Study data were obtained from the VA’s National Registry for Depression (NARDEP) which was developed by the VA’s Serious Mental Illness Treatment Research and Evaluation Center (SMITREC) in Ann Arbor, Michigan. The study was approved by the Institutional Review Board of the Veterans Affairs Ann Arbor Health System.
Patient Population/Observation Days
The study sample consisted of all 887,859 patients in NARDEP who received either two depression diagnoses or a diagnosis of depression and an antidepressant fill between April 1, 1999 and September 30, 2004. Patients were excluded if they received any diagnoses of bipolar I, schizophrenia, or schizoaffective disorder during this period. Depression diagnoses were identified using the ICD-9 codes: 296.2x, 296.3x, 296.90, 296.99, 298.0, 300.4, 311, 293.83, 301.12, 309.0, or 309.1.
Observation-days for the main study analyses began on the dates of patients’ first treatment event of interest (discharge from a psychiatric hospitalization, new antidepressant start, other start, or dose change) and continued until 60 weeks following their last treatment event of interest, their date of death as indicated in the National Death Index (NDI), or the end of the study period (September 30, 2004), whichever came first.
Treatment Events: Psychiatric Hospitalizations, New Antidepressant Starts, Other Starts and Dose Changes
We ascertained dates of discharge from psychiatric hospitalizations, defined as hospitalizations with a primary or principal psychiatric discharge diagnosis of ICD-9 codes 290.x – 319.x or hospitalizations with bed section codes indicating a psychiatric stay.
Patients were considered to have received an antidepressant medication if they filled a prescription within the VA system for any of the following: amitriptyline, amoxapine, atomoxetine, bupropion, citalopram, clomipramine, desipramine, doxepin, duloxetine, escitalopram, fluoxetine, fluvoxamine, imipramine, isocarboxazid, maprotiline, mirtazapine, nefazodone, nortriptyline, paroxetine, phenelzine, protriptyline, sertraline, tranylcypromine, trazodone, trimipramine, and venlafaxine. As in prior studies, we considered trazodone, mirtazapine, amitriptyline, and nortriptyline to have been used as antidepressants rather than for other purposes only if the doses were ≥ 300 mg/day, ≥ 15 mg/day, ≥ 75 mg/day, or ≥ 25 mg, respectively.
A new antidepressant start was defined as a fill of an antidepressant medication that occurred after a “clean period” of ≥ 6 months without any antidepressant fills. Other antidepressant starts were fills that occurred during ongoing treatment, either without a clean period or with a short clean period (< 6 months). Otherstarts included: 1) antidepressant switches, defined as the discontinuation of one antidepressant followed by the initiation of a second in ≤ 30 days; 2) combination treatment, the addition of a second antidepressant to ongoing treatment with the first agent, and 3) and other starts with < 6 months free from all antidepressant use. We did not include “restarts” of the same antidepressant medication within a 6 month period as an “other start” of interest, as this might be due to short gaps in use due to incomplete adherence or to a restart of a previously tolerated medication. Antidepressant starts that occurred in the 6 months prior to cohort entry were all considered to be other starts and used to categorize the initial days following cohort entry, as described below. A significant change in antidepressant dose was defined as a ≥ 50% difference in total dose between two consecutive fills of a specified antidepressant occurring within 6 months. Because we had access only to data on outpatient medication fills (including discharge medications from inpatient stays), we identified new and other antidepressant starts that occurred during inpatient stays at the time of the patient’s discharge, by comparing pre and post hospital medications.
Risk Periods Following Treatment Events
We classified patient observation days into discrete “risk periods” based on their proximity to treatment events of interest occurring during the study period or, for patients newly entering the cohort, by proximity to events in the 6 months prior to study entry. Patient-days were classified by whether they fell within one of five sequential 12-week periods following each type of treatment event (new antidepressant starts, other starts, dose changes, or discharges from inpatient psychiatric stays). Days were classified as being within: 1) 1–84 days (12 weeks) following the treatment event, 2) 85–168 days (13–24 weeks) following the event, 3) 169–252 days (25–36 weeks) following the event, 4) 253–336 days (37–48 weeks) following the event, or 5) 337–420 days (49–60 weeks) following the event.
In our primary analyses, when a patient was observed to have a second occurrence of a specific treatment event of interest (e.g., a second “new antidepressant start” following a first “new antidepressant start”), observation days that occurred subsequent to the second event were “reset” and classified based on proximity to the second event. We also constructed a variable for the cumulative number of specific treatment events to date (1, 2, ≥3), allowing us to take into account the effect of additional occurrences of specific events in the analyses. In these analyses which considered each event separately, a patient who was discharged from the hospital AND who also had a new antidepressant start would have contributed exposure time to analyses that examined suicide rates following psychiatric hospitalizations and to separate analyses examining suicide rates following new antidepressant starts. However, the vast majority of antidepressant starts occurred on an outpatient basis.
For sensitivity analyses, we constructed several additional variables. First, we constructed a variable that categorized observation days only in relation to the first occurrence of a treatment event, regardless of whether there were additional occurrences of the same type of event. Secondly, recognizing that specific treatment events (such as a new antidepressant start) may be followed by other treatment events of interest (a dosage change), we constructed a variable that allowed us to examine whether the recency of “any treatment event” corresponded with higher suicide risks. For this variable, observation days were categorized based on the time from the date of the most recent occurrence of “any antidepressant treatment event” (new start, other start, or dose change) or from the most recent occurrence of “any treatment event” (new antidepressant start, other start, dose change, or psychiatric hospitalization). A patient who was discharged from the hospital AND had a new antidepressant start identified upon discharge would have contributed time in the 5 sequential 12 week periods following this common date of “any treatment event”.
To identify suicides, we submitted National Death Index Plus
queries for cohort patients who had a date of death in the VA Beneficiary Identification and Records Locator System (BIRLS) Death File during the study period. The NDI is considered the “gold standard” of US mortality databases.(Cowper et al., 2002
) BIRLS data have a sensitivity of 87%–96.5% and a specificity of 94% for deaths when compared to NDI data.(Cowper et al., 2002
; Dominitz et al., 2001
) Because BIRLS data may be less sensitive when patients are seen only as outpatients or do not have a service-connected disability, we also initiated NDI searches for cohort patients who did not use VA services in the year following the study period, even if there was no date of death in VA data. This process resulted in a comprehensive assessment of death among cohort patients.
We first identified deaths due to suicide using ICD codes in the NDI data file that specified suicide (ICD-10 codes X60–X84, Y87.0). Because underreporting is a concern, in sensitivity analyses we used a broader definition of suicide, assuming that deaths due to “events of undetermined intent” (ICD-10 codes Y10–34, Y87.2, Y89.9) were also suicides.(Speechley and Stavraky, 1991
We completed descriptive statistics for the patient sample, using frequencies or means as appropriate. Suicide rates were calculated based on the number of suicides observed and the person-years of observation for each of the five sequential 12-week periods following hospital discharge or antidepressant treatment events. For time periods following each type of treatment event, rates were calculated across the combined occurrences of this event.
To test if the suicide rates differed significantly across the sequential 12-week periods following treatment events, we used piecewise exponential models that allowed suicide risks to vary across time periods.19 A Poisson regression model was used to fit the piecewise exponential models, 20 with generalized estimating equations to allow for correlation within patients when multiple episodes of treatment events were included in the analyses.21 Relative risks were calculated after adjusting for patient age, gender, race, ethnicity, marital status, diagnosis of a substance use disorder, post traumatic stress disorder, and service connection. Data were entered by intervals of 12-week periods for time-fixed covariates of gender, marital status, race/ethnicity, comorbid PTSD and substance use and for time-varying covariates such as psychiatric hospitalization status. The models also adjusted for the cumulative numbers of the specific treatment event occurrences to date (1, 2, ≥3).
Models that examined suicide risks following each antidepressant event type also included a time-varying covariate for psychiatric hospitalization, to adjust for potentially different suicide risks associated with a prior hospitalization. The log-likelihood ratio test was used to test for the overall differences in suicide risks due to time-periods. When suicide risks between sequential time periods were compared, we adjusted p-values to correct for multiple comparisons and protect the overall alpha level at 0.05.
We also used the likelihood ratio test to determine whether there significant differences in suicide rates across age-groups in treatment periods following each type of treatment event.
In sensitivity analyses, we compared suicide risks during sequential treatment periods, following an approach used in prior work in which observation days were categorized only from the first occurrence of a specific treatment event (such as a new antidepressant start).22,23 We also conducted analyses in which we examined suicide risks in time periods following “any treatment event” and “any antidepressant treatment event”.