Your model should make sense and be of practical utility. And these are the criteria to choose variables - not only the variables, but also the entire functional structure of the model (which variable goes in what transformation and in what interaction with which other variable(s)).

If you don't have the intellectual background to formulate a sensible or useful model, significance (p-values) is a very bad guide. It is way better to build the model by excatly what it is made for, and check its performance on that task using a test set of data. You will see that having a lot of possible different models will need a hell lot of data to come to a conclusion. There is no free lunch. But if you do have lots and lots of data, you can spare out a test sets and use several training sets to fit all thse different models and compare their performance on the test sets.

Article Step away from stepwise

https://www.lexjansen.com/pnwsug/2008/DavidCassell-StoppingStepwise.pdf

There are lots of criteria, but they depend on what you are using the model for. It will be useful if you provide more details, but I predict most commentators will urge you not to rely that much on the p-values of individual effects after conditioning on all others).

Your model should make sense and be of practical utility. And these are the criteria to choose variables - not only the variables, but also the entire functional structure of the model (which variable goes in what transformation and in what interaction with which other variable(s)).

If you don't have the intellectual background to formulate a sensible or useful model, significance (p-values) is a very bad guide. It is way better to build the model by excatly what it is made for, and check its performance on that task using a test set of data. You will see that having a lot of possible different models will need a hell lot of data to come to a conclusion. There is no free lunch. But if you do have lots and lots of data, you can spare out a test sets and use several training sets to fit all thse different models and compare their performance on the test sets.

Article Step away from stepwise

https://www.lexjansen.com/pnwsug/2008/DavidCassell-StoppingStepwise.pdf

You should choose the variable which has p.value less than 0.05 . these variable are significance

I guess my prediction above was wrong, someone did recommend just using p-values! Dan loses a point!

But Daniel Wright gets 10 points for being correct about not using p-values. People at this late date are still confused between predictive criteria and explanatory critera. You must be clear on what you want to do with your final model, predict something or explain something. Modeling is different in these two situations. Variable selection methods are based on predictive criteria. See:

https://www.google.com/search?q=prediction%2C+explanation+and+variable+selection+in+regression&oq=prediction%2C+explanation+and+variable+selection+in+regression&aqs=chrome..69i57j33.40425j1j9&sourceid=chrome&ie=UTF-8

Forget stepwise regression. It is a fancy form of p-value selection which doesn't work. You might find the attached interesting. Best, David Booth

Usually, the variables having a p-value of more than 0.05 are dropped from the model. But it entirely depends on your need. If you decide to have a higher cutoff point then you should keep those variables. It is better not to drop both the variables at one go, but drop the one having a higher p-value, and then observing the p-values of the rest variables, and then taking decision accordingly.

Use your p-values to remove variables that are not significant. To test the goodness of fit between the full model and the reduced model use robust model adequacy measures like resistant R², resistant predicted-R², median square prediction error, etc.

You can use only those p-values which are less than. 05,it means, these variables are significant.

just using p-values vs. thinking about a sensible model: 4 : 3 up to now.

Doesn't lookgood for us, Daniel and David. The world seems much simpler than we think.

Be strong Jochen Wilhelm , now you can say 4:4.

Best regards,

roberto

Jochen Wilhelm You are correct. The model is all....5 to 4... One has to think and not act blindly based on a number. It's the difference between being efficient and being effective...

Dear Alan F Rawle ,

based on Jochen Wilhelm bet:

just using p-values vs. thinking about a sensible model

and you answer, it is 4 to 5 .

Still not so good! Will see

Roberto Molteni Yes, I'm in the sensible model group. There is no scientific alternative is there? The 'usefulness' (or better uselessness) of p-values has been well described in the (recent) literature.

This is (link to) the article I was trying to recall:

'It’s time to talk about ditching statistical significance'

https://www.nature.com/articles/d41586-019-00874-8

https://www.nature.com/articles/d41586-019-00857-9

From the ASA statement (point 2):

'P-values do not measure the probability that the studied hypothesis is true, or the probability that the data were produced by random chance alone'.

Thank you all for your kind support :)