A joint mean-correlation modeling approach for longitudinal zero-inflated count data (Q2180256): Difference between revisions

Let \((Y_{i1},Y_{i2},\dots, Y_{im_i})^{\intercal}\), \(i\in\mathbb{N}\), \(\sup_{i\in\mathbb{N}} m_i<\infty\) be a system of longitudinal measurements. Let \[ \mathbb{P}(Y_{ij}=y)=\begin{cases} p_{ij}+(1-p_{ij})\mathbb{P}(W_{ij}=0)\ \text{ if }\ y=0,\\ (1-p_{ij})\mathbb{P}(W_{ij}=y)\ \text{ if } \ y\geq 1,\end{cases} \] where \(W_{ij}\) follows a negative binomial distribution \[ \mathbb{P}(W_{ij}=w)=\frac{\Gamma(w+1/\tau)}{w!\Gamma(1/\tau)}\left(\frac{1}{1+\tau\lambda_{ij}}\right)^{1/\tau}\left(\frac{\tau\lambda_{ij}}{1+\tau\lambda_{ij}}\right)^w,\ w\in\mathbb{N}_0 \] with a positive shape parameter \( \tau\). In addition, suppose that the distribution function of \((Y_{i1},Y_{i2},\dots, Y_{im_i})\) follows the Gausian copula representation \[ \mathbb{P}(Y_{i1}\leq y_{i1},Y_{i2}\leq y_{i2},\dots, Y_{im_i}\leq y_{im_i})=\Phi_{m_i}((z_{i1},z_{i2},\dots, z_{im_i}),R_i), \] where \(z_{ij}=\Phi_1^{-1}(\mathbb{P}(Y_{ij}\leq y_{ij})\) and \(R_i\) is the correlation matrix. The authors of the paper develop the maximum likelihood estimator for vector \(\boldsymbol{\theta}=(\boldsymbol{\beta}^{\intercal},\boldsymbol{\gamma}^{\intercal},\boldsymbol{\alpha}^{\intercal}, \tau)^{\intercal} \) by supposing that \(\log(\lambda_{ij})\) can be expressed by \(\boldsymbol{\beta}\), \(\mathrm{logit}(p_{ij})\) depends on \(\boldsymbol{\gamma}\), and correlation matrices \(R_i\) can be expressed by \(\boldsymbol{\alpha}\).