Yes. That can happen.
Hi!
Does it make sense that odds ratio are getting better (higher) and more significant when moving from a univriate to a multivariate regression model?
Yes. That can happen.
I don't have emotions and sometimes that makes me very sad.
Please describe you situation in more detail.
So you have a dataset looking at a binary dependent variable with (lets say) a binary independent variable (X1). Now you conduct multiple logistic regression controlling for additional variables (X2..Xk) and the estimates for X2 are increasing?
Describe what you did for univariate procedure. Was it simple logistic regression with just X1 in the model?
I wanted to see if they were using 2x2 table notation first then logistic or if they may have flipped the exposure and treatment groups in the former approach. Also, I was curious about the confidence intervals on the logistic measure. Also, were the new covariates significant?
Dason, I am not calling you a lying robot, but explain an example where the OR would increase. It seems data would become multi-dimensional and possibly sparse for groupings.
The examples are essentially equivalent to thinking about slopes in multiple regression. Can you think of an example where the slopes for one predictor increase if you add another variable?
I don't have emotions and sometimes that makes me very sad.
Doobdoob (10-20-2015)
That is how I initially tried to think about it. But must be too tire, because I am not picturing a scenario where one covariate is having a bigger effect when controlling for another variable unless they were associated in some way, which may affect the standard error.
Educate me...
Allowing A LOT of correlation and making one of the predictors binary is probably the easiest way to think about this.
I don't have emotions and sometimes that makes me very sad.
Doobdoob (10-20-2015)
Hi,
It's my first time using this forum - Thanks for the prompt response!
Shortly, I have a group of patients all underwent Cardiac CT. I'm looking to see weather a history of shift work is associated with worse results on CT (Calcium score >100)
When comparing the demographics of those with a history of shift work vs. those without - The shift work group is significantly younger and with more %males.
Now, I did univariate logistic regression with the dependent variable being Calcium score >100. The results were:
Age OR=1.1 (P=0.000)
Male gender OR 1.79 (P=0.073)
Diabetes OR=1.1 (P=0.7)
Hx of shift work OR=1.66 (P=0.108)
Other variables I tested for (smoking, hypertension etc. did not yield anything near significant)
Multyvariate logistic regression model (with age, male, shift work)
Age OR=1.1 (p=0.000)
Male gender OR=3.6 (P=0.01)
Hx of Shift work OR=2 (P=0.046)
Since those with history of shift work are younger, I thought it made sense that the OR got better when adjusting for age. But male sex also correlates with shift work - am I double dipping....?
Thanks!
You should test for interactions between these covariates. Do you know how to do that?
Not sure. Pearson correlation with SPSS?
Tweet |