The lower rates of self-harm in those prescribed lithium may be due either to improved mood stabilization compared with other treatments or specific effects on impulsive aggression and risk taking. The similarity of the negative association between lithium use and unintentional injury and that between lithium use and self-harm supports the latter hypothesis because there is little reason to expect that lower rates of depressive symptoms would reduce unintentional injury.²⁸ Also, there is scant evidence that lithium is superior to quetiapine in preventing depressive episode relapse.⁷

Our study had notable strengths and limitations. This study uses a large, nationally representative sample to examine rates of fatal and nonfatal self-harm and unintentional injury. The use of EHRs to capture those episodes of self-harm managed entirely in primary care, as well as those admitted to secondary care, captures the true burden of self-harm morbidity both in the community and hospital presentation. Therefore, rates of recorded self-harm in our study were slightly higher than those in previous cohort studies.^10,62 As in any analysis of health records, this study would have missed episodes of untreated and unreported self-harm: only population survey methods could capture these episodes, and estimates generated in this manner are prone to response biases. The number of suicides was low, so we were unable to examine differences in rates between drugs. It is possible that misclassification or nonrecording of suicides occurred. However, similarities with other cohorts suggest that this limitation is not a major problem,^10,11 and there is no evidence to indicate that misclassification of cause of death would be differential by drug. Exposure time was foreshortened because of both left truncation (eg, practices joining THIN later than 1995) and right censoring (eg, switch to or addition of another drug, patients leaving the primary care practice, or dying of nonsuicide causes). This censoring was equally distributed by exposure drug.

Potential biases relating to selection into the study should have been avoided by the use of contemporaneous, representative medical records, as well as information bias minimized by the use of THIN’s detailed and well-recorded prescribing data as the exposure. However, exposure to the study drug is approximated through prescriptions issued and may not reflect how the patient used the medication. It is possible that erratic adherence is more likely for drugs other than lithium (because lithium is more closely monitored through regular blood tests) and may have contributed to lithium’s perceived superiority. However, these patients had no more physician contacts than those taking other medication, all individuals had to fill more than 1 prescription during their follow-up period (suggesting drug adherence), and other longitudinal cohort studies have not shown differential adherence.⁶³ Because people taking lithium tended to be older, suicides could have occurred in this group before the start of follow-up, thus reducing the observed rate relative to other treatment groups. However, this occurrence should not be the case in the matched analysis.

Through PS adjustment and PS matching, we attempted to account for potential confounding, including confounding by indication, and it is reassuring that both analyses produce similar results. Despite the numerous variables included in the PS, it is possible that residual confounding remained. It may be that important sociodemographic or clinical factors were not captured by the score, and we cannot confirm balance of unobserved covariates.^64,65 Notably, detailed information on educational level and individual socioeconomic status is lacking from the database. However, although these covariates are likely to be associated with self-harm, unintentional injury, and suicide, they should not be associated with treatment allocation. For these (or any) unmeasured covariates to have an important effect on the results, they would have to be strongly associated with exposure and outcome and be independent of covariates included in the PS.^66,67

Previously, it has been shown that a combination of lithium and valproate was associated with the lowest rate of suicide attempt.¹⁶ This group (and other combinations) were excluded from our study because we wanted to examine the association with monotherapy; in fact, treatment with this combination was rare despite recommendations for its use.⁶⁸