I have some difficulties in understanding the notation of multilevel regression models. Let's consider, for example, a varying intercept and varying slope model with just one level-I predictor. We have:

Model of level I: $y_{ij}=\beta _{0j}+\beta _{1j}x_{ij}+\epsilon _{ij}$
Model of level II:
$\beta _{0j}=\gamma _{00}+u_{0j}$
$\beta _{1j}=\gamma _{10}+u_{1j}$
Combined model: $y_{ij}=\gamma _{00}+\gamma _{10}x_{ij}+u_{0j}+u_{1j}x_{ij}+\epsilon _{ij}$

$\gamma _{00},\gamma _{10}$ are fixed

$u_{0j}\sim N(0,\sigma _{\beta_{0}}^{2})$
$u_{1j}\sim N(0,\sigma _{\beta_{1}}^{2})$
$\epsilon _{ij} \sim N(0,\sigma _{\epsilon}^{2})$

$\epsilon _{ij} \bot u_{0j}$
$\epsilon _{ij} \bot u_{1j}$

$\beta _{0j}\sim N(\gamma _{00},\sigma _{\beta_{0}}^{2})$
$\beta _{1j}\sim N(\gamma _{10},\sigma _{\beta_{1}}^{2})$
$Cov(\beta _{0j},\beta _{1j})= \rho \sigma _{\beta_{0}}\sigma _{\beta_{1}}$

$Var(y_{ij} |x_{ij})=\sigma _{\beta_{0}}^{2}+x_{ij}^{2}\sigma _{\beta_{1}}^{2}+2\rho \sigma _{\beta_{0}}\sigma _{\beta_{1}}x_{ij}+\sigma _{\epsilon}^{2}$
$Var(y_{ij} |x_{ij})=\sigma _{y_{i}}^{2}$

Now my question is the following: which of these notations are correct and why?
1) $y_{ij}\sim N(\beta _{0j}+\beta _{1j}x_{ij}, \, \sigma _{\epsilon}^{2})$
2) $y_{ij}\sim N(\beta _{0j}+\beta _{1j}x_{ij}, \, \sigma _{y_{i}}^{2})$
3) $y_{ij} | \beta _{0j},\beta _{1j}\sim N(\beta _{0j}+\beta _{1j}x_{ij}, \, \sigma _{\epsilon}^{2})$
4) $y_{ij}\sim N(\gamma _{00}+\gamma _{10}x_{ij},\,\sigma _{\epsilon}^{2})$
5) $y_{ij}\sim N(\gamma _{00}+\gamma _{10}x_{ij},\,\sigma _{y_{i}}^{2})$
The fact is that I don't now if at the individual level I should consider $\beta _{0j}$ and $\beta _{1j}$ fixed or not. So I don't know if I should add to the within variance the between variance or not.
Thank you.


This is how I would denote this hierarchical model,

$$y_{ij} \,\vert\, \alpha_{j},\beta_{j} \sim \textsf{N}\left( \alpha_j+\beta_j x_{ij},\sigma^2 \right)$$ $$\alpha_j\sim \textsf{N}\left( \gamma,\tau^2 \right)$$ $$\beta_j \sim\textsf{N}\left( \kappa,\phi^2 \right)$$ where $\gamma,\tau,\kappa,\phi $ are given constants. I changed notation since I think this looks cleaner without all the subscripts and rids any idea of interdependence. If there is covariance, then this might work better, $$y_{ij}\,\vert\,\vec{\beta}\sim\textsf{N}\left( \vec{\beta}\mathbf{X},\sigma^2 \right)$$ $$\vec{\beta}\sim\textsf{MVN}\left( \vec{\gamma},\Sigma \right)$$

where $\vec{\beta},\vec{\gamma}$ are given $2\times1$ vectors, and $\Sigma$ is a given $2\times2$ matrix.

Unless you've fitted the model, you cannot hold the top level fixed since it is conditional on the lower levels.

  • $\begingroup$ +1, but why $\textsf N$ and not $\mathcal N$? $\endgroup$ – amoeba Apr 7 '15 at 14:52
  • $\begingroup$ $y_{ij} | \alpha_j,\beta_j\sim N(\alpha_j+\beta_j x_ij,\sigma^2)$. * Is that the model : $y_{ij}=\alpha_j+\beta_j x_ij+\epsilon_{ij}$ , where $\epsilon_{ij}\sim N(0,\sigma^2)$ ? * Why is $y_{ij} | \alpha_j,\beta_j$ instead of $y_{ij} | x_{ij} $ as i usually seen in linear regression model ? $\endgroup$ – ABC Apr 8 '15 at 1:23
  • $\begingroup$ $\alpha_j\sim \textsf{N}\left( \gamma,\tau^2 \right)$. * Is that the model: $\alpha_j=\gamma+u_{\alpha_j}$ , where $\gamma$ is fixed and $u_{\alpha_j}\sim N(0,\tau^2)$ ? $\endgroup$ – ABC Apr 8 '15 at 1:28
  • $\begingroup$ $\beta_j\sim \textsf{N}\left( \kappa,\phi^2 \right)$. * Is that the model: $\beta_j=\kappa+u_{\beta_j}$ , where $\kappa$ is fixed and $u_{\beta_j}\sim N(0,\phi^2)$ ? $\endgroup$ – ABC Apr 8 '15 at 1:31
  • $\begingroup$ $y_{ij}\,\vert\,\vec{\beta}\sim\textsf{N}\left( \vec{\beta}\mathbf{X},\sigma^2 \right)$ , so that number(3) in the question is correct , which is 3) $y_{ij} | \beta _{0j},\beta _{1j}\sim N(\beta _{0j}+\beta _{1j}x_{ij}, \, \sigma _{\epsilon}^{2})$. Isn't it ? $\endgroup$ – ABC Apr 8 '15 at 1:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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