Statistical significance is probably not a useful criterion in this case.

How about an effect size? Perhaps r-squared for the model or a change in r-squared when adding or removing a term? Perhaps a partial eta-squared for each term?

Strictly speaking, p-value does not depend on the sample size if H0 is true, so p < 0.05 [or selected threshold] always mean « statistically significant », whatever the sample size is. Which does not mean « interesting » or even « meaningful ». What happens is that your null hypothesis is never true in reality, and you manage to see it with such high sample sizes.

Which mean you must think about your null hypothesis and what you really want to demonstrate. And adapt consequently your test procedure.

Without any information about your particular context and questions, one cannot give serious advice. You should construct a collaboration with a local statistician who is familiar with your field of research and can suggests or develop suited methods.

For instance, you may use Salvatore's suggestion of effect size estimation (but with so high sample size, the confidence interval will be so small that you may wonder what it really means in a context of false model), you may investigate the power (1 - type II error) - but it will be very high if you have a small p-value -, you may think about what kind of differences are of interest and adapt equivalence testing approaches, you also may wonder about the predictive power of your model and so on. You can also decide to use a more stringent cut-off for significance than 0.05. Note also that if you make several tests, multiplicity correction can lower a lot the cut-off, so you may end-up with non-significant p after multiplicity correction even with so high sample sizes (imagine you make the model for 15000 differents genes, for instance, then you must have roughly p < 0.05/15000 to be « sure » of significance)

Once again: no serious answer can be given without details.

Thank you for help. I have sufficiently explained this to me.

I can handle it now.

In this case, the statistical significance is the same in the small samples, which is statistically significant whenever the probability value is smaller than the moral level of 5%. If it is meant that there are 20000 variables and each of them has a number of explanatory variables, then in this case we have a general multivariate General multivariate In this case, the regression is significant if the probability value is below the moral level