Statwing's approach to statistical testing

A user-friendly guide to logistic regression

What is logistic regression?

Logistic regression estimates a mathematical formula that relates one or more input variables to one output variable.

For example, let’s say you run a lemonade stand, and you’re interested in which types of customers tend to come back. Your data includes an entry for each customer, their first purchase, and whether they came back within the next month for more lemonade. Your data might look like this:

Return  Customer age Sex Temp. at first purchase Lemonade color Pant length
Didn’t 21 Male 24 Pink Shorts
Returned 34 Female 20 Yellow Shorts
Returned 13 Female 25 Pink Pants
Didn’t 25 Female 27 Yellow Dress
etc. etc. etc. etc. etc. etc.

You think that Customer age (an input or explanatory variable) might impact Return (an output or response variable). Logistic regression might yield this result:

At age 12 (the lowest age) likelihood of return being Returned is 10%
For every additional year of age, Return is 1.1 times more to be Returned

This bit of knowledge is useful for two reasons.

First, it allows you to understand a relationship: older customers are more likely to return. This insight might lead you to bend your advertising towards older customers, since they’ll be more likely to become repeat customers.

Second, and relatedly, it can also help you make specific predictions. If a 24-year old customer walks by, you could estimate that if they bought some lemonade, there is a 26% chance they’d later become a return customer.

Understanding multiplication of odds

Note that if we said Returned was “1.5 times more likely” in some situation than in another, we’re doing the following:

Odds were    1 : 9       Also written 1/(1+9) = 10%.
The “odds for” (the 1) is multiplied by 1.5
Now             1.5 : 9       Also written 1.5/(1.5+9) = 14%

Another example, this time of going from 50% likelihood to something 3 times as likely:

ODDS WERE    1 : 1      ALSO WRITTEN 1/(1+1) = 50%.
THE “ODDS FOR” (the left-side 1) IS MULTIPLIED BY 3
NOW               3 : 1       ALSO WRITTEN 3/(3+1) = 75%

Now we’ll walk through the process of creating this regression model.

 

 

Preparing to create a regression model

 

1. Think through the theory of your regression

Once you’ve chosen a response variable, Revenue, hypothesize how various inputs may be related to it. For example, you might think that higher Temperature at first purchase will lead to higher likelihood of Returned, you might be unsure about how Age will affect Return, and you might believe that Pants (vs. shorts) is affected by Temperature but doesn’t have any impact on your lemonade stand.

The goal of regression is typically to understand the relationship between several inputs and one output, so in this case you’d probably decide to create a model explaining Return with Temperature and Age (also said as “predicting Return from Temperature and Age, even if you’re more interested in explanation than actual prediction).

You probably wouldn’t include Pants in your regression. It might be correlated with Return because both are related to Temperature, but it doesn’t come before Return in the causal chain, so including it would confuse your model.

 

2. “Describe” all variables that could be at all useful for your model (you can do this in 3 clicks)

Note which have a shape like this…

Screen Shot 2014-11-08 at 9.21.57 PM

…where most of the data is in the first few bins of the histogram. Those variables will require special attention later.

 

3. “Relate” all the possible explanatory variables to the response variable.

Statwing will order the results by the strength of the statistical relationship. Take a look and get a feel for the results, noting which variables are related to Revenue, and how.

 

4. Start building the regression.

Building a regression model is an iterative process. You’ll cycle through three stages as many times as necessary:

 

 

The three stages of building a regression model

 

Stage 1: Add or subtract a variable

Stage 2: Assess the model

Stage 3: Modify the model accordingly

 

Stage 1: Add or subtract a variable.

One by one, start adding in variables that your previous analyses indicated were related to Revenue (or add in variables that you have a theoretical reason to add). Going one by one isn’t strictly necessary, but it makes it easier to identify and fix problems as you go along, and helps you get a feel for the model.

Let’s say you start by predicting Revenue with Temperature. You find a strong relationship, and you assess the model and find it to be satisfactory (more details in a minute).

Return <– Temperature

You then add in Lemonade color and now your regression model has two terms, both of which are statistically significant predictors. Like this:

Revenue <– Temperature & Lemonade color

Then you add Sex, and the model results now show that Sex is statistically significant in the model, but Lemonade color no longer is. Typically you’d remove Lemonade color from the model. Now we have:

Revenue <– Temperature & Sex & Lemonade color

That is, if you know the sex of the customer, knowing what color of lemonade they ordered doesn’t give you any more information about whether they’ll be a return customer.

You might investigate and discover that women tend to pick yellow lemonade more than men, and women are more likely to return. So initially it looked choosing yellow made a customer more likely to return, but in fact Lemonade color is only related to Return through Sex. So when you include Sex in the regression, Lemonade color “drops out” of the regression.

Interpreting regression results take a good deal of judgment, and just because a variable is statistically significant doesn’t mean it’s actually causal. But by carefully adding and subtracting variables, noting how the model changes, and always thinking about the theory behind your model, you can tease apart interesting relationships in your data.

 

Stage 2: Assess the model

Every time you add or subtract a variable, you should assess the model’s accuracy by looking at its r-squared (R2), AICc, and any alerts from Statwing. Every time you change the model, compare the new r-squared, AICc, and diagnostic plots to the old ones to determine whether the model has improved or not.

R-squared (R2)

The numeric metric for quantifying the model’s prediction accuracy is known as r-squared, which falls between zero and one. A zero means that the model has no predictive value, and a one means that the model perfectly predicts everything.

For example, the data represented on the left will lead to a much less accurate model than the data on the right. Imagine trying to draw a line through the scatterplot; you could almost completely separate blue (Returned) from red (Didn’t) on the right side, but on the left side it’d be hard to do so.

That is, the right side has a high r-squared, if you know Temperature and Age you can Returned vs. Didn’t quite easily. The left side has a low-to-medium r-squared, if you know Temperature and Age you have a pretty good guess as to whether it will beReturned vs. Didn’t , but there will be a lot of errors.

Logistic classification r-squared

There’s no fixed definition of a “good” r-squared. In some settings it might be interesting to see any effect at all, while in others your model might be useless unless it’s highly accurate.

Any time you add a variable, r-squared will go up, so achieving the highest possible r-squared isn’t the goal; rather, you want to balance the model’s accuracy (r-squared) with its complexity (generally, the number of variables in it).

AICc

AICc is a metric that balances accuracy with complexity—greater accuracy leads to better scores, added complexity (more variables) leads to worse scores. The model with the lower AICc is better.

(Note that the AICc metric is only useful for comparing AICcs from models that have the same number of rows of data and the same output variable).

Alerts

From time to time Statwing will suggest ways to improve your model. For example, Statwing may suggest that you take the logarithm of a variable (details on what that means).

Confusion Matrix and Precision-Recall Curve

The confusion matrix and the precision-recall curve are also useful tools for understanding how accurate your model is. And if you want to make predictions based on your model these tools will help you do so. They’re not strictly necessary for getting a good understanding of what your model is telling you, so we put them in a different section about the confusion matrix and precision-recall curve

 

Stage 3: Modify the model accordingly

If your assessment of the model found it to be satisfactory, you’re either done, or you can go back to Stage 1 and enter more variables.

If your assessment finds the model lacking, you’ll use Statwing’s alerts to fix the issues.

As you modify the model, continually note the changing r-squared, AICR, and residual diagnostics, and decide whether the changes you’re making are helping or hurting your model.

 

 

That’s it, that’s the flow of creating a regression model. Enjoy!