Increase in anticholinergic burden from 1990 to 2015: Age‐period‐cohort analysis in UK biobank

The use of prescription drugs with anticholinergic properties has been associated with multiple negative health outcomes in older people. Moreover, recent evidence suggests that associated adverse effects may occur even decades after stopping anticholinergic use. Despite the implicated importance of examining longitudinal patterns of anticholinergic prescribing for different age groups, few such data are available.

of drugs. 5 This is especially pertinent in the case of anticholinergic compounds, which are commonly prescribed and whose side effects are well documented. 1 Anticholinergic burden in older adults is associated with reduced physical and cognitive ability, 6,7 impaired ability to perform activities of daily living, 8 increased risks of falls, 9 dementia 10 and mortality. 11 The association with dementia has been observed even when the anticholinergic exposure occurred decades prior to diagnosis. 10 Several tools to assess inappropriate prescribing have been developed in the last few decades. 12 Subsequently, inappropriate prescribing in older people declined from 45.5% to 40.8% between 2006/2007 and 2009/2010 in the United States, 13 and from 32.2% to 28.3% between 1996 and 2005 in the UK. 14 However, older adults remain exposed to anticholinergic drugs, 15 the prevalence of which has remained stable in the United States, 16 but has increased by 3% in the UK from 1995 to 2010 17 and by 12.5% in Finland from 2007 to 2017. 18 While some studies have found associations between anticholinergic use and demographic factors, 16,17,19 these variables are rarely examined in detail. Moreover, it is not known whether these potential group differences persist over time.
The study of temporal changes of prescribing practices with indepth assessment of age-period-cohort (APC) effects necessitates longitudinal designs. Cross-sectional studies 15 or repeated crosssectional studies 17,18 have explored the extent of anticholinergic use in European countries, 17,18 but the last year of sampling in the UK was in 2010. 17 Moreover, they either lack longitudinal data or rely on participants from a relatively limited geographic area and within a narrow age range. In this paper, we address those limitations by using a large national sample from UK Biobank to characterise longitudinal prescribing patterns of anticholinergic drugs in 1990-2015.

| Hypotheses
We based our hypotheses on previous cross-sectional studies in the UK that showed greater polypharmacy 4 and anticholinergic burden 17 in 2010 when compared to 1995, and increased polypharmacy in older individuals. 4 We hypothesised that anticholinergic burden increased as a function of both period and age. Additionally, we hypothesised that anticholinergic burden was higher in women and in less educated individuals, as had been reported before. 15,17

| Sample
UK Biobank is a prospective study of >500 000 participants aged 37-73 years, recruited in 22 assessment centres throughout the UK in 2006-10. 20 To ensure a representative sample in the given age range, eligible participants for the study were identified through general practice registers and invited by post. The assessments consisted of touch-screen questionnaires, computer-assisted interviews, measures of physical function and the collection of blood, saliva and urine. Primary care prescriptions were available for $230 000 participants to May 2017 for Scotland, to September 2017 for Wales and to August 2017 for England. The data were provided to UK Biobank by regionspecific data providers and include, among other information, the dates of prescriptions, names of drugs prescribed and drug codes. The latter include BNF codes provided by the British National Formulary, which provides prescribing guidance on medicines (https://www.bng. org/), Read v2-and CTV3-codes provided by the Terminology Reference Data Update Distribution (TRUD) service (https://isd.digital.nhs. uk/), and dmd + d codes provided by the National Health Service (NHS) (https://www.nhs.uk/). The drug code systems are used as dictionaries for medicines.

| Assignment of anticholinergic burden and drug class
Several resources allow for the identification of drugs with anticholinergic properties and provide a score of anticholinergic potency for each drug. These anticholinergic burden scales derive the lists of drugs from different sources, utilise different methods to assign the scores and validate the resulting tools in different populations and on different outcome measures. Previous studies have compared various existing anticholinergic scales and have generally reported poor overlap among them. [21][22][23] For the purposes of our study, we identified What is already known about this subject • Anticholinergic burden has been associated with reduced physical and cognitive ability, and an increased risk of dementia and all-cause mortality. Extant epidemiological studies in Europe suggest an increase in anticholinergic prescribing over time, but focus on limited geographic areas, utilise cross-sectional designs or focused on individuals in older age, despite the potential importance of anticholinergic exposure throughout life.

What this study adds
• We performed an age-period-cohort analysis of changes in anticholinergic burden in >220 000 participants from UK Biobank, using GP electronically prescribed longitudinal data from 1990 to 2015. Anticholinergic burden in the UK has increased across several age groups and classes of prescription drugs. The increase was related to both ageing of the underlying sample as well as periodrelated changes in prescribing. multiple scales 24-33 from a systematic review 34 ; apart from two 30,33 all had a four-point (0-3) scoring system of anticholinergic potency, where a lower score corresponds to lower anticholinergic potency (Supporting Information Table S1). We derived a meta-scale (Supporting Information Table S2) by calculating the mean anticholinergic burden across all nine original scales that had rated a drug. Thus, scales that scored a drug (even if that score was zero) were included in the computation for that drug, while scales that did not score the drug were not. All prescriptions of medicines with ophthalmic, otic, nasal or topical routes of administration were assigned an anticholinergic score of zero, as has been done before. 29,31,34,35 For prescription entries that did not list any drugs (ie, for which the column indicating the name of the drug was blank), drug codes were used to supplement them. A series of steps was taken to exclude incomplete data or low number of individuals (Supporting Information  (2), lowers bloodglucose (3) and is a biguanide (4). Not all classes were equally represented in the sample. To allow for comparability of frequency of occurrence, we classified anticholinergic drugs into classes that do not all correspond to the same level in the ATC hierarchy (see number in parentheses): "drugs for acid disorders" (3), "analgesics" (2), "antidepressants" (3), "antithrombotic drugs" (2), "cardiovascular drugs" (1), "drugs for diabetes" (2), "gastrointestinal drugs" (2), "psycholeptics" (2), "respiratory drugs" (1) and "urological drugs" (3). A final class of "other drugs" was constructed that contained drugs that primarily due to their low prevalence individually contributed relatively little to the total anticholinergic burden. These included anticonvulsants, antibiotics, anti-Parkinsonian drugs, corticosteroids, immunosuppressants, anti-inflammatory drugs, muscle relaxants and anti-diarrhoeal drugs.

| Statistical approach
To enable longitudinal analyses, the original format of the data was transformed into two different formats that reflected for each participant the monthly and yearly anticholinergic burden, respectively.
These period-based anticholinergic burden scores were calculated by summing the anticholinergic burden of all prescriptions in that period (Supporting Information Figure S3 and Text S1). When individual-level data are collected longitudinally, changes can be due to age, period or cohort effects. 36 Because the three effects are colinear (age = period cohort), they cannot all be included in a regression analysis, as holding two terms constant keeps the third term constant as well. 37 While there have been attempts to estimate the unique contributions among the three effects, 38,39 no solution has been widely adopted. Hidden assumptions can have a strong effect on the interpretation of the APC effect, 40 and in our analysis we make the following assumptions. First, age is probably positively associated with anticholinergic burden due to the positive association between polypharmacy and age. 4 Second, we assume that birth cohort does not play a role in the above association, ie, we are only interested in whether a potential longitudinal change in anticholinergic burden is due to the participants' age or due to changes in prescribing practices over time. For the analysis of APC effects, we ran three models, excluding one of the APC terms at a time (ie, its effect was assumed to be zero). Thus, anticholinergic burden was modelled as a function of either period and cohort (period-cohort model), age and cohort (age-cohort model) or age and period (ageperiod model). This three-model approach represents the same processthe change in anticholinergic burden in the samplefrom three different perspectives and allows for an appraisal of possible drivers of the observed trend. For example, assuming that the effect of birth cohort is zero, positive effects for both period and cohort in the period-cohort model, and a positive effect of period, but a negative effect of age in the age-period model demonstrates that anticholinergic burden (a) increases with time across cohorts, (b) is higher in younger cohorts in a given period, (c) decreases with age and (d) is higher in recent periods across age groups. This would suggest that the anticholinergic burden increased with the time period but decreased with age. We additionally computed the above models by fitting separate intercepts and slopes: for the period-cohort and agecohort models separate intercepts and slopes for each cohort, and for the age-period model separate intercepts and slopes for each period.
For analyses of lifestyle and demographic factors, we fitted tobit linear models 41 to average monthly anticholinergic burden, adjusting for sex, education, physical activity, social deprivation, region, smoking, body mass index (BMI), frequency of alcohol consumption and age at assessment. Tobit models are models of censored regression, where the values that fall either above or below a certain value are censored.
In our analysis, tobit models were censored from below at 0, effectively simulating zero inflation. For models with random effects, we used generalised linear mixed models (R package glmmTMB 42 ); for all other models, we used Tobit regression (R package censReg). Due to the relative infrequency of anticholinergic drugs, anticholinergic burden was right-skewed and models were adjusted for zero inflation. A sensitivity analysis was conducted that was limited to the period from 2000 to 2015. This was done due to the relatively low level of ascertainment in the sample before that period (Supporting Information Figure S2).

| Nomenclature of targets and ligands
Key protein targets and ligands in this article are hyperlinked to corresponding entries in http://www.guidetopharmacology.org, the common portal for data from the IUPHAR/BPS Guide to PHARMA-COLOGY, and are permanently archived in the Concise Guide to PHARMACOLOGY 2019/20. 46

| RESULTS
The 220 867 participants were born between 1938 and 1969 (Supporting Information Figure S4). Individuals were being added to the database of prescriptions throughout the sampling period , but the demographic structure of the sample (Table 1) remained relatively stable over time. However, it is unclear how demographic variables changed within individuals over time.  between 3-and 9-fold from 1990 to 2015. Most anticholinergic prescriptions were for antidepressants, which accounted for 32.5% of the total anticholinergic burden (Table 3 and Figure 1B). The anticholinergic burden for each drug class increased with time ( Figure 1C).

| Age-period-cohort analysis
In the basic period-cohort model, anticholinergic burden was positively associated with period and negatively associated with cohort.
In the basic age-cohort model, anticholinergic burden was positively associated with age and with cohort. In the basic age-period model, anticholinergic burden was positively associated with age and with period. The same trends were observed in the basic-adjusted and fully adjusted models (Supporting Information Table S3). These results indicate that greater anticholinergic burden relates to both ageing and later period. That is, in a given period, older individuals experience a higher anticholinergic burden than younger individuals in the same period. Moreover, in recent periods, individuals will experience a higher anticholinergic burden than individuals of the same age did in the past.  Table S4). The proportion of drugs with different anticholinergic potencies remained stable over time (Supporting Information Figure S5). Thus, the increase in anticholinergic burden was likely due to a general increase in anticholinergic prescribing, rather than a relative increase in the prescribing of stronger anticholinergic drugs.
When the change in anticholinergic burden was plotted for each drug class separately (Supporting Information Figure S6

| Anticholinergic burden and demographic factors
Higher anticholinergic burden was associated with female sex, lower educational attainment, greater deprivation, higher BMI, less frequent alcohol consumption and lower physical activity, and was greater in Scotland and Wales than in England (Table 4).
Examining each drug class separately, most effects remained (Supporting Information Table S6). However, anticholinergic burden due to antithrombotic drugs, cardiovascular drugs and drugs for diabetes was higher in males than in females. Moreover, regional differences in anticholinergic burden strongly depended on drug class.
Deprivation was transformed into a binary categorical variable, with the median ($2.2) across all participants defining the groups. For

F I G U R E 2 APC analysis with basic mixed models with random intercepts and slopes (left) and associations between slopes and different levels of predictors (right). (A)
The period-cohort model with cohort as a random effect, (B) the age-cohort model with cohort as a random effect and (C) the age-period model with period as the random effect region, sex, education and deprivation, we then plotted anticholinergic burden as a function of period for different levels of predictor variables. Supporting Information Figure S8 illustrates the association between the above predictors and anticholinergic burden.

| Sensitivity analyses
When the observation period was restricted to prescriptions after 1999, the trends above were again observed for all models except for when polypharmacy was used as covariate (Supporting Information   Tables S7-10, Text S2 and Figure S9). There, period was negatively associated with anticholinergic burden in the period cohort and in the age-period model. Age was negatively associated with anticholinergic burden in both the age cohort and the age-period model. Birth cohort was positively associated with anticholinergic burden in the periodcohort model, but negatively associated with anticholinergic burden in the age-cohort model. Thus, when accounting for polypharmacy, the sensitivity analysis supports an age-related decrease in anticholinergic burden, but does not support a period-related increase in anticholinergic burden.

| DISCUSSION
In a large longitudinal study of prescription drugs with anticholinergic properties, we showed that the anticholinergic burden in the UK is increasing, and older individuals continue to have the highest anticholinergic burden. Age-related increases in anticholinergic burden can be explained by polypharmacy in older adults. Indeed, when accounting for polypharmacy and period, anticholinergic burden decreases with age, possibly demonstrating proportionate deprescribing of anticholinergic drugs in older age. We also find associations between higher anticholinergic burden and various demographic and lifestyle factors, including female sex, less education and greater socioeconomic deprivation.

| Anticholinergic burden over time
Anticholinergic burden increased in all APC models. Throughout time periods and across birth cohorts, ageing was associated with greater anticholinergic burden. Moreover, across age groups and birth cohorts, anticholinergic burden has increased in recent years. Finally, at a given age, later-born cohorts experienced a greater anticholinergic burden than earlier-born cohorts, while in a given period, later-born cohorts experienced a smaller anticholinergic burden than earlier-born cohorts.
Because of the collinearity of age, period and cohort (age = period cohort), they cannot all be included in a regression analysis, as holding two terms constant keeps the third term constant as well. 37 Some argue that the APC problem cannot be completely resolved 40 and that results from APC-based models should be based on well-founded and clearly communicated assumptions. In the present paper we assumed no cohort effects and predicted anticholinergic burden to increase with ageing. Based on current knowledge on polypharmacy and anticholinergic burden, the following conclusions can be drawn from our results. First, due to increased multimorbidity and polypharmacy in older individuals, 4 age contributed to the trend.
When intercept and slope were modelled separately in mixed models with random effects, cohort was negatively associated with the slope, suggesting not only a greater anticholinergic burden, but also a more The increases in anticholinergic burden could be related to an increase in general polypharmacy and not an increase in specifically anticholinergic prescribing. Indeed, when the models were adjusted for the number of prescriptions, the changes in anticholinergic burden were greatly diminished. Furthermore, earlier-born individuals exhibited a lower anticholinergic burden across periods and across age groups than those born later. Moreover, across age groups, anticholinergic burden was higher in later periods than in earlier periods.
While correcting for polypharmacy had no effect on the trend of the age-cohort model, it changed the direction of birth cohort and age in the period-cohort model and the age-period model, respectively.
Later-born individuals exhibited a higher anticholinergic burden, and this burden was positively associated with period, but negatively asso-

| Demographic-and lifestyle factors
Anticholinergic use has been linked with some demographic and lifestyle factors. 16,17,19 In our study, female sex, lower education, higher socioeconomic deprivation, higher BMI, lower frequency of alcohol consumption, lower physical activity and being prescribed in Scotland or Wales (compared to England) were associated with a higher anticholinergic burden. Certain groups do require a greater number of medications but medical professionals may prescribe more to certain groups, independent of underlying medical conditions.
Interestingly, greater alcohol consumption was associated with decreased anticholinergic burden. Individuals who take many medications may reduce their alcohol consumption to reduce the risk of drug interactions or to reduce the impact of existing disease.

| Strengths and limitations
The present study used a very large, well-characterised sample and utilised primary care electronic prescription data over a wide period.

| CONCLUSION AND FUTURE DIRECTIONS
Prescribing drugs involves balancing their medicinal value with potential harms. Moreover, exhaustive longitudinal studies are required to fully determine all their effects. However, besides well-documented side effects, 48 exposure has been linked to an increased frequency of falls, 9 reduced physical, cognitive and functional ability, 6-8 and increased risks of dementia 10 and all-cause mortality. 11 Thus, anticholinergic drugs ought to be prescribed sparingly and the use of alternatives strongly considered. An understanding of temporal prescribing trends in a population may help to guide prescribing and stimulate further research. Our work represents an overview and future studies should describe prescribing trends and their relationship to age groups, and demographic and lifestyle characteristics in greater detail.
There is also evidence of differences between drug classes in the association between anticholinergic burden and health outcomes. 10 Identifying distinct anticholinergic trends for individual drug classes for different groups could help to further improve prescribing guidelines. Additionally, future work should attempt to identify the causes for the increase in anticholinergic prescribing, and more precisely quantify the potential implications for important life outcomes, including brain and cognitive health, and dementia. Finally, decreases in potentially inappropriate prescribing have been reported even when the same population experienced increases in polypharmacy and in anticholinergic use. 49 Thus, increases in anticholinergic burden should not be considered in isolation, but in the context of other prescribing practices.