Validation of key behaviourally based mental health diagnoses in administrative data: suicide attempt, alcohol abuse, illicit drug abuse and tobacco use

Three random samples of patients were selected, each stratified by geographic region, gender, and year of cohort entry, from a VHA depression treatment cohort from April 1, 1999 to September 30, 2004. The first sample was selected from patients who died by suicide, the second from patients who remained alive on the date of death of suicide cases, and the third from patients with a new start of a commonly used antidepressant medication. Four variables were assessed using administrative codes in the year prior to the index date: suicide attempt, alcohol abuse/dependence, drug abuse/dependence and tobacco use.Specificity was high (≥ 90%) for all four administrative codes, regardless of the sample. Sensitivity was ≤75% and was particularly low for suicide attempt (≤ 17%). Positive predictive values for alcohol dependence/abuse and tobacco use were high, but barely better than flipping a coin for illicit drug abuse/dependence. Sensitivity differed across the three samples, but was highest in the suicide death sample.Administrative data-based diagnoses among VHA records have high specificity, but low sensitivity. The accuracy level varies by different diagnosis and by different patient subgroup.In many administrative data-based studies, most variables, including primary outcomes, are based on the International Classification of Diseases (ICD), clinical modification codes (ICD-9 or ICD-10). Although the criteria for these diagnoses codes are clearly delineated, the primary purpose of these codes in many health systems is for billing. However, these codes are often used in health services research.Administrative codes may have issues with sensitivity, specificity or accuracy when used for research purposes. Some level of financial incentive exists for the clinicians and billing clerks to note all relevant diagnoses (e.g., slightly higher billing revenues for the clinician or the system). However, even if each diagnosis or condition of interest covered in an encounter is fai