I'm reading Model selection and inference: Facts and fiction by Leeb & Pötscher (2005) (link), in this paper they look at an example in linear regression:

Let $$Y_i = \alpha x_{i1}+\beta x_{i2}+\epsilon_i \qquad \epsilon_i \stackrel{d}{=}N(0,\sigma^2)$$ They denote the full, unrestricted, model as $U$ (where $\beta \not = 0$) and the restricted as $R$ (when $\beta = 0$). The least squares estimator $\hat \beta(U)$ can be calculated for the unrestricted model (it's estimator 'is' zero in the restricted model $\hat \beta(R)=0$). To decide whether to choose for the unrestricted model the following test statistic is used $$\left| \dfrac{\sqrt{n}\hat\beta(U) }{\sigma_\beta} \right| > c \qquad \text{for a certain cutoff point } c>0$$

Then they state:

This is a traditional pretest procedure based on the likelihood ratio, but it is worth noting that in the simple example discussed here it coincides exactly with Akaike's minimum AIC rule in the case $c=\sqrt{2}$ and Schwarz's minimum BIC rule if $c=\sqrt{\ln n}$

I don't see why this is the case, I have learned the following as definition of the AIC and BIC statistics: $$\text{AIC}_p = n\ln SSE_p - n\ln n + 2p \qquad \text{BIC}_p=n\ln SSE_p - n\ln n + p\cdot \ln n$$

Can anyone point to the connection between the statement and the definition?


I've learned OLS through Applied Linear Statistical Models by Kutner et all, there they define SSE as the sum of square errors or $\text{SSE}_p = \sum_i (Y_i-\hat y_i)^2$ in the model with $p$ parameters. Here when $p=1$ then $M_0=R$, when $p=2$ then $M_0 = U$.

I've looked at your answers but I don't follow yet. I'll try to explain the problem further.

If I look at AIC, then model $U$ would be chosen if $AIC_2 < AIC_1$, writing this out results in $$n\ln \text{SSE}_2 - n\ln n +2\cdot 2 < n\ln\text{SSE}_1 - n\ln n +2$$ or equivalently $$n\ln \dfrac{\text{SSE}_1}{\text{SSE}_2} > 2$$

I don't see why the left part should equal $\dfrac{n\hat \beta(U)^2}{\sigma^2_b}$.

  • $\begingroup$ It would be helpful to fully define your notation, e.g., $SSE_p$, so that we can establish equivalence of your and my expressions. $\endgroup$ – Christoph Hanck Apr 26 '17 at 8:39
  • $\begingroup$ Also, is $\sigma_b^2=\sigma_\beta^2$? $\endgroup$ – Christoph Hanck Mar 16 '19 at 9:05
  • $\begingroup$ The last two expressions are only equal asymptotically, as the expression for the LR part only obtains (see stats.stackexchange.com/questions/276192/…) when we estimate the error variance. $\endgroup$ – Christoph Hanck Mar 16 '19 at 10:47

In my answer here I show that in a case like the present one, in which we test nested models against each other, the minimum AIC rule selects the larger model (i.e., rejects the null) if the likelihood ratio statistic $$ \mathcal{LR}=n[\log(\widehat{\sigma}^2_1)-\log(\widehat{\sigma}^2_2)], $$ with $\widehat{\sigma}^2_i$ the ML error variance estimates of the restricted and unrestricted models, exceeds $2K_2$. Here, $K_2$ is the number of additional variables in the larger model. In your case, $K_2=1$, corresponding to $x_{i2}$. Thus, select the larger model if $\mathcal{LR}>2$.

Now, in the present linear regression framework, the absolute value of the $t$-statistic $$|t|=\left| \dfrac{\sqrt{n}\hat\beta(U) }{\sigma_\beta} \right|$$ is simply the positive square root of the LR-statistic.

(Actually, this in general only holds asymptotically, as we have $t^2=F$, the $F$- or Wald-statistic, which is in general not numerically identical to $\cal{LR}$ in finite samples. Leeb and Pötscher however assume that $\sigma^2$ is known, which, as is shown here, restores exact numerical equivalence of Wald, LR and score statistics in this setup.)

Hence, going with the larger model according to the mininum AIC rule when $\mathcal{LR}>2=c$ corresponds to rejecting when the t-statistic exceeds $\sqrt{c}$.

It is worth pointing out that this implies that, in this case, the AIC rule is nothing but a hypothesis test at level $\alpha=0.157$, as (the LR statistic being $\chi^2_1$ under the present $H_0$ of the smaller model being the correct one)

> 1-pchisq(2,df = 1)
[1] 0.1572992


> 2*pnorm(-sqrt(2))
[1] 0.1572992

Solving the equation $1.96=\sqrt{\ln n}$ for $n$ gives that BIC would be of the same size as a test at the 5%-level at $n\approx46$.

It does not seem to be a general result that AIC corresponds to a liberal nested hypothesis test. For example, when $K_2=8$, AIC is equivalent to rejecting when $\mathcal{LR}>16$, which, under the null, has probability

> 1-pchisq(2*8,df = 8)
[1] 0.04238011

In fact, the probability tends to zero with $K_2$:

enter image description here

| cite | improve this answer | |
  • $\begingroup$ Thanks for the answer, I've added some further explanation, I don't see the equivalence yet... $\endgroup$ – dietervdf Apr 26 '17 at 16:50
  • $\begingroup$ @dietervdf, I worked out the answer in a separate thread (linked to in my edited answer), as I believe the equivalence may be of some independent interest. $\endgroup$ – Christoph Hanck Apr 27 '17 at 8:10
  • $\begingroup$ @dietervdf, please let me know if the answer is incomplete. $\endgroup$ – Christoph Hanck May 11 '17 at 8:45
  • $\begingroup$ I have to look at it in greater detail, I'll try to accept soon $\endgroup$ – dietervdf May 11 '17 at 9:25
  • $\begingroup$ @dietervdf, anything still missing? $\endgroup$ – Christoph Hanck Mar 16 '19 at 8:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.