Let us assume a mixed logit model with a binary dependent variable $y_{i, t } $ that is explained by a fixed effect matrix X and a simple random intercept for each individual $i$

$y_{i,t}^* = x_{i,t}'\beta +\alpha_i + \epsilon_{i,t} \hspace{35pt} \alpha \sim N(0, \sigma_{\alpha}) \hspace{17pt} where \hspace{17pt}\begin{array}{ll} 1 & \widetilde{y}_{ij,t}^{*} \geq 0 \\ 0 & \, \widetilde{y}_{ij,t}^{*} < 0 \\ \end{array} $

As far as I know, when we want to estimate the fixed and random effects via maximum likelihood, we have to integrate out the random effects over every individual so that the log likelihood becomes

$ LL(\beta) = \sum_{i=1}^N \log \left(\int_{-\infty}^{\infty} \prod_{t=1}^{T_i}\left(\left(\frac{e^{x_{i,t}'\beta +\alpha_i}}{1 + e^{x_{i,t}'\beta +\alpha_i}}\right)^{y_{i,t}} \left(\frac{1}{1 + e^{x_{i,t}'\beta +\alpha_i}}\right)^{1 - y_{i,t}} \frac{1}{\sqrt{2\pi\sigma_\alpha^2}}e^{-\frac{\alpha_i^2}{2\sigma_\alpha^2}}\right) d \alpha_i \right)$

Let us now assume a time-dynamic Bradley-Terry model. The probability that player $i$ wins a match against player $j$ at time $t$ is then given by

$p_{ij,t} = .P(y_{ij,t} = 1| \lambda_{i,t},\lambda_{j,t}) = \frac{e^{\lambda_{i,t} - \lambda_{j,t}}}{1 + e^{\lambda_{i,t} - \lambda_{j,t}}} \label{eq:second_BTm}$


$\lambda_{i,t} = x_{i,t}'\beta$

can be though of the skill level of player $i$ at time $t$, which is explained by a player-specific covariate vector $\beta$.

While this is the common formulation of the Bradley-Terry model, it is also possible to write this in the latent variable representation (as in the case of the first model)

$\widetilde{y}_{ij,t} = \left\{ \begin{array}{ll} 1 & \widetilde{y}_{ij,t}^{*} \geq 0 \\ 0 & \, \widetilde{y}_{ij,t}^{*} < 0 \\ \end{array} \right. \\ \widetilde{y}_{ij,t}^{*} = \sum_{l=1}^{k}\beta_l x_{il,t} - \sum_{l=1}^{k}\beta_l x_{jl,t} + \epsilon_{i,t} - \epsilon_{j,t} \hspace{20pt} \epsilon_{q,t} \sim Logistic(0, \frac{\pi^2}{3}) \hspace{20pt} q = i,j$

Since the model is time dynamic, the log likelihood is given by

$ LL(\Theta) = \sum_{t = 1}^{T}\sum_{(i,j)\in I_t}\log \left(y_{ij,t}p_{ij,t} + (1 - y_{ij,t})(1 - p_{ij,t}) \right) \\ I_t = \{(i,j): \text{a match between player} \hspace{3pt} i \hspace{3pt} \text{and} \hspace{3pt} j \hspace{3pt} \text{is played at time} \hspace{3pt} t \} $

The reason the likelihood needs to be formulated in a time-dynamic fashion is that some of the covariate values are constantly updated in random-walk process, as proposed by Gorgi, Koopman and Lit (2018): https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3110555

Let us now add a player-specific time-constant random intercept $\alpha_i$ to account for player specific heterogeneity, so that we obtain

$\lambda_{i,t} = x_{i,t}'\beta +\alpha_i \hspace{35pt} \alpha \sim N(0, \sigma_{\alpha})$

This can also be written in terms of the latent variable representation

$\widetilde{y}_{ij,t} = \left\{ \begin{array}{ll} 1 & \widetilde{y}_{ij,t}^{*} \geq 0 \\ 0 & \, \widetilde{y}_{ij,t}^{*} < 0 \\ \end{array} \right. \\ \widetilde{y}_{ij,t}^{*} = \sum_{l=1}^{k}\beta_l x_{il,t} - \sum_{l=1}^{k}\beta_l x_{jl,t} + u_{i,t} - u_{j,t} \hspace{35pt} u_{q,t} = \alpha_q + \epsilon_{q,t} \hspace{35pt} \\ \epsilon_{q,t} \sim Logistic(0, \frac{\pi^2}{3}) \hspace{35pt} q = i,j$

What form does the log-likelihood now take? Do we still have to integrate out the random effects over every player $i$ or rather over every player pair $i,j$?

Thank you very much in advance!


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.