Presume we would like to model the dynamics of two related variables as a bi-variate Normal, while also accounting for the effect of other covariates (via regression).

E.g. I would like to model heights of various pairs of siblings, while also accounting for their gender. Then, let

  • $Y_1, Y_2$ - heights of two siblings (at an adult age, when they can't grow taller)
  • $x_1, x_2$ - their respective genders

I would like to both account for dependence between $Y_1$ and $Y_2$, while including $x$ as a regression covariate. Assuming normality for the time-being, there are two approaches of capturing that dependence I'm thinking about:

  1. Introducing appropriate covariance structure inside the bivariate normal framework: \begin{equation} \begin{pmatrix} Y_{1,i} \\ Y_{2,i} \end{pmatrix} \sim N(\begin{pmatrix} \alpha + \beta x_{1,i} \\ \alpha + \beta x_{2,i} \end{pmatrix}, \sigma^2 \begin{pmatrix} 1 & \rho \\ \rho & 1\end{pmatrix}), \ \ i =1,\dots, n , \end{equation} $i$ indicates $i^{th}$ pair of siblings, $\beta$ - gender effect coefficient (same for both siblings)

  2. Just explicitly include sibling's height as a covariate with a coefficient, without having to specify correlation parameter in the covariance structure (hence, no $\rho$), as follows: $$ \begin{pmatrix} Y_{1,i} \\ Y_{2,i} \end{pmatrix} \sim N(\begin{pmatrix} \alpha + \beta x_{1,i} + \gamma y_{2,i} \\ \alpha + \beta x_{2,i} + \gamma y_{1,i} \end{pmatrix}, \sigma^2 \begin{pmatrix} 1 & 0 \\ 0 & 1\end{pmatrix}), \ \ i =1,\dots, n,$$ where $y_{1,k}, y_{2,k}$ are the observed values of a "sibling's height" variable, which we explicitly use as fixed covariates (!! see Question #2 below !!) on the right-hand side.


  1. Could someone please explain both theoretical and practical differences between these two attempts to capture bi-variate dependence, while including covariate info? Most importantly: how would my interpretations of covariate effect $\beta$ differ (e.g. how would dependence between $Y_1 \& Y_2$ play into interpretation of gender effect $\beta$ in approach #1)?

  2. Is approach # 2 even legitimate, given that I'm using inherently random responses $Y_1,Y_2$, as fixed covariates on the right-hand side? What issues/peculiarities does that aspect introduce theoretically?


Your Answer

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

Browse other questions tagged or ask your own question.