Heart disease is the leading cause of death for both men and women. According to the Centers for Disease Control and Prevention, as of 2013, about 715,000 heart attacks occur in the United States each year.1 The cost of cardiovascular diseases in the United States, including health care expenditures and lost productivity from deaths and disability, is estimated to be more than $444 billion each year.

Many businesses have seen their cost for health insurance grow beyond bounds.

Negotiations with insurance companies can depend, in part, on the demographics of a company’s workforce. To understand the relationship between costs and other variables, companies must understand how factors such as age and sex can help predict health costs. Because diseases and therapies vary greatly, it is best to build a separate model for each disease group.

And, of course, heart disease is near the top of the list as an important condition to predict. Models for health care costs can help in predicting and controlling costs. The data file Heart attack charges holds records for 12,844 patients admitted to hospitals in the Northeast United States with a diagnosis of acute myocardial infarction (heart attack; AMI). The data are available from Medicare records for public use. These data provide a good example of the kinds of models we can build to understand and predict health care expenses.

The variables available in this dataset are:

Charges: the total hospital charges in dollars.

Age: in years.

Sex: coded M for males, F for females.

DRG: the Diagnosis Related Group, which groups together patients with similar management. In this dataset there are three different DRGs.

121: patients with AMIs and cardiovascular complications who did not die.

122: patients with AMIs without cardiovascular complications who did not die.

123: patients with AMIs who died.

LOS: hospital length of stay in days.

Died: 1 for patients who died in hospital and a 0 otherwise.

Build a model to predict Charges. Here are some things to consider in building the model:

1. Should Charges be re-expressed? Examine displays and find a suitable re-expression.

2. Should LOS be re-expressed? Are there unusual features of the distribution that may deserve special attention?

3. Some of these records may hold errors or otherwise be extraordinary. Make appropriate displays and set aside records that may not be reliable.

4. How do men and women differ? Consider, for example, the distribution of Age for male and female patients. Should you introduce Sex as a variable in your model? How will you do that?

5. Can you use DRG in your regression model as it is presented here, or should it be recoded in some way? Diagnosis raises similar questions. If our purpose is to predict costs based on demographic information, should these variables be predictors in our model?

6. Ultimately, how well can you predict Charges? Consider not just the R2, but also the se. (Be sure to transform the se back to dollars if you are working with a transformed version of Cost.) Discuss how well your model predicts health care costs due to heart attack.