- Apply logistic regression to model the probability of an event as the response variable, both with and without nusiance variables.
- Understand the log-odds of an event and how it is used in logistic regression.
- Apply one-sample tests for a mean.
Looking back through your notes, you’ll notice that we have not yet worked with an inference problem where the response (DV) variable is categorical and the effect (IV) variable is continuous. To do this, we will make use of logistic regression.
If you are only concerned with inference, logistic regression is very straightforward and virtually unchanged from linear regression. Let’s take the cars dataset one last time and look at whether a car has been awarded a “Green Energy Award” as a function of its fuel efficency in the city:
The p-value here tells us the probability of observing an effect at least as strong as shown in the data if there is no change in the probability of a green award as a function of the city fuel efficency.
We can also extend this to a categorical variable (and it becomes an F-test):
And we can include any number of nusiance variables into the model:
So far this is fairly straightforward, and most of the time this is the most import part (inference) of logistic regression in applications.
How does logistic regression actually work and what are the point estimates telling us? Let’s look at these two questions now.
Assume for a moment that we have a variable p that tells us the probability of observing a particular event. The odds of this event is the ratio of p divided by 1-p (the probability of not observing p):
This quantity is often used in gambling contexts. Notice that an odds of 1 indicates that the probability is equal to 0.5; it is just as likely to observe the event as it is to not observe the event.
The log odds is simply the (natural) logarithm of the odds:
Logistic regression tries to model the logs odds of an event. Specifically, it tries to build a model that look like:
For some variable X. Why model the log odds and not the probability itself? The probability can only be between 0 and 1, whereas the log odds can range over and real number. The right-hand side above could take on any value (depending on the inputs X), and so it is much more mathematically sound to model the log odds rather than the probability directly. The formulation given here also a high-degree of symmetry about what class we pick as the event (specifically, if we flip 0 and 1 in the input, the values of the slopes and intercepts simply switch signs).
Looking again at the logistic regression of the green award as a function of the city variable, how do we interpret the parameter value?
We say that the logs odds increases by 0.3892 for every unit of increase in the city miles per gallon fuel efficency of the car. Is this very satisfying? Probably not, because there is not way to directly map this statement into things about probabilites. It is just what we are stuck with and the best that we can do.
There is one final type of test that we have not seen yet that is fairly common in applied statistics. It is actually one of the easiest tests to understand and is often the first type of test introduced in a statistics course. I choose to do it last because while the theoretical concepts are easy to understand I find it harder to explain why the test might be useful.
Specifically, consider the hypothesis test for a variable x defined by:
For some value of the parameter mu. This at first looks a lot like the tests we have already done, but it involves testing the exact value of mean variable’s mean rather than looking at the difference between the means of two variables. This is called a one-sample test.
The two most common one-sample tests are: the one-sample t-test (parametric) and the
Wilcoxon test (non-parametric). To run this with the tmodels package, use the
tmod_t_test_one and just put the number
1 on the right-hand-side. There
are no independent variables in this test; the
1 serves as just a place-holder.
The point estimate gives the mean of the city variable and the confidence interval
gives a range of guesses for the mean (NB: In my experience, the confidence intervals
are actually more common and useful with one-sample tests than the p-value). We can
change the value of mu that defines the null hypothesis by inputing the parameter
The non-parametric test works similarly: