Skip to main content

Table 1 Description of simulation scenarios for combination construction

From: Using ordinal outcomes to construct and select biomarker combinations for single-level prediction

Data-generating model

K

Training sample size

Prevalences

Biomarker distributions

Parameters

Non-proportional odds

3

200, 400, 800, 1600

P(D=1)=0.1,0.5

(X|D=1)∼N(0,2I2)

μ∈{–1,0,1,2,3}

   

P(D=K)=0.05,0.3

(X|D=2)∼N(μ,2I2)

 
    

(X|D=3)∼N(2,2I2)

 
 

5

200, 400, 800, 1600

P(D=1)=0.1,0.5

(X|D=1)∼N(0,2I2)

μ∈{–1,0,1,2,3}

   

P(D=K)=0.05,0.3

(X|D=2)∼N(0.5,2I2)

 
    

(X|D=3)∼N(1,2I2)

 
    

(X|D=4)∼N(μ,2I2)

 
    

(X|D=5)∼N(2,2I2)

 

Proportional odds

3

200, 400, 800, 1600

P(D=1)=0.1,0.5

X1∼N(1,0.25)

(β1,β2)∈{(1,2),

   

P(D=K)=0.05,0.3

X2∼N(1,0.25)

(1,1.5),(−1,1)}

 

5

200, 400, 800, 1600

P(D=1)=0.1,0.5

X1∼N(1,0.25)

(β1,β2)∈{(1,2),

   

P(D=K)=0.05,0.3

X2∼N(1,0.25)

(1,1.5),(−1,1)}

  1. When K=5, P(D=2)=P(D=3)=P(D=4). For the proportional odds data-generating model, logit{P(D≤k|X1,X2)}=α k +β1X1+β2X2. I2 is the two-dimensional identity matrix