Hypothesis Testing in GLMs

Likelihood-Based Testing

In a simple Generalised Linear Model (GLM) with one parameter, there are three common methods for testing our parameter using the likelihood function.

Suppose we want to test:

\[H_0: \theta = \theta_0\]

The three likelihood-based methods are:

  • Wald Test
  • Score Test
  • Likelihood Ratio Test

In this post, we will focus on the Wald Test and Score Test.

Simple GLM Model - With One Parameter

Wald Test

The Wald Test is based on the distance between our estimated parameter $\hat{\theta}$ and the hypothesised value $\theta_0$.

The Wald statistic is defined as:

\[W = \frac{(\hat{\theta} - \theta_0)^2} {\widehat{\operatorname{Var}}(\hat{\theta})}\]

When $H_0$ is true,

\[W \sim \chi^2_1 \qquad \text{as } n \rightarrow \infty.\]

Notice that:

  • $\sqrt{W} = Z$, thus $Z \sim \mathcal{N}(0,1)$, so $Z$ can be used as the test statistic.
  • A small $W$ means that the distance between $\hat{\theta}$ and $\theta_0$ is small relative to the uncertainty in $\hat{\theta}$. Therefore, we have less evidence against $H_0$ for a given significance level $\alpha$.
  • A large $W$ means that $\hat{\theta}$ is far from $\theta_0$ relative to its uncertainty, providing stronger evidence against $H_0$.

Generally, in a simple GLM we may be interested in testing a coefficient using:

\[H_0: \beta_1 = 0\]

against an alternative such as:

\[H_a: \beta_1 > 0.\]

Proof

We know that the second-order Taylor series expansion around $x=a$ is:

\[f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2}f''(a)(x-a)^2.\]

Taking the second-order Taylor series expansion of the log-likelihood around $\theta=\hat{\theta}$, where $\hat{\theta}=\hat{\theta}_{MLE}$, we obtain:

\[\ell(\theta;\mathbf{y}) \approx \ell(\hat{\theta};\mathbf{y}) + \ell'(\hat{\theta})(\theta-\hat{\theta}) + \frac{1}{2}\ell''(\hat{\theta})(\theta-\hat{\theta})^2.\]

Since $\hat{\theta}$ is the maximum likelihood estimate,

\[\ell'(\hat{\theta})=0.\]

Therefore,

\[\ell(\theta;\mathbf{y}) \approx \ell(\hat{\theta};\mathbf{y}) + \frac{1}{2}\ell''(\hat{\theta})(\theta-\hat{\theta})^2.\]

Let $\theta=\theta_0$. We obtain:

\[2\left( \ell(\theta_0;\mathbf{y}) - \ell(\hat{\theta};\mathbf{y}) \right) \approx \ell''(\hat{\theta}) \left(\theta_0-\hat{\theta}\right)^2.\]

Recall that the observed information is:

\[J(\theta)=-\ell''(\theta).\]

Therefore,

\[2\left( \ell(\hat{\theta};\mathbf{y}) - \ell(\theta_0;\mathbf{y}) \right) \approx J(\hat{\theta}) \left(\hat{\theta}-\theta_0\right)^2.\]

Since

\[\widehat{\operatorname{Var}}(\hat{\theta}) \approx \frac{1}{J(\hat{\theta})},\]

we obtain:

\[2\left( \ell(\hat{\theta};\mathbf{y}) - \ell(\theta_0;\mathbf{y}) \right) \approx \frac{ \left(\hat{\theta}-\theta_0\right)^2 }{ \widehat{\operatorname{Var}}(\hat{\theta}) }.\]

The expression on the right-hand side is precisely the Wald statistic:

\[W = \frac{ \left(\hat{\theta}-\theta_0\right)^2 }{ \widehat{\operatorname{Var}}(\hat{\theta}) }.\]

This demonstrates the relationship between the Wald statistic and the local curvature of the log-likelihood around the maximum likelihood estimate.

Score Test

The Score Test works by normalising the score function and examining its value at $\theta_0$.

Recall that the score function is:

\[U(\theta) = \frac{\partial \ell(\theta;\mathbf{y})}{\partial\theta}.\]

When calculating our parameter estimates, we find the maximum likelihood estimate $\hat{\theta}$ by solving:

\[U(\hat{\theta})=0.\]

The Score Test uses this idea but asks a slightly different question: what is the slope of the log-likelihood at the hypothesised value $\theta_0$?

If

\[U(\theta_0)\approx0,\]

then $\theta_0$ is likely to be near $\hat{\theta}$, meaning that we have less evidence against $H_0$.

If $U(\theta_0)$ is far from zero, then the likelihood is still increasing or decreasing substantially at $\theta_0$, suggesting that $\theta_0$ is not close to the maximum likelihood estimate.

The Score statistic is:

\[S = \frac{U(\theta_0)^2} {I(\theta_0)},\]

where $I(\theta_0)$ is the Fisher information evaluated at $\theta_0$.

When $H_0$ is true,

\[S \sim \chi^2_1 \qquad \text{as } n\rightarrow\infty.\]

Notice that:

  • $\sqrt{S}=Z$, thus $Z\sim\mathcal{N}(0,1)$.
  • The Score Test only requires quantities evaluated at $\theta_0$.
  • Unlike the Wald Test, we do not necessarily need to know the unrestricted estimate $\hat{\theta}$.
  • If $U(\theta_0)$ is close to zero, then the slope of the log-likelihood at $\theta_0$ is close to zero, providing less evidence against $H_0$.

Proof

Take a first-order Taylor expansion of the score function around $\theta=\theta_0$:

\[U(\theta) \approx U(\theta_0) + U'(\theta_0)(\theta-\theta_0).\]

Evaluate this when $\theta=\hat{\theta}$:

\[U(\hat{\theta}) \approx U(\theta_0) + U'(\theta_0)(\hat{\theta}-\theta_0).\]

Since $\hat{\theta}$ is the maximum likelihood estimate,

\[U(\hat{\theta})=0.\]

Therefore,

\[0 \approx U(\theta_0) + U'(\theta_0)(\hat{\theta}-\theta_0).\]

Recall that:

\[U'(\theta) = \ell''(\theta) = -J(\theta).\]

Therefore,

\[U(\theta_0) \approx J(\theta_0)(\hat{\theta}-\theta_0).\]

For sufficiently large samples, the observed information and Fisher information are approximately equivalent:

\[J_n(\theta_0) \approx I_n(\theta_0).\]

We also know that:

\[\operatorname{Var}(\hat{\theta}) \approx \frac{1}{I_n(\theta_0)}.\]

Since $\theta_0$ is a constant,

\[\operatorname{Var}(\hat{\theta}-\theta_0) = \operatorname{Var}(\hat{\theta}).\]

Using

\[U_n(\theta_0) \approx I_n(\theta_0)(\hat{\theta}-\theta_0),\]

we can take the variance of both sides:

\[\begin{aligned} \operatorname{Var}\left[U_n(\theta_0)\right] &\approx I_n(\theta_0)^2 \operatorname{Var}(\hat{\theta}-\theta_0) \\ &= I_n(\theta_0)^2 \operatorname{Var}(\hat{\theta}) \\ &\approx I_n(\theta_0)^2 \frac{1}{I_n(\theta_0)} \\ &= I_n(\theta_0). \end{aligned}\]

Therefore,

\[\operatorname{Var}\left[U_n(\theta_0)\right] \approx I_n(\theta_0).\]

Under $H_0$, we also have:

\[\mathbb{E}[U_n(\theta_0)]=0.\]

We can therefore standardise the score function:

\[Z_S = \frac{ U_n(\theta_0) }{ \sqrt{I_n(\theta_0)} } \overset{H_0}{\approx} \mathcal{N}(0,1).\]

Squaring the standardised statistic gives:

\[S = Z_S^2 = \frac{ U_n(\theta_0)^2 }{ I_n(\theta_0) } \overset{H_0}{\approx} \chi^2_1.\]

Wald Test vs Score Test

The Wald Test and Score Test approach the same hypothesis from two different perspectives.

The Wald Test examines the distance:

\[\hat{\theta}-\theta_0.\]

It asks: How far is our estimated parameter from the value specified under the null hypothesis?

The Score Test examines:

\[U(\theta_0).\]

It asks: How steep is the log-likelihood at the value specified under the null hypothesis?

Therefore, the Wald Test focuses on the distance between $\hat{\theta}$ and $\theta_0$, while the Score Test focuses on the slope of the likelihood at $\theta_0$.

Under standard regularity conditions and as the sample size becomes large, both tests are asymptotically related and should generally lead to similar conclusions.

Conclusion

For a simple GLM with a single parameter, likelihood-based hypothesis testing provides several ways of determining whether the data provide evidence against:

\[H_0:\theta=\theta_0.\]

In this post, we explored two of these methods.

The Wald Test measures how far $\hat{\theta}$ is from $\theta_0$ relative to the uncertainty associated with $\hat{\theta}$.

The Score Test, on the other hand, evaluates the slope of the log-likelihood at $\theta_0$. If the slope is close to zero, then $\theta_0$ may be close to the maximum likelihood estimate.

Both approaches ultimately produce test statistics that asymptotically follow a $\chi^2_1$ distribution under $H_0$.

The third likelihood-based approach is the Likelihood Ratio Test, which compares the likelihood under the null hypothesis with the maximised likelihood. This will complete the three main approaches to likelihood-based hypothesis testing.