Dependent variable selection in phylogenetic generalized least squares regression analysis under Pagel's lambda model

This study demonstrates that swapping dependent and independent variables in phylogenetic generalized least squares (PGLS) regression can yield inconsistent conclusions and proposes using Pagel's lambda or Blomberg's K as superior criteria for selecting the dependent variable when causal relationships between traits are unclear.

The Gist

The Big Problem: The "Which Came First?" Dilemma

Imagine you are a detective trying to solve a mystery about two suspects, Trait A and Trait B. You know they are related—when one changes, the other tends to change too. But you don't know who is the mastermind and who is the sidekick. Did A cause B? Or did B cause A? Or did they just grow up together?

In biology, scientists use a high-tech tool called PGLS (Phylogenetic Generalized Least Squares) to study these relationships. Think of PGLS as a super-advanced calculator that looks at the family tree of species (like a giant genealogy chart) to figure out if two traits are truly connected, ignoring the fact that cousins share DNA.

The Catch: To use this calculator, you have to tell it which trait is the "predictor" (the cause) and which is the "outcome" (the effect).

The authors of this paper discovered a weird glitch: If you swap the roles, the calculator sometimes gives you a completely different answer.

• Scenario 1: You say "Trait A predicts Trait B." The calculator says, "Yes! They are definitely linked!"
• Scenario 2: You say "Trait B predicts Trait A." The calculator says, "Nope, that's just a coincidence."

This is like asking a judge, "Is the suspect guilty?" and getting a "Yes" if you ask it one way, but a "No" if you ask it the other way. That's a problem for science!

The Investigation: Running a Simulation Lab

To figure out why this happens and how to fix it, the researchers (Chen, Guo, and Niu) built a virtual laboratory.

1. The Setup: They created 16,000 fake evolutionary histories (family trees) with 100 species each.
2. The Experiment: They invented two fake traits for these species. In some cases, the traits were strongly linked; in others, they were weakly linked.
3. The Test: They ran the PGLS calculator on these fake data, swapping the roles of the traits over and over again.

The Result: They found that swapping the roles caused conflicting results about 13% of the time. When the link between the traits was weak (lots of "noise" or randomness), the calculator got very confused and gave different answers depending on how you set it up.

The "Golden Standard": How to Know the Truth

Since they were working with fake data, the researchers knew the absolute truth. They could look at the "branches" of the fake family tree and see exactly how the traits changed over time. This gave them a "Golden Standard"—a way to know which answer was actually correct.

They realized that the confusion happened because the calculator was trying to guess how much of the trait's history was written in the family tree (this is called phylogenetic signal).

• The Analogy: Imagine two people, Alice and Bob.
  • Alice has a very strong family tradition (high phylogenetic signal). Her traits are passed down strictly from her ancestors.
  • Bob is a rebel. He changes his traits based on whatever is happening in the moment, ignoring his family history (low phylogenetic signal).

The researchers found that if you try to predict Bob's behavior based on Alice's, the math gets messy. But if you use Alice (the one with the strong family tradition) as the starting point, the math works much better.

The Solution: The "Strongest Signal" Rule

The paper tested seven different ways to decide which trait should be the "predictor" and which should be the "outcome." They compared things like:

• Which model fits the data best?
• Which one has the lowest p-value?
• Which one has the highest "R-squared"?

The Winner: They found that three specific tools were the best at picking the right direction:

1. Pagel's Lambda (λ)
2. Blomberg's K
3. The estimated Lambda (λ̂)

The Simple Rule: Always pick the trait with the stronger "family tradition" (higher phylogenetic signal) to be the dependent variable (the outcome).

Think of it like this: If you are trying to understand why a car is moving, you should look at the engine (the strong, consistent force) rather than the wind blowing against the windshield (the noisy, unpredictable force). By letting the "stronger" trait drive the analysis, the results become consistent, no matter which way you look at it.

Why This Matters

Before this paper, scientists might have been getting different results just because they guessed the wrong direction for their variables. This could lead to wrong conclusions about how evolution works.

The Takeaway:
When you don't know if Trait A causes Trait B, or vice versa, don't just guess. Check which trait has a stronger connection to the family tree. Use that one as your anchor. It's like choosing the most stable leg of a wobbly table to stand on; it keeps the whole analysis from tipping over.

This doesn't mean we know the true cause in biology, but it ensures that our statistical tools give us the most reliable, consistent answer possible.

Technical Summary

1. Problem Statement

Phylogenetic Generalized Least Squares (PGLS) is a standard method for analyzing evolutionary associations between traits while accounting for phylogenetic non-independence. However, PGLS requires designating one trait as the dependent variable (response) and the other as the independent variable (predictor), implying a causal direction. In many biological contexts, the causal relationship between two traits is unknown or debated.

The authors identified a critical flaw: swapping the dependent and independent variables in PGLS analyses (using Pagel's $\lambda$ model) can lead to inconsistent and conflicting conclusions. Specifically, a correlation might be statistically significant ($p < 0.05$) in one configuration but non-significant ($p \ge 0.05$) in the other, or the sign of the regression coefficient might change. This ambiguity undermines the robustness of evolutionary inferences when causal directions are unclear.

2. Methodology

The study employed a combination of empirical re-analysis and extensive simulation to investigate the issue and propose a solution.

• Empirical Data: The authors re-analyzed a dataset of 262 bacterial species (traits: minimal doubling time, CRISPR spacer numbers, optimal growth temperature, prophage numbers) from a previous study (Liu et al., 2023). They performed PGLS regressions in both directions ($X_1 \sim X_2$ and $X_2 \sim X_1$) to quantify the frequency of conflicting outcomes.
• Simulation Design:
  • Structure: 16,000 simulations were conducted on binary trees with 100 terminal nodes.
  • Evolutionary Models: Two scenarios were tested:
    1. "BM & BM + Norm": Trait $X_1$ evolved via Brownian Motion (BM); $X_2 = X_1 + \epsilon$ (where $\epsilon$ is Gaussian noise).
    2. "Norm & Norm + BM": Trait $X_1$ was drawn from a normal distribution; $X_2 = X_1 + \epsilon$ (where $\epsilon$ evolved via BM).
  • Variance Gradients: The variance of the noise term ($\sigma^2$) was varied from $10^{-4}$ to $1024$ to create a gradient of correlation strengths (from strong to weak).
  • Analysis: For each simulation, PGLS regressions were run in both directions using Pagel's $\lambda$ model.
• "Golden Standard" for Accuracy: Since the true evolutionary correlation is known in simulations, the authors established a ground truth. They calculated the correlation of trait changes along phylogenetic branches ($\Delta X / L$) using Pearson or Spearman correlation. This served as the benchmark to determine if a PGLS result was "correct."
• Criteria Evaluation: Seven potential criteria were evaluated to determine which one best predicts the correct dependent variable in cases of conflicting PGLS results:
  1. Log-likelihood (LLK)
  2. Akaike Information Criterion (AIC)
  3. Coefficient of Determination ($R^2$)
  4. P-value of the regression coefficient
  5. Pagel's $\lambda$ (phylogenetic signal of the traits)
  6. Blomberg's $K$ (phylogenetic signal of the traits)
  7. Estimated $\hat{\lambda}$ (the $\lambda$ parameter estimated within the PGLS model)

3. Key Results

• Prevalence of Conflicts:
  • In the empirical bacterial dataset, swapping variables led to conflicting conclusions (significant vs. non-significant) in 26.3% of trait pairs.
  • In simulations, conflicting results occurred in 12.9% of cases. The frequency of conflict was highest when the correlation between traits was moderate (noise variance was intermediate) and decreased when correlations were extremely strong or extremely weak.
• Impact on Parameter Estimation:
  • The estimated $\hat{\lambda}$ in PGLS is heavily influenced by the choice of the dependent variable.
  • When two traits have similar phylogenetic signals, swapping variables yields similar $\hat{\lambda}$ values.
  • When traits have strikingly different phylogenetic signals, swapping variables leads to drastically different $\hat{\lambda}$ estimates and conflicting significance results. The model tends to estimate $\hat{\lambda}$ close to the phylogenetic signal of the dependent variable.
• Evaluation of Selection Criteria:
  • The authors compared the seven criteria against the "Golden Standard" in the 2,058 simulations where the two regression directions produced conflicting results.
  • Superior Criteria: Pagel's $\lambda$, Blomberg's $K$, and the estimated $\hat{\lambda}$ performed equally well and significantly outperformed the other four criteria (LLK, AIC, $R^2$, p-value).
  • Selecting the trait with the higher phylogenetic signal (higher $\lambda$ or $K$) as the dependent variable resulted in the correct model in 84.3% of the conflicting cases (1,734 out of 2,058).
  • Criteria based on model fit (LLK, AIC, $R^2$, p-value) performed no better than random selection in distinguishing the correct model.

4. Key Contributions

1. Identification of a Systematic Bias: The study definitively demonstrates that PGLS results under Pagel's $\lambda$ model are not symmetric; the choice of dependent variable is not arbitrary and can fundamentally alter biological conclusions.
2. Establishment of a Selection Rule: The paper proposes a robust, data-driven rule for variable selection: When causal direction is unknown, designate the trait with the stronger phylogenetic signal (higher Pagel's $\lambda$ or Blomberg's $K$) as the dependent variable.
3. Performance Benchmarking: By using simulated data with a known "golden standard," the authors quantified the accuracy of PGLS. They showed that while PGLS has an inherent accuracy limit (~84.57%), applying the phylogenetic signal criterion maximizes this accuracy (achieving ~82.55% overall accuracy in the study's context).
4. Clarification of Terminology: The authors emphasize that in PGLS, "dependent" and "independent" should not be interpreted literally as "cause" and "effect" when causality is unknown. Instead, the "dependent" variable is simply the response variable that yields the most statistically robust estimate of the correlation.

5. Significance

This paper addresses a critical methodological gap in phylogenetic comparative methods. For researchers studying trait evolution where causality is ambiguous (e.g., co-evolution of morphological traits, host-parasite interactions), the findings provide a practical protocol to avoid spurious conclusions.

• Practical Application: Researchers can calculate Pagel's $\lambda$ or Blomberg's $K$ for their traits before running PGLS. By assigning the trait with the higher signal as the dependent variable, they can ensure more reliable and reproducible inference without needing to run bidirectional analyses and guess which is correct.
• Theoretical Insight: The results highlight that PGLS parameter estimation (specifically $\lambda$) is sensitive to the regression direction when phylogenetic signals differ, suggesting that the model's assumption of residual covariance structure is better satisfied when the dependent variable drives the phylogenetic structure of the residuals.
• Future Directions: The study notes that while the proposed rule improves accuracy, PGLS still fails in ~15% of cases even with optimal variable selection, indicating a need for further methodological improvements in phylogenetic regression models.