A lightning introduction to linear mixed models for designed experiments: part 3

Welcome to the third part of our four-part blog series aimed at providing you with a brief introduction to linear mixed models. This part focuses on how to assess the importance of fixed model terms.

The fixed model

Estimates of fixed effects (BLUEs)

The estimates of the fixed effects used with REML are often called BLUEs, which is an acronym for Best Linear Unbiased Estimates. Being unbiased means that the expected value of the estimator is equal to the true parameter value. In this context, ‘best’ means that these estimates have minimum variance within the class of unbiased estimators, conditional on the estimated variance components.

Assessing fixed model terms

We usually wish to investigate the contribution of individual terms within the fixed model in explaining patterns of response, and this is achieved using Wald tests. If the set of fixed terms is non-orthogonal (i.e., independent from each other), it is necessary to consider both incremental and marginal forms of these statistics; incremental statistics reflect the change in fit on sequentially adding individual terms into the fixed model, and marginal statistics reflect the change on dropping individual terms from the full model. On dropping terms, we respect marginality, for example, we do not drop a term if it is involved in any higher-order interactions.

For a single effect, the Wald statistic is equivalent to the square of the t-statistic. It is calculated by dividing the estimated effect by its standard error. For a set of effects corresponding to a model term, a Wald statistic can be thought of as the sum of squares of the effects weighted by their variance-covariance matrix. The asymptotic reference distribution for the Wald statistic is a chi-square distribution with degrees of freedom (df) equal to the df of the term. This distribution ignores variation associated with estimation of the variance components. Hence, some caution is therefore required in the use of Wald tests, which tend to be overly optimistic, i.e., to give false positive results more often than would be expected. However, the approximation improves greatly as the residual df associated with the estimated variance-covariance matrix increases.

To avoid this problem, various methods exist to convert, or adjust, Wald statistics into a form that can be comparable to an F distribution, where the denominator df gives a measure of uncertainty in the variance estimation. The most popular method was introduced by Kenward and Roger in 1997. This method divides the Wald statistic by the df of the treatment term to give the form of an F statistic. This statistic is compared to an F distribution with numerator df equal to those of the treatment term, and the method provides an estimate for the denominator df. For balanced designs, F tests based on the Kenward-Roger method will be exactly the same as F tests based on the variance ratios in an ANOVA table.

The table of marginal Wald or approximate F-tests can be used to implement a process of backwards selection for the fixed model, refitting the linear mixed model each time a term is dropped.

The next blog in this series will discuss how to check the assumptions of the linear mixed model and assess goodness of fit.

Reference

Kenward, M.G., Roger, J.H. (1997). Small sample inference for fixed effects from restricted maximum likelihood. Biometrics 53:983-997.

Picture of The VSNi Team
The VSNi Team

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *

Related articles

How much of the variation in your field trial comes from the treatments, and how much comes from...
What should you do when count data are more variable than your model expects? Vanessa Cave and David...
Precision breeding can deliver scientifically impressive traits, but practical value depends on more than technical success. Owen McIntosh...