Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.
Commentary on the paper by Schwartz (p. 956)
The case-crossover design has become an accepted method for investigating acute health consequences of transient exposures, or “triggers”. This design, originally presented in 1991 by Maclure,1 can be thought of as a variant of the conventional case-control study with individual (pairwise) matching. The principal difference is that a case-crossover study only includes cases, where each serves as his own referent. This feature has a compelling methodological appeal in that it allows for efficient control of potential confounders that are fixed (for example, genetic traits) or change minimally or slowly over time (for example, socioeconomic status), and are difficult or impractical to measure. Who better to provide control for confounding than the case himself?
To date, the case-crossover design has been applied most often to investigate environmental exposure effects in studies of air pollution, as an alternative to time series analyses. The paper by Schwartz2 in this issue exemplifies this application, and also illustrates the utility of the case-crossover method for controlling for seasonal and long term air pollution time trends. There have also been several applications to occupational studies of acute outcomes, primarily injuries, such as a recent investigation of risks for acute hand injuries associated with working with malfunctioning tools, accelerated work pace, and distractions.3 Further applications of this design for investigating acute workplace injuries and illnesses are clearly warranted, as other authors have stated.4,5
There are some important practical and methodological considerations for designing and implementing case-crossover studies of workplace hazards. Obviously, data on acute health outcomes need to be recorded and retrievable for this design to be successful. Some workplaces routinely record acute events, such as injuries that require medical treatment. Data reporting is sometimes performed to comply with government regulations, and some industries also maintain illness and injury reporting systems for medical benefits and other internal purposes. Case-crossover analyses of routinely tabulated outcomes, linked with relevant historical exposure data, can then be undertaken fairly expeditiously without large resource expenditures. The validity of these analyses will nonetheless require accurate classification of health outcomes and exposures in order to minimise bias. Data on acute health effects other than injuries, such as symptoms or transient changes in physiological functions, will require specially designed data collection protocols, typically involving detailed, repeated assessments of workers’ health and exposure status. This may entail physical examinations or repeated physiological measurements, which will ordinarily be resource intensive and can impose severe limitations on study size. Exposure assessment requirements for transient exposure changes can also engender substantial costs and logistical complexity, beyond what ordinarily are encountered in standard cross-sectional or longitudinal studies. In particular, personal level continuous exposure monitoring or repeated biomonitoring may be necessary for accurate characterisation of short term environmental changes.
There have been some notable advances in case-crossover design methodology since the initial applications in the early 1990s, yet some issues have not been fully resolved. Among these, unbiased selection of the aetiologically relevant and representative time periods for the “case” and “control” intervals, respectively, has probably received greatest attention. The case interval is generally defined as the time period immediately preceding or in close temporal proximity to the health event. Allowance for a lag period between the case interval and the event will be important when the health outcome of interest is a delayed effect of exposure. The more difficult choice is the selection of the correct control period, which should represent the case’s expected exposure profile, much as the exposure experience of a control in a case-control study is assumed to reflect the norm. There is now fairly wide recognition that limiting selection of control periods to times preceding the event can be biased when there are secular exposure trends. Air pollution research provided the impetus for this concern, but it is not hard to imagine the same type of bias occurring in an occupational study. Improved workplace ventilation, introduction of personal protective equipment, and changes in production are all possible sources of secular exposure trends. Accordingly, bi-directional control interval selection, in which control periods are selected from times before and after the event, has been shown to be a reasonable strategy for reducing such bias.6 The validity of this approach requires the assumption that the case event will not influence subsequent exposure. This assumption is certainly likely to hold true for a study of daily counts of mortality or hospital admissions related to daily changes in air pollution (which cannot be modified easily or rapidly), but may not apply in some occupational settings. To illustrate, consider the possible consequences of a limb amputation in a specific workshop on subsequent work and safety enforcement practices. Analogous health event influenced exposure modifications might be instituted following recognition of outbreaks of acute illness or symptom reports. Other nuances of control selection may involve matching case and control times on season, day of week, or work shift if there is prior evidence that these time related factors are potential confounders. Of course, care should be taken to avoid over-matching on an exposure related time factor.
The case-crossover design can offer a versatile approach for studying a range of work related health outcomes. In addition to applications that address illness and injury aetiology, this method may also prove to be advantageous for occupational health surveillance, at the population-wide level and in specific workforces. Newly hired workers without past exposure to agents of concern should be especially valuable groups for case-crossover studies designed for industry specific surveillance programmes. Future, expanded use of this design in occupational epidemiology should be encouraged.
Commentary on the paper by Schwartz (p. 956)