# Finding maximum likelihood

This is a model that is used to model soccer scores, so $i$ and $j$ are, respectively, home and away teams. Random variables $(x,y)$ are the goals scored by the home and away teams, respectively. Parameter $\lambda$ is a known mean goals scored by the home team and $\mu$ is the mean goals scored by the away team. I have managed to fix all the other parameters except for $\rho$, which I have to estimate via MLE.

$$Pr(X_{i,j}=x, Y_{i,j}=y)=\tau_{\lambda, \mu}(x,y)\frac{\lambda^x \text{exp}(-\lambda)}{x!}\frac{\mu^y\text{exp}(-\mu)}{y!}$$ where $$\lambda=\alpha_{i}\beta_{j}\gamma$$ $$\mu=\alpha_{j}\beta_{i}$$ and $$\tau_{\lambda,\mu}(x,y)=\left\{\begin{array}{cc} 1-\lambda\mu\rho &\text{if x=y=0,} \\ 1+\lambda\rho &\text{if x=0,y=1,}\\ 1+\mu\rho &\text{if x=1,y=0,}\\ 1-\rho &\text{if x=y=1,}\\ 1 &\text{otherwise}\end{array} \right.$$

Based on the above equations, all the parameters $(\lambda, \mu, \alpha, \beta, \gamma)$ are known constants.

So, now, the problem that I am having is that I have no clue on how to estimate $\rho$ using the maximum likelihood function since a piece-wise equation is involved.

Also, it will be great if anyone can do this using R.

• it is actually a Poisson distribution attached at the back, maybe that will clarify the problem. Sorry. – SooBin Oct 31 '13 at 16:07
• This is a model that i am using to model soccer scores, so i and j are home and away team that are playing respectively. x and y are the goals scored by the home and away team respectively. I have managed to fix all the other parameters except for p which i have to estimate it via MLE. – SooBin Oct 31 '13 at 16:12
• yeah, because like what you said, the probability distribution will be strange if i dont add that in. I was thinking that it was not necessary since it will be ignored when differentiating as it is independent of p. Sorry Zen. – SooBin Oct 31 '13 at 16:16
• I've extensively edited your question for syntax and to add LATEX rendering for your mathematics. Please check that I have not accidentally introduced any errors. – Sycorax Nov 1 '13 at 2:45
• I don't see anything "piecewise" about the $\tau$ at all: $\tau_{\lambda, \mu}(x,y)=\rho (-\lambda \mu (1-x) (1-y)+\lambda (1-x) y+\mu x (1-y))+1.$ Besides, $x$ and $y$ are data, not parameters, so all that matters is how $\tau$ depends on the parameters $\rho, \mu, \lambda$, and it's a simple (multilinear) polynomial in them (and linear in $\rho$). – whuber Nov 1 '13 at 20:03

If I understand it, your data is $\{(i_m,j_m,x_m,y_m) : m=1,\dots,n\}$, in which, for the $m$-th match, $i_m$ is the index of the home team, $j_m$ is the index of the away team ($i_m\neq j_m$), $x_m$ is the number of goals scored by the home team, and $y_m$ is the number of goals scored by the away team.
If we have $t$ teams disputing the $n$ matches, the likelihood $L(\rho,\gamma,\alpha_i,\beta_i, i=1,\dots,t)$ is proportional to $$\prod_{m=1}^n \tau_{\lambda_m,\mu_m}(x_m,y_m)\,\lambda_m^{x_m} \,\mu_m^{y_m}\,e^{-(\lambda_m+\mu_m)} \, ,$$ in which $\lambda_m = \alpha_{i_m} \beta_{j_m}\gamma$ and $\mu_m = \alpha_{i_m} \beta_{j_m}$. This likelihood is a function of $2(t+1)$ parameters. You will probably need to implement some conjugate gradient method to solve this constrained optimization problem. I wonder what happens if we put some priors on the parameters and MCMC the posterior with a random walk Metropolis. You said that you know the values of the parameters, except for $\rho$ (and I wonder how...). Hence, your first step is to write an R function that computes the product above for each value of $\rho$. Then, plot it and use the optimizer to see what happens.