judgment—or, in the case of statistics, without sound theoretical reasoning and appropriate data.

Connection to Advanced Theory — Modern Extensions

The foundational methods discussed so far have spawned sophisticated extensions that increasingly appear in the behavioral health literature and, occasionally, on the EPPP. Understanding how the basic techniques connect to their advanced counterparts strengthens your interpretive toolkit and prepares you for emerging trends in psychological research.

Foundational Methods and Their Modern Extensions
Foundational MethodAdvanced ExtensionKey Difference / Application
Pairwise meta-analysisNetwork meta-analysis (NMA)Compares multiple treatments simultaneously using both direct and indirect evidence. Useful when no single RCT compares all available interventions (e.g., ranking 8 antidepressants).
EFA / CFA (correlated factors)Bifactor modelsModel a general factor (e.g., psychopathology 'p' factor) alongside group-specific factors. Disentangles general from domain-specific variance in symptom measures.
Path analysis (observed variables)Full SEM with latent variablesCorrects for measurement error by using multiple indicators per construct. Provides more accurate path coefficients than path analysis with observed composites.
Cross-sectional SEMLatent growth curve modeling (LGCM)Models individual trajectories of change over time using repeated measures. Estimates latent intercept (starting level) and slope (rate of change) as latent variables.
Traditional causal assumptionsDirected Acyclic Graphs (DAGs)Graphical framework for formalizing causal assumptions and identifying confounders, colliders, and mediators before running statistical models. Increasingly used in epidemiology and behavioral health.

For the EPPP, focus primarily on demonstrating competence with the foundational methods and their standard output (effect sizes, factor loadings, path coefficients, fit indices). However, awareness of these extensions—particularly SEM and mediation/moderation analysis—is increasingly expected. The critical interpretive principle remains constant across all these methods: statistical models test the plausibility of theoretical claims, but they do not prove causation in the absence of experimental control and temporal ordering.

Practice Problems

PROBLEM 1CONCEPTUAL
A researcher reports that a meta-analysis of 15 studies on mindfulness-based stress reduction yielded a pooled Cohen's d = 0.45 with I² = 82%. What does the I² value tell you, and what is the most appropriate next analytic step?
PROBLEM 2BASIC CALCULATION
In an exploratory factor analysis of a 20-item anxiety questionnaire, the first three eigenvalues are 6.4, 2.1, and 0.9. Using the Kaiser criterion (eigenvalue > 1), how many factors should be retained? What percentage of the total variance does the first factor account for?
PROBLEM 3INTERMEDIATE
A CFA of a two-factor model of emotional intelligence yields the following fit indices: χ²(34) = 89.5, p < .001; CFI = .93; RMSEA = .09; SRMR = .07. Does the model fit the data well? Justify your answer by evaluating each index against conventional cutoffs.
PROBLEM 4APPLIED
A researcher tests a mediation model: Childhood Adversity (X) → Emotion Dysregulation (M) → Adult Depression (Y). The path from X to M yields a = 0.42 (p < .001); the path from M to Y yields b = 0.35 (p = .002); and the direct path from X to Y drops from c = 0.38 (total effect) to c' = 0.23 (p = .04) when M is included. Calculate the indirect effect and determine whether this represents full or partial mediation.
PROBLEM 5CRITICAL THINKING
A colleague presents a cross-sectional SEM in which a latent 'Social Support' variable causes a latent 'Resilience' variable, which in turn causes a latent 'Well-Being' variable. The model fits well (CFI = .97, RMSEA = .04). Your colleague concludes that social support causally increases resilience, which causally enhances well-being. Identify at least three methodological problems with this causal conclusion, even given good fit.

Summary — Advanced Analysis for the EPPP

This lesson covered three indispensable analytic families in behavioral health research. Meta-analysis synthesizes findings across multiple studies by computing a weighted mean effect size, evaluating heterogeneity (Q and I²), conducting moderator analyses when variability is high, and assessing publication bias via funnel plots and statistical tests. Factor analysis identifies latent constructs from observed data: EFA explores structure when theory is underdeveloped, while CFA tests a priori models using fit indices (CFI ≥ .95, RMSEA ≤ .06, SRMR ≤ .08). Key decisions include the number of factors to retain, choice of rotation (orthogonal vs. oblique), and the threshold for meaningful factor loadings (≥ |.40|).

Causal modeling—including path analysis and structural equation modeling (SEM)—tests theorized networks of direct and indirect effects. SEM integrates a measurement model (CFA) with a structural model (path analysis) and evaluates overall model fit. The critical caveat is that causal arrows are theory-driven: good model fit does not prove causation, especially with cross-sectional data. Distinguish mediators (mechanism variables on a causal chain) from moderators (variables that change the strength or direction of an effect). For the EPPP, be prepared to interpret forest plots, evaluate fit indices, decompose direct and indirect effects, and critically appraise causal claims made from statistical models.

Varsity Tutors • EPPP: Part 1, Knowledge • Advanced Analysis — Interpret meta-analysis, factor analysis, and causal modeling findings