I'm working on some economic stuff and the objective is to conduct a panel data analysis. I assumed the following data-generating process: \begin{equation} y_{it} - y_{i,t-1} = \eta z_{i,t-1} + \gamma_i x_{i,t-1} + \beta_i y_{i,t-1} + \delta_t + \alpha_i + u_{it} \end{equation} where $y_{it} \equiv log (Y_{it})$. Thus, $x_{it}$ and $z_{i,t}$ are two vectors of observable predictors, $\delta_t$ denotes time specific fixed-effects, $\alpha_i$ denotes individual unobserved heterogeneity, with $i=1,2,\ldots,N$. I must allow for arbitrary correlation between $\alpha_i$ and my predictors.
A general fixed-effects estimator with individual specific slopes is not feasible in this case because having the lagged dependent variable among the predictors would violate the strict exogeneity assumption. I need an instrument for $y_{i,t-1}$. The GMM should be the natural solution but I cannot find an estimator for dynamic panels allowing for heterogeneity in the slopes, i.e., $\beta$ and $\gamma$ vary across cross-sectional units $i$. Any suggestion?