I have kinetic data measured under several treatments (one experiment per treatment) which can be fitted by several nonlinear two-parameter models, how can I compare whether there there are differences in response that vary by treatment?

Comparing model fits. My initial thought was to use an F-test to compare relative residual sum-of-squares (RSS) across different data sets for each model, but if I understand correctly, this measures RSS relative to differences in degrees of freedom, so is more suitable for comparing a subset of a model relative to the full model for the same data set. Is there a similar test that would be appropriate in this case (to compare the same model applied to different data sets rather than different models with varying number of parameters to the same data set)?

Comparing regression coefficients. I found this reference which suggests to use a t-test where the test statistic is calculated by aggregating the standard errors of the nonlinear regression coefficients ($\hat{\beta}$ is the coefficient estimate and $\hat{s}_e$ is its standard error; subscripts 1 and 2 denote estimates for treatments 1 and 2):

$t = \frac{\hat{\beta}_1 - \hat{\beta}_2}{\sqrt{\hat{s}_{e,1}^2+\hat{s}_{e,2}^2}}$

with the combined degree of freedom being the sum of the original two. Is this a standard approach and is there a good literature/textbook reference?


To compare models you need to have replicates at each level of the predictor. This allows you to partition the SS(residual) into SS(lack of fit) plus SS(pure error). Ideally SS(LOF)->zero. You can test this using an F-test.(https://en.wikipedia.org/wiki/Lack-of-fit_sum_of_squares). If your models all have two parameters then that with lowest F will be the best choice. It may be that the theory you rest a model on is desirable or perhaps you are just interested in minimizing the SS(residual).The important thing is that you have replicates. Without them you cannot partition the SS(residual).


You can check the following link http://www.graphpad.com/guides/prism/6/curve-fitting/index.htm?reg_interpreting_comparison_of_mod.htm

  • $\begingroup$ Welcome to the site, @Curie. This isn't quite an answer by CV standards. We are trying to build a permanent repository of statistical information in the form of high-quality questions & answers. As such, 1 thing we worry about is linkrot. It's fine to have a link as a reference, but not as a standalone answer. Would you mind expanding this to provide an explanation? $\endgroup$ – gung - Reinstate Monica Sep 10 '13 at 16:07

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.