Validation of a Single Screening Question for Problem Drinking

Patients

Of 3616 injured persons presenting for care to one of the participating emergency centers during times when research staff were present, 579 were excluded, because the injury occurred more than 48 hours before, because of mental status changes (chronic or acute), or because the injury occurred in a controlled environment. Of 3037 eligible patients on covered shifts, 12.2% declined to participate, and 15.4% were missed, either because their injuries were severe enough to preclude an interview in the emergency department (8.1%) or because research staff were busy with other interviews (7.3%); 2199 persons were interviewed during covered emergency department shifts from February 1998 through January 2000.

Some injured patients were missed because of the severity of their injuries. Therefore, we recruited additional patients who had been admitted to the hospital during times not covered in the emergency department by study staff. These interviews were conducted from June 1999 through March 2000. A total of 618 were identified: 52 refused (8.4%); 69 were too severely injured (11.2%); and 139 were missed (22.5%), leaving 358 who were interviewed.

We combined these 2 groups (from covered and noncovered emergency department shifts) and excluded those from noncovered shifts who had minimal effect on the results presented here. Of those 2557 interviews, 40 were with patients who had been injured and interviewed before. We excluded these from analysis, leaving 2517 individual patients. Table 1 shows basic demographic data and prevalence of alcohol problems. Seven patients (0.3% of interviews) did not answer the single problem-drinking screening question, and 13 (0.5%) did not complete the calendar-based review of recent drinking.

Statistical Analysis

We used bivariate analysis to calculate sensitivity and specificity, and confidence interval (CI) analysis¹⁷ to determine 95% CIs. We used the c statistic from logistic regression to calculate the area under the receiver-operating characteristic (ROC) curve¹⁸ in bivariate and multivariate models. We used the formula in Hanley and McNeil1⁹ to calculate 95% CIs around the area under the ROC curve.*

Results

For detecting problem drinking, the single question (with “within the last 3 months” considered positive) had a sensitivity of 86% and a specificity of 86% Table 2. Sensitivity was higher in men; specificity was higher in women. The question was better in detecting hazardous drinking than alcohol use disorder and more effective in whites than in African Americans. In this study with a prevalence of problem drinking of 35%, the positive and negative predictive values were 77% and 92%. In a clinical setting with a prevalence of 15% (the national prevalence of past-month heavy drinking⁴), positive and negative predictive values would be 52% and 97%. Likelihood ratios18 are provided in Table 3.

Breath or blood alcohol levels were obtained for 2335 patients; 257 had some alcohol detected, and 139 had a level of 0.22 mmol/L (0.1 g/dL) or greater. Of 1493 patients without hazardous drinking or an alcohol use disorder only 49 had any alcohol detected, a specificity of 96%. However, the sensitivity of alcohol testing was only 24% (198 with a positive breath or blood alcohol level out of 842 with problem drinking). Interestingly, 27 of the patients with an alcohol level higher than 0.22 mmol/L (indicative of problem drinking) were negative by the criterion standards (22 of the 27 screened positive with the single question).

Tobacco use is common among patients with alcohol use disorders²⁰ and was associated with problem drinking in our study. Almost half the patients in this study used tobacco in some form; 1013 were cigarette smokers, and 139 used only other forms of tobacco. Current tobacco use had a sensitivity 65% and a specificity of 64% in identifying problem drinking.

The single screening question had 4 answer options (never, >12 months ago, 3-12 months, and within the last 3 months). Using the 3 cut points defined by those answer options, the area under the ROC curve²¹ for identifying problem drinking was 0.90 (95% CI, 0.88-0.91). With a past-year alcohol use disorder as the diagnostic criterion, the area under the ROC curve was 0.81 (95% CI, 0.79-0.83). Entering the single question, sex, age (continuous or ordinal), and tobacco use as independent variables, the area under the ROC curve increased minimally to 0.91 Figure 1. Entering race/ethnicity, injury severity score,²² or alcohol level into the model had essentially no effect. Considering only current drinkers (n=1535), the single question had a sensitivity for detecting problem drinking of 86% and a specificity of 69%, and the area under the ROC curve was 0.79 (95% CI, 0.77-0.82).

Discussion

A single question about recent heavy drinking has clinically useful sensitivity and specificity for detecting recent hazardous drinking and current alcohol use disorders. In response to “When was the last time you had more than X drinks in one day?”, an answer within the past 3 months has a sensitivity and a specificity of 86%. An answer of “never” with a sensitivity of more than 99% essentially rules out problem drinking. The question has less utility in African American patients, but works equally well in women and men.