Measuring Severity from Claims Data

This page was last updated on Friday, January 27, 2017.  This page is based on the following publication:  Alemi F, Walters SR. “A mathematical theory for identifying and measuring severity of episodes of care.” Quality Management in Health Care. 15.2 (Apri-Jun 2006): 72-82. Read►

Introduction

Because of lack of randomization it is difficult to attribute the patient outcomes to quality of care as opposed to patient’s severity of illness.  This note shows how severity of patient’s illness can be measured through extant data in your records.  A recent patent application allows researchers to estimate severity of episodes of illness from data on time of diagnoses and nature of diagnoses.

Procedures for Calculation of Severity

The procedure described here assume that one has access to a large claims file containing a measure of patient outcomes (typically charges or paid charges) and measures of illness (typically diagnoses as coded in International Classification of Diseases, ICD).  The data can come from only clinic visits, both clinic and hospital visits or only hospital visits.  The data are assumed to provide longitudinal information on the same patient.  There are several distinct steps in creation and testing of the severity index:

1. Clean the data to remove contradictory information
2. Organize subsets of the data to allow easier statistical analysis of very large databases
3. Estimate severity of each diagnosis
4. Estimate severity of multiple diagnoses of a patient
5. Evaluate the accuracy of the estimates

Each of these steps is further clarified in the following.

Clean the Data

The first step is to clean the claims data.  There are a number of obvious problems in claims data:

1. First, check for impossible negative numbers (e.g. negative charges or negative length of stay).  Negative charges often reflect denial of charges or denial of money paid for charges.  These charges must be matched with the original paid amount and the balance reported. There are fewer occasions in which a negative length of stay shows, so it may be best for the dependent variable.
2. Second, if charges are used as the dependent variable, pay attention that there are a number of occasions that charges for the same patient at the same time period are paid in different occasions. As a consequence of this situation, some entries in the claims file have very low charges.  All charges for the same hospitalization must be combined.
3. Third, there are a number of occasions of zero charges or zero length of stay.  Zero charges in claims file often indicate a correction of disease classification or other minor corrections.  The cleaning of data should note the change in diagnosis over the two claims and make the correction in the case.  Zero length of stay is a different matter.  It may indicate outpatient bills.  Since the reason for these cases are not clear, one may drop these cases from analysis.
4. Fourth, many diagnosis codes include procedures.  These constitute overlapping information and the procedure codes should be eliminated to avoid the overlap.
5. Fifth, the setting of the diagnosis matters.  Clinic and hospital visits will have radically different cost structure for addressing the same diagnosis.  The diagnosis codes should be modified to indicate whether the setting was in a hospital or a clinic.  This is typically done by adding an “H” to the diagnosis code.
6. Some hospitals report Dead on Arrival as a hospital stay of 0 days.  Others do not report these cases as a hospital admission.  A consistent approach should be used.  One could, for example, exclude from analysis patients who have died and have stayed for 0 days.
7. There are also errors in claims files that correspond to data entry errors (e.g. use of ICD code that has no valid interpretation).  These cases should be deleted from the analysis.
8. Finally, note that diagnoses codes may be added with leading zero or added zeros at the end.  Make sure that these variations in data analysis do not affect your select queries.

The enclosed video shows how data from Healthcare Utilization Project of Agency for Healthcare Quality and Research can be prepared for the analysis.  The key in this preparation is to select appropriate cases across all 15 diagnosis fields included in these files.

Organize Subset of Data

The statistical analysis of very large databases is problematic by its sheer size.  Since some visits involve multiple diagnoses, it is important to discern which of several codes should be rated as higher in severity.  Imagine two cases one in which the patient presents with Myocardial Infarction and Congestive Heart Failure and another where the patient presents with Myocardial Infarction and shock.  In trying to find out the severity of myocardial infarction it is important to separate out the effect of shock or the effect of congestive heart failure that accompanied the infarction.  One way to do so is through analysis of variance.  While conceptually the analysis is simple, practically it is difficult to conduct an analysis of variance with more than 10,000 variables (one variable is needed for each diagnosis).  Typically, dividing the diagnoses into several sets of related codes and conducting separate Analysis of Variance for each set is one possibility. 

To reduce the size of the data, statistical analysis is done separately for each diagnosis.  Since there are more than 10,000 diagnoses, this leads to significant reduction in data and enables easier analysis of variance.  To organize data into subset, all cases in which a particular diagnosis has occurred is selected.  The statistical analysis, described in next step, is carried out and the results are fed back into subsequent analysis.

Estimate Severity of Each Diagnosis

Assigning a severity score to each diagnosis requires the re-organization of the data.  Typically, up to 10 diagnoses are listed for each case in the claim data file.  Different diagnoses are listed in the same field titled primary, secondary, etc.  This must be reorganized so that each diagnosis is listed as a separate field.  When the diagnosis is present a value of 1 is assigned and when absent a value of 0 is assigned.  Table one shows how typical claims data are organized.

 The revised layout of this table looks as in Table 2, where the presence of each diagnosis is indicated by an indicator variable assuming the value of 1 when present and 0 otherwise.

 Once the data is organized in this fashion, several methods of analysis can be used to assess the severity associated with each diagnoses:  (1) Analysis of Variance, (2) Multi-Linear utility Modeling, and (3) Averaging. 

(1) Analysis of Variance

A regression analysis is done so that the dependent variable is amount paid, the independent variables are whether the patient was discharged alive and all other diagnosis within the data subset and the last variable in the regression is the target diagnosis:

In the above equation, the parameters α, β, … are estimated from the data.  The variables ICDxxx, ICDyyy, … reflect all diagnoses that is present in the subset and ICDTarget reflects the diagnoses used to organize the subset.  The value of coefficient δ is referred to relative severity score for the target diagnosis.  The use of analysis of variance is often impractical because of the large number of diagnoses present in most data.

(2) Multi-Linear Utility Modeling

In this approach, a Multi-Attribute Utility model is fitted to the measure of overall severity of a case, e.g. mortality or length of stay.  The severity of case x, u(x), is assumed to be proportional to mortality or length of stay and is calculated as: