Is univariate analysis a must before binary logistic regression

#1
Could I enter all my independent variables directly for a binary logistic regression

OR
is it necessary that I should run a one way anova for each of the independent variables first and select only the significant ones for the binary logistic regression with the dependant variable

Please help
 

Mean Joe

TS Contributor
#2
You don't need to do univariate analysis first.

One thing to note is that even if a variable is not significant in univariate analysis, it may still have a place in a logistic regression.

is it necessary that I should run a one way anova for each of the independent variables first and select only the significant ones for the binary logistic regression with the dependant variable
FYI there are issues with throwing in all the significant ones.

Just be reasonable in how many variables you put into your regression.
 
#5
You don't need to do univariate analysis first.

One thing to note is that even if a variable is not significant in univariate analysis, it may still have a place in a logistic regression.



FYI there are issues with throwing in all the significant ones.

Just be reasonable in how many variables you put into your regression.
What would be the maximum number of variables that we could put in the regression. My study has eight variables
 

gianmarco

TS Contributor
#6
You don't need to do univariate analysis first.

One thing to note is that even if a variable is not significant in univariate analysis, it may still have a place in a logistic regression.

I agree with you. Yet, I am struck by the quantity of academic works (e.g., PhD dissertations) in which the inclusion of predictors into regression (either, linear, logistic, ect) is based on the univariate analysis on each single predictor against the DV. I have seen this, for example, in more than a couple of PhD dissertations from a famous University in Europe.

Gm
 

gianmarco

TS Contributor
#7
In my logistic regression modelling, I have followed this:

Peduzzi et al, "A simulation study of the number of events per variable in logistic regression analysis" (https://www.ncbi.nlm.nih.gov/pubmed/8970487)

The minumum number of observations should be:
N = 10 k / p
where p be the smallest of the proportions of negative or positive cases, and k the number of covariates
 

hlsmith

Omega Contributor
#8
I think the unvariate analysis component may get propagated due to two reasons:

-historically before big computing, it was probablyly easier to run a bunch of uni's then multiple regression;

-experimental studies will check covariate balance between the two treatment groups, I bet many believe this is a form of doing a bunch of uni's, which it is actually something different.
 
#9
Has this univariate first come as a part of purposeful selection of variables from a whole lot of variables. But I saw in one tutorial that they use p <0.25 to select those variables from univariate for multivariate. So out of my eight variables with a sample of 310 should I do this purposeful selection and can I use p value of less than 0.25 or should I enter all variables?