Clinical Categorical vs. Regression Based: Understanding classification system fundamentals

Increasingly, policymakers are focusing on outcomes as the ultimate goal of health care delivery. We’re pleased to see this trend and we have developed classification systems to help achieve an outcomes-based focus.

All classification systems categorize patients into groups. But not all classification systems are the same. In this and future blogs, we’ll look at the clinical distinctions in the approaches to case mix and classifications systems.

Nearly all classification systems have these characteristics:

a. Severity of illness categories

b. Categorical or regression modeling

c. Recognition of frail, rare and expensive conditions or illnesses

d. For each dependent variable of interest (readmissions, complication, etc.), one, and only one, clinical classification system applies to all individuals

Let’s focus on a critical distinguishing feature in the classification systems created by the 3M HIS Clinical and Economic Research group. Most other classification systems use regression based models. At 3M, we use clinical categorical models, and we’re nearly exclusive in that approach.

Defining clinical categorical and regression

Now, you’re probably wondering what we mean by clinical categorical and regression. Let’s define these models.

A categorical clinical model consists of discrete cells driven by clinical rules defined by clinicians. The clinicians are informed by data so the cells are both clinically similar and useful to predict or to define an outcome of interest. It may take an entire year with clinicians to define the clinical classification of the dependent variable. Then, using patient data, cases are classified into mutually exclusive categories based on the patients’ underlying health status and applicable demographic factors (i.e., age). Categorical clinical models are defined by listing the specific combinations of clinical and demographic variables that are used in the category assignment. DRGs are an example of a categorical clinical model, which are intended to predict resource use. 3M’s Potentially Preventable Readmissions (PPRs) is an example of a categorical clinical model that separates subsequent admissions within a time interval from an initial (index) admission to determine if subsequent admissions are potentially preventable or not.

A regression based model is a statistical technique (formula) that uses a set of independent variables to produce a numerical score that predicts the value of a dependent variable – a variable which is an outcome of interest. The relationship between the clinical and demographic factors is defined by a mathematical formula developed using regression techniques instead of a clinical categorization. When a regression model is estimated using a data set, the result is a set of coefficients (one for each of the independent variables) which can then be used to compute a score based on the values of the same independent variables (but from a different data source) to predict the same outcome variable. That is, the output of the statistical model is a formula where the coefficients in the formula which were estimated for each of the independent variables are then used to produce a score that estimates the value of the dependent variable. Hierarchical Condition Categories (HCCs) are an example of a regression based model that uses inpatient and outpatient diagnoses to estimate resource use.

Regression based models are limited

There’s a reason why 3M uses a clinical categorical model. Regression based models are inexpensive to develop and quick to estimate, but they have limitations. The mathematical analysis of relationships is driven by the specific database used, and the underlying nature of a relationship is obscured from review. This can result in the need for a different model for each population; for example, HCCs have separate versions for commercial, Medicaid and Medicare. With a clinical categorical model, the clinical relationship between different conditions in the classification system is the same for all populations (for example, a diabetic with eye deterioration from the diabetes is always assigned to the same severity of diabetes).

Another regression based model is the risk adjustment applied to the CMS All-Cause Readmission. Since it’s regression based, the relationship between clinical conditions has to be re-specified when this system is applied to Medicaid and when new data bases are examined. Initially the CMS All-Cause Readmission model paid virtually no attention to patients with significant mental health conditions. This changed to a certain extent over time with new data bases. However, a clinical categorical model always considers significant mental health conditions as a major risk factor for readmissions. Our data analyses showed major mental health conditions to be important from day one.

Clinical judgment is the key

In 3M’s practice, clinical judgment is formed by testing clinical hypotheses against one or more data sets to categorize a medical condition. In the final analysis, clinical judgment always overrides the data as the latter may have collection challenges or simply not be precise enough.

In our next blog, we’ll discuss the advantages of the clinical categorical model.

Richard Averill, MS, is the Director of Clinical and Economic Research for 3M Health Information Systems.

Norbert Goldfield, MD, is Medical Director for 3M Health Information Systems.

2 responses to “Clinical Categorical vs. Regression Based: Understanding classification system fundamentals

  1. Pingback: What a Readmission Risk Score Doesn’t Tell You | 3M Health Information Systems

  2. Pingback: Quality, Prices, and Health System Consolidation |

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s