A substantial literature exists documenting risk of injury from drinking prior to the event based on emergency department (ED) studies. These studies have primarily used a case–control study design in which noninjured patients serve as controls (Cherpitel, 1993a, 2007) or a case-crossover design in which patients serve as their own controls (Maclure, 1991; Mittleman et al., 1993), based on either their drinking during a predetermined period such as the previous day or the previous week (pair-matched) (Borges et al., 2006b; Vinson et al., 2003) or on their usual frequency of drinking over a period of time (Borges et al., 2006a). Despite the fact that these ED studies have used standardized protocols and instrumentation, with strict probability sampling of patients around the clock (Cherpitel, 2009), the magnitude of relative risk (RR) estimates for injury has varied across studies (Cherpitel, in press). Because no study has compared RR of injury, based on case–control estimates and case-crossover estimates using different approaches (pair-matched and usual frequency), in the same sample of ED patients, the impact on RR estimates of variation in study design and sampling is unknown.
The usual frequency approach to case-crossover analysis has typically generated larger risk estimates for injury due to alcohol consumption than either the pair-matched approach (Borges et al., 2004; Vinson et al., 1995) or the case–control approach (Ye et al., 2013; Zeisser et al., 2013) and has been attributed to recall bias, particularly among infrequent drinkers who underestimate their usual consumption (Stockwell et al., 2008; Ye et al., 2013). Additionally, recall bias in alcohol consumption using the pair-matched approach has generated mixed results.
To explore the impact of study design and sampling frame on RR of injury related to drinking prior to injury, estimates based on case–control analysis and case-crossover analysis, using both the last-week pair-matched approach and the usual frequency approach, are compared in a sample of injured and noninjured patients arriving at the ED on weekend evenings. Control-crossover analysis is also provided to determine biases in the case-crossover design, estimates from which may be used to apply adjustments to case-crossover estimates. These findings are important for informing the calculation of improved estimates of alcohol attributable fraction due to injury morbidity and ongoing work on comparative risk assessment in the global burden of disease (Rehm et al., 2010). Findings are also important in establishing safe drinking guidelines (Butt et al., 2011; Chikritzhs et al., 2011) and for determining economic costs related to excessive alcohol consumption (Bouchery et al., ).
- Top of page
- Materials and Methods
Table 1 shows the demographic and drinking characteristics of the injured and noninjured patients. Injured were significantly more likely to be male and under 30 years of age compared with noninjured (p < 0.001). Injured patients were also more likely to be white, but less likely to be married and less likely to have a lower household income. About 38% of injured patients reported drinking during the 6-hour period prior to injury, compared with 17% of noninjured patients before their illness event (p < 0.001). Injured patients also reported a higher volume of drinking during this time than noninjured among those who reported drinking. Injured patients were also more likely to report drinking during the same 6-hour period the previous week (30%) than noninjured (23%), but no significant difference was found in volume consumed during this time. Among injured patients, 21% sustained injuries related to a fall, 14% to sports or recreation, 11% to assault, 5% to a vehicular crash, and the remainder to a mixture of other types and causes.
Table 1. Demographic and Drinking Characteristics Between Injury and Noninjury Emergency Department Patients (%)
| ||Injury (n = 281)||Noninjury (n = 515)|
|50 and up||13.2||26.3|
|High school graduate or less||25.3||30.9|
|College graduate or higher||34.4||30.7|
|Any drinking before injury/illness||38.2||17.4***|
|Volume before event, if drinking|
|7 drinks or more||41.0||23.5|
|Any drinking same time last week||30.0||22.9*|
|Volume, if drinking|
|7 drinks or more||32.4||25.2|
Table 2 shows the ORs for injury from drinking, using case–control analysis and case-crossover analysis (unadjusted and adjusted using control-crossover analysis) based on both last week's drinking and usual frequency of drinking over the last 30 days.
Table 2. Odds Ratios (ORs) of Injury Associated with Acute Drinking Estimated from Case–Control and Case-Crossover Methods
|Adjusted ORs||2.7 (1.9, 3.8)†|
|Model 1: any versus none|
|1–2 drinks versus none||1.9 (1.1, 3.3)**|
|3–6 drinks versus none||3.0 (1.7, 5.1)†|
|7 or more versus none||3.9 (2.2, 7.0)†|
|Model 1: any versus none||1.6 (1.0, 2.6)**|
|1–2 drinks versus none||1.0 (0.5, 1.9)|
|3–6 drinks versus none||1.9 (1.0, 3.6)***|
|7 or more versus none||2.7 (1.2, 5.8)**|
|Model 1: any versus none||0.5 (0.3, 0.8)***|
|1–2 drinks versus none||0.6 (0.4, 1.1)|
|3–6 drinks versus none||0.4 (0.2, 0.8)**|
|7 or more versus none||0.4 (0.2, 0.9)**|
|Adjusted case-crossover ORs|
|Model 1: any versus none||3.2 (1.7, 6.0)†|
|1–2 drinks versus none||1.6 (0.7, 3.7)|
|3–6 drinks versus none||4.6 (1.7, 12.3)***|
|7 or more versus none||7.1 (2.2, 22.9)***|
|Model 1: any versus none||10.7 (8.0, 14.3)†|
|Model 1: any versus none||4.8 (3.7, 6.3)†|
|Adjusted case-crossover ORs|
|Model 1: any versus none||2.2 (1.5, 3.3)†|
Case–Control Versus Pair-Matched Case-Crossover Determination of Risk
Although both are significant, the OR estimated from case–control analysis (2.7) was greater than that estimated from case-crossover analysis based on last week's drinking (1.6). The control-crossover pair-matched estimate (OR = 0.5) suggests that the OR estimated from case-crossover analysis is an underestimate, as, if no biases were present, the control-crossover estimated would be expected to be 1. When the control-crossover pair-matched estimate is used to “adjust” the case-crossover estimate, the case-crossover estimate is increased to 3.2 (1.6/0.5), which is now nearer the 2.7 case–control estimate.
Both the case–control and case-crossover analysis demonstrated a dose–response relationship for drinking 1 to 2, 3 to 6, and 7 or more drinks. When the case-crossover estimates are adjusted from respective control-crossover estimates, as the volume consumed prior to injury increases, so does the proportional increase in the ORs, reaching nearly twice the risk of injury compared with case–control analysis at 7 or more drinks during the 6 hours prior to injury, and over twice the risk compared with unadjusted case-crossover analysis.
Usual Frequency Case-Crossover Determination of Risk
The OR for injury risk base on case-crossover usual frequency of drinking (OR = 10.7) was substantially greater than that from either case–control or case-crossover pair-matched analysis, but when adjusted from control-crossover usual frequency analysis (OR = 2.2), was quite similar to the case–control estimate (OR = 2.7), and less than the adjusted case-crossover pair-matched estimate (OR = 3.2).
- Top of page
- Materials and Methods
Comparing risk of injury based on case–control analysis with that based on pair-matched case-crossover analysis, using drinking during the same time the previous week as the control period, a higher OR was found for case–control analysis (2.7 vs. 1.6). Medical patients in the ED may not be the best controls for injured patients, however, because some conditions treated in the ED, such as liver cirrhosis, are directly linked to long-term chronic heavy drinking and would result in an underestimate of risk of injury due to drinking. Conversely, medical patients may choose not to drink due to health problems, the so-called “sick quitter” effect (Shaper, 1990) or may be less likely to recall alcohol use prior to a medical problem than an injury, both of which would result in an overestimation of risk of injury from drinking.
Additionally, when using medical patients as controls, while some known (e.g., demographic) characteristics which may influence the alcohol–injury relationship can be controlled, others, such as dispositional characteristics (risking taking/impulsivity and sensation seeking), for example, are more difficult to account for, and may be expected to either confound an observed association between alcohol and injury or, alternatively, modify the relationship (Cherpitel, 1993b, 1999; McLeod et al., 2003).
Control-crossover analysis demonstrated that case-crossover analysis resulted in an underestimation of risk of injury, because the OR from control-crossover analysis was less than 1, as expected if no bias was present. When the pair-matched control-crossover estimate was used to “adjust” the case-crossover estimate, the OR increased to 3.2, and was larger than that derived from case–control analysis. Interestingly, although underestimation of risk across different levels of consumption prior to injury did not appear to vary markedly in control-crossover analysis, these adjustments, when applied to the pair-matched case-crossover dose–response analysis, resulted in substantial increases in risk of injury for each volume level, and was most marked for those consuming 7 or more drinks prior to injury, suggesting that such adjustments may be more important at higher levels of consumption.
Unlike case–control analysis, case-crossover analysis controls potential confounding of the alcohol–injury relationship from stable within-person risk factors (e.g., demographic and dispositional characteristics and usual substance use patterns); however, other potential biases are possible when using the pair-matched design, depending on the control period selected. For example, when using the same 6-hour period the week prior as the control, bias can occur when the case and control periods differentially fall at times when the likelihood of drinking may vary (e.g., holidays or other celebrations); similarly, using the same 6-hour period the previous day as the control, the likelihood of drinking may vary by day of the week (e.g., Friday evening vs. Saturday evening).
The OR for risk of injury from case-crossover usual frequency analysis was substantially larger (10.7) than that from either case–control (2.7) or pair-matched case-crossover (1.6) analysis, as found on separate samples in prior studies (Borges et al., 2004; Vinson et al., 1995; Zeisser et al., 2013). When the case-crossover usual frequency estimate was “adjusted” based on control-crossover analysis, however, the estimate was slightly less (OR = 2.2) than the case–control estimate (OR = 2.7) and less than the “adjusted” pair-matched case-crossover estimate (OR = 3.2).
A prior analysis comparing risk of injury from the case–control approach and the case-crossover usual frequency approach in 15 ED studies across 7 countries, all using representative samples of patients, found the 2 estimates were similar after adjusting the case-crossover estimate from control-crossover analysis (OR = 2.08 in case–control and 2.11 in adjusted case-crossover analysis), with both estimates similar to the respective estimates found here (Ye et al., 2013).
Because sampling in the present study occurred on weekend evenings, analysis based on the usual frequency of drinking resulted in inflated ORs for both case-crossover (10.7) and control-crossover (4.8) analysis, underscoring that sampling issues are important to consider in case-crossover analysis based on the usual frequency approach.
When control-crossover estimates are used to adjust for case-crossover estimates, the assumption is made that the same magnitude of bias exists for both case and control samples, and this may be a potential issue given differences in demographic characteristics between the injured and noninjured patients observed in Table 1. For example, although the control-crossover estimates were similar between men and women (ORs = 0.5 and 0.6, respectively), they were different across age groups (ORs = 0.7, 0.2 and 1.0, for <30, 30 to 49 and 50+, respectively) (not shown). While age and gender were controlled in these analyses, other uncontrolled demographic differences between cases and controls may have influenced study findings, and this incomparability issue has not been fully addressed in the literature on the case-crossover method.
Another potential source of bias in this study comes from missing data on last-week drinking which was used as the control period exposure. About 11% of the injured patients and 9% of the noninjured did not answer this question, while only about 2% of both samples had missing data on drinking before the event. Among those with missing data on the last-week drinking question, 45% of the injured reported alcohol use before the event, while 25% of the noninjured reported drinking prior to their illness, and neither of these were significantly different from their respective rate of reporting drinking prior to the event reported in Table 1, suggesting results here may not be affected greatly by the missing data.
Last, both the case-crossover and case–control approaches are subject to potential bias due to contextual factors that may be critical in determining risk of injury, and such factors were not taken into account in these analyses. Such environmental factors or activities may be independently correlated with both alcohol use and injury risk, and bias may occur if context is not represented equally in the case and control period comparisons (Watt et al., 2006), an issue that has been virtually ignored in empirical research (Stockwell et al., 2002).
- Top of page
- Materials and Methods
Risk of injury based on case–control analysis in the ED may result in either underestimates or overestimates, and while case–control estimates cannot be relied on as the “gold standard,” data from noninjured controls are important for applying adjustments derived from control-crossover analysis to case-crossover estimates, particularly important at higher levels of consumption. When using the pair-matched case-crossover approach, an average OR based on multiple matching of control periods, using both the previous day and the previous week, would appear optimal to circumvent biases related to the differential likelihood of drinking during predetermined time periods.
Future research on risk of injury from drinking should include representative samples of patients in comparing different approaches to risk estimation at various volumes of consumption prior to injury, controlling for context where possible. Such research is important for improved estimates of alcohol attributable fraction for injury morbidity to inform comparative risk assessment for determining the global burden of disease, establishing safe drinking guidelines, and determining economic costs related to excessive alcohol consumption.