Random Slopes With the Identity Link

Why the Identity Link Makes Slopes Direct

A random slope allows the association between a predictor and the conditional response mean to differ across clusters. With the identity link, the linear predictor is itself the conditional mean, so a slope has a direct response-scale interpretation: it describes the change in conditional mean response for a one-unit increase in the predictor. This tutorial develops that interpretation for fixed and cluster-specific slopes.

The interpretation is conditional: it holds for the specified cluster’s random effects and for the other covariates in the model held fixed. A slope describes a model-based association, not necessarily a causal effect. In particular, an identity-link slope is not an odds ratio or rate ratio; it is an additive change in the response mean.

Here, \(\beta_0\) is the fixed intercept and \(b_{0j}\) is cluster \(j\)'s deviation from it. The random intercept and slope are commonly modeled jointly as mean-zero random effects, such as \((b_{0j},b_{1j})^\mathsf{T}\sim N(\mathbf{0},\Sigma_b)\), independently across clusters. The mean-zero assumption makes \(\beta_1\) the mean of the cluster-specific slopes in the random-effects population. A particular cluster's slope can be above or below that mean.

The conditional response distribution must also be appropriate for the outcome, and the identity-link conditional mean must be valid over the predictor range being used. For a Gaussian response model, a common specification is \(Y_{ij}\mid b_{0j},b_{1j},x_{ij}\sim N(\mu_{ij},\sigma^2)\), with observations conditionally independent given the random effects and covariates. Other response distributions require their own assumptions. The linear mean can produce values outside a response's allowable range, so an identity link is not automatically suitable for every outcome.

From the Model to a Change in Mean

At a fixed value of the random effects, the model can be written as a line in \(x_{ij}\). Its intercept for cluster \(j\) is \(\beta_0+b_{0j}\), and its slope is \(\beta_1+b_{1j}\). To see why the slope is a response-scale change, compare the conditional means at two predictor values \(x_a\) and \(x_b\), keeping the cluster and all other modeled quantities unchanged.

The intercept terms cancel. For a one-unit increase, \(x_b-x_a=1\), so the change in conditional mean is exactly \(\beta_1+b_{1j}\). For a change of \(k\) predictor units, it is \(k(\beta_1+b_{1j})\). The fixed slope \(\beta_1\) gives the corresponding one-unit change for the fixed-effects line, which is the mean cluster slope when the random slopes have mean zero.

The units matter. If \(Y\) is measured in millimetres and \(x\) in years, the slope is millimetres per year. A negative slope means the conditional mean decreases as \(x\) increases; it does not mean the response itself must be negative. The intercept is the conditional mean at \(x=0\), so its substantive interpretation depends on whether zero is meaningful and within a useful range of the predictor.

Reading Fixed and Cluster-Specific Slopes

The fixed slope and cluster-specific slope answer related but distinct questions. The fixed slope describes the change associated with one predictor unit on the fixed-effects line. The cluster-specific slope adds that cluster's random deviation. It is therefore incorrect to interpret \(\beta_1\) as the slope for every cluster unless the random slope deviation for each cluster is zero. Conversely, a cluster-specific slope is not a separate fixed coefficient: it combines the common fixed slope with the cluster's deviation.

For example, if \(\beta_1=-0.40\) score points per week and a cluster has \(b_{1j}=0.15\) score points per week, then that cluster's slope is \(-0.25\) score points per week. The positive random deviation makes the slope less negative; it does not make it positive. Always add the signed quantities, retaining their units.

When random effects have mean zero and their distribution does not depend on the predictor, averaging the identity-link conditional mean over the random effects gives the fixed-effects mean: $$ E_{b}\{\mu_{ij}\mid x_{ij}\}=\beta_0+\beta_1x_{ij}. $$ This follows because \(E(b_{0j})=E(b_{1j})=0\), so \(E(b_{0j}+b_{1j}x_{ij})=0\) at a fixed \(x_{ij}\). Thus the fixed slope is also the population-average slope under these conditions. This equality concerns the mean response; it does not imply that every cluster has that slope or that the conditional response distribution is the same across clusters.

Assumptions, Diagnostics, and Common Pitfalls

The slope calculations are algebraic consequences of the identity-link mean model, but using them as scientific descriptions depends on the model being credible. In particular, the conditional mean should be approximately linear in the predictor over the range interpreted, the random-effects specification should represent relevant cluster differences, and the response distribution and dependence assumptions should be suitable for the data. If the model includes additional fixed or random terms, the interpretation must hold those terms at specified values; an omitted interaction, for example, can make a single slope an inadequate summary.

Useful checks include plotting observed responses and fitted conditional means against the predictor by cluster, checking whether the relationship is systematically curved, and examining residuals for patterns or changing spread. Compare observed cluster trajectories with model-based trajectories where appropriate. A pattern of curvature, substantial systematic departures, or implausible predicted means suggests the linear identity-link mean may not describe the data well. Check also whether conclusions rely on extrapolation beyond observed predictor values.

The identity link imposes no automatic restriction that fitted means be positive or lie within a bounded scale. If the outcome is nonnegative, binary, or otherwise constrained, a fitted conditional mean outside its valid range signals a serious problem for that use of the model. A random slope can make predictions especially problematic at extreme predictor values. Do not treat a plausible fitted mean within the observed range as proof that all assumptions hold; diagnostics and knowledge of the measurement scale both matter.

• Common mistake: treating the fixed slope as every cluster's slope. The incorrect reasoning is that \(\beta_1\) is the slope for all observations. The correction is to add the cluster's deviation: \(\beta_1+b_{1j}\).
• Common mistake: interpreting the slope as a multiplicative effect. The correction is that an identity-link slope is an additive change in conditional mean, in response units per predictor unit.
• Common mistake: ignoring the size of a predictor change. A slope is a one-unit change; for a \(k\)-unit increase, multiply it by \(k\), provided the linear mean remains appropriate across that interval.
• Common mistake: confusing fitted means with observed outcomes. A slope describes how the conditional mean changes. Individual observations can differ from that mean because of response variation.