Analysis of Recurrent Event Data under the Case-Crossover Design with Applications to Elderly Falls
By Xianghua Luo, Ph.D.
Assistant Professor
Division of Biostatistics
University of Minnesota
School of Public Health
See Luo X and Sorock GS, Analysis of recurrent event data under the case-crossover design with applications to elderly falls. Statistics in Medicine, Published online in Wiley InterScience (www.interscience.wiley.com) DOI: 10.1002/sim.3171
Background
The case-crossover design is a method to answer the question, ¿Did the patient do anything unusual just before the onset of the disease?¿ For example, ¿Did you consume any alcoholic beverage over the 2-day period before a gout attack?¿
Since first proposed by Maclure1, this method has been proved to be useful for studying the effects of transient exposures on short-term risk of diseases. Several characteristics of the case-crossover design distinguish itself from other useful designs. First, in a case-crossover study, only data on people who have the disease are required (case only). Second, this design allows each subject to serve as his/her own control such that the association between disease and risk factors can be estimated by comparing the exposure of risk factors prior to the disease onset (case time) with that in a reference time (control time). Third, subject-level confounding variables, such as gender, are readily controlled because the matched case-control pair is from the same person. Because of the same reason, the case-crossover design is only aimed to identify triggers of disease rather than who is at the highest risk of disease.
Single Event versus Recurrent Event
The original design was proposed for single event data where each subject experiences the studied disease for only once. However, in many applications, repeated or recurrent events are frequently encountered. Examples include elderly falls, gout attacks and sexually transmitted infections. In such situations, the within-subject dependence among recurrent events has to be taken into account in the data analysis.
Methods
We reviewed three existing conditional logistic regression-based approaches for recurrent event data under the case-crossover design. A simple approach is to use only the first event for each subject; however, we would expect loss of precision in estimation because this method ignores the recurrent nature of the data and wastes all the second and later events. The other two reviewed methods rely on the assumption that the recurrent events within the same subject are independent conditionally on a set of subject-level covariates.
We also proposed two new methods. One method adjusts the traditional logistic regression using a computer-based resampling technique. The other method adjusts the formulation of estimation in the conditional logistic regression with a weight assigned to each event. The weight chosen is the reciprocal of the total number of events of each subject. For example, when an individual contributes five outcome vents to the data, each of the five events contributes information with one-fifth of the weight of an individual who contributes just one outcome event. Let¿s call the second method ¿weighted method¿. It was proved that the two new methods are approximately equivalent in certain cases.
We also proposed a weighted Mantel-Haenszel method for situations when the only risk factor is a binary (yes/no) variable.
Elderly Falls Data
We illustrated the application of the reviewed and proposed methods in the elderly falls data, which was collected from a total of 158 nursing home residents at three Maryland and Florida nursing homes. The scientific question of interest was to determine whether certain medication change was associated with risk of falls. Medication change record was retrieved for patients during 1-2 days before versus 8-9 days before each fall date. Two types of medication use change were of particular interest: central nervous system (CNS) medication and non-CNS medication.
We applied both the existing methods and proposed methods to the data. All methods found that significantly elevated fall risk followed CNS medication changes and no significant effect of non-CNS medication changes on fall risk was observed. The proposed methods showed satisfactory point and interval estimates of the relative risk.
To summarize, our proposed methods enriched the toolbox for analyzing recurrent event data under the case-crossover design and were shown to possess more flexibilities than the existing methods in the situations with unknown correlation structures among recurrent events.
1Maclure M. The case-crossover design: a method for studying transient effects on the risk of acute events. American Journal of Epidemiology 1991; 133:144-53.