In 2005, amidst growing concern that the use of antiepileptic drugs (AEDs) increases the risk of suicidality, the Food and Drug Administration (FDA) identified 11 AEDs for further analysis: carbamazepine, divalproex, felbamate, gabapentin, lamotrigine, levetiracetam, oxcarbazepine, pregabalin, tiagabine, topiramate, and zonisamide (Food and Drug Administration, 2008). As a result of the analysis the FDA concluded that patients treated with AEDs had nearly twice the risk of suicidal behavior or ideation compared with placebo (0.43 vs. 0.22%), with a stratified, adjusted odds ratio (OR) of 1.8 [95% confidence interval (CI): 1.24–2.66] for the association. The FDA determined that the results were consistent among individual drugs as well as across all drug subgroups. Among the trials with a psychiatric indication, there were 5.7 events per 1000 placebo patients and 8.5 per 1000 drug patients (OR=1.51; 95% CI: 0.95–2.45). In the gabapentin trials, there were two events among the 2903 active drug patients, and one among the placebo patients, resulting in an adjusted OR of 1.57 (95% CI: 0.12–47.66).
As a result of their analysis, the FDA determined that all AEDs present an increased risk of suicidality, regardless of mechanism or indication. The agency issued safety alerts and mandated that AED manufacturers include a warning in their product label and develop medication guides for patients that include the warning (Food and Drug Administration, 2009). These actions resulted in pointed criticisms of the FDA’s study methodology, including claims that the results were affected by selection bias, bias in adverse event ascertainment, confounding by previous suicidality, and potential heterogeneity in treatment mechanism (Hesdorffer and Kanner, 2009 ; Ferrer et al., 2014). Gibbons et al. (2010), in particular, reinforced these criticisms, noting that the FDA analysis required the exclusion of any study with zero suicide/self-harm events, and thus for gabapentin, in particular, this reduced the number of viable trials from 49 to 3.
The response to such concerns, in part, was a closer examination of the association between AED use and suicidality by various researchers. Patorno et al. (2010) used a proportional hazards (PH) model to evaluate the risk of suicide attempt or self-harm (SA/SH), completed suicide, or violent death for patients on AEDs compared with a reference treatment. Their model, adjusted for age, sex, year, extensive comorbidities, and concomitant medication, showed an increased risk of suicidality, with a hazard ratio (HR) of 1.44 (95% CI: 1.13–1.83). A subanalysis showed that among patients with a diagnosed mood disorder the adjusted HR was 2.0 (95% CI: 1.43–2.79).
Collins and McFarland (2008) studied 12 662 Oregon Medicaid patients diagnosed with bipolar disorder (BD). A Cox PH model was used to compare completed suicides or emergency department visits due to SA/SH between patients treated with AEDs to those treated with lithium. The PH model was adjusted for comorbid physical and mental illness, concomitant use of antidepressants, antipsychotics, age, sex, and year of diagnosis. Patients who took gabapentin had an adjusted HR of 1.6 (P=0.2) for SA/SH and 2.6 (P<0.001) for completed suicide.
Pugh et al. (2012) studied the association between AED use and suicide-related events in a cohort of older bipolar veterans. Suicide-related events were measured using ICD-9 codes, with analysis limited to new prescriptions. A propensity score-adjusted Cox PH model showed that, relative to patients with no AED exposure, patients taking any AED had a substantially increased risk of suicidality (HR=3.9; 95% CI: 2.93–5.19). Patients taking gabapentin also showed an increased risk (HR=2.56, 95% CI: 1.96–4.16).
Gibbons et al. (2009) examined the association between SA and single-drug therapy for BD for the 11 AEDs examined by the FDA. The study utilized a database of 47 918 bipolar patients who were classified by monotherapy with one of the 11 AEDs, lithium therapy, or no therapy. In contrast with the results of previous studies, Gibbons et al. (2009) reported no difference between the SA rates for patients treated with AEDs versus the rate among patients treated with neither an AED nor lithium. The authors reported an SA rate for both groups of 13 events per 1000 person-years (PY). For patients treated with gabapentin, the authors reported 61 SA events per 1000 PY before treatment, versus 13 per 1000 PY after treatment. On the basis of these results, the authors concluded that gabapentin was protective for SA in BD patients.
At the present time, the literature presents a conflicting image of the association between gabapentin and suicidality in patients with BD. This apparent conflict prompted a re-examination of the data presented in the Gibbons analysis, and the results of this re-examination are presented in the current study.
Patients and methods
Data for this study came from the PharMetrics Patient-Centered Database, a patient-level, a de-identified database of inpatient and outpatient medical and pharmaceutical administrative claims from more than 100 health plans across the USA. The data were originally compiled by PharMetrics Inc. (Danbury, Connecticut, USA) for the Gibbons’ bipolar cohort study (Gibbons et al., 2009) and was acquired by the current authors (W.M.L., M.D.F.) during the discovery process in a lawsuit against Pfizer, in which Gibbons provided expert witness testimony. The data comprise 47 918 bipolar patients drawn from medical claims between 2000 and 2006. Inclusion criteria included continuous enrollment in the same healthcare plan for 1 year both before and after bipolar diagnosis. The data set includes demographics (age, sex), date of bipolar diagnosis, pharmaceutical records including dates for prescriptions of lithium and AEDs, as well as concomitant medications (antipsychotics, antidepressants, and anticonvulsants), comorbid diagnoses, and dates of SA/SH.
We identified all patients who began a new prescription of gabapentin or lithium. For our purposes, ‘new’ refers to the first instance of a gabapentin or lithium prescription during the observation period. Lithium was chosen as the comparison medication as it is considered the standard of care for BD. From the date of the initial prescription, patients were followed for up to 1 year for one of the following outcomes: SA/SH (defined by the presence of a diagnosis of any ICD-9 code in the range of E950–E959, which indicate an intention of self-harm); addition of another AED; switching between lithium and gabapentin; or discontinuation of the prescription as defined by a gap of more than 30 days between the end of one prescription period and the beginning of a new prescription (i.e. we accounted for lower-than-prescribed usage by allowing a period of up to 30 days to exist between prescriptions, as measured by the days’ supply). The exposure risk window was extended by 30 days from the end of the last prescription period for any patient censored because of medication discontinuation. Any patient with an SA/SH event or concomitant AED prescription on the same day as the initial treatment prescription, or under the age of 18 was excluded from the analysis.
Statistical analysis plan
The gabapentin and lithium treatment groups were assessed for a number of SA/SH events and total PY of study time. Comorbid conditions considered included a diagnosis of cancer, HIV, pain disorder, epilepsy, schizophrenia, major depressive disorder (MDD), and other psychological disorders as defined by ICD-9 codes (Table 1). Other confounders included concomitant medications (antipsychotics, antidepressants, and anticonvulsants, identified by the National Drug Code directory), age, sex, and pre-existing SA/SH diagnosis. Concomitant medications and comorbid diseases were classified in two ways: (i) existing before the index prescription, and (ii) concurrent with the study period. The incidence of SA/SH was evaluated with Fisher’s exact test. Incidence rates were evaluated by Poisson regression with a log-time offset. Confounding variables were evaluated using Fisher’s exact test for categorical variables and t-tests for continuous variables. Crude and adjusted HRs were evaluated with Cox PH models. Final multivariable Cox PH models were determined by stepwise selection, with the following criteria: entry P=0.20, exit P=0.05.
In addition to the primary analysis, we carried out a propensity score-matched (PSM) analysis to mediate baseline differences between the treatment groups. We first used a stepwise logistic regression model (entry P=0.20, exit P=0.05) to estimate the probability of receiving either gabapentin or lithium on the basis of pre-existing conditions and prescriptions for each patient in the study. Gabapentin patients were then matched to lithium subjects using the Greedy-5 algorithm (Parsons, 2001). We analyzed the propensity-matched data with the above-detailed methods.
For all analyses, P values less than 0.05 were considered significant. All analyses were performed using SAS 9.4 SAS Institute, Inc., Cary, North Carolina, USA. This study was approved by the Oregon Health and Science University Institutional Review Board (IRB00012073).
We identified 5522 patients who initiated a new prescription for gabapentin or lithium. Of these, 2421 (43.8%) were treated with gabapentin and 3101 (56.2%) with lithium (Table 2). On average, gabapentin patients were older (43.5 vs. 40.6 years, P<0.001), and more likely to be female (67.3 vs. 58.7%, P<0.001). Gabapentin patients were more likely to have comorbid epilepsy, pain disorder, MDD, other psychological disorders, and HIV, whereas lithium patients were more likely to have comorbid schizophrenia. Accordingly, gabapentin patients were more likely to have concomitant prescriptions for antidepressants (78.9 vs. 61.2%, P<0.001), and anticonvulsants (18.7 vs. 14%, P<0.001); lithium patients were more likely to have concomitant antipsychotics (35.4 vs. 27.9%, P<0.001). There was no difference in the risk of previous SA/SH attempt (1.3% among gabapentin users vs. 1.4% among lithium users, P=0.82).
Study patients contributed a total of 2337.1 PY to the analysis (Table 2). Gabapentin patients contributed 915.8 PY versus 1421.3 PY in the lithium cohort. On average, gabapentin patients had 138.1 days of follow-up time, versus 167.3 days for lithium patients (P<0.001). There were 37 SA/SH events, 21 (56.8%) in the gabapentin group and 16 (43.2%) in the lithium group (P=0.13). The unadjusted incidence rate was 22.9 per 1000 PY for gabapentin patients versus 11.3 per 1000 PY for lithium patients (P=0.03). The crude HR was 1.96 (95% CI: 1.02–3.76). Stepwise regression resulted in a model that included treatment group, age, previous SA/SH, and a concurrent diagnosis of other psychological disorders. The adjusted HR was 2.3 (95% CI: 1.2–4.5).
Propensity score-matched analysis
For the analysis, we first fit a logistic regression model including all of the pre-existing comorbid conditions and pre-existing medications and used stepwise selection to estimate the probability of each patient having been prescribed gabapentin. The final model included age, sex, pre-existing use of antidepressants and anticonvulsants, and previous diagnoses of pain disorder, other psychological disorders, HIV, and cancer. We were able to match 2079 lithium patients to 2079 gabapentin patients, retaining 85.9% of our original gabapentin patients. After matching, there were no differences between the two groups with respect to age, sex, previous SA/SH events, or pre-existing medications and comorbidities (Table 3). In addition, there were no differences with respect to concurrent epilepsy, schizophrenia, HIV, cancer, or use of anticonvulsants. However, the groups remained significantly different with respect to concurrent diagnoses of pain, MDD, and other psychological disorders, and use of antidepressants and antipsychotics (Table 4).
These 4158 patients contributed 1753.8 person-years of study time, with an average of 139.5 and 168.5 days for the gabapentin and lithium groups, respectively (P<0.001) (Table 5). There were 31 SA/SH events during follow-up: 20 (64.5%) in the gabapentin group, and 11 (35.5%) in the lithium group (P=0.15). The crude SA/SH rates were 25.2 per 1000 PY for gabapentin and 11.5 per 1000 PY for lithium (P=0.04). The crude HR was 2.1 (95% CI: 1.01–4.41). After stepwise selection, the model included treatment, age, previous SA/SH, and comorbid diagnoses of MDD and other psychological disorders. The resulting HR was 2.1 (95% CI: 1.02–4.5).
Our analysis showed a statistically significant positive association between the new use of gabapentin and the risk of suicidality and self-harm. Specifically, we found that, after adjusting for demographics, comorbid diagnoses, concomitant medications, and a history of SA/SH, bipolar patients treated with gabapentin have approximately twice the risk of SA/SH as patients treated with lithium (HR=2.3; 95% CI: 1.2–4.5). Additional analysis by PSM supported our conclusions.
Our results provide a stark contrast with findings of Gibbons et al. (2009), who, using the same data set, reported an 85% decrease in suicidality among BD patients who took gabapentin [the authors reported an event rate ratio of 0.15 (95% CI: 0.05–0.47)]. Although our findings represent the first reanalysis of the data relied on by Gibbons et al. (2009), they are not the first published criticism of the study methods. Patorno (2010) described several sources of error and potential flaws in the Gibbons analysis. These included: the study design excluded completed suicides, as the patients were required to have a full year of continuous medical coverage after their initial bipolar diagnosis or their initial gabapentin prescription; the analyses suffered from immortal time bias (in which there is a period of time during which the outcome of interest cannot occur); and the protective effect shown can be attributed to suicidal activity as a trigger for the prescription of medication in BD patients (Patorno, 2010). Ferrer et al. (2014) also noted potential problems, including exposure misclassification, selection bias, outcome misclassification, confounding by indication, and conflict of interest because the analysis was performed for litigation.
Our analysis is limited by fundamental, irreconcilable issues with the data set. The data set was constructed specifically for the Gibbons study; we were limited by their inclusion criteria, which restricted data to patients with two uninterrupted years of health insurance coverage. There is no way to account for the outcomes of those patients who lost coverage, nor the rates of completed suicide, as any completed suicide would have been excluded by the ‘continuous insurance coverage’ definition. These limitations likely result in an underestimation of the true effect of gabapentin, because of the fact that a history of suicidality is the strongest predictor of future completed suicide. Therefore, we expect that the increase in SA/SH observed here would translate into a greater rate of completed suicide among patients taking gabapentin. The Gibbons study also dictated which concomitant medications were to be included and we could not account for anything other than antidepressants, anticonvulsants, and antipsychotics. Although we would have preferred to control for more concomitant medications (e.g. pain medications and benzodiazepines), it is likely that these are correlated with the comorbid conditions for which we accounted.
Because this was an observational study of insurance claims data, we have no way of measuring the severity of illness. We attempted to mitigate this constraint by using a comparably medicated control group. The requirement that some treatment is used (i.e. lithium), and the demonstrated comparable premedication SA/SH rates helped ensure that the gabapentin patients were not more severely ill than the comparison group. Indeed, the gabapentin patients may have had less severe illness as they were disproportionately less likely to be prescribed an antipsychotic compared with their rates of schizophrenia, and they may have been taking gabapentin for reasons completely unrelated to their BD (i.e. they were not considered ill enough to warrant medication for their BD illness). We feel confident that the patients in the gabapentin arm were no more severely ill than those in the lithium arm for these reasons, and because gabapentin, in the cases where it was specifically prescribed for the treatment of BD, was done so as a monotherapy, rather than as an adjuvant therapy.
We have no way of accounting for any nonmedical treatments, such as psychotherapy, though there is no reason to believe that such treatments would be differentially distributed. We cannot be certain that the treatment groups had equivalent levels of medical complexity, as the gabapentin group had significantly higher rates of baseline comorbid conditions such as MDD, HIV, and pain disorder. However, we hesitate to draw a conclusion about which group was more medically complex. For instance, we note that the lithium group had higher rates of schizophrenia and potentially untreated or undertreated epilepsy, pain disorder, MDD, and other psychological disorders. In addition, our PSM subgroup analysis resulted in well-matched groups with respect to pre-existing conditions and resulted in the same direction and strength of association as the full cohort.
Patients were not randomized to the treatment groups, and therefore differences in SA/SH rates may be because of other, unmeasured differences between those who were prescribed gabapentin and those who were prescribed lithium. However, because of correlations between known and unknown confounding conditions and medications, our use of the PSM analysis likely had the added benefit of balancing the groups with respect to these unknown confounders.
We cannot be certain of patient adherence to the medication, though our definition of ‘continuous use’ minimizes the amount of potentially misclassified study time. In addition, in recent years a black market for gabapentin has developed (Smith et al., 2010 ; Schifano, 2014). Patients with comorbid drug abuse disorder may have used gabapentin recreationally or sold their medication rather than using it as prescribed. There is no way for us to control for this. However, with respect to the early-2000s to mid-2000s time period of the data used in our analysis, we could not find evidence to support such a prevalent black market in that period. We, therefore, believe it was not a significant source of bias in our study.
Finally, because the data were drawn from insurance claims, there may be nondifferential reporting of comorbid illnesses. For instance, it is plausible that a patient may have had a diagnosis before the study period, but it was not noted again until after their index prescription. Thus, the patient would be noted to have a concurrent diagnosis, but not a pre-existing diagnosis. However, there is no reason to believe that this would disproportionately affect one arm over the other.
Despite these limitations, we feel this study presents an accurate, rigorous examination of the data available. In so far as statistical methods allow, we attempted to mediate the potential impact of the above-mentioned shortcomings of the data. We examined patients who initiated a new, monotherapy prescription for gabapentin or lithium. Thus, we were able to isolate the potential effects of each medication. This is of particular concern, given the fact that the FDA has determined that AEDs as a class of medications increase suicidality. Our conclusions in terms of gabapentin are therefore insulated from the potential effects of other AEDs. Our use of Cox PH models accounts for actual exposure time as accurately as possible. Our models accounted for many comorbid conditions and concomitant medications, as well as a history of suicidality. Finally, and perhaps most importantly, we measured events from the index prescription date, rather than comparing SA/SH events before medication to those following medication, which likely produces misleading results because of the increase in suicidal behaviors in the months preceding initial medication (and in the case of the Gibbons study, results in a protective effect for gabapentin). Our PSM analysis further controlled for the effect of these potential confounders and our results remained consistent.
Considering that gabapentin has no demonstrable efficacy in controlling BD symptoms, and has been specifically noted as not effective and not recommended by leading experts, we may question why gabapentin continues to be used as a treatment for BD at all. Patients with BD have a significantly higher risk of suicidality and completed suicide, with ~25–50% of patients with BD attempting suicide in their lifetime. The annual rate of completed suicide in the bipolar population is 1%; more than 66 times greater than the rate in the general population (estimate 0.015%) (Fountoulakis et al., 2009). It is imperative that the situation is not exacerbated with the use of medications that can induce suicidality. Whether the increased risk showed in our study is because of untreated disease or due to a physiological effect of gabapentin, we feel the evidence supports that gabapentin increases the risk of suicidality in BD patients. However, for the FDA-approved indications, the risks associated with not receiving treatment may outweigh the risks associated with gabapentin (Mula et al., 2013). Therefore, it is of the utmost importance to closely monitor any patient taking gabapentin.
Conflicts of interest
M.D.F. serves as a consulting expert in litigation involving pharmaceutical products. W.M.L. provides analytical services for Forensic Research & Analysis, which in turn provides services for litigation involving pharmaceutical services. For the remaining authors, there are no conflicts of interest.
Collins JC, McFarland BH (2008). Divalproex, lithium
among Medicaid patients with BD. J Affect Disord 107:23–28.
Ferrer P, Ballarin E, Sabate M, Vidal X, Rottenkolber M, Amelio J, et al (2014). Antiepileptic drugs and suicide
: a systematic review of adverse effects. Neuroepidemiology 42:107–120.
Food and Drug Administration (2008). Statistical review and evaluation: antiepileptic drugs and suicidality. Washington, DC: U.S. Department of Health and Human Services.
Fountoulakis KN, Gonda X, Siamouli M, Rihmer Z (2009). Psychotherapeutic intervention and suicide
risk reduction in bipolar disorder
: a review of the evidence. J Affect Disord 113:21–29.
Gibbons RD, Hur K, Brown CH, Mann JJ (2009). Relationship between antiepileptic drugs and suicide
attempts in patients with BD. Arch Gen Psychiatry 66:1354–1360.
Gibbons RD, Hur K, Brown CH, Mann JJ (2010). Gabapentin
attempts. Adv Pharmacoepidemiol Drug Saf 19:1241–1247.
Hesdorffer DC, Kanner AM (2009). The FDA alert on suicidality and antiepileptic drugs: fire or false alarm? Epilepsia 50:978–986.
Mula M, Kanner AM, Schmitz B, Schachter S (2013). Antiepileptic drugs and suicidality: An expert consensus statement from the Task Force on Therapeutic Strategies of the ILAE Commission on Neuropsychobiology 54:199–203.
Parsons LS (2001). Reducing bias in a propensity score matched-pair sample using greedy matching techniques. In: Proceedings of the twenty-sixth annual SAS users group international conference; 22-25 April 2001; Cary, NC: SAS Institute Inc.
Patorno E (2010). Letters: antiepileptic medications and the risk of suicide
. JAMA 304:522–523.
Patorno E, Bohn RL, Wahl PM, Avorn J, Patrick AR, Liu J, et al (2010). Anticonvulsant medications and the risk of suicide
, attempted suicide
, or violent death. JAMA 303:1401–1409.
Pugh MJV, Copeland LA, Zeber JE, Wang CH, Amuan ME, Mortensen EM, et al (2012). Antiepileptic drug monotherapy exposure and suicide
-related behavior in older veterans. J Am Geriatr Soc 60:2042–2047.
Smith BH, Higgins C, Balacchine A, Kidd B, Bannister J (2010). Substance misuse of gabapentin
. Br J Gen Pract 62:406–407.
Schifano F (2014). Misuse and abuse of pregabalin and gabapentin
: cause for concern? CNS Drugs 26:491–496.
Keywords:Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.
bipolar disorder; gabapentin; lithium; self-harm; suicide