Robustness of the linear mixed model to misspecified error distribution

Hélène Jacqmin‐Gadda, Solenne Sibillot, Cécile Proust‐Lima, Jean‐Michel Molina, Rodolphe ThiébautComputational 2006 · DOI

TL;DR

This paper investigates how robust linear mixed models (LMMs) are to violations of their assumptions about error distribution, specifically when errors are non-Gaussian, heteroscedastic, or correlated. The study finds that LMMs are surprisingly robust to non-Gaussian and heteroscedastic errors, but inference can be impaired when errors are correlated, especially for models with only a random intercept.

What Problem Does This Paper Address?

Linear mixed models (LMMs) are widely used for analyzing longitudinal data, where repeated measurements are taken on the same subjects over time. A key assumption in LMMs is that the errors (the part of the outcome not explained by the model) are normally distributed, independent, and have constant variance (homoscedastic). However, in real-world data, these assumptions are often violated. For instance, errors might be skewed, have different variances for different subjects or over time, or be correlated due to the repeated measurements. While LMMs are known to be somewhat robust to violations of the random effects distribution, the robustness of fixed effects inference to misspecified error distributions (especially heteroscedasticity and correlation) is less understood. This paper aims to fill this gap by systematically evaluating the impact of these misspecifications on LMM inference through simulation studies.

Methods

The authors conducted a comprehensive simulation study to assess the robustness of LMMs. They generated data under various scenarios where the error distribution was misspecified in different ways:

  • Heteroscedastic errors: This included cases where the error variance depended on the conditional expectation of the outcome, on time, or on a covariate.
  • Correlated errors: An autoregressive structure was used to simulate correlated errors.
  • Non-Gaussian errors: Various non-Gaussian distributions were employed, including bimodal symmetric mixtures, asymmetric mixtures, Gamma, Exponential, and g-and-h distributions.

They compared two types of LMMs: a random intercept (RI) model and a random intercept and slope (RIS) model. The primary metric for evaluating robustness was the coverage rate of 95% confidence intervals (CIs) for the fixed effects. Ideally, these coverage rates should be close to the nominal 95%. They also examined the ALBI clinical trial data to illustrate potential model misspecifications and to inform the parameter choices for the simulation study.

Datasets

The ALBI clinical trial data was used to illustrate model misspecifications and to guide the parameter choices for the simulation study. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Surprising Robustness to Non-Gaussian and Heteroscedastic Errors: The study found that LMMs, particularly the more complex random intercept and slope model, are remarkably robust to non-Gaussian and heteroscedastic errors. This suggests that transformations of the outcome variable to achieve normality or constant variance might not always be necessary, simplifying interpretation.
  • Correlation is the Main Culprit: The most significant impact on inference was observed when errors were correlated. The random intercept model showed poor performance in these scenarios, with coverage rates dropping substantially.
  • Random Slope Matters: The inclusion of a random slope in addition to a random intercept significantly improved the robustness of the model, especially when errors were correlated. This highlights the importance of adequately modeling the individual variability in change over time.
  • Covariate-Dependent Variance is Tricky: Heteroscedasticity that depends on a covariate, particularly when interacting with time, can bias the variance estimates of the slope parameters. However, the inference for the treatment effect (difference in slopes) remained robust.

Key Results

  • Heteroscedasticity: For most heteroscedasticity scenarios, coverage rates remained close to 95%, especially for the RIS model. Even when variance increased with time or depended on a covariate, the RIS model showed good robustness.
  • Correlated Errors: For correlated errors, the RIS model generally maintained coverage rates above 91%, while the RI model's performance was severely compromised, with some coverage rates dropping below 80%.
  • Non-Gaussian Errors: Even with severely skewed and heavy-tailed error distributions, coverage rates were generally good (above 92.7%) and parameter estimates remained unbiased.

Limitations and Future Work

  • Focus on Fixed Effects: The study primarily focused on the robustness of inference for fixed effects. While it was noted that variance component estimates and random effects can be biased when the covariance structure is misspecified, a detailed investigation of their robustness was not the main focus.
  • Specific Correlation Structure: The simulation for correlated errors only considered an autoregressive structure. Other types of correlation structures might yield different results.
  • Moderate Sample Sizes: The simulations were conducted with moderate sample sizes (N=50 and N=200). The behavior of LMMs with very small or very large sample sizes under these misspecifications was not explored.

The authors suggest that while graphical evaluation of residuals is useful, it may not always distinguish between misspecification of the error structure and the fixed part of the model. They recommend evaluating the sensitivity of results to various assumptions, which can be challenging with standard software. Future work could explore a wider range of correlation structures and investigate the robustness of random effects estimates and individual predictions more thoroughly.

Summary updated

BibTeX
@inproceedings{jacqmingadda2006robustness,
  author = {Hélène Jacqmin‐Gadda and Solenne Sibillot and Cécile Proust‐Lima and Jean‐Michel Molina and Rodolphe Thiébaut},
  title = {{Robustness of the linear mixed model to misspecified error distribution}},
  booktitle = {Computational},
  year = {2006},
  doi = {10.1016/j.csda.2006.05.021},
}

Loading related content...

Figure