Original articles
Analytic strategies for recurrent events in epidemiologic studies: background and application to hospitalization risk in the elderly

https://doi.org/10.1016/S0895-4356(99)00137-7Get rights and content

Abstract

Due to the intraindividual dependence, specific analytic strategies are needed to assess risk factors for recurrent events. Although well established in the biostatistics literature, applications of these techniques are almost nonexistent in the field of epidemiology. The authors applied four different regression approaches for recurrent events (logistic, Poisson, and two different Cox proportional hazards regressions) to derive rate ratios of hospitalizations for various prognostic factors in a cohort of 2424 frail elderly. Over a median follow-up of 670 days, 3299 hospitalizations were observed in 1564 persons. Estimated rate ratios were similar in all four approaches and virtually identical in three. With all methods, confidence intervals of the rate ratios were considerably wider than with naive Poisson regression neglecting intraindividual dependence of events. Appropriate analysis of recurrent events is feasible with minor modifications of multivariable models familiar to epidemiologists and should no longer be neglected in epidemiologic research. In our setting, Poisson regression was the most convenient approach.

Introduction

Recurrent events, such as asthma attacks, migraines, low back pain, sickness leaves, or hospitalizations, are common outcomes in longitudinal clinical and epidemiologic studies. The main methodologic challenge in the context of recurrent events is how to deal with the dependence of events within a person. Different conceptual frameworks have been applied to deal with this problem in longitudinal studies. These include measures of proportion 1, 2, as well as counts 3, 4, 5, 6 and survival time 7, 8, 9, 10 approaches. These methods are rarely encountered in the medical or epidemiologic literature 2, 11, 12, 13, 14, however, and they are only briefly mentioned in a recent standard textbook of epidemiologic methodology [15]. The goal of this article is to bring recent methodologic developments in this important field to the attention of epidemiologists who may not be familiar with the pertinent biostatistical literature.

Section snippets

Review of methods

In the proportion approaches, the follow-up period is divided into fixed time intervals and cumulative incidences of events are calculated in all of these time intervals. Results of multiple time intervals are combined adjusting for the dependence of events within a person [e.g., by use of the Generalized Estimating Equations (GEE) method], as recommended by Zeger and Liang [1]. The width of these fixed time intervals should be chosen to make multiple events in a single time interval unlikely.

Background

In Germany, financial aid was newly provided in 1991 to severely handicapped individuals to cover outpatient nursing care. The aim of these benefits was to help handicapped individuals to maintain life in the community setting and to prevent admission to long-term care facilities. Benefits were provided by the system of health insurance contingent on an in-depth medical examination.

The cohort assessed in this article consisted of 2424 residents of the region of Augsburg (southern Germany) aged

Discussion

In this article, we provide an illustration of the application of several existing options for the analysis of recurrent events in epidemiological studies. We extend the recent illustration of the application of different analytic strategies modeling proportions [2] to the modeling of rates and include methods of count data and survival time approaches in our comparison.

All analytic approaches need some preparation of the data. The relative ease of adapting available programs for person-time

References (32)

  • R.G. Cumming et al.

    Methodologic issues in the study of frequent and recurrent health problems. Falls in the elderly

    Ann Epidemiol

    (1990)
  • S.L. Zeger et al.

    Longitudinal data analysis for discrete and continuous outcomes

    Biometrics

    (1986)
  • F.B. Hu et al.

    Comparison of population-averaged and subject specific approaches for analyzing repeated binary outcomes

    Am J Epidemiol

    (1998)
  • P.F. Thall

    Mixed Poisson likelihood regression models for longitudinal interval count data

    Biometrics

    (1988)
  • T.A. Stukel

    Comparison of methods for the analysis of longitudinal interval count data

    Stat Med

    (1993)
  • R.J. Glynn et al.

    Estimating the variance of standardized rates of recurrent events, with application to hospitalizations among the elderly in New England

    Am J Epidemiol

    (1993)
  • T.A. Stukel et al.

    Standardized rates of recurrent outcomes

    Stat Med

    (1994)
  • L.J. Wei et al.

    Regression analysis of multivariate incomplete failure time data by modeling marginal distributions

    J Am Stat Assoc

    (1989)
  • P.K. Andersen et al.

    Cox's regression model for counting processesa large sample study

    Ann Stat

    (1992)
  • D.Y. Lin

    Cox regression analysis of multivariate failure time datathe marginal approach

    Stat Med

    (1994)
  • D. Clayton

    Some approaches to the analysis of recurrent event data

    Stat Methods Med Res

    (1994)
  • E.S. Fisher et al.

    Hospital readmission rates for cohorts of medicare beneficiaries in Boston and New Haven

    N Engl J Med

    (1994)
  • R.J. Glynn et al.

    Ways of measuring rates of recurrent events

    BMJ

    (1996)
  • D.R. Hoover

    Extension of the life table to repeating and changing events

    Am J Epidemiol

    (1996)
  • P. McCullagh et al.

    Generalized Linear Models

    (1989)
  • Cited by (45)

    • Heart Failure

      2016, International Encyclopedia of Public Health
    • Patient-level compared with study-level meta-analyses demonstrate consistency of D-dimer as predictor of venous thromboembolic recurrences

      2013, Journal of Clinical Epidemiology
      Citation Excerpt :

      The first empiric demonstrations of good fitting of Poisson regression compared with Cox regression in a single study on time-to-event data dates back many years [26,27]. Successive Poisson models, including fixed or random effects, were proposed as alternative meta-analytic approaches to pool rates and ratios calculated in follow-up studies and were often applied for studies with recurrent events [28,29]. When applied to recurrent event scenarios, Poisson and Cox analyses theoretically would answer two different clinical questions: the first looks at number of recurrences over a period, whereas the latter looks at time to (first) recurrence.

    • Inconsistent results in meta-analyses for the prevention of falls are found between study-level data and patient-level data

      2011, Journal of Clinical Epidemiology
      Citation Excerpt :

      One field where this is certainly the case is the prevention of accidental falls in hospitals. It has previously been argued that the rate of falls (number of falls by a patient divided by the length of time they are observed) is of greatest importance as an outcome in this field [9–12]. This is because considering only patients who fall or do not fall ignores the number of falls had by patients who had more than one (each of which having its own deleterious consequences).

    • Quantifying the magnitude of risk for balance impairment on falls in community-dwelling older adults: a systematic review and meta-analysis

      2010, Journal of Clinical Epidemiology
      Citation Excerpt :

      And, although the OR is a valid measure of association in its own right, when an outcome is as common as falls in older adults, the OR overestimates the RR. Finally, logistic regression may not produce a risk estimate that is as informative as other regression-based methods that take into account the recurrent nature of falling [74,75]. Balance impairment is associated with increased fall risk in community-dwelling older adults with overall summary risk values of RR 1.42 (1.08, 1.85) and OR 1.98 (1.60, 2.46).

    • Comparison of Regression Models for the Analysis of Fall Risk Factors in Older Veterans

      2009, Annals of Epidemiology
      Citation Excerpt :

      This is despite the fact that nearly twenty years ago, Cumming et al. (11) highlighted the public health importance of studying the rate of falls, which involves analyzing all falls occurring in a study, rather than studying the outcome of “fallers.” Stürmer et al. (12) discussed a number of analytic strategies to deal with recurrent event data, and pointed out that these techniques can now be implemented using standard statistical packages. In the context of falling among older adults, Robertson et al. (13) compared the Andersen-Gill extension of the Cox proportional hazards (PH) model, the marginal model, and the negative binomial regression model, and concluded that the negative binomial approach had advantages over survival analysis.

    View all citing articles on Scopus
    View full text