I'm using a poisson fit; is the residual deviance = $ \chi^{2}$, and residual deviance / residual degrees of freedom = $\chi^{2}_{reduced}$?

Does this method provide a valid tool comparable to a reduced $\chi^{2}$ for goodness of fit?


This is in relation to a radioactive decay experiment. The data is made up of 50 points, taken at 18 seconds intervals, and are discrete counts. I am looking for a way to test the goodness of fit. I have used Mathematica to plot and fit with Generalised Linear Model with exponential family set to Poisson, fitting $y=Ae^{-\lambda t}$. The graph produced:

enter image description here

The 'Reduced $\chi^{2}$' on this graph is actually 'residual deviance / dof' I just haven't changed the label yet.

However there's another way to process the data also, which involves plotting a histogram of the response locations on the detector, fitting a gaussian on them, then using the area under the peak as the points for the data set, then plot and use NonLinearModel to fit $y=Ae^{-\lambda t}$, and the goodness of fit being $\chi^{2}$ / dof. enter image description here

I was wondering how comparable these two methods are for goodness of fit. Can i reasonably compare one to the other directly?

Also, the Mathematica Stack Exchange answer that explained the fit that should be used can be found here

  • $\begingroup$ Are you saying you are trying to perform a $\chi^2$ type fit on data using a Poisson model? $\endgroup$
    – Mister Mak
    Apr 8 '21 at 16:05
  • $\begingroup$ Also note reduced $\chi^2$ is not sufficient for goodness of fit, you need both the $\chi^2$ and the number of degrees of freedom. $\endgroup$
    – Mister Mak
    Apr 8 '21 at 16:06
  • $\begingroup$ More that I want something that is a parameter for goodness of fit for a poisson model, somewhat akin to a chi^2 or chi^2 itself. I just wasn't sure if residual deviance would be that thing and if it was identical to chi^2 or similar or not $\endgroup$
    – Epideme
    Apr 8 '21 at 16:07
  • $\begingroup$ True... Is Chi^2 / dof = Reduced Chi^2? That is what I had understood that to be $\endgroup$
    – Epideme
    Apr 8 '21 at 16:08
  • $\begingroup$ Can you explain your fit a bit more clearly? $\endgroup$
    – Mister Mak
    Apr 8 '21 at 16:09

If the fluctuations are of Poisson type, and too small to be Gaussian, then the standard weighted-least-squares estimator $\sum \frac{(data - model)^2}{(model fluctuations)^2}$ is not guaranteed to follow a $\chi^2$ distribution.

The good news is that if the fluctuations are of Poisson type, you may be able to use a likelihood method with a saturated data set to recover a $\chi^2$: https://www.sciencedirect.com/science/article/abs/pii/0167508784900164


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.