Associations of Benzodiazepines, Z-Drugs, and Other Anxiolytics With Subsequent Dementia in Patients With Affective Disorders: A Nationwide Cohort and Nested Case-Control Study

Patients were citizens of Denmark with a first-time hospital contact due to an affective disorder from January 1996 through December 2015. A total of 268,218 inpatients and outpatients were identified in the Danish National Patient Registry (10) using ICD-10 codes F30.0–F39.9. A total of 22,677 individuals under age 20 or with missing information pertaining to their birth date were excluded from this patient cohort. The cohort was followed from 1969 until 2016 for a dementia diagnosis or prescriptions for dementia medication as documented in the Danish Psychiatric Central Research Register, the Danish National Patient Registry, or the Danish National Prescription Registry. During the follow-up, 19,852 patients with incident dementia were identified. Among these, 10,076 had a diagnosis before or at study entry and were excluded from the main analyses.

Based on this information, a case-control study was nested within the cohort using risk set sampling generated by the procedure sttocc in Stata. Briefly, we randomly selected control subjects among members of the cohort who remained dementia free and matched them on follow-up time with case subjects (N=9,776) who reached the same age at the time of dementia diagnosis in a 1:4 ratio (N=39,104). They represented a subsample of the followed person-time (Figure 1).

The study was approved by the regional Danish Data Inspection. Individual-level consent was not required, and all data were anonymized.

Dementia was defined as a first inpatient or outpatient hospital discharge diagnosis of dementia (ICD-8 codes 290.00–290.99; ICD-10 codes F00.00–F03.9 and G30.0–G30.9) as documented in the Danish Psychiatric Central Research Register and the Danish National Patient Registry. The validity of dementia diagnoses obtained from the Danish National Patient Registry has been assessed in two studies (11, 12) reporting that 70% and 83%, respectively, of dementia cases diagnosed by an external rater according to ICD-10 criteria conformed with the diagnosis in the register, with the lowest validity in younger patients. We supplemented our definition of dementia with data on individual patient refills of at least one redeemed prescription of an acetylcholinesterase inhibitor identified by the Anatomic Therapeutic Chemical (ACT) classification system codes (N06D) in the Danish National Prescription Registry. This register holds information on all prescribed and redeemed drugs sold at Danish pharmacies since 1995 (13).

All refills of prescriptions for medication were identified in the Danish National Prescription Registry, where the quantity dispensed for each prescription is expressed by the defined daily dose. We used ACT codes to define benzodiazepines (N05BA, N05CD, and N03AE01), Z-drugs (N05CF), and other anxiolytics (N05BB01, N05BE01, and N03AX16). Benzodiazepines and Z-drugs were also classified into short- and medium-acting (N05BA04, 06, 12; N05CD04–09, 11; and N05CF01–03) and long-acting (N03AE01; N05BA01–03, 05, 08, 09, 11, 16; and N05CD01–03), as recommended by other investigators (3, 14). To assess dose dependence, we calculated the number of prescriptions, the sum of all prescribed defined daily doses registered in the Danish National Prescription Registry, and the intensity of use as the cumulative amount divided by the duration of use. Because benzodiazepine use can be either chronic or episodic, we assigned each prescription an exposure period of 105 days based on an analysis of waiting-time distributions used in previous Danish register-based studies (15). Because we found no evidence for specific cut-points for cumulative use, it was categorized on the basis of the distribution of exposure. In addition, we modeled restricted cubic splines with four knots equally spaced over the variables to examine whether the results were influenced by the cut-points chosen for exposure categories.

We selected the following covariates, assumed to be associated with both benzodiazepine use and dementia risk, from the Danish National Patient Registry and the Danish National Prescription Registry: depression subtype, year of diagnosis, history of alcohol or mixed substance abuse, diabetes or cardiovascular disease, and prescriptions for antipsychotic (ACT code N05A) and antidepressant (ACT code N06A) medications. Data on the highest achieved educational level were obtained from the Danish Population Education Register and categorized as low, middle, high, and unknown. Information on gender, age, and marital status was obtained from the Danish Civil Registration System.

The ACT codes used for definitions of exposure, outcome, and covariates are listed in Table S1 in the online supplement.

For comparison with previous studies, we analyzed our data using both a cohort design and a nested case-control design. The nested case-control design is suitable for analysis of time-varying exposures, such as benzodiazepines. In the cohort study, we analyzed the associations between any use of benzodiazepines (overall and divided according to type of medication (benzodiazepines, Z-drugs, and other anxiolytics and short- or medium-acting and long-acting drugs) and subsequent dementia by using Cox proportional hazards regression (hazard ratios and 95% confidence intervals), with age as an underlying time scale. Person-years of follow-up were accumulated from age at study entry (date of the first registered affective disorder) and terminated at age at first registration of dementia or death or at the end of follow-up (October 31, 2016), whichever came first. Because the mechanism linking benzodiazepines to the development of dementia is unknown (6), we used different approaches to model benzodiazepine exposure. For comparison with previous studies, we first analyzed benzodiazepine exposure (any use) at study entry (baseline) and further evaluated any interactions between the different types of benzodiazepines using the likelihood-ratio test. Second, we included benzodiazepines initiated during follow-up. Therefore, to account for immortal time bias, benzodiazepine exposure was entered as a time-varying variable in these analyses. This measure of benzodiazepine exposure included current and former (prevalent) users as well as new (incident) users, and we evaluated the effect of timing of benzodiazepine use on dementia by defining a supplementary time-varying exposure matrix including both pre- and postdiagnostic benzodiazepine use, categorized into four categories: no prior postdiagnostic use (no lifetime use, reference group), prediagnostic but no postdiagnostic use (former use), pre- and postdiagnostic use (continuing use), and postdiagnostic but no prediagnostic use (new use). Finally, we analyzed the associations with cumulated use of benzodiazepines at baseline as exposure. Long-term users may represent a selected group of patients. To address this potential bias, we repeated the analyses for benzodiazepine use cumulated 2 years before study entry, which may reveal more immediate effects (e.g., whether benzodiazepines influence the course of presymptomatic dementia). However, prodromal symptoms of dementia may be diagnosed as depression. Therefore, to account for reverse causality (protopathic bias), we delayed follow-up for 2 years using the stsplit option in Stata to divide the follow-up time into 0–2 years and >2 years. Physicians may interpret cognitive deficits in patients treated with benzodiazepines as side effects and thus be less likely to diagnose these patients with dementia the first time after hospital admission for an affective disorder. In our analyses, any such bias may be accounted for by the 2-year latency period. The assumption of proportional hazards was assessed by inspection of cumulative hazard plots and trends in scaled Schoenfeld residuals. The analyses for any benzodiazepine use and cumulated benzodiazepine use were repeated in a nested case-control design, where odds ratios for benzodiazepine and related drug use estimated from 1995 until 2 years before the index date (date of dementia diagnosis or control assignment) were calculated using conditional logistic regression. The odds ratio in the nested case-control analysis is largely equivalent to the hazard ratio in the cohort study. We accounted for confounding by selection of patients who received benzodiazepines by using multiple adjustment in models with all covariates mentioned above included. The analyses were performed using Stata, release 15.

We performed two sensitivity analyses to explore the robustness of our results. First, the exclusion of patients who have a history of the event under study (dementia in this study) before study entry may cause bias in a prevalent user design (16). This exclusion may have resulted in a reference population representing patients who were less likely to develop dementia during the period of exposure. We explored this potential selection by comparing premorbid use of benzodiazepines among the 9,671 patients who developed dementia between 1995 and study entry with benzodiazepine use among control subjects with no dementia at the same age and time before inclusion. Second, because mortality in this cohort was considerable and appeared to be related to exposure (see Table S2 in online supplement), competing mortality risk may have influenced the risk estimates, since patients who die are no longer at risk of developing dementia. Thus, to account for competing mortality risk, we used the Fine-Gray model for the cumulative incidence of dementia to estimate subdistribution hazard ratios for benzodiazepine use.

All data were obtained from administrative registers.

Any use of benzodiazepines, Z-drugs, or other anxiolytics was not associated with subsequent dementia after adjustment for sociodemographic and clinical variables in the cohort analysis or the nested case-control design (Table 2). However, in the cohort study, benzodiazepine use was associated with lower risk of dementia during the first 2 years after study entry. The estimates for benzodiazepine use or long-acting drug use were slightly attenuated when adjusting for any use of Z-drugs or short-acting drugs, respectively. In addition, any use of benzodiazepines did not appear to influence the effect of Z-drugs on dementia. Similarly, use of long-acting drugs did not modify the association of short-acting drugs with dementia (Figure 2; also see Table S6 in the online supplement). Second, in our analyses evaluating the timing of benzodiazepine use, continued and new users showed a reduction in dementia rates compared with patients with no lifetime use, whereas the association was closer to unity among former users (see Table S7 in the online supplement).

In the cohort analysis, number of prescriptions and cumulated dose of benzodiazepines or Z-drugs at baseline were not associated with rates of dementia (Table 3). Similar results were obtained for recent prolonged use. In the nested case-control design, where prescriptions were counted from 1995 until 2 years before the index date, the odds of developing dementia were slightly higher among patients with the lowest rate of benzodiazepine use compared with patients with no lifetime use (cumulated defined daily dose [DDD]: odds ratio=1.08, 95% CI=1.01, 1.15), while those with highest use had lower odds (cumulated DDD: odds ratio=0.83, 95% CI=0.77, 0.88). This pattern was observed for all drug types (benzodiazepines, Z-drugs, and long-acting and medium- and short-acting drugs). The estimates for drug intensity (DDD/duration) were similar to those for cumulated dose. The results from exposure modeled as a continuous variable with restricted cubic splines are presented in Figure S1 in the online supplement. These curves revealed no indication of any dose-response relation.

Case subjects who were diagnosed with dementia before study entry appeared to have the same rate of benzodiazepine use as control subjects. Thus, the odds ratio for dementia among patients with any benzodiazepine use before study entry was 0.93 (95% CI=0.87, 1.00) (see Table S8 in the online supplement). The mean age of the cohort at baseline was 73.2 years (SD=11.7), and 28.9% (N=68,048) of the patients died during the follow-up period (see Table S2 in the online supplement). However, the associations reported above did not differ when the presence of competing mortality risk was considered (see Table S9 in the online supplement).

The results from previous studies on the association between benzodiazepine use and dementia have been inconsistent. Some studies have demonstrated that benzodiazepine use is a risk factor for developing dementia (14, 17), whereas other studies, in line with ours, found no association (18, 19), including no protective association (20, 21). In a recent meta-analysis, the authors reported a 38% increased odds of dementia among users of benzodiazepines; however, the heterogeneity of the included studies was high (98%), and the 95% prediction interval of 0.58–3.25 indicated low certainty (7). The effect of elimination time has also been examined. Some investigators have found higher risk of dementia with use of long-acting compared with short-acting benzodiazepines (14, 22, 23), while others have found no difference in effect (17) or lack of effect (19). In terms of evidence regarding any association of Z-drugs specifically with dementia, findings are mainly restricted to subanalyses in benzodiazepine studies, which suggest a risk of dementia similar to that seen with benzodiazepines (4). However, few studies have accounted for simultaneous use of benzodiazepines (21). In the present study, we explored both the independent effect of each drug type (benzodiazepines, Z-drugs, and long-acting and medium- and short-acting drugs) on dementia risk and any interactions. This revealed that use of one drug type did not explain or modify the effect of other drug types. Studies on dose-response relationships with benzodiazepine use have been equally heterogeneous, with two case-control studies demonstrating a dose-response effect (14, 23), one study revealing an increased risk with a smaller cumulative dose only (24) (in line with our results), and three studies showing no trend (17, 19, 21). Our results are consistent with those from a Swiss case-control study (21) and a cohort study conducted in the United States using computerized pharmacy data (24). These two studies showed no association between benzodiazepines or Z-drugs and dementia. In line with our results, the Swiss study suggested a protective effect of the highest cumulative dose when adjusting for depression measured by antidepressant use (21). Most previous studies also accounted for reverse causality by introducing a latency period (lag time) before the outcomes were counted in order to account for the potential bias that may occur when benzodiazepines are used to treat prodromal symptoms of dementia. In our study, benzodiazepine use was associated with lower risk of dementia the first 2 years after study entry, which may indicate that physicians are less likely to make a diagnosis of dementia in patients taking benzodiazepines shortly after they are diagnosed with an affective disorder. However, the largest methodological limitation in previous studies was confounding by indication, which may explain the weak association between benzodiazepine use and dementia revealed in some studies. Most previous studies adjusted for depression, but less than half adjusted for anxiety, and none, to our knowledge, adjusted for depression severity or alcohol abuse.

One strength of this study is the sample size and use of nationwide population-based registers in a country with free access to health care, which provided a large, unselected patient population. The Danish personal identification numbers allowed individual patient data to be linked with different registers and complete follow-up information for dementia outcomes with positive predictive test values >70% to be obtained (11, 12). Analyses were conducted for patients with a first-time hospital admission for an affective disorder, a group assumed to be more homogeneous regarding psychiatric comorbidity, and this restriction was assumed to reduce confounding by indication. We also accounted for confounding by adjustment for several sociodemographic variables, comorbidity, and comedication. Because benzodiazepines and Z-drugs were frequently used in this patient population, it was not possible to account for confounding in a propensity-score-matched study, since there was not a sufficient number of nonexposed individuals for matching, and a simple adjustment including a propensity score would not have added to the multiple regression performed (25, 26). We intended additionally to use other anxiolytics as a negative control exposure because these drugs may have confounding similar to that for benzodiazepines but no assumed shared pathogenesis. Thus, if the association between other anxiolytics and dementia was due to confounding, we would expect to find a similar association between the negative control exposure and dementia. Unfortunately, this analysis was hampered by considerable contemporary use of benzodiazepines. Cohort entry was defined at hospital contact. This may have limited the risk of detection bias, which may hamper studies in which patients who are not exposed to benzodiazepines or Z-drugs may have less contact with health professionals and thus be at lower risk of being diagnosed with dementia. Information on the use of benzodiazepines, Z-drugs, and other anxiolytics was obtained from nationwide registers, which limits misclassification of exposure. Benzodiazepines may be used sporadically over longer periods, and we cannot exclude the possibility that some patients who received prescriptions for benzodiazepines before 1995 were misclassified as not exposed or as new users. However, any such misclassification would need to be substantial in order to move our results away from unity. Our benzodiazepine measure was based on prevalent users, and among these, patients who remained under treatment were those who tolerated the treatment well, but we also showed that the exclusion of patients with dementia before study entry did not appear to introduce any selection that may have influenced our estimates. However, we cannot exclude the possibility that the apparent positive effect of long-term benzodiazepine use was due to depletion of susceptibility (24). We accounted for competing mortality risk. This is especially relevant when examining outcomes in elderly patients or for conditions in which mortality is more commonly observed, because the competing risk event may substantially alter the probability of the occurrence of dementia, precluding its onset. However, to our surprise, our analyses showed that this did not appear to explain the results from our cohort study.

If these results are not due to selection bias as discussed above, they suggest that benzodiazepines are not a risk factor for dementia, and in fact there may be a protective effect of long-term benzodiazepine use on dementia risk in patients with affective disorder. This may be, as suggested, attributed to an inhibitory effect of benzodiazepines on excitotoxic or other pathological lesions or symptoms (6). Anxiety and insomnia are frequent in these patients and are also associated with risk of dementia. Thus, benzodiazepines may also exert a beneficial effect through treatment of these comorbidities.

In summary, many older people and patients with severe anxiety and sleeping problems benefit from using benzodiazepines and Z-drugs as recommended for short-term treatment. The adverse effects of benzodiazepines on cognition and studies reporting elevated risk with any use of benzodiazepines have fueled the fear of dementia among patients and clinicians. This study demonstrated that notwithstanding other possible adverse short-term or long-term effects, there is not sufficient evidence that benzodiazepines or Z-drugs increase the risk of dementia.