Can I set the breakpoint of piecewise regression by theoritic reference?

#1
I want to prove different regression coefficient for SBP on BMI according to BMI categories.

So I want to use piecewise regression.

BMI is categorized by WHO guideline which means medicaly proven breakpoint. (<18.5 , 18.5-23, 23-25, 25-30, >30)

Can I use these breakpoint ?

Or do I have to find breakpoint statistically? Like by using segment package in R.
 

ondansetron

TS Contributor
#2
I want to prove different regression coefficient for SBP on BMI according to BMI categories.

So I want to use piecewise regression.

BMI is categorized by WHO guideline which means medicaly proven breakpoint. (<18.5 , 18.5-23, 23-25, 25-30, >30)

Can I use these breakpoint ?

Or do I have to find breakpoint statistically? Like by using segment package in R.
No, breakpoints in clinical medicine are generally nonsensical despite their widespread use and "acceptance". BMI is a continuous variable and patients in the arbitrary WHO guidelines have very different risk profiles, even within the same category. SBP likely doesn't "jump" between categories of BMI and remain flat within categories of BMI, so the categories aren't good practice.
 
#3
No, breakpoints in clinical medicine are generally nonsensical despite their widespread use and "acceptance". BMI is a continuous variable and patients in the arbitrary WHO guidelines have very different risk profiles, even within the same category. SBP likely doesn't "jump" between categories of BMI and remain flat within categories of BMI, so the categories aren't good practice.
Thank you for your reply.
I have SBP&BMI data for 6 years.
I want to show SBP & BMI relationship increase when BMI category is increasing.
In addition, want to show yearly increasing SBP& BMI relationship in each BMI category.
Then is there any appropriate way of analysis that shows above?
 

ondansetron

TS Contributor
#4
Thank you for your reply.
I have SBP&BMI data for 6 years.
I want to show SBP & BMI relationship increase when BMI category is increasing.
In addition, want to show yearly increasing SBP& BMI relationship in each BMI category.
Then is there any appropriate way of analysis that shows above?
For the first just model SBP as a function of the raw BMI values (possible with a transform to make SBP nonlinear in BMI (SBP doesn't increase infinitely with BMI).

For "want to show yearly increasing SBP&BMI..."? Can you reword this, I'm not quite sure I understand what you mean.
 
#5
For the first just model SBP as a function of the raw BMI values (possible with a transform to make SBP nonlinear in BMI (SBP doesn't increase infinitely with BMI).

For "want to show yearly increasing SBP&BMI..."? Can you reword this, I'm not quite sure I understand what you mean.
dfdfdfdff.png
This is my result.
In this table I didn't perform piecewise regression.
Instead of pieciwise regression, I performed 40 (8*5) simple linear regression (SBP on BMI) for all combination of year * BMIgroup.
I thought direct comparison of these coefficient is not statistical , because each group is different and independent each other.
Nevertheless, the result shows that the larger the BMI category, the larger the coefficient, and as the year passes, the coefficient increases.
So to prove these two things, I considered piecewise regression.
 

Karabiner

TS Contributor
#6
Heebyungkoh said:
want to show SBP & BMI relationship increase when BMI category is increasing.
In addition, want to show yearly increasing SBP& BMI relationship in each BMI category.
Then is there any appropriate way of analysis that shows above?
SBP = b0 + b1*BMI +b2*BMI² +b3*year + b4*(year*BMI) + e

With kind regards

Karabiner
 
Last edited:

hlsmith

Not a robit
#9
@Heebyungkoh - I may agree that I am unsure if another interaction term may be needed, you would think so, but I am also unsure. But if you step back, perhaps the data should be visualized before ever thinking about adding the quadratic term. If there is a non-linear relationship and weird looking residual patterns, then explore the quadratic, otherwise why incorporate it without viewing data.