At first glance it appear that you need something like family =binomial in your specification. Without that it knows not what kind of regresion to do. I hope this helps. May the force be with you, David Booth

Do you have any NAs in your data?

Is there a reason that you did not do use lm, like:

lm( FS_POLL_CONTRIB ~ PRECIP_fs*PPI_fs + PRECIP_fs*GPI_fs + PRECIP_fs*WEI_fs)

This shouldn't affect the model though. I am assume the reason is that you have missing data and haven't told R how to deal with them.

David Eugene Booth it doesn't shows but it has the family gaussian. Daniel Wright it doesn't have any missing data. I have only one value for each variable, being n = 8, when I delete one predictor variable it runs, but when I have all the variables it gives me just NAs.

Does your data call for Gaussian or should it be something else? Could you describe the dependant variable in detail? Good Luck, David

Oh, it is just the standard errors missing, not the estimates. You have too many things things being estimated for the sample size. You need a larger sample. Can you give the results of your power analysis.

Daniel Wright the problem is that the sample is indeed small and lower than the one from the power analysis and I can't collect more due to the current crisis. When I remove one of the predictor variables, the model works. How can I determine which variable to remove so that the model runs? Which selection do I use so that the removal is objective?

David Eugene Booth the data is growth expressed as the diferrence of height between one month and after one month from 0.6 cm (the lowest) and 19 cm (the highest).

Daniel Wright from the last post this looks like a job for adaptive lasso unless he is getting close to the n=p situation which usually doesn't have enoufg information. Here's the link to get him going: https://www4.stat.ncsu.edu/~boos/var.select/lasso.adaptive.html

Best, David Booth

Diego Centeno-Alvarado , I assume the standard errors are large even when you exclude a few terms. This likely means the statistics are not that useful.

David Eugene Booth , why do you think is better adaptive lasso than standard lasso?

Diego Centeno-Alvarado because adaptive lasso has an oracle property that says you will get the optimal predictor for the original variables included. I have attached some of our work that you may find helpful in this regard in addition to what I sent to Daniel Wright This will work also if p>n but for n approx = p probably not based on a few recent tests we conducted. HTH Best wishes, D. Booth

You have 8 values to estimate 8 coefficients. That will always give you a perfect fit with no standard errors. As Daniel said, as long as you cannot considerably increase the sample size, evaluating such a model is quite worthless.

David's suggestion to use lasso is possibly not such a good advice, if the model as it is makes sense (so the coefficients do have a desired meaning in the context of your research). Lasso is a methods of changing the model to something else. This "something else" may give a more parsimonious description of the data, but it may fail to give a model with coefficients that are interpretable in the context of your research.

Hi Diego Centeno, you cannot estimate as much parameters, in your case 8, including the intercept, as the number of data points which in your case is n=8. You need at least n=9 to estimate 8 parameters in your model. If that is not the case you wont have enough degrees of freedmon for the error term in you regression.