In GLM, \(g(\mu_i)=\sum_j \beta_jx_{ij}\) and likelihood equations are \[ \sum_i \frac{(y_i-\mu_i)x_{ij}}{\text{Var}(Y_i)}\frac{d\mu_i}{d\eta_i}=0,\mbox{ }\mbox{ } j=0,1,\ldots,p. \]

Also, note that ML estimates depend on the distribution of \(Y_i\) only through \(\mu_i\) and \(\text{Var}(Y_i)=V(\mu_i)\).

  

  

Definition

A function \(h(Y,\beta)\) is an unbiased estimating function if \[ E[h(Y;\beta)]=0 \mbox{ }\mbox{ for all }\beta. \]

  

Remark(Quasi likelihood approach)
  1. Use model \(g(\mu_i)=\sum_j \beta_jx_{ij}\) and variance function \(V(\mu_i)\) but do not assume distribution for \(Y_i\).

  2. To allow overdispersion, take \(V(\mu_i)=\phi V^*(\mu_i)\), where \(V^*(\mu_i)\) is variance function for common model such as \(V^*(\mu_i)=\mu_i\) for count data.

  3. Use estimating equations \(S(\beta)=0\) even if they do not correspont to likelihood equations for distribution in exponential family. Here, \(S(\beta)\) is the quasi score functions such that \[ S(\beta)=(S_0(\beta), S_1(\beta)\ldots,S_p(\beta)), \] where \[ S_j(\beta)=\sum_i \frac{(y_i-\mu_i)x_{ij}}{\phi V^*(\mu)}\frac{d\mu_i}{d\eta_i}\stackrel{\text{set}}{=}0,\mbox{ }\mbox{ } j=0,1,\ldots,p. \]

  

Remark
  1. Quasi-score function \(S(\beta)\) is regarded as derivative of quasi log-likelihood function (need not correspond to a proper distribution).

  2. QL estimators have similar properties as ML estimators (McCuallagh, 1983). Solution \(\hat\beta\) of \(S(\beta)=0\) is asymptotically normal with covariance matrix \(V=(X'WX)^{-1}\), where \(W\) is diagonal with \(w_i=\left(\frac{d\mu_i}{d\eta_i}\right)^2/\text{Var}(Y_i)\).

  3. If \(g(\mu_i)=\sum_j \beta_jx_{ij}\) is correct, then \(\hat\beta\rightarrow\beta\), regardless of whether variance function is correctly specified (White, 1984).

  4. For \(\text{Var}(Y_i)=\phi V^*(\mu_i),\) \(\phi\) drops out of estimating equations, usually estimated by approximate method of moments.

    For example, note that \[ X^2=\sum_i\frac{(Y_i-\hat\mu_i)^2}{\hat{\text{Var}}(Y_i)}= \sum_i\frac{(Y_i-\hat\mu_i)^2}{\phi V^*(\hat\mu_i)}\stackrel{\text{approx}}{\sim}\chi^2_{N-p}, \]

    where \(X^2\) is Pearson statistic. Then \[ E\left\{ \sum_i\frac{(Y_i-\hat\mu_i)^2}{\phi V^*(\hat\mu_i)} \right\}\approx N-p \\ \implies E\left\{ \sum_i\frac{(Y_i-\hat\mu_i)^2}{ V^*(\hat\mu_i)} \right\}/(N-P) \approx \phi. \\ \] Thus, we have \[ \hat\phi= \frac{X^2}{N-p}= \sum_i\frac{(Y_i-\hat\mu_i)^2}{ V^*(\hat\mu_i)}/(N-P). \]

    The \(X^2\) is Pearson statistic is used for testing fit for GLM with variance function \(V^*(\mu_i)\)(Wedderbern, 1974).

  

Example
  1. Quasi-likelihood for counts:

    For poisson GLM, \(V^*(\mu_i)=\mu_i\) does not allow overdispersion. In QL alternative, \(V(\mu_i)=\phi\mu_i\), where typically expect \(\phi>1\). Then, we have \[ S_j(\beta)= \sum_i \frac{(y_i-\mu_i)x_{ij}}{\phi \mu_i}\frac{d\mu_i}{d\eta_i}=0, \mbox{ }j=0,1,\ldots,p. \]

    Note that \(\hat\beta\) is the same as \(\hat\beta\) for poisson GLM: \(\text{Cov}(\hat\beta)= (X'WX)^{-1}\) with \(w_i=\left(\frac{d\mu_i}{d\eta_i}\right)^2/\text{Var}(Y_i)\). For log link, \(w_i=\left(\frac{d\mu_i}{d\eta_i}\right)^2/\text{Var}(Y_i)=b''(\theta_i)^2/\phi\mu_i=\mu_i/\phi\).

    So, \(\text{Cov}(\hat\beta)=\phi\times\text{Cov}\) for poisson GLM. Also,

    \[ \hat\phi=\frac{X^2}{N-p}, \] where \(X^2=\sum_i \frac{(y_i-\hat\mu_i)^2}{\hat\mu_i}.\)

  

  1. Quasi-likelihood for Binomial overdispersion: Let \(n_iY_i\sim \text{Bin}(n_i, \pi_i)\), where \[ \text{Var}(Y_i)=\phi V(\mu_i)= \frac{\pi_i(1-\pi_i)}{n_i}, \mbox{ } \phi=1. \] To allow overdispersion, take \(V(\mu_i)=V(\pi_i)=\phi \frac{\pi_i(1-\pi_i)}{n_i}\). Then, we can derive \(\text{Cov}(\hat\beta)=\phi(X'WX)^{-1}\) where \(W=\text{diag}(w_1,\ldots,w_n)\) with \(w_i=\left(\frac{d\mu_i}{d\eta_i}\right)^2/V(\mu_i)\).

  

  

정리

  

back