Sample selection bias in the emergency room: an examination of the role of alcohol in injury.

Abstract

Injury location, injury cause and patient drinking patterns were used to predict blood alcohol content (BAC) and self-reported drinking before injury using emergency room (ER) data. Models estimating both BAC and self-reported drinking among emergency room injury patients were used; the ER sample was also compared to an injured sample from the general population. Data were from three of six communities participating in the project "Preventing Alcohol Trauma: a community trial". ER data were collected from nine hospitals on Friday and Saturday nights between 6 p.m. and 2 a.m. on alternate weekends from June 1992 to December 1995. Telephone survey data were collected between April 1992 and March 1996. Drinking measures included drinking frequency, drinks per occasion, and variance. Other measures involved injury time, location, and type; drinking before and after injury; and age, race, gender, education, marital status and household income. Model estimation corrected both for selection bias and censoring of the dependent measure. The results indicate: (1) ER populations tended to be female, less well educated, non-white, poor and younger; (2) there were significant selection bias effects in the ER sample; (3) assaults were more likely to involve drinking than other injury types; (4) drinking patterns were significant non-linear predictors of alcohol involvement; and (5) self-reported drinking before injury was both a sensitive and specific indicator of measured BAC. Assaults uniquely involve the use of alcohol and selection bias may threaten ER study validity.