I am confused by lmer's formula syntax for mixed models.

I found a very useful "cheat sheet" in an answer to a related question.

Unfortunately, there is no example of a fairly common class of model, the intercept/slope as outcomes model, and I do not understand how it might be formulated.

Specifically, I do not know how to formulate a model such as this one:

Level 1

$Y_{ij} = β_{0j} + β_{1j}(X_{ij}) + r_{ij}$

Level 2

$β_{0j} = γ_{00} + γ_{01}W_j + u_{0j}$

$β_{1j} = γ_{10} + γ_{11}W_j + u_{1j}$

where i ranges over individuals, j over groups, and $W_j$ is a group-level predictor of the level 1 slope and intercept. Combined this creates a cross-level interaction:

$Y_{ij} = (γ_{00} + γ_{01}W_j + u_{0j}) + (γ_{10}+γ_{11}W_j + u_{1j})X_{ij} + r_{ij}$

By pattern matching on the answer to this question the correct formulation might be something like:

Y ~ 1 + X + W:X + (1 + X | Group)

but I really don't understand why. Would someone please explain the correct way to model this, and in particular how it might inform me about the semantics of lme4's formula?


1 Answer 1


It's just a little algebraic manipulation. Take the fixed effects as an example: These are two different ways of writing the same fixed effects model:

$$ Y \sim \mathcal{N}(\mathbf{X}\beta, \sigma^2)$$


$$ Y = \mathbf{X}\beta + \epsilon; \quad \epsilon \sim \mathcal{N}(0, \sigma^2).$$

They are equivalent parameterizations that lead to OLS as the MLE for $\beta$.

When random effects have a non-zero mean, it is the same thing to write them as a sum of fixed and random effects with a zero mean.

The formula object you write does not give rise to the random slopes/random intercepts growth model in your earlier expressions. X + X:W is a 2 degree of freedom expression which has no fixed effect for $W$, yet your 2 level model has a random intercept which varies in $W$. Rather use this formula:

~X*W + (1+X|Group)

If you go through the algebra you will find:

  • The coefficient (Intercept) is an estimate of $\gamma_{00}$
  • the coefficient to $X$ is an estimate of $\gamma_{10}$
  • the coefficient to $W$ is an estimate of $\gamma_{01}$
  • The coefficient to $X:W$ is an estimate of $\gamma_{11}$

for the fixed effects. The random effects are subtler. The variance in the group level controls for the between-group confounders $W$ and so is lower than the actual unconditional group level variance. That is the point of adjusting for these variables anyway.

  • The random effect for Intercept is an estimate of the variance of $u_{0j}$
  • The random effect for $X$ is an estimate of the variance of $u_{1j}$
  • The residual variance is an estimate of the variance of $r_{ij}$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.