The National Health and Nutrition Examination Surveys (NHANES) are large national surveys conducted by the National Center for Health Statistics that collect information about risk factors and health outcomes of representative samples of Americans. The NHANES are the basis for national studies and goals, such as some of those in Healthy People 2010, a national health report designed to identify significant preventable threats to health and to establish national goals to reduce these threats (10).
The estimates presented in this article are based on data from the NHANES I and NHANES III. The NHANES I, fielded in 1971–1975, is unique in that it had a longitudinal followup, (NHANES I Epidemiologic Followup Study [NHEFS]), which tracked participants' health for 2 decades, allowing baseline risk factors to be linked to subsequent outcomes. The NHEFS was used as the basis for our simulation model. The hospitalization submodel of the simulation model was applied to data for adults from NHANES III, fielded in 1988–1994, because the prevalence of risk factors in that survey better represents current trends in adults (e.g., more obesity, and fewer smokers).
The hospitalization submodel consists of projection equations based on the NHEFS that link annual hospital admissions for each person to a comprehensive set of risk factors measured at baseline. The submodel subdivides men and women ages 45–74 into 4 groups (men ages 45–64 and 65–74 and women 45–64 and 65–74), and uses separate regressions for each group. The NHEFS excluded persons older than 74 at baseline. For each person (i) within an age-sex group in each year of followup (t), logged hospital admissions during the year were related to the individual's baseline characteristics using the following equation:
Ln admissionsit = f (risk factors)
where risk factors equal arthritis, current smoking, former smoking, ln systolic blood pressure, overweight, or other risk factors discussed below. The equations are negative binomial count regressions fitted with STATA statistical software, version 3.1 (Stata Corporation, College Station, TX), (regressions available on request from the author). Count regressions are closely related to hazard models. They model the number of events during a period of time (in this study, a year), while hazard models focus on the duration of time between events. Count models are particularly appropriate for data such as annual hospital admissions because they are built on probability distributions that assume that the dependent variable takes on only integer values and that zeroes and low numbers are common (11). The negative binomial is more flexible than the other commonly used count model, the Poisson, because it does not require the mean and variance of the dependent variable to be equal; it contains the Poisson as a special case. Estimates based on the hospitalization submodel have been published in 2 previous reports (12, 13).
The submodel is fitted to data for the full 2 decades of followup and can project admissions over this period, adjusting for mortality. The estimates presented here do not use the ability of the submodel to project events over time. All estimates were made for the baseline period, before anyone had died. Thus the estimates refer to the full NHANES III population in the baseline years, 1988–1994, and compare alternative scenarios for that period.
The literature on determinants of heart disease, cancer, and stroke was reviewed to identify the risk factors used in the regressions. The focus was on clinical risk factors, specifically, those that would be determined during a visit to a physician. Hospital admissions were related to all clinical risk factors that have been shown to be statistically significantly related to disease and/or death in multiple studies: age, race, smoking, systolic blood pressure, overweight and underweight, laboratory test results (serum albumin, serum cholesterol), exercise, alcohol consumption, diet (fiber, fish/shellfish, fruits/vegetables), and 9 groups of chronic conditions (12, 14, 15). The effect of each risk factor, including arthritis, smoking, and blood pressure, was estimated after controlling for all other risk factors. The measurement of the 3 risk factors focused on in the study (arthritis, current smoking, and hypertension) is described below (12, 14, 15).
A dichotomous variable for arthritis was defined, based on each participant's answer to the question in the NHANES I medical history interview, “Has a doctor ever told you that you have had arthritis?” Because the question did not distinguish between osteoarthritis and rheumatoid arthritis, the answer encompassed both.
Only half of NHANES I participants were asked about their smoking habits at baseline; however, at the first NHEFS followup in 1982–1984, the National Center for Health Statistics retrospectively collected this information for the other half of the participants. Retrospective information was used only when baseline information was missing. Machlin et al (16) reported that, for participants ages 45–74 with both baseline and retrospective data, and for the 3 categories of smokers used here (current, former, never), the 2 sources matched for 89% of subject respondents and 83% of proxy respondents (proxies were used primarily for people who had died). Smoking was represented by 2 dichotomous variables in the regressions, one for participants who were smokers at baseline, and one for participants who reported being former smokers. Never smokers were the comparison group.
Participants' systolic blood pressure was measured only once during their physical examinations. Because a single measurement can be atypical, national guidelines recommend that a diagnosis of hypertension be based on several measurements taken at different visits (17). However, because the NHEFS is longitudinal (individuals were followed for 2 decades after baseline and hospitalizations were recorded), our estimating equations self-correct for this problem. Individuals who had an unusually high single baseline pressure measurement would naturally have had the number of hospitalizations during followup that corresponded to their true lower pressures. The equations linking hospitalizations to baseline pressures thus correct for regression to the mean because persons whose pressures were overstated by a single reading contributed a smaller number of hospitalizations during followup than those with genuinely elevated pressures. Systolic blood pressure was a continuous variable in the estimating equations. All continuous risk factors were entered in log form because logs gave a better fit.
The equations used in this analysis, and those used in the study by Russell et al (13), were based on the full NHEFS followup through 1992. These equations project the sample experience well, tracking both the higher admissions of the 1970s and early 1980s, and the drop in admissions due to managed care in the later 1980s and early 1990s (18). Although the model's projections for 1988–1992 exceeded observed admissions, data from the benchmark National Hospital Discharge Survey showed that they accurately represented the longer-term trends beyond the early 1990s, especially the return to rising annual admission rates among people ages 65 and older (19).
The hospitalization submodel was used to estimate the impact of arthritis, smoking, and hypertension on the number of hospital admissions in adults ages 45–74 in the NHANES III. The cohort dataset includes all NHANES III respondents in this age group who were examined by a physician (n = 6,265). Data were used on all sample persons because 84% of the respondents had complete information available, 9% of the respondents had missing data for only one risk factor, and only 0.21% (13 individuals) were missing data for more than 5 risk factors. Missing values were replaced with the mean of the risk factor for the age-sex group.
Arthritis, smoking habit, and systolic blood pressure were determined in the same manner in both the NHANES I and NHANES III (except that all NHANES III adults were asked about smoking at baseline). Because the NHANES I measured systolic pressure only once (see above), the first NHANES III measurement was used in the analysis to maintain consistency with the projection equations.
Baseline estimates of annual hospital admission rates were calculated by entering observed baseline values of the risk factors for the 6,265 NHANES III adult respondents ages 45–74 into the model equations, and calculating admissions for the baseline period. Population attributable-risk estimates were then calculated for 3 scenarios. In each scenario, 1 risk factor was changed in the baseline period, while all other risk factors remained at observed baseline values. These scenarios followed the standard definition of population attributable risk, which measures the amount of the health problem that could have been prevented if the risk factor had never existed (20). The differences between the baseline estimates and each of the attributable-risk estimates represent number of hospital admissions attributable to arthritis, current smoking, and hypertension, respectively. Simulations were run separately for all persons ages 45–74 and overweight persons ages 45–74. The scenarios are described in detail below, and the same procedures were followed for both groups.