0
$\begingroup$

I am performing a model selection analysis with some variations of the same 3-equation dynamical system of ODEs:

$$ \dot{x} = f(x,y,z)\\ \dot{y} = g(x,y,z)\\ \dot{z} = h(x,y,z) $$

The particularity of the problem is that I don't have the same number of data points for each variable. Let's call $n_x$ the number of observations for variable $x$, $n_y$ for $y$ and $n_Z$ for $z$. I obtain the best fit via minimizing least squares, which for this problem would be

$$ SSR = SSR_x + SSR_y+SSR_z=\sum_{i=1}^{n_x}(x_i-\hat{x}_i)^2+\sum_{i=1}^{n_y}(y_i-\hat{y}_i)^2+\sum_{i=1}^{n_z}(z_i-\hat{z}_i)^2 $$ where the hat denotes data point. So I fit each model and compare quality of fit with the Akaike Information Criterion, which, following this question, can be computed from the SSR as

$$ AIC=2K+n\ln RSS, $$ where $K$ is the number of parameters and $n$ is the number of observations. where , and the +1 comes from including the variance of the statistical model as an estimated parameter (I guess we can drop this since it's a constant). My issue is of course what to plug in $n$. My guess is that, since the likelihood is additive, we could simply split in three terms as

$$ AIC=2K+\sum_{i\in{x,y,z}} n_i \ln RSS_i, $$

Does this make sense or am I missing something?

$\endgroup$
5
  • $\begingroup$ Hi and welcome to SE! A few comments: the exact definition of the AIC is $2K-2\log L$. In the special case where model errors are i.i.d. and normally distributed, then the AIC becomes $2K+n \log(RSS/n)$ (see here for the BIC: en.wikipedia.org/wiki/…). $K$ should include all parameters of your model, including the error variance, there is no need to add a term for the variance of the statistical model. $\endgroup$ Commented Sep 14, 2022 at 14:47
  • $\begingroup$ Yes! Thank you very much. K would in this case then be the number of model parameters plus 3? For the three error variances, one for each variable. $\endgroup$
    – amr95
    Commented Sep 15, 2022 at 7:34
  • $\begingroup$ There is no need to "add" anything to $K$. This is the number of parameters of the model, including its error variance, which does not change whether you use the likelihood or the RSS. $\endgroup$ Commented Sep 15, 2022 at 7:46
  • $\begingroup$ Excellent, thank you. What about the sum? Does it make sense? $\endgroup$
    – amr95
    Commented Sep 16, 2022 at 8:24
  • $\begingroup$ Can you update your question so as to add more details about your model? Especially, what are the three variables you are mentioning, what are the three ODEs, and what are your observations? Thanks! $\endgroup$ Commented Sep 16, 2022 at 14:58

1 Answer 1

0
$\begingroup$

It is hard to give a specific answer without further details about your model, but here is a general answer from the textbook by Burnham and Anderson (Model selection and multimodel inference):

enter image description here

Then it depends on how the 3 sub parts of your model that you mentioned are related. If all 3 parts are independent, it would make sense to compute a separated AIC for each of them.

$\endgroup$
1
  • $\begingroup$ Thank you for your additional response, and sorry for the late reply. I have updated the question, I hope now it's clearer what my doubt is. $\endgroup$
    – amr95
    Commented Sep 22, 2022 at 12:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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