# comparing two odds ratio using two different variables for the same subjects

#### Alex C

##### New Member
For the same group of subjects I have two weight variables, one that is reported by the subject (weight1), and one that represents the weight that is appropriately measured with a scale after the first answer (weight2).

If I do two logistic regression:

logistic disease weight1
logistic disease weight2

I have the first OR = 1.037 [1.023, 1.051] and the second OR= 1.069 [1.053, 1.087], is it ok to conclude that using one variable instead of the other will result in a significant difference?

#### hlsmith

##### Not a robit
So you ran a simple logistic regression model twice with two different versions of the same variable and the respective ORs are listed above, correct?

#### Alex C

##### New Member
yes exactly that, I use the same outcome variable but two different versions of the exposure variable in two distinct models. The same subjects are used for both models.

#### hlsmith

##### Not a robit
Do both variables use the same units? You are trying to make the significance statement based on the confidence intervals? Which I believe would be incorrect for multiple reason. One the non-overlap of CIs wouldn't imply that.

#### hlsmith

##### Not a robit
Do both variables use the same units? You are trying to make the significance statement based on the confidence intervals? Which I believe would be incorrect for multiple reason. One the non-overlap of CIs wouldn't imply that.

#### Alex C

##### New Member
Yes both independent variables have the same units and represent the weight of a person. We first simply ask the person how much they weight. Then, after they gave their answer, we measure it with a scale. The idea is that the first variable (weight reported by the subject) is potentially not accurate. The second variable (weight measured with a scale after) is the correct measure. The goal is to assess that if we only ask the weight of a subject without measuring it, then it could lead to different result for further analysis (in my example it would lead to OR = 1.037 [1.023, 1.051] instead of OR= 1.069 [1.053, 1.087].

#### hlsmith

##### Not a robit
In many studies, there can be a bias correction based on a validations set. So I ask people their weight and get both values in a validation study and make a correction based on the differences. These validation corrections and studies are popular in epidemiology.

I would imagine plotting overlapping ROC curves for the models would be interesting along with actually reporting information on the measurement error (tendancies on differences). Another approach would be creating a scatterplot. In order to get different results based on the two variables, I would imagine the differences are not random, but differential. This should be examined.