1
$\begingroup$

I have a set of nutrient fluxes data and I would like to know which environmental drivers explains the fluxes. I used DistLM and the marginal test showed that none of the independent variables were significantly different (p<0.05) however one did come close (C:N ratio, p = 0.07). The sequential tests for the stepwise model showed that after adding C:N ratio to the model, 2 other variables contributes significantly to the model. Is the stepwise still valid even though the C:N ratio was not significant in the marginal test?

$\endgroup$
1
$\begingroup$

I don't think stepwise regression is valid irrespective of whether the C:N ratio was significant marginally. It may help to read my answer here: Algorithms for automatic model selection.

If you want a valid hypothesis test / p-value, you need to know what variables you want to test a-priori, not fish for them. It is fine to think of this as exploratory research, draw an independent sample and test those variables a-priori in the new sample.

If you want to assess the out of sample predictive utility of your model, you could try cross validation and nest the entire stepwise selection within the training folds, then get the predictive error of the selected model (which will probably differ iteration by iteration) from the test fold. This will give you a sense of how much distortion the stepwise selection process does.

| cite | improve this answer | |
$\endgroup$

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.