y**3 发帖数: 267 | 1 I was given a linear model which contains patients demographics such as
height weight age etc. Heihgt and weight were treated indenpendent and
significant. But the validation data I was given has BMI instead of Height
and weight. Is it statistically sound if I calculate the coefficient for BMI
using those of height and weight? |
c********h 发帖数: 330 | 2 No...
Better ask for height and weight or refit the model with BMI. |
y**3 发帖数: 267 | 3 thanks for the reponse! Can you explain why? I guess the interaction term of
height and weight is insgnificant. Can we assume height and weight are
independnet, and calculate the parameter for BMI
Unfornately I dont have the original dataset. SO I can refit the original
model using BMI. And The goal is to validate the original model estimated
coefficients |
c********h 发帖数: 330 | 4 1. Height and weight are probably not independent. I would expect they're
positively correlated. Whether they are correlated or not is not a direct
conclusion for the insignificance of the interaction term. But this is not
the key issue.
2. You can simply try some easy examples.
y = a + b*x + c*z
y = d + g*(x/z)
I don't see any possible way to get g out of b and c. BMI is sort of the
ratio between height and weight. If your original model uses log of height
and log of weight, then it is possible. |
y**3 发帖数: 267 | 5 Thanks!
Yes,it makes sense!
I tried out another data set from a different product. The fitted BMI
coefficient is close to the one calculated from those of height and weight,
but all the other estimated coefficient also changed dramatically |