August 2, 2019

Interpreting Logistic Regression Coefficients - Odds Ratios

Logistic Regression

The Logisitc Regression is a generalized linear model, which models the relationship between a dichotomous dependent outcome variable \(y\) and a set of independent response variables \(X\).

However, to get meaningful predictions on the binary outcome variable, the linear combination of regression coefficients models transformed \(y\) values. The transformations is done via the Logit, which basically is the natural logarithm of the odds, also called Logit: \(log(\frac{p(x)}{1 - p(x)})\), where \(p(x)\) is the probability that \(y=\). The log odds are modeled as a linear combinations of the predictors and regression coefficients: \(\beta_0 + \beta_1x_i\)

The complete model looks like this:

\(Logit = ln(\frac{p(x)}{1 - p(x)}) = \beta_0 + \beta_1x_i\)

This equation shows, that the linear combination models the Logit and model coefficients \(\beta\) descibe the influence of the predictors \(X\) on the logit, e.g. a one unit increase in \(x_i\) changes the Logit by the amout of \(\beta_i\). Unfortunatly, we do not have a reasonable intuition about the Logit and this makes it hard to interpret the \(\beta\)-coefficients.

Often, the regression coefficients of the logistic model are exponentiated and interpreted as Odds Ratios, which are easier to understand than the plain regression coefficients.

So the odds ratio tells us something about the change of the odds when we increase the predictor variable \(x_i\) by one unit. In the following two sections, First, I will present a mathematial expression to show that exponentiated betas are actually the odds ratio and secondly, I will give an illustrative example in R.

Odds Ratios

In this section we first present the Logit and then move on to show that the exponentialted regression coefficients can be interpreted as Odds Ratios.

Logit

To beginn with the Logit it is defined, as explained in the introduction, as the natual logarithm of the odds.

Odds are the ratio of the probability that the outcome variable will be 1 \(p(Y=1)\), also considered as the proabability of success, over the proabability that it will be 0 \(p(Y=0)\), sometimes considered as the probability of failure. The probabilty can also be expressed as \(p(Y=0) = 1-p(Y=1)\).

So the ratio of the probability of success over the probability is simply:

\(odds=\frac{p(Y=1)}{1-(Y=1)}\)

To be more concise, let us first define \(\pi\) which is the models’s predicted probability that \(Y=1\), so \(\pi=p(Y=1)\) and we wil l simpy write:

\(odds(\pi)=\frac{\pi}{1-\pi}\)

And the Logits is simply the natual logarithm of this odds and this Logits is modeled as a linear model \(\beta_0 + \beta_1x_i\).

As explains, we do not have an intuition about the Logits. Thats why we want to predict values that are easier to understand, i.e. the odds.

To do so, we apply the exponential function to both sides of our expression

\(Logit(\pi)=ln(\frac{\pi}{1-\pi)}) = \beta_0 + \beta_1x_i\)

which gives us

\(\frac{\pi}{1-\pi} = e^{\beta_0 + \beta_1x_i}\).

Beta Coefficients

Now that we know what the Logit is, lets move on to the interpretation of the regression coeffcients.

To do so, let us initially define \(x_0\) as an value of the predictor \(X\) and \(x_1=x_0 + 1\) as the value of the predictor variable increased by one unit.

When we plug in \(x_0\) in our regression model, that predicts the odds, we get:

\(\frac{\pi_0}{1-\pi_0} = e^{\beta_0 + \beta_1x_0}\) which is the predicted odds of \(X = x_0\).

When we plug in \(x_1\) we get:

\(\frac{\pi_1}{1-\pi_1} = e^{\beta_0 + \beta_1x_1}\) which is the predicted odds when \(X = x_1\) or equivalently \(X = x_0 + 1\).

In the last formula we can now replace \(x_1\) with \(x_0 + 1\) to get:

\(\frac{\pi_1}{1-\pi_1} = e^{\beta_0 + \beta_1(x_0 + 1)}\)

by mulitpling \(\beta_1(x_0 + 1)\) in the exponent we get:

\(\frac{\pi_1}{1-\pi_1} = e^{\beta_0 + \beta_1x_0 + \beta_1}\)

In this situation, we can apply the rule for exponential expressions that additions in the exponent can be rewritten as a mutliplicative expression. According to this rule \(a^{b+c} = a^b\times a^c\), which in our example gives:

\(\frac{\pi_1}{1-\pi_1} = e^{\beta_0 + \beta_1x_0} \times e^{\beta_1}\).

Since we previously defined \(\beta_0 + \beta_1x_i\) as \(\frac{\pi_0}{1-\pi_0}\), the odds where \(X=x_0\), we can now replace it in the equation in the formula above to get:

\(\frac{\pi_1}{1-\pi_1} = \frac{\pi_0}{1-\pi_0} \times e^{\beta_1}\).

All we have to do in this situation is to rearrange the equation by dividing both sides with \(\frac{\pi_0}{1-\pi_0}\) to show that \(e^{\beta_1}\) actually is the odds ratio which describes the relationship between odds when we increase X by 1 unit.

\(e^{\beta_1} = \frac{\frac{\pi_1}{1-\pi_1}}{\frac{\pi_0}{1-\pi_0}}\).

Odds Ratios in R

In this section, I will demonstrate in R, that the exponentiated regression coefficient of a logistic regression is actually the odds ratio.

Please consider the comments in the code for further explaination.

# 1. simulate data
# 2. calculate exponentiated beta
# 3. calculate the odds based on the prediction p(Y=1|X)
#
# Function takes a x value, for that x value the odds are calculated and returned
# Beside the odds, the function does also return the exponentiated beta coefficient
log_reg <- function(x_value) {
  # simulate data, the higher x the higher the probability of y=1
  set.seed(256)
  X <- c(rnorm(50, mean = 10, sd = 2), rnorm(50, mean = 14, sd = 2))
  y <- c(rep(0, 50), rep(1, 50))
  
  plot(y ~ X) 
  model <- glm(y ~ X, family = "binomial")
  beta_1 <- coef(model)[2]
  exp_beta_1 <- exp(beta_1) # exponentiate beta_1 to get odds ratios
  
  # predict pi_i, the probabilty that y=1 given the x value
  pi <- predict(model, newdata = data.frame(X = x_value), type = "response") 
  
  odds <- pi / (1 - pi)
  
  list(odds = odds, exp_beta = exp_beta_1)
}

x_value <- 14
x1 <- log_reg(x_value)
x2 <- log_reg(x_value + 1)

# Show the exponentiated beta coefficients
show(x1$exp_beta) # exp beta
##        X 
## 2.682433
odds_ratio <- x2$odds / x1$odds # calculate odds ratio
show(odds_ratio) # odds ratio
##        1 
## 2.682433
# Is the exponentiated beta the same as the calculated odds ratio?
round(odds_ratio, 6) == round(x1$exp_beta, 6)
##    1 
## TRUE

The R-code above demonstrates that the exponetiated beta coefficient of a logistic regression is the same as the odds ratio and thus can be interpreted as the change of the odds ratio when we increase the predictor variable \(x\) by one unit.

In this example the odds ratio is 2.68. This means that the odds increase by 2.68 when we increase the predictor variable \(x\) by one unit.

In a situatuation where the predictor variable is lower that 1, the odds decrease, i.e. the proabability of \(p(Y=1)\) also gets lower as \(x\) increases.

Outlook

As I demonstrated in this post, a way to interpret the regression coefficients of a logistic regression is to exponentiate the coefficient and view it as the change in the odds. That means the exponentiated beta is the odds ratio.

We did so, because the plain, non-exponentiated betas represent the change of the Logit and it is hard to get an intuition what that actually means.

However, we also should keep in mind, that the odds ratio, bypass the interpretation of hard to understand Logits, the odds ratio is itself a hard to understand concept and therefore not easy to understand.

While the odds ratio bypass the interpretation of hard to understand Logits and the odds ratio may be easier to interpret, their meaning is often not easy to understand.

We can overcome this problem by presenting representative values and its predicted probabilites by the logistic model, since probabilites are easier to understand than odds ratios.

© Jens Hüsers 2018

Powered by Hugo & Kiss.