Suicide risk in Veterans Health Administration patients with mental health diagnoses initiating lithium or valproate: a historical prospective cohort study

Demographic, inpatient and outpatient mental and nonmental health treatment records, and outpatient pharmacy prescription data was obtained from the VHA’s National Psychosis and Depression Registries [18] (linked, de-identified healthcare databases of all VHA patients since 1997 with at least one psychotic or depressive disorder diagnosis). This study involved no prospective enrollment of human subjects. Treatment exposure, covariates, and outcome were characterized by information previously recorded during use of the VHA by patients. As with other large healthcare database outcome studies, no informed consent was obtained. This study was approved by the Institutional Review Boards of the Bedford and Ann Arbor Veterans Affairs Medical Centers.

Incident users (> 6 months of no lithium or valproate use) with recent VHA utilization (past year and a previous year) were identified among all patients with mood or psychotic disorders within the past 30 days receiving at least one outpatient prescription for lithium or valproate from April 1999 to December 2008. These broad diagnostic inclusion criteria (Additional file 1: Appendix 1) were chosen to maximize statistical power, given the comparatively few suicides expected, even in a large cohort, over a fixed one-year follow-up period, and to facilitate the evaluation of lithium and valproate as broadly useful suicide preventatives. Prior research suggests that any effectiveness of lithium against suicide death is not restricted to patients with bipolar disorder [9],[19]. Individuals with potentially nonpsychiatric indications for treatment were excluded (epilepsy, migraine headache, or neuropathy diagnoses in the past 30 days; dementia medication use in the past 180 days; cancer, dementia, skull fracture or traumatic brain injury diagnosis in the past year; traumatic brain injury treatment, home care, or hospice care in the past year; or any nursing home residence or inpatient rehabilitation in the past 2 years). Patients were also excluded if they had started their mood stabilizer on an “as needed” basis (as indicated by receiving an initial prescription designated “prn”) or both mood stabilizers simultaneously (Figure 1).

Receipt of lithium or valproate was determined by outpatient prescription fills. The primary analysis examined all individuals initiating either lithium or valproate treatment and followed these individuals until end of follow-up (365 days), suicide death, or death from other causes (i.e., an “intent-to-treat” analysis). As clarification, our study did not examine primary and secondary outcomes per se (the focus consistently being on suicide death), but we did perform primary (intent-to-treat) and secondary analyses of this outcome. Additional secondary analyses examined briefer follow-up times and/or stratified follow-up time by whether individuals were still receiving initial treatment. Individuals were identified as “still receiving initial treatment” if they had not switched to or added the other treatment, nor discontinued the initial treatment (experienced a ≥ 15 day gap between outpatient prescriptions, adjusted for early refills). All other follow-up time was classified as occurring during the period when individuals had “stopped/modified” initial treatment. Since this stopper/modifier group included individuals subsequently resuming either treatment, we also analyzed suicide death risks for individuals over 0–180 days which had stopped initial treatment and not resumed either treatment before suicide, other mortality, or the end of follow-up. Risks observed after treatment discontinuation may reflect risks related to discontinuation of the medication (e.g., “rebound” mania or depression), but may also reflect differences in baseline risk between the treatment groups (i.e., confounding) or differences in selection occurring during follow-up [17].

Date and cause of death (suicide) was obtained from National Death Index files [20] for 1999–2009 using previously established definitions (International Classification of Diseases, Tenth Revision, codes X60–X84, Y87.0, and U03) [21].

An extensive set of 934 baseline covariates was derived (Table 1 and Additional file 1: Appendix 2) from VHA databases reflecting demographics, diagnoses, general VHA mental and nonmental health healthcare utilization [16], hospitalizations, clinic use, diagnostic testing, current and recent prescriptions (Additional file 1: Appendix 3), diagnosed suicidal behavior, injuries, regional (state-level) suicide risk (Additional file 1: Appendix 2), and prior mood stabilizer treatment. Data in these categories were often modeled over several time periods or with flexible forms (multiple indicator variables). This approach follows the general aims of “high-dimensional” propensity score methodology [16],[22], but did not include automated variable generation or selection. Instead, a number of covariates in each domain (Table 1) were generally included, since the full determinants of suicide risk are not well understood, although covariates with a substantial association with treatment exposure (i.e., with which treatment a patient receives) were individually evaluated and removed if they were judged unlikely to be confounders. (Inclusion of variables that are substantially associated with treatment exposure, but which are not confounders, can actually increase confounding from uncontrolled confounders) [23]-[26].

The propensity score was calculated by logistic regression (c statistic = 0.69). We then 1:1 matched patients initiating lithium and valproate using propensity score calipers of 0.03 (0.2 standard deviations of the propensity score logit) [27],[28], achieving 98.7% matching of lithium-initiated patients. Balance in the prevalence of covariates between the treatment groups was assessed using standardized differences (Table 2). Standardized differences are equivalent to Cohen’s d effect sizes, with a difference of ≥ 0.10 often considered as indicating significant imbalance [15].

For the analyses of the intent-to-treat cohort and of individuals still receiving initial treatment, we used conditional logistic regression, whereas for individuals stopping or modifying treatment, ordinary logistic regression was used since matching was not rigorously preserved.

Several additional analyses were conducted, including comparing the prevalence between the treatment groups of diagnostically-coded suicidal ideation (V62.84, a code for suicidal ideation first introduced in 2005) in the 30 days prior to treatment initiation among the patients who potentially could receive this diagnosis (the < 50% of the sample initiating treatment in 2005 or later). We also compared the suicide risk associated with the treatment groups prior to matching, conducted a Cox regression analysis, and conducted a sensitivity analysis matching the sample with an alternative propensity score. All analyses were performed using SAS, version 9.3, except the standardized difference calculations (Microsoft Excel 2007).

Several study limitations should be noted. Data limitations include gaps in VHA prescription records for inpatients and a lack of information about inpatient and outpatient care received outside the VHA, and potential errors in the measurement of covariates. Analytic limitations included the absence of rematching/reweighting of patients during follow-up. This limitation precluded analysis of whether patients experiencing suicidal ideation or behavior during treatment were preferentially discontinued off one of the two treatments, and of whether differences existed during follow-up between the treatment groups in either the initiation of, or persistence with, other psychiatric medications. This study examined typical care, rather than being restricted to monotherapy, unlike some recent studies [34],[36]. Given that other psychiatric medications may influence the risks of suicide death [37],[38], future studies should consider approaches which reweight patients during follow-up, such as marginal structural models (Additional file 1: Appendices 6A and 8). However, numerous classes of psychiatric medications prescribed at and before lithium or valproate initiation were very closely balanced between the two treatment groups, thus likely producing close similarity in concomitant medications, at least early during follow-up.

In addition, serum medication levels would provide information separate from prescription data about medication persistence, and might enhance future analyses if available. Study findings might have been influenced by the considerable diagnostic heterogeneity of this patient cohort, although each individual diagnosis was closely balanced in prevalence among lithium and valproate recipients and almost 90% of our sample had mood disorder diagnoses (Additional file 1: Appendix 1). Our focus upon suicide mortality (comprehensively documented nationwide, even for patients who leave VHA care) improved outcome ascertainment compared to nonfatal suicidal behavior, but unfortunately limited statistical power. This difference in outcome ascertainment can be substantial: separate studies have estimated that diagnosed nonfatal suicidal behavior may underestimate actual nonfatal suicidal behavior by up to 6-fold to 10-fold [39]-[41]. This expected underestimate of suicidal behavior has led us to focus our initial investigation, reported here, upon suicide deaths. Similarly, the V-code diagnosis of suicidal ideation we used to assess the potential for residual confounding (Table 5) likely captures only a small fraction of suicidal ideation. However, use of these codes was intended only as an aid to assess the potential presence and direction of residual confounding. For this purpose, diagnoses of suicidal ideation, even if infrequently coded, can provide valuable information, as has been the case in other studies [42].

Propensity score methods may also potentially inadvertently amplify any remaining confounding [23]-[26]. As pointed out above, such confounding could result from unmeasured differences in mental illness severity or in suicidal risks factors such as suicidal thinking, impulses, or intent. Such amplification is exacerbated if variables are included that are substantially associated with treatment (i.e., lithium or valproate initiation) but not outcome (i.e., suicide death) [25],[26]. How much of a bias is typically produced is controversial [25],[43]. Nevertheless, the differences between treatment groups in diagnosed suicidal ideation at baseline suggests some unaddressed confounding may remain in this study, and our particular implementation of propensity score matching included extensive variables in some domains (e.g., medical diagnoses) in which only a subset of variables may have been strongly related to suicide risk. (Harris and Barraclough, however, found that 90% of the medical diagnoses they reviewed were significantly associated with suicide death risk [44]). To reduce the potential for amplification of residual confounding, we actively evaluated the plausibility of confounding for those covariates with particularly substantial associations with treatment. Furthermore, a sensitivity analysis targeting this concern by removing a large number of covariates with the weakest apparent associations with suicide death from the propensity score produced only modest effect estimate changes (Additional file 1: Appendix 7). Most importantly, the direction of change in the 0–365 day effect estimates from the unmatched (OR = 1.45) to the matched sample (cOR = 1.22) strongly suggests that overall confounding was most likely reduced by the propensity score matching, not amplified (Additional file 1: Appendix 6B). Therefore, any amplification of residual confounding appears to be sufficiently minor that the propensity score methodology still produced an important reduction in overall confounding. When comparing this study to other studies, however, it is important to note that amplification of residual confounding could have potentially enhanced, to at least a slight degree, an apparent bias in this study against observing a protective association between lithium treatment and suicide death.

Generalizability of this study's findings to non-VHA patients, to patients with the excluded medical conditions (e.g., cancer, head injury, or seizures), to cohorts with differing rates of treatment discontinuation, or to patients that are treated for longer than one year is uncertain. Nevertheless, our findings generally agree with recent studies. With one exception [45], recent nonrandomized studies of suicide or suicidal behavior risk have observed nonsignificant (and typically modest) differences between lithium and valproate specifically, or lithium and anticonvulsants in general [34],[36],[46]-[49]. Recently, a small but methodologically-rigorous trial focused on suicidal behavior prevention [50] observed only nonsignificant differences in suicidal behavior between lithium and valproate. Results from this trial (involving 2.5 years of follow-up) and the BALANCE trial [51] (involving 2 years of follow-up) were combined in the recent randomized trial meta-analysis [8], which estimated that lithium treatment was associated with a nonsignificant reduction in nonfatal suicidal behavior compared to valproate (OR 0.64, 95% CI 0.30-1.36, p = 0.24). Thus, the study reported here and the recent trials have both observed nonsignificant intent-to-treat differences between lithium and valproate in suicide death or suicidal behavior. The central estimate for lithium’s effect size, however, did differ in direction between this study (nonsignificantly increased risk of suicide death) and the two trials (nonsignificantly decreased risk of suicidal behavior). This difference might simply be due to chance, residual confounding (possibly augmented by residual confounding amplification), differences in follow-up time (one year versus 2–2.5 years), or differences in outcome (suicide death versus suicidal behavior). However, this difference could also reflect a “two-sided” nature to lithium’s association with the risk of suicide death/suicidal behavior. That is, some degree of decreased suicide/suicidal behavior may be associated with active lithium treatment (thus contributing more greatly to the intent-to-treat estimates from the trials, which had much higher treatment persistence rates), combined with some degree of increased risk associated with lithium discontinuation (which would thus contribute more greatly to the intent-to-treat estimates in our study). A possible “two-sided” association between lithium and the risk of suicide death would be also consistent with the significant differences in baseline suicidal ideation diagnoses between the treatment groups observed in this study (suggesting some residual confounding and thus a greater benefit to active lithium treatment than indicated) and the timing of the emergence of significant risks after lithium discontinuation (suggesting risks associated with discontinuation itself) (Additional file 1: Appendix 6D).

We suspect, although this supposition is inherently somewhat speculative, that the most likely interpretation of our findings includes both phenomena: some genuine association between lithium discontinuation and a short-term increase in the risk of suicide death (compared to valproate discontinuation), along with some residual confounding biasing against lithium. (This residual confounding would imply that active lithium treatment was associated with a larger decrease in suicide risk than estimated in this study). This interpretation would imply that it is difficult to determine whether initiation of lithium in our VHA cohort was associated with a net reduction or increase in suicide risk over the first year of treatment. Our data does suggest, however, that the overall balance of suicide-related risks and benefits from lithium initiation might be increased if greater adherence to initial lithium treatment could be achieved. Research investigating how different health systems could achieve this goal should be a high priority. Possible options include group psychoeducation, which is supported by two clinical trials [52],[53]. One of these trials included psychoeducation with other treatment enhancements including detailed illness and treatment histories for each patient, follow-up within 2 weeks of hospital discharge, and a post-discharge “settling in” group [52]. A recent review identified five evidence-based strategies to boost mood stabilizer adherence, including psychoeducational, cognitive-behavioral, interpersonal, and family therapy approaches, as well as systematic care models [54].

Our findings clearly indicate a need for further research of lithium’s association with suicide death or suicidal behavior (Additional file 1: Appendix 8). Until such research resolves the degree to which the increased risks of suicide death observed in patients discontinuing lithium relate specifically to lithium discontinuation (rather than confounding), prudence suggests patient and provider education about the possible risks of treatment discontinuation and close monitoring of patients discontinuing lithium (and potentially valproate [36]) when feasible. Such monitoring is already recommended to limit mood episode recurrences [55]. As discussed above, healthcare systems, providers, and patients should also strive to maximize persistence with lithium treatment once initiated. Finally, when discontinuation does occur, there may be value to gradually discontinuing lithium when appropriate, given that gradual discontinuation has been observed to substantially reduce the risk of mood episode relapse [56],[57].

These recommendations concerning the need to minimize lithium and valproate discontinuation, and any risks resulting from this discontinuation, may have relevance well beyond the VHA. The substantial rates of lithium and valproate discontinuation observed here (median time to discontinuation of approximately 90 days) may not be particularly unusual. The observed rates of discontinuation fall within the range provided by the rates of lithium discontinuation reported for a U.S. health maintenance organization (median time to discontinuation of approximately 72 days) [58] and for patients with mood disorders in Denmark being managed primarily by general practitioners, rather than psychiatrists (median time to discontinuation of approximately 181 days) [59].