Multivariate logistic regression analysis from patient data

#1
I would like to conduct a multivariate regression analysis on patient data using essentially the following setup:

Dependent: a continuous variable
Independent: several categorical and continuous variables.

So I want to check whether any of the independent variable(s) predict the continuous dependent variable using the regression model.

The data is most likely non-linear. What options do I have for the type of regression I can carry out?
 
Last edited:
#2
Data can't be nonlinear - relationships can be.

If you suspect strongly nonlinear relationships, two standard ideas are to add quadratic and possibly cubic transformations of the IVs or to use restricted cubic splines.
 

hlsmith

Less is more. Stay pure. Stay poor.
#3
The focus should not be on the linearity of the variables themselves, but the error terms generated by the model. If the error terms (residuals) are skewed, then you attempt to transform variables.
 
#4
Thank you Peter and hlsmith.

The meaning was the relationship would likely be non-linear, not the data - I apologize.

What is a sample regression equation that is similar to my situation?
 

Karabiner

TS Contributor
#5
I would like to conduct a multivariate regression analysis
I suppose you mean multiple regression (many independent vars,
1 dependent var), not multivariate (multiple dependent vars).

Dependent: a continuous variable
Independent: several categorical and continuous variables
How many are there? And how large is your sample size?

So I want to check whether any of the independent variable(s) predict the continuous dependent variable using the regression model.
If you want to know if *any* of your variables predicts the DV,
you just perform a bunch of simple regressions (each with 1
predictor variable). What you'll get from multiple regression ist
much more complicated than that, since supposedly your variables
are correlated and "overlap" with respect to their abilty to predict
the DV. So some variables may seemingly be not predictive, but
only because what they can explain from the DV is already
explained by other, "overlapping" predictors. Other variables
who were non-predictive in a simple regression may become
predictive in a multiple regression model, because of the presence
of other variables in the model (suppressor effect).

The data is most likely non-linear. What options do I have for the type of regression I can carry out?
What do you actually mean by likely non-linear (relationships)? What do you expect?

With kind regards

K.
 
#6
I suppose you mean multiple regression (many independent vars,
1 dependent var), not multivariate (multiple dependent vars).

How many are there? And how large is your sample size?


If you want to know if *any* of your variables predicts the DV,
you just perform a bunch of simple regressions (each with 1
predictor variable). What you'll get from multiple regression ist
much more complicated than that, since supposedly your variables
are correlated and "overlap" with respect to their abilty to predict
the DV. So some variables may seemingly be not predictive, but
only because what they can explain from the DV is already
explained by other, "overlapping" predictors. Other variables
who were non-predictive in a simple regression may become
predictive in a multiple regression model, because of the presence
of other variables in the model (suppressor effect).


What do you actually mean by likely non-linear (relationships)? What do you expect?

With kind regards

K.
Continuous dependent variable: Depression severity

Factors (Independent variables): RLS prevalence, age, age of diagnosis, gender, disease duration, H&Y score, and UPDRS-III score.

I'm sorry I meant conducting a regression analysis where all the variables are considered together, not like the case of a simple regression.

Sample size:

Control group: n = 70
Cases: n = 115
N = 185.

I expect that the relationships are non-linear.
 
#7
If you have a treatment group and a control group, then that has to be one of your independent variables.

If you suspect nonlinear relationships, you can use splines of the continuous IVs. But I'd start with bivariate scatterplots of the DV and each continuous IV and parallel boxplots of the DV and the categorical IVs.