We conducted a secondary analysis of the prospective cohort study known as the Survey on Assets and Health Dynamics among the Oldest Old [AHEAD; 19, 20]. AHEAD is a national, omnibus health and retirement longitudinal data source of Medicare-eligible older adults administered by the Survey Research Center at the University of Michigan. AHEAD is a prospective cohort study in which subject interviews have been conducted about every two years since 1993. The survey questions field a wide array of information including: demographics, cognitive performance, physical and functional health, Medicaid eligibility, family structure, care-giving, and out-of-pocket costs for health and social services. Human subjects approval for our study was provided by the AHEAD Restricted Data Use Committee (# 2003-006), the University of Iowa IRB (# 2003-03008), and the Centers for Medicare and Medicaid Services (Data Use Agreement # 14807). Two sampling frames were used in creating AHEAD--a 1992 multi-stage household screening process, and a supplemental sample of persons ≥80 years old from the Medicare Master Enrollment File. Baseline interviews were conducted in 1993 with 7,447 participants ≥70 years old (response rate = 80.4%).
Sample
We created our analytic sample by linking baseline AHEAD interviews to Medicare inpatient, outpatient, and carrier claims for calendar years 1991-2005. Among the 7,447 older adults who completed the baseline AHEAD interviews in 1993 and 1994, we excluded 802 from this study because their AHEAD data could not be linked to their Medicare claims. We also excluded 604 participants enrolled in managed Medicare during the two years prior to baseline because these plans did not provide comparable data to the fee-for-service Medicare plans in which all AHEAD participants were enrolled. The 530 AHEAD participants who required proxy respondents at the baseline interview also were excluded because they did not complete cognitive and psychosocial protocols that measured risk factors included in our analysis. In this study, the number of AHEAD participants with linked Medicare claims data totaled 5,511 men and women (74.0% of the original AHEAD sample). In previous work [21], we developed propensity scores to address the potential sample selection bias based on our exclusion criteria but found that such adjustments did not have a substantial impact on the result. In this study, we could not use these adjustments because our unit of analysis was the ED episode rather than the individual. Nonetheless, based on our previous work, we assumed the bias would be minimal.
Medicare Claims
Three Medicare standard analytic files contain data on the provision of care in the ED--the outpatient claims files, the carrier claims files, and the inpatient claims files. ED services provided by physicians employed as hospital staff are submitted by the hospital as outpatient claims, and these claims reflect both the professional (i.e., physician effort) and technical (i.e., lab testing) components of ED care. Physicians who are either self-employed or part of a larger, hospital affiliated physician group submit their ED claims to a designated Medicare carrier. Therefore, ED claims in either the outpatient or carrier files are easily identified with CPT evaluation and management codes 99281-99285. This approach to identifying ED claims previously was used by the IOM, and accounts for over 80% of all Medicare expenditures for ED services [3].
We also could have identified visits to the ED with inpatient claims and from other outpatient and carrier claims that did not include CPT codes 99281-99285. For example, 5,854 ED claims for our study sample were uniquely found in the inpatient claims files (i.e., they did not have a corresponding outpatient or carrier claim with a CPT code 99281-99285). This occurred because Medicare statutes dictate that when a patient who presents to the ED subsequently is admitted to the hospital, the services provided in the ED must be "rolled" into the inpatient claim under bill values of 111-119 or revenue center codes (RCCs) of 0450-0459. We also could have used claims with CPT codes 99291 and 99292 to identify ED care that was provided to older persons who were critically ill or critically injured [22]. Our initial review of the sample data revealed these two codes appeared on 1,872 separate claims. However, we chose not use any of these claims to construct additional episodes of ED care because the episodes likely included services provided outside of an ED in locations such as a coronary care unit, intensive care unit, or respiratory care unit. This did not necessarily reflect a comparable construct of ED use. Moreover, these claims did not allow us to construct a measure of service intensity comparable to the information provided by CPT codes 99281-99285 in the outpatient and carrier claims.
ED Episodes
In the absence of prior work to define ED episodes from Medicare claims [6–10], we developed a bundling algorithm reflecting Medicare billing policy to identify ED episodes. For the outpatient files, we bundled claims for which the "from" and "through" dates overlapped or were within 3 days, consistent with Medicare policy requiring outpatient claims to be bundled if they occur within 72 hours [22, 23]. For the carrier files, we bundled claims with overlapping dates or those that were within 1 day of each other. This was necessary because Medicare claims have date but not time stamps, and therefore it is possible for a late-night ED encounter to carry over into the next calendar day. We then bundled the outpatient and carrier claims with overlapping dates and defined them as belonging to the same ED episode. We recognized that bundling claims over a consecutive three-day period may underestimate the actual number of episodes given that some individuals may enter and complete an ED episode on one day and then return to the ED on the next day. Therefore, we identified the number of episodes in which all claims were filed in a one day period from those in which claims spanned a two or three day period.
Measures of ED Episodes
We used two approaches to measure each ED episode in terms of severity and intensity. Our first approach relied on a modified-NYU algorithm. Originally, Billings et al. [24] created an algorithm (i.e., the NYU algorithm) to classify the severity of ED care by using the ICD9-CM diagnostic codes as identified in the ED. Using the diagnostic information, Billings and his colleagues calculated the probability that an ED claim fell into one of four categories: 1) non-emergent (NE); 2) emergent, primary care treatable (EPCT); 3) ED care needed, preventable/avoidable (EDCNPA); and 4) ED care needed, not preventable avoidable (EDCNNPA; http://wagner.nyu.edu/chpsr/index.html).
NE cases are those in which the patient's initial complaint, presenting symptoms, vital signs, medical history, and age indicated that immediate medical care was not required within 12 hours. The EPCT cases are those in which emergent care was required within 12 hours, though the presenting problem did not require continuous observation and no procedures were performed or resources used (i.e., a CT scan or lab work) that were not available in a primary care setting. The EDCNPA cases indicate that emergency department care was required, but the emergent nature of the condition was potentially preventable/avoidable if timely and effective ambulatory care had been received. Finally, EDCNNPA cases are those in which emergency department care was required and ambulatory care treatment could not have prevented the condition.
Since administrative records do not contain adequate information to make absolute determinations as to the appropriate category, the original NYU algorithm assigns probabilities that a visit falls into each of the four above categories, yielding four probability estimates. In developing this algorithm, Billings et al. did not classify visits to the ED that involved trauma, alcohol, drug-related, or psychiatric diagnoses. They also did not categorize bladder and urinary tract infections, and other infrequent diagnostic conditions.
Wharam et al. [17] simplified Billings et al.'s approach by generating a single measure of ED severity based on the summation of probabilities. Their modified-NYU algorithm, which we used in this study, defined a severe visit (modified-NYU rating = 3) as one in which the probability that the ED was needed was ≥75% (e.g, EDCNNPA + EDCNPA ≥0.75). ED episodes were defined as non-severe (modified-NYU rating = 1) if the probability that ED care was needed was ≤ 25% (i.e. EDCNNPA + EDCNPA ≤ 0.25). ED episodes that did not meet either criteria were defined as indeterminate severity (modified-NYU rating = 2).
Our second measurement relied on the actual CPT codes. As mentioned earlier, all physician claims for professional services provided in the ED receive one of five CPT codes [18], which reflect the intensity of services provided to the patient. The lowest CPT code (i.e., 99281) indicates that a minimal number of services were provided, services that could have been provided in alternative settings (i.e., physician offices or other outpatient settings). The highest code (i.e., 99285) reflects the amount of care provided during a severe or life-threatening emergency situation. In this study, claims were classified as highest intensity if they received a 99285 code; claims receiving 99281-99284 codes were classified as lower intensity.
We initially identified 38,311 claims for ED visits among our study sample. However, 12,731(33.2%) of these included diagnoses of trauma (n = 7,836), alcohol (n = 21), drug-related (n = 4), psychiatric diagnoses (n = 503), and 4,367 other diagnoses that were not classifiable using the modified-NYU algorithm. As such, these claims were removed from further analysis. Moreover, as we bundled the remaining claims into episodes and measured the severity and intensity of each ED episode, we occasionally identified more than one modified-NYU or CPT code within a single episode. This was not problematic when those multiple codes were concordant, which was the case for 96% of the episodes with multiple modified-NYU codes, and for 82% of the episodes for which there were multiple CPT codes. When the multiple modified-NYU codes or CPT codes were not concordant, we selected the codes reflecting the highest level of severity or intensity to define the episode. Figure 1 presents the data sources and corresponding number of ED claims excluded and included in our analysis.
Analysis
We plotted the annual trends in the per capita number of ED episodes among the study sample from 1991 to 2005, and examined trends pertaining to the two measures of ED episodes (i.e., modified-NYU severity & CPT intensity) and tested for their concordance over the 15-year observation period. Then, following on work completed by Ballard et al. [25], we observed which classification approach had greater predictive validity relative to being hospitalized. Hospitalization following the ED episode (i.e., admission from the ED) was defined as a binary outcome and regressed onto the modified-NYU code 3 (using 1 & 2 as the pooled reference) and CPT code 99285 (using 99281-99284 as the pooled reference).
We then regressed hospitalization onto a model that included several other risk factors. Because 4,171 (76%) of the 5,511 AHEAD participants had ≥1 ED episode, and these ED episodes were clustered, we incorporated generalized estimating equation (GEE) methods using an exchangeable correlation structure [26]. Several model development strategies were employed, including forced entry of all potential risk factors, as well as forward, backward, and stepwise selection. All risk factors statistically significant in one or more of those approaches were retained in the final model.
Our initial variable selection for the model was consistent with Andersen's widely used behavioral model of health services use [27]. We chose age, race, sex, marital status, living alone, and the importance of religion as demographic factors. Socioeconomic characteristics included education, income, and the number of health insurance policies. Health behaviors were represented by smoking status, alcohol use, and body mass index. Functional status assessment included a 5-item activity of daily living (ADL) index, a 5-item instrumental ADL (IADL) index, a single self-reported health question, and an 8-item version of the Center for Epidemiologic Studies Depression rating scale [CESD-8; [28]]. Cognitive assessment was evaluated using the Telephone Interview to Assess Cognitive Status [TICS; [29]], as well as delayed word recall. Disease history was tapped by self-reports of having been told by a physician that one has angina, arthritis, cancer, diabetes, heart attack, heart disease, stroke, hip fracture, hypertension, lung disease, or a psychiatric condition. Geographic factors included population density, Census region, the number of hospital beds and physicians per 1,000 persons in the county of residence (imported from the Area Resource File and based on the participant's place of residence at baseline), and perceived neighborhood safety. Health services use included the number of visits to a physician and hospital episodes in the previous year.
Because the preponderance of these risk factors for hospitalization were measured at baseline, we created an additional measure which captured the change in an individual's self-rated health from the baseline interview to the interview that preceded the first ED episode. We sorted the measures of change in self rated health into four groups: those individuals with no change in self rated health from the baseline to the interview prior to the first ED episode (780 out of 4,171 respondents); those with decline in self reported health (676 respondents); those with increases in self reported health (408 respondents), and a group in which only one assessment of self-rated health was provided at baseline (2,307 out of 4,171). To capture secular trends and access differentials we also used Medicare claims to account for when the ED episode occurred during our analytic period, and whether it occurred on a weekend or holiday.