The 131st Annual Meeting (November 15-19, 2003) of APHA

Abstract #57067

Charity G. Moore, PhD¹, Judith A. Shinogle, PhD², Janice C Probst, PhD², and Andrew Lawson, PhD¹. (1) Department of Epidemiology and Biostatistics, University of South Carolina, Health Sciences Building, Sumter Street, University of South Carolina, Columbia, SC 29208, (803) 777-7353, cgmoore@gwm.sc.edu, (2) Health Administration, University of South Carolina, School of Public Health, College of Pharmacy, CLS Rm#311E, Columbia, SC 29208

Medicaid claims from 1997 to 1999 for any recipient with a diagnosis of diabetes (n=3577) were obtained from a state database to compare differences in utilization of emergency departments (ED) and prescription fillings by six race (AA/White/Other) and residence (rural/urban) groups. The average follow time for these individuals was 2.37 years. The number of ED visits over the period ranged from 0 to 33, 74.2% with no ED visits and 16.1% with one. The number of prescriptions filled ranged from 0 to 77 scripts, 5.8% with no scripts filled and a median of 16 scripts. Four models were used to assess the race/residence effect: Poisson regression, zero-inflated Poisson regression (ZIP), negative binomial regression (NB), and zero-inflated negative binomial regression (ZINB). The degree of differences (relative risk and significance) among the four models depended on the percentage of zero values. Overall, the NB and ZINB models demonstrated better fit. The questions arise: What percentage of zero visits makes a difference in the results? At what point does over-dispersion in the data make a difference in the results? These four models will be compared in simulations to answer these questions along with comparisons of assumptions and a discussion of what drives no utilization. In addition, we propose investigating availability of these models (i.e. software) to public health researchers encountered with count data. In summary, not adjusting for over-dispersion can lead to potentially false conclusions about health care utilization behavior among persons with the same disease and the same health care coverage.

Learning Objectives:

At the conclusion of the session, the participant (learner) will be able to
• Identify event data and different distributions that arise from event data.
• Identify 4 models for use in analyzing event data such as health care utilization.
• Evaluate appropriateness of models when comparing event data across independent variables.

Keywords: Utilization, Biostatistics

Presenting author's disclosure statement:
I do not have any significant financial interest/arrangement or affiliation with any organization/institution whose products or services are being discussed in this session.