I am reading a scientific paper that fits a numerical model to a timeseries (observations). The numerical model has a few free parameters, some of them linear and some are not, which the authors wish to fit to the observations.

The numerical model produces a timeseries for each set of input parameters. The authors use regression (I think, they don't explicitly say so) to produce a $\chi^2$ value for each model run. Then, they report the parameters that minimize the $\chi^2$ value as those that best fit the observed timeseries.

So, just as an example, if the observations timeseries are represented by $O(t)$, then the model produces $i={1..N}$ timeseries $M_i(t,a_i,b)$, where $a_i$ is a linear model parameter that changes for each timeseries and $b$ is a nonlinear parameter which the authors choose to hold constant and do not attempt to fit, but is a free parameter (so in theory it may be changed). The parameter $a_i$ producing model timeseries $M_i$ that has the lowest $\chi^2$ is selected as best fit parameter.

This confuses me, I admit. Here are my questions:

  1. As I mentioned, the authors do not explicitly say which method they used, but do report $\chi^2$ values. Is there any possibility they did not use regression? Is $\chi^2$ minimization used in any other goodness of fit method?
  2. If regression is the only option, can the authors even do that? As I mentioned, some of the parameters are nonlinear with the function. I mean, the model is linear with the one fit parameter $a$, but the for the other parameter $b$ (which they hold constant and do not attempt to fit) it is nonlinear. I am only mentioning it, since for the best of my knowledge you cannot trust $\chi^2$ for nonlinear regression since the number of degrees of freedom is unknown. I am not sure if this holds in this case, as they treat $b$ as a constant.
  3. If this approach does indeed make sense: the authors compare the $\chi^2$ of a few different models to say which model best fits the data (that is, after they found the best fit parameter $a_i$ for each timeseries). How do you estimate the errors in this case? Using cross validation? (they did not produce any errors with their $\chi^2$s).

I would appreciate any help.


Your Answer

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

Browse other questions tagged or ask your own question.