The case-crossover design was introduced to evaluate the effect of transient exposures on the risk of acute outcomes.1,2 One application has been in the studies of emergency room patients to estimate the risk of injury related to alcohol consumption. Unlike traditional case-control studies using noninjury emergency room patients3 or community samples4 as controls, the case-crossover design compares injury patients’ drinking before the event with their own alcohol intake during an earlier control period. Two commonly used approaches to assess exposures for the control period are the usual frequency of drinking (eg, within the last year)5–8 and drinking during one or several matched prior intervals (eg, the same time 1 week before injury).6,9–11
Results from previous studies, however, suggest that risk estimates produced using the usual-frequency case-crossover method might be biased upward when compared with other methods. For example, the usual-frequency method generated larger estimates than the matched interval case-crossover method in the two studies using both approaches.6,9 A usual-frequency case-crossover analysis on emergency room data in 16 countries produced a pooled odds ratio (OR) of 5.7 (95% confidence interval [CI] = 4.0–8.0) for injury related to any drinking,7 substantially larger than the pooled OR of 1.6 (1.2–1.9) from a five-country case-control study using noninjury patient controls.12
This study examines potential bias using the usual-frequency case-crossover method compared with the case-control design for estimating the risk of injury from drinking. A case-crossover analysis on control (ie, noninjury) patients (called the control-crossover approach) is also conducted for comparison. Because the excess risk with control-crossover is expected to be zero, the approach has been widely used in case-crossover studies as a validity check and to adjust biased estimates.9,13–18
Our analyses are based on data from the Emergency Room Collaborative Alcohol Analysis Project,12,19 including 15 studies (each typically covering a city or region) in seven countries. Data from probability samples of emergency room patients admitted for injury or illness were collected using a similar methodology.20 Using data from various regions around the world, findings from these analyses are more robust, with risk estimates examined in various socioeconomic and cultural conditions thought to be associated with alcohol use and injury, as well as varying levels of emergency room utilization.
Cases and controls were defined as patients who reported their primary reason for visiting the emergency room as either injury or noninjury illness. Injury and noninjury patients were compared in case-control analysis, whereas only injury patients were used in the case-crossover analysis and only noninjury patients in the control-crossover analysis. For the injury and noninjury samples, exposure to acute drinking was defined as consumption of any alcoholic beverage by the patient during the 6-hour period before either the injury or the illness event that led to their emergency room visit.
In case-control analysis, adjusted ORs were estimated from unconditional logistic regressions comparing acute drinking between injury and noninjury patients, controlling for sex and age. For the usual-frequency case- or control-crossover analysis, risk estimates were generated comparing acute drinking of the injury or noninjury patients before the event to their usual frequency of drinking. The case- or control-crossover analysis can be considered as a stratified self-matched case-control analysis with the OR generated from the Mantel-Haenszel estimator for dichotomous exposure21 (any drinking vs. none). Maclure1 showed that for the usual-frequency case-crossover analysis, the OR can be derived by dividing the total expected unexposed periods of the acutely exposed cases by the total expected exposed periods of the acutely unexposed cases. The expected exposed periods were obtained from subjects’ usual frequency of drinking during last 12 months. To be consistent with the acute drinking assessment, a 6-hour period was also used as the effect period for each occasion of any alcohol use in the control period. The expected unexposed periods were derived by subtracting the expected number of exposed periods from the total number of possible 6-hour periods each year, defined as 365×3, which excluded one 6-hour sleeping period each day.22
The control-adjusted case-crossover estimates were obtained by dividing ORs from the case-crossover by those from the control-crossover analysis. In addition to study-specific estimates, we estimated the pooled overall effect using a meta-analytic method23 of weighted average, taking into account the standard errors of the study-specific estimates. Only random-effect estimates are reported in Table 1, given that the homogeneity test yielded very small P values.
As shown in Table 1, the usual-frequency case-crossover method produced ORs at least 1.5 times as large as the corresponding estimates from case-control analyses for 11 of the 15 emergency room studies, and 2 times the case-control results in 8 studies. The pooled OR was 4.7 (95% CI = 2.6–8.5) from case-crossover analysis compared with 2.1 (1.6–2.7) from case-control analysis. This trend with case-crossover estimates was also seen in the control-crossover analysis. ORs were >1.5 for 13 studies and >2 for 9 studies. The control-crossover pooled OR was 2.2 (1.8–2.8), substantially larger than the expected estimate of 1. After adjusting for the case-crossover estimate, based on the control-crossover estimate, the pooled adjusted case-crossover OR was 2.1 (1.5–3.1)—very close to 2.1 (1.6–2.7) from case-control analysis. The across-study Pearson’s correlation between the case-control and the unadjusted and adjusted case-crossover estimates (logarithm of ORs) improved from 0.75 to 0.84.
To examine the potential source of upward bias in the case-crossover estimates, we performed separate control-crossover analyses for subsamples at various drinking-frequency levels. For each frequency level, with all studies combined, the control-crossover ORs were quite similar to the ratio between the observed and expected exposure prevalence (Table 2). ORs generally decreased monotonically, from the lowest to the highest usual-frequency level. The largest bias was found among those reporting the least frequent drinking (OR = 15 for one to five times last year), whereas the estimate for “daily drinkers” was close to 1. Control-crossover analyses were also performed for each of the six countries separately (Italy was dropped because of a small sample size), although with a slightly modified usual-frequency measure reduced to four categories. As shown in the Figure, clear monotonic relationships between the control-crossover estimates and usual frequency of drinking were observed for four of the six countries. One country exhibited no relationship (Spain, with log ORs close to zero for each usual-frequency level) and one appeared to be variable but with log ORs all above zero (Poland).
Our comparisons between the case-control and case-crossover analysis and the control-crossover estimates show that the usual-frequency case-crossover method apparently overestimates the risk of injury related to drinking. In particular, the observed monotonic relationship between control-crossover estimates and usual-frequency levels (where larger biases were associated with less frequent drinking) may be most plausibly explained by recall bias. Survey estimates assessing annual alcohol volume, particularly estimates derived from usual-frequency and quantity measures, have consistently been found to account for only a fraction of per capita consumption from sales data.24 Our results suggest that the bias is more likely among persons with a less consistent drinking pattern, who may thus have more difficulty recalling their drinking over time, particularly over an assessment period as long as 12 months.
One possible alternative explanation for the larger case-crossover estimates observed is that, among noninjury emergency room patients, some illnesses (eg, myocardial infarction)25 may be related to alcohol use before the event. To the extent this is true, estimates from the control-crossover analyses represent at least some elevated risk of noninjury illness associated with acute drinking, and case-control analyses may thus underestimate the risk of injury to some extent. However, the clear monotonic relationship between usual-frequency and control-crossover estimates is the stronger evidence for recall bias than an effect of drinking on noninjury illness.
The cross-study pooled estimate from case-control analysis is virtually the same as that from case-crossover analysis after adjusting for potential bias suggested by control-crossover analysis (OR = 2.1 for both). Not all study-specific results are similar, however. For example, adjusted case-crossover ORs in the three Mexican studies are all larger than case-control estimates, suggesting some inconsistencies not caused by random error. There has been a debate regarding whether the two designs are comparable, with the case-control study asking “why is the event happening to me versus another person,” whereas the case-crossover analysis asks “why now versus another time.”26 Conversely, both case-control and case-crossover designs can be linked to an underlying cohort study,21,27 and hypothetical examples have been constructed to show the equivalence of estimates from the two approaches if they follow the same assumed data-generating process.27 While the results from the two approaches might not necessarily converge, large discrepancies between their estimates suggest bias; here, recall bias seems a plausible candidate. This potential bias is not restricted to the usual-frequency case-crossover design for estimating the alcohol-injury relationship; it would potentially apply to any exposure based on self-reported data over a long retrospective window.
1. Maclure M. The case-crossover design: a method for studying transient effects on the risk of acute events. Am J Epidemiol. 1991;133:144–153
2. Mittleman MA, Maclure M, Tofler GH, Sherwood JB, Goldberg RJ, Muller JE. Triggering of acute myocardial infarction by heavy physical exertion. Protection against triggering by regular exertion. Determinants of Myocardial Infarction Onset Study Investigators. N Engl J Med. 1993;329:1677–1683
3. Cherpitel CJ. Alcohol and injuries: a review of international emergency room studies. Addiction. 1993;88:923–937
4. Stockwell T, McLeod R, Stevens M, Phillips M, Webb M, Jelinek G. Alcohol consumption, setting, gender and activity as predictors of injury: a population-based case-control study. J Stud Alcohol. 2002;63:372–379
5. Borges G, Cherpitel C, Mittleman M. Risk of injury after alcohol consumption: a case-crossover study in the emergency department. Soc Sci Med. 2004;58:1191–1200
6. Borges G, Cherpitel CJ, Mondragón L, Poznyak V, Peden M, Gutierrez I. Episodic alcohol use and risk of nonfatal injury. Am J Epidemiol. 2004;159:565–571
7. Borges G, Cherpitel CJ, Orozco R, et al. Acute alcohol use and the risk of non-fatal injury in sixteen countries. Addiction. 2006;101:993–1002
8. Cherpitel CJ, Ye Y, Moskalewicz J, Swiatkiewicz G. Risk of injury: a case-crossover analysis of injured emergency service patients in poland. Alcohol Clin Exp Res. 2005;29:2181–2187
9. Vinson DC, Mabe N, Leonard LL, et al. Alcohol and injury. A case-crossover study. Arch Fam Med. 1995;4:505–511
10. Vinson DC, Maclure M, Reidinger C, Smith GS. A population-based case-crossover and case-control study of alcohol and the risk of injury. JStud Alcohol. 2003;64:358–366
11. Borges G, Cherpitel C, Orozco R, et al. Multicentre study of acute alcohol use and non-fatal injuries: data from the WHO collaborative study on alcohol and injuries. Bull World Health Organ. 2006;84:453–460
12. Cherpitel CJ, Bond J, Ye Y, Borges G, Macdonald S, Giesbrecht N. A cross-national meta-analysis of alcohol and injury: data from the Emergency Room Collaborative Alcohol Analysis Project (ERCAAP). Addiction. 2003;98:1277–1286
13. Marshall RJ, Wouters S, Jackson RT. A case-crossover analysis of a case-control study of alcohol consumption and coronary events: the effects of exposure definition and the use of control data. J Epidemiol Biostat. 2000;5:367–373
14. Hallqvist J, Möller J, Ahlbom A, Diderichsen F, Reuterwall C, de Faire U. Does heavy physical exertion trigger myocardial infarction? A case-crossover analysis nested in a population-based case-referent study. Am J Epidemiol. 2000;151:459–467
15. Suissa S. The case-time-control design. Epidemiology. 1995;6:248–253
16. Suissa S. The case-time-control design: further assumptions and conditions. Epidemiology. 1998;9:441–445
17. Greenland S. A unified approach to the analysis of case-distribution (case-only) studies. Statistics in Medicine. 1999;18:1–15
18. Hernández-Díaz S, Hernán MA, Meyer K, Werler MM, Mitchell AA. Case-crossover and case-time-control designs in birth defects epidemiology. Am J Epidemiol. 2003;158:385–391
19. Cherpitel CJ, Ye Y, Bond J. Alcohol and injury: multi-level analysis from the emergency room collaborative alcohol analysis project (ERCAAP). Alcohol Alcohol. 2004;39:552–558
20. Cherpitel CJSGiesbrecht N, Gonzales R, Grant M, Österberg E, Room R, Rootman I, Towle L eds. A study of alcohol use and injuries among emergency room patients. In: Drinking and Casualties: Accidents, Poisonings and Violence in an International Perspective.. 1989 New York Tavistock/Routledge:288–299
21. Rothman KJ, Greenland S Modern Epidemiology. 19982nd ed Philadelphia, PA Lippincott-Raven Publishers
22. Maclure M, Mittleman MA. Should we use a case-crossover design? Annu Rev Public Health. 2000;21:193–221
23. Warren FC, Abrams KR, Golder S, Sutton AJ. Systematic review of methods used in meta-analyses where a primary outcome is an adverse or unintended event. BMC Med Res Methodol. 2012;12:64
24. Rehm J. Measuring quantity, frequency, and volume of drinking. Alcohol Clin Exp Res. 1998;22(2 suppl):4S–14S
25. Gerlich MG, Krämer A, Gmel G, et al. Patterns of alcohol consumption and acute myocardial infarction: a case-crossover analysis. Eur Addict Res. 2009;15:143–149
26. Maclure M. ‘Why me?’ versus ‘why now?’—differences between operational hypotheses in case-control versus case-crossover studies. Pharmacoepidemiol Drug Saf. 2007;16:850–853
Copyright © 2013 Wolters Kluwer Health, Inc. All rights reserved.
27. Greenland S. Confounding and exposure trends in case-crossover and case-time-control designs. Epidemiology. 1996;7:231–239