# confused by this statement on bivariate and multivariate

#### runnyeggsham

##### New Member
Hi, I am stuck on this statement:

Variables were included in the multivariate analysis if they had a P value of 0.10 in the bivariate analysis.

I thought P value of 0.05 or lower meant that the findings were signifcant and we disregard anything above that. No? Please explain. I am just taking my first stats class.:shakehead

#### Dason

There is nothing necessarily magic about using an alpha of .05. It just means that you're ok with false rejecting the null hypothesis (when it's true) 5% of the time (on average). If somebody used an alpha of .10 it just means that they're willing to be a little more liberal.

#### ledzep

##### Point Mass at Zero
I thought P value of 0.05 or lower meant that the findings were signifcant and we disregard anything above that
P-values are not be all or end all. I would strongly discourage religiously following the 0.05 as the p-value cutoffs to test significance. 5% level of significance is just a line in the sand and a rough guide. It should always be interpreted practically. For example, for me, p-value of 0.04 and 0.07 are the same sort of significance (p-value between 0.01 and 0.09 are always treated with caution). So, I wouldn't disregard a variable with p value of 0.07 or even 0.09.
If p<0.001 means there is a strong signal.

For univariate analysis, 5% is the generally accepted significance level.

Generally, when building a model, the significance level are relaxed a bit. 10% or even 20% are acceptable. This is because the significance of a varible might change in the presenece/absence of other variables in the model. In order not to loose this signal, we relax the significance level.

#### runnyeggsham

##### New Member
Oh, I thought it was pretty much written in stone that 5% was the rule of thumb and anything above was not as significant. Why wouldnt bivariate or multivariate also use 5% as the generally accepted sig level? Thank you all.

#### Dason

Oh, I thought it was pretty much written in stone that 5% was the rule of thumb and anything above was not as significant.
Only for those that don't really know what they're doing.

#### runnyeggsham

##### New Member
"Variables were included in the multivariate analysis if they had a P value of 0.10 in the bivariate analysis."

So getting back to the original question. The person is considering a variable significant if the P value is 10% or lower. Why would someone do a bivariate analysis between a risk and exposure for each risk involved, say for example, smoking, drinking, age, socio economic status, education. And then do a multivariate analysis for each exposure that was found to be significant?

Wouldnt doing a multivariate analysis on ALL of the exposures in the beginning SAVE an extra step of not doing the bivariate analysis?

Bivariate and multivariate test the same things and gives the same info except that bivariate deals with 2 variables, while multivariate deals with 3+ variables. Correct?

#### Karabiner

##### TS Contributor
And then do a multivariate analysis for each exposure that was found to be significant?
Multivariate: because one wants to know how good the variables jointly predict the
dependent variable, while taking into account the interrelatedness of the predictors.

Include only variable with p < 0.10: Seemingly they wished to make the analysis
less complex by reducing the number of variables beforehand. And/or maybe they
wanted to circumvent the problem that ususally one would demand > 20 or so cases
for each predictor variable in the analysis. So with a small sample it looks desirable
to have only a few variables (the pre-testing does not solve the problem, though, since
it introduces bias not accounted for in subsequent the multiple regression model, but
often those who use the pre-selection procedure are not aware of this).

Wouldnt doing a multivariate analysis on ALL of the exposures in the beginning SAVE an extra step of not doing the bivariate analysis?
That's the alternative. One may run short of subjects and/or the model is too complex,
but nevertheless it is an alternative.

Kind regards

K.