Max Rohde - Sensitivity of odds-ratios calculated on dichotomized variables to inclusion criteria

Code

library(tidyverse)
set.seed(123)

Motivation

In this document, we will show how calculating an odds ratio based on a dichotomized continuous predictor variable can be manipulated by changing the range of the predictor variable that was sampled (i.e, study inclusion criteria), whereas a logistic regression model that uses the continuous values of the predictor will produce a stable estimate.

Scenario

Assume that we are interested in a disease where the incidence varies with age.

We will assume as the true model a simple relationship where the probability of developing the disease is a linear function of age. The below plot shows this relationship.

Code

# True model of disease probability is a linear function of age
p_disease <- function(age){
  0.25 + 0.0075*age
}

# Plot true model
tibble(age = seq(20, 80, length.out=2), prob = p_disease(age)) %>%
  ggplot() +
  aes(x=age, y=prob) +
  geom_line() +
  labs(title = "True probability of having the disease",
       x="Age",
       y="P(Disease)") +
  theme_bw()

We decide to sample subjects from the population and record if they have the disease. For simplicity, assume we sample patients uniformly within a given age range. We will show that dichotomizing age at a cutpoint is not a good idea, and can lead to estimates that can be greatly affected by the chosen age range to be sampled.

To dichotomize the predictor variable, let’s compare the incidence of disease among old (age > 50) and young (age < 50) patients and calculate an odds ratio, instead of using age as a continuous variable. The below simulation shows the results of two scenarios. As a comparison, we also fit a logistic regression using continuous age.

First, we sample 10,000 subjects with ages between 40 and 60. Second, we sample 10,000 subjects with ages between 20 and 80. We show that the choices of inclusion criteria has a large effect on the odds ratio comparing odds of disease between young and old subjects, but the estimates provided by logistic regression are unchanged.

Simulation

Sample from ages 40 to 60

Code

# Draw 10,000 patients uniformly between 40 and 60
ages <- runif(10000, min=40, max=60)

# Calculate true probabilities for each patient
probs <- p_disease(ages)

# Generate data where each patient has `probs` probability of having the disease
data <- map_dbl(probs, ~sample(c(0,1), size=1, prob=c(1-.x, .x)))

Code

# Put simulation data into a data frame
df <- tibble(age=ages, prob=probs, disease=data)

# Dichotomize at age = 50
df$old <- (df$age > 50)

head(df)

age	prob	disease	old
45.75155	0.5931366	1	FALSE
55.76610	0.6682458	1	TRUE
48.17954	0.6113465	0	FALSE
57.66035	0.6824526	1	TRUE
58.80935	0.6910701	1	TRUE
40.91113	0.5568335	1	FALSE

Code

table(df$disease, df$old)

   
    FALSE TRUE
  0  2081 1614
  1  2976 3329

Code

odds_ratio <- (3329 / 1614) / (2976 / 2081)

odds_ratio

[1] 1.442279

Code

# Fit logistic regression model using continuous age
glm(disease ~ age, family=binomial(), data=df)


Call:  glm(formula = disease ~ age, family = binomial(), data = df)

Coefficients:
(Intercept)          age  
   -1.23211      0.03547  

Degrees of Freedom: 9999 Total (i.e. Null);  9998 Residual
Null Deviance:      13170 
Residual Deviance: 13080    AIC: 13080

Sample from ages 20 to 80

Code

ages <- runif(10000, min=20, max=80)
probs <- p_disease(ages)
data <- map_dbl(probs, ~sample(c(0,1), size=1, prob=c(1-.x, .x)))
df <- tibble(age=ages, prob=probs, disease=data)
df$old <- (df$age > 50)

table(df$disease, df$old)

   
    FALSE TRUE
  0  2400 1179
  1  2625 3796

Code

odds_ratio <- (3796 / 1179) / (2625 / 2400)

odds_ratio

[1] 2.943705

Code

glm(disease ~ age, family=binomial(), data=df)


Call:  glm(formula = disease ~ age, family = binomial(), data = df)

Coefficients:
(Intercept)          age  
   -1.15573      0.03586  

Degrees of Freedom: 9999 Total (i.e. Null);  9998 Residual
Null Deviance:      13040 
Residual Deviance: 12220    AIC: 12230

Conclusion

We see that when sampling from ages 40 to 60, the dichotomization approach estimated an odds ratio of 1.44 compared to an odds ratio of 2.94 when sampling from ages 20 to 80.

In contrast, when sampling from ages 40 to 60, the logistic regression estimated a regression coefficient for age of 0.0355 compared to a very similar value of 0.0359 when sampling from ages 20 to 80.

--- title: Sensitivity of odds-ratios calculated on dichotomized variables to inclusion criteria description: A short simulation example showing why dichomization of continuous variables can lead to wrong conclusions. author: Max Rohde date: 01/16/2022 image: preview.png code-fold: show --- ```{r} library(tidyverse) set.seed(123) ``` ## Motivation In this document, we will show how calculating an odds ratio based on a dichotomized continuous predictor variable can be manipulated by changing the range of the predictor variable that was sampled (i.e, study inclusion criteria), whereas a logistic regression model that uses the continuous values of the predictor will produce a stable estimate. ## Scenario Assume that we are interested in a disease where the incidence varies with age. We will assume as the true model a simple relationship where the probability of developing the disease is a linear function of age. The below plot shows this relationship. ```{r} # True model of disease probability is a linear function of age p_disease <- function(age){ 0.25 + 0.0075*age } # Plot true model tibble(age = seq(20, 80, length.out=2), prob = p_disease(age)) %>% ggplot() + aes(x=age, y=prob) + geom_line() + labs(title = "True probability of having the disease", x="Age", y="P(Disease)") + theme_bw() ``` We decide to sample subjects from the population and record if they have the disease. For simplicity, assume we sample patients uniformly within a given age range. We will show that dichotomizing age at a cutpoint is not a good idea, and can lead to estimates that can be greatly affected by the chosen age range to be sampled. To dichotomize the predictor variable, let's compare the incidence of disease among old (age > 50) and young (age < 50) patients and calculate an odds ratio, instead of using age as a continuous variable. The below simulation shows the results of two scenarios. As a comparison, we also fit a logistic regression using continuous age. First, we sample 10,000 subjects with ages between 40 and 60. Second, we sample 10,000 subjects with ages between 20 and 80. We show that the choices of inclusion criteria has a large effect on the odds ratio comparing odds of disease between young and old subjects, but the estimates provided by logistic regression are unchanged. ## Simulation ### Sample from ages 40 to 60 ```{r} # Draw 10,000 patients uniformly between 40 and 60 ages <- runif(10000, min=40, max=60) # Calculate true probabilities for each patient probs <- p_disease(ages) # Generate data where each patient has `probs` probability of having the disease data <- map_dbl(probs, ~sample(c(0,1), size=1, prob=c(1-.x, .x))) ``` ```{r} # Put simulation data into a data frame df <- tibble(age=ages, prob=probs, disease=data) # Dichotomize at age = 50 df$old <- (df$age > 50) head(df) ``` ```{r} table(df$disease, df$old) ``` ```{r} odds_ratio <- (3329 / 1614) / (2976 / 2081) odds_ratio ``` ```{r} # Fit logistic regression model using continuous age glm(disease ~ age, family=binomial(), data=df) ``` ### Sample from ages 20 to 80 ```{r} ages <- runif(10000, min=20, max=80) probs <- p_disease(ages) data <- map_dbl(probs, ~sample(c(0,1), size=1, prob=c(1-.x, .x))) df <- tibble(age=ages, prob=probs, disease=data) df$old <- (df$age > 50) table(df$disease, df$old) ``` ```{r} odds_ratio <- (3796 / 1179) / (2625 / 2400) odds_ratio ``` ```{r} glm(disease ~ age, family=binomial(), data=df) ``` ## Conclusion We see that when sampling from ages 40 to 60, the dichotomization approach estimated an odds ratio of 1.44 compared to an odds ratio of 2.94 when sampling from ages 20 to 80. In contrast, when sampling from ages 40 to 60, the logistic regression estimated a regression coefficient for age of 0.0355 compared to a very similar value of 0.0359 when sampling from ages 20 to 80.