A two-part model to estimate inpatient, outpatient, prescribing and care home costs associated with atrial fibrillation in Scotland

Objective This study aimed to estimate global inpatient, outpatient, prescribing and care home costs for patients with atrial fibrillation using population-based, individual-level linked data. Design A two-part model was employed to estimate the probability of resource utilisation and costs conditional on positive utilisation using individual-level linked data. Settings Scotland, 5 years following first hospitalisation for AF between 1997 and 2015. Participants Patients hospitalised with a known diagnosis of AF or atrial flutter. Primary and secondary outcome measures Inpatient, outpatient, prescribing and care home costs. Results The mean annual cost for a patient with AF was estimated at £3785 (95% CI £3767 to £3804). Inpatient admissions and outpatient visits accounted for 79% and 8% of total costs, respectively; prescriptions and care home stay accounted for 7% and 6% of total costs. Inpatient cost was the main driver across all age groups. While inpatient cost contributions (~80%) were constant between 0 and 84 years, they decreased for patients over 85 years. This is offset by increasing care home cost contributions. Mean annual costs associated with AF increased significantly with increasing number of comorbidities. Conclusion This study used a contemporary and representative cohort, and a comprehensive approach to estimate global costs associated with AF, taking into account resource utilisation beyond hospital care. While overall costs, considerably affected by comorbidity, did not increase with increasing age, care home costs increased proportionally with age. Inpatient admission was the main contributor to the overall financial burden of AF, highlighting the need for improved mechanisms of early diagnosis to prevent hospitalisations.


Article summary
Strengths and limitations of this study  Costs are estimated through an incidence-based approach using patient-level morbidity records.
 Sufficient follow-up time is used to capture all relevant costs to generate a contemporary estimate of health and care home costs related to AF.
 Scotland offers a robust record linkage system, where administrative patient-level health data are routinely collected.
 Data on primary care consultations were not available for linkage at a national level, however the impact this might have on overall costs is expected to be small.
 The potential risk of AF going undiagnosed and clinical miscoding of morbidity records may lead to an underestimation of the AF cohort and associated costs.  of the adult population, and rises to 6% among those aged 65 years or over [1]. In an ageing population, AF has a substantial impact on the economic burden of the healthcare system.
A number of cost analyses on estimating the economic burden of AF exist. The majority of these studies are of selective cohorts of AF patients, based on data sourced from administrative database [2][3][4], health insurance databases [2,[5][6][7], hospital records [8,9] and surveys [10].
Many of these studies included relatively young patients -those aged 18-20 years or older [2,[4][5][6][8][9][10], or those under the age of 65 years [7]. However, the prevalence of AF increases significantly with age, and is most affected by patients who are older than 50 years; in patients under the age of 50, AF is often associated with structural heart disease, hyperthyroidism, or alcohol excess [11]. Hence, inclusion of younger patients and the exclusion of older patients may result in imprecise cost estimates.
There is a lack of generalisable studies based on large national population datasets that examine the total and the distribution of costs associated with AF [12]. The aim of this study was to quantify the inpatient, outpatient, prescribing and care home costs associated with AF over a five-year period. Using record-linkage of national datasets from Scotland, we also examined the distribution of costs that are attributable to AF.

Methods
Cost analyses or cost of illness studies typically adopt either the prevalence or incidence based approaches [13]. In the context of AF, the prevalence based approach determines costs attributable to all cases of AF in a given year, while the incidence based approach determines costs of new cases of AF in a given time period. In the present study, costs were estimated with an incidence-based approach. Records 01 (SMR01). These records contain all general acute admissions, categorized as inpatients or day cases, discharged from non-obstetric and non-psychiatric specialties [14].

Data
Incident AF events (ICD10 code I48) were identified using all six diagnostic positions in SMR01, with a look back period of five years to minimise double counting. After checking for data entry errors and removal of duplicate records, an initial AF cohort consisting of 279,883 individuals hospitalised with a diagnosis of AF or atrial flutter was identified. For the purpose of this analysis, only patients aged 50 years or older were included. Based on clinical advice and the evidence that prevalence and incidence of AF typically increase exponentially from 50 years onwards [11], the analysis including 50 years age group at the lowest range would be inclusive of all patients potentially at risk of AF. The final dataset for analysis consisted of 272,716 patients. Individual-level data linkage was then carried out with outpatient clinic attendance (Outpatient Attendance Scottish Morbidity Records 00; SMR00), the prescribing information system (PIS), care home census and mortality records (National Records for Scotland, NRS). Records from SMR00 include information on new and follow up outpatient appointments for any clinical specialty [15]. The PIS database includes prescribing records for all medicines and their associated costs, which are prescribed and dispensed by community pharmacies, dispensing doctors and a small number of specialist appliance suppliers [16]. The quality of PIS data is guaranteed by an electronic data capture, and it passes several stages of quality control before and after data are submitted [17]. The care home census combines the former Residential Care Home Census (run by the Scottish Government) and the Private Nursing Homes Census (run by ISD Scotland). Items reported in the care home census include discharge dates to care home residency such as NHS and private nursing homes, as well as an indication on whether nursing care is required [16].
Patients were followed up for five years following incident AF event in terms of their healthcare resource use, care home admissions and mortality. Since AF is often a precursor of stroke and cardiovascular conditions, an estimation of costs for a period of five years post AF event would allow us to fully capture costs associated with an AF patient.

Costing
Inpatient care costs were obtained from the latest (2013/2014) Scottish National Tariff (SNT), a list of standard average prices based on Healthcare Resource Groups (HRGs) [17,18]. The SNT uses HRG4 for grouping clinically similar treatments that use similar levels of healthcare resources. After defining a total cost per episode, the total cost for a continuous inpatient stay (CIS) was calculated. A CIS describes the entire duration of an inpatient stay from the date of admission to the date of discharge and can consist of several episodes in different specialties. Since the SNT is based on spells of care (inpatient stay within the same specialty) rather than individual inpatient episodes or a CIS, a CIS was partitioned into spells when a change in specialty occurred [17].
If within a CIS, two or more episodes were in the same specialty, only the highest incurred cost was taken into account, and the remaining episodes were replaced with a zero cost. Outpatient costs were obtained by assigning outpatient specialty costs, to outpatient attendances [17]. Unit costs were specific to whether the outpatient attendance took place at a consultant or nurse led clinic [15].
The cost of each prescription dispensed per patient was obtained from PIS [19]. Firstly, the price per unit was obtained by dividing the item price by the pack size. Secondly, the total number of items dispensed was obtained by multiplying the number of items dispensed by the number of instalments. Care home costs, obtained from the care home census, were based on length of stay or residency. Care home residency was established from care home census records, reporting admission to a care home like structure [16]. An average of care home charges for long stay residents was calculated using information on whether nursing care was provided or not. The average weekly care home charge was expressed per day, so that only the effective days spent in a care home were costed.

Econometric model
Healthcare expenditure data are typically characterised by: i) a significant proportion of zerocost observations for individuals who have not utilised any healthcare resources in a given time period, and ii) akewed distribution for positive costs. A two-part model was used [20,21]. In order to account for the skewed nature of cost data, generalised linear models (GLMs) were used. These were compared against ordinary least squares regression (OLS) and log transformed OLS by means of the Akaike Information Criterion (AIC), which measures goodness of fit. When comparing the different models, GLM reported the lowest AIC indicating the best fit for the given set of data. A user-written STATA programme "glmdiagnostic.do" [20], performing four different tests simultaneously, was used to identify the most appropriate distributional family and link function.

Econometric model covariates
The two-part model adjusted for age, sex, year of inpatient admission, socio-economic status, urban-rural classification, health board, comorbidities and mortality. These covariates are considered to be the main confounders that have an effect on costs incurred by an AF population. We controlled for age because AF and associated comorbidities are age-related conditions, and may have an impact on the overall costs. We also assumed costs to vary between males and females, in particular those for care home residency. Variation in healthcare utilisation and associated costs and care home residency by socio-economic status is controlled for using the Scottish Index of Multiple Deprivation (SIMD). The SIMD reflects areas of multiple deprivation ranked from the most to the least deprived and expressed as quintiles where the most and the least deprived areas are represented by 1 and 5 respectively [22]. In Scotland, there are 14 regional health boards responsible for the provision of healthcare [23]. Hence, potential differences in healthcare utilisation and prescribing costs, may reflect variation in clinical practice and prescribing behaviour rather than the ability of patients to access care. Patients living in urban areas may have easier access to care compared to patients living in more remote areas, which is controlled for including the 8-fold classification measuring rurality [24].
Patients with one or more comorbidities are expected to incur significantly higher costs than those with none. We accounted for this by including the Charlson comorbidity index, where 1 indicates the absence of comorbidities, 2 the presence of only a single comorbidity and 3 the presence of more than one comorbidity [25]. Two interaction terms between age and comorbidities, and mortality and SIMD were included in the econometric model. Intuitively, a relationship of direct proportionality between age and comorbidities suggests that the level of comorbidities increases, as patients get older. Similarly, the socio-economic status may significantly influence the rate of socio-economic inequalities in mortality [26].

Ethics statement
The authors state that no ethical approval was needed.

Econometric modelling results
Regression results for both modelling parts are presented in Table 2. Overall, an inversely Ushaped association between age and the likelihood of utilising any health or social care services was observed -a gradual increment in the likelihood in resource use with advancing age up to 80 years, when compared with the reference group (50-54 years), while patients 80 years or older showing a decreased probability of utilising healthcare services. However, this association was not observed in the second modelling part model, estimating costs conditional on having incurred positive costs, where a statistically significant gradient between age and costs indicated increasing costs as the cohort ages. The use of health or social care services and associated costs also increased significantly for patients living in the most deprived areas, when compared with patients living in areas with the lowest level of deprivation. The effect of socioeconomic status on healthcare utilisation was also measured for those who are alive at the end of the five-year follow-up period through an interaction term between SIMD and mortality, but no statistically significant effect was found.  (Table I, please see online supplement).
For patients with comorbidities, the probabilities of utilising healthcare services were 49.5% (one comorbidity) and 78.6% (two or more comorbidities) greater than the probability for those with no comorbidities. Although healthcare utilisation increased with the number of comorbidities, the interaction term between age and comorbidities indicated that as patients get older the use of healthcare services on average is lower for patients with one or more comorbidities than those with none. The decrease in healthcare utilisation by age is more pronounced in patients with more comorbidities than in those with only one comorbidity. The difference in healthcare costs between comorbidity categories indicated that in the presence of one or more comorbidities, on average healthcare costs decrease as patients get older.

Cost estimates
The estimated mean annual cost per AF patient was £3861 (95% CI: £3842-£3880). The estimated total costs and distribution of costs according to sex are shown in Table 3. While there is little difference between the total costs and the distribution of costs for inpatient, outpatient and prescription costs, the difference seems more pronounced when comparing the care home component of costs (5% of total costs among males vs 7% of total costs among females). The average annual cost per AF patient by age and for each health or care home sector is shown in Figure 1. This, combined with the likelihood for people living in the most deprived quintile having longer inpatient stays due to a lack of support at home, may explain the difference in inpatient care utilisation between patients from the most and the least deprived areas, with associated costs being higher for the former group. As AF is more likely to affect the elderly, AF related costs were expected to increase with age. As health deteriorates with age, older age groups are assumed to make greater use of healthcare services, and therefore incur higher costs than younger age groups. However, age was found to have a modest impact on overall healthcare costs, being fairly consistent across age groups. This finding is in line with existing evidence indicating that healthcare expenditure depends not only on patients' calendar age, but is also significantly associated with remaining lifetime [27]. Any observed correlation between healthcare expenditure and age may therefore be attributable to the fact that the proportion of patients who are at the end of their lives is substantially greater in older rather than younger age groups [27]. On the other hand, comorbidity had a considerable effect on the overall cost, increasing significantly in patients with more than one comorbidity.
However, the decrease in healthcare utilisation by age is more pronounced in patients with more comorbidities than in those with only one comorbidity. Decreasing inpatient and outpatient costs for the oldest patients were offset by increasing care home costs, in particular for women. Indeed, the main cause for higher overall costs incurred by women is attributable to the higher likelihood for elderly women to reside in care homes. Interestingly, care home contribution to the overall costs was noticeably lower for patients with multiple comorbidities than for those with none or one comorbidity. This may suggest that sicker patients are more likely to be in hospital than in a care home.
To date, only one single study published in 2004 has estimated the cost of AF in Scotland; the authors estimated the cost of AF in 1995/1996, and projected these to the year 2000 [28].
Previous work has focussed on a 12-months follow-up, which seems limited in order to capture all healthcare resource utilisation for AF patients. Our study offers a longer follow-up and a contemporary estimate of healthcare costs related to AF including all relevant care settings.
Many of these studies included relatively young patients -those aged 18-20 years or older [2,[4][5][6][8][9][10], or those under the age of 65 years [7]. However, the prevalence of AF increases significantly with age, and is most affected by patients who are older than 50 years; in patients under the age of 50, AF is often associated with structural heart disease, hyperthyroidism, or alcohol excess [11]. Hence, inclusion of younger patients and the exclusion of older patients may result in imprecise cost estimates.
There is a lack of generalisable studies based on large national population datasets that examine the total and the distribution of costs associated with AF [12]. The aim of this study was to quantify the inpatient, outpatient, prescribing and care home costs associated with AF over a five-year period. Using record-linkage of national datasets from Scotland, we also examined the distribution of costs that are attributable to AF.

Methods
Cost analyses or cost of illness studies typically adopt either the prevalence or incidence based approaches [13]. In the context of AF, the prevalence based approach determines costs attributable to all cases of AF in a given year, while the incidence based approach determines costs of new cases of AF in a given time period. In the present study, costs were estimated with an incidence-based approach. Records 01 (SMR01). These records contain all general acute admissions, categorized as inpatients or day cases, discharged from non-obstetric and non-psychiatric specialties [14].

Data
Incident AF events (ICD10 code I48) were identified using all six diagnostic positions in SMR01, with a look back period of five years to minimise double counting. After checking for data entry errors and removal of duplicate records, an initial AF cohort consisting of 279,883 individuals hospitalised with a diagnosis of AF or atrial flutter was identified.
For the purpose of this analysis, only patients aged 50 years or older were included. Based on clinical advice and the evidence that prevalence and incidence of AF typically increase exponentially from 50 years onwards [11], the analysis including 50 years age group at the lowest range would be inclusive of all patients potentially at risk of AF. doctors and a small number of specialist appliance suppliers [16]. The quality of PIS data is guaranteed by an electronic data capture, and it passes several stages of quality control before and after data are submitted [17]. The care home census combines the former Residential Care Home Census (run by the Scottish Government) and the Private Nursing Homes Census (run by ISD Scotland). Items reported in the care home census include discharge dates to care home residency such as NHS and private nursing homes, as well as an indication on whether nursing care is required [16].
Patients were followed up for five years following incident AF event in terms of their healthcare resource use, care home admissions and mortality. Since AF is often a precursor of stroke and cardiovascular conditions, an estimation of costs for a period of five years post AF event would allow us to fully capture costs associated with an AF patient. Costing Since the SNT is based on spells of care (inpatient stay within the same specialty) rather than individual inpatient episodes or a CIS, a CIS was partitioned into spells when a change in specialty occurred [17]. If within a CIS, two or more episodes were in the same specialty, only the highest incurred cost was taken into account, and the remaining episodes were replaced with a zero cost. Outpatient costs were obtained by assigning outpatient specialty costs, to outpatient attendances [17]. Unit costs were specific to whether the outpatient attendance took place at a consultant or nurse led clinic [15].
The cost of each prescription dispensed per patient was obtained from PIS [19]. Firstly, the price per unit was obtained by dividing the item price by the pack size. Secondly, the total number of items dispensed was obtained by multiplying the number of items dispensed by the number of instalments. Care home costs, obtained from the care home census, were based on length of stay or residency. Care home residency was established from care home census records, reporting admission to a care home like structure [16]. An average of care home charges for long stay residents was calculated using information on whether nursing care was provided or not. The average weekly care home charge was expressed per day, so that only the effective days spent in a care home were costed. The tariffs used for costing account for inflation, therefore further cost adjustment was not needed.

Econometric model
Healthcare expenditure data are typically characterised by: i) a significant proportion of zerocost observations for individuals who have not utilised any healthcare resources in a given time period, and ii) a skewed distribution for positive costs. A two-part model was used [20,21].
In the first part of the model, the probability of using a healthcare service in a given time period In order to account for the skewed nature of cost data, generalised linear models (GLMs) were used. These were compared against ordinary least squares regression (OLS) and log transformed OLS by means of the Akaike Information Criterion (AIC), which measures goodness of fit. When comparing the different models, GLM reported the lowest AIC indicating the best fit for the given set of data. A user-written STATA programme "glmdiagnostic.do" [20], performing four different tests simultaneously, was used to identify the most appropriate distributional family and link function.

Econometric model covariates
The two-part model adjusted for age, sex, year of inpatient admission, socio-economic status, urban-rural classification, health board, comorbidities and mortality. These covariates are considered to be the main confounders that have an effect on costs incurred by an AF population. We controlled for age because AF and associated comorbidities are age-related conditions, and may have an impact on the overall costs. We also assumed costs to vary between males and females, in particular those for care home residency. Variation in healthcare utilisation and associated costs and care home residency by socio-economic status is controlled for using the Scottish Index of Multiple Deprivation (SIMD). The SIMD reflects areas of multiple deprivation ranked from the most to the least deprived and expressed as quintiles where the most and the least deprived areas are represented by 1 and 5 respectively [22]. In Scotland, there are 14 regional health boards responsible for the provision of healthcare [23]. Hence, potential differences in healthcare utilisation and prescribing costs, may reflect variation in clinical practice and prescribing behaviour rather than the ability of patients to access care.
Patients living in urban areas may have easier access to care compared to patients living in more remote areas, which is controlled for including the 8-fold classification measuring rurality [24].
Patients with one or more comorbidities are expected to incur significantly higher costs than those with none. We accounted for this by including the Charlson comorbidity index, where 1 indicates the absence of comorbidities, 2 the presence of only a single comorbidity and 3 the presence of more than one comorbidity [25]. Two interaction terms between age and comorbidities, and mortality and SIMD were included in the econometric model. Intuitively, a relationship of direct proportionality between age and comorbidities suggests that the level of comorbidities increases, as patients get older. Similarly, the socio-economic status may significantly influence the rate of socio-economic inequalities in mortality [26].

Sensitivity analyses
In order to ascertain whether mortality had an impact on overall AF related healthcare costs, average annual cost per patient by age and for each health or care home sector, was estimated for patients who were alive and those who were dead at the end of the five-year follow-up period. The two econometric models (Equation IV and V, please see online supplement) followed the same structure of the model described in the previous section and used for the main analysis; however, those models were not adjusted for mortality.

Ethics statement
The authors state that no ethical approval was needed.

Patients and public involvement
There was no patients or public involvement

Econometric modelling results
Regression results for both modelling parts are presented in Table 2. Overall, an inversely Ushaped association between age and the likelihood of utilising any health or social care services was observed -a gradual increment in the likelihood in resource use with advancing age up to  (Table I, please see online   supplement).
For patients with comorbidities, the probabilities of utilising healthcare services were 49.5% (one comorbidity) and 78.6% (two or more comorbidities) greater than the probability for those with no comorbidities. Although healthcare utilisation increased with the number of comorbidities, the interaction term between age and comorbidities indicated that as patients get older the use of healthcare services on average is lower for patients with one or more comorbidities than those with none. The decrease in healthcare utilisation by age is more pronounced in patients with more comorbidities than in those with only one comorbidity. The difference in healthcare costs between comorbidity categories indicated that in the presence of one or more comorbidities, on average healthcare costs decrease as patients get older.
Full details of regression results for patients who were alive and those who were dead at the end of the five-year follow-up period are presented in the supplementary online material (Table   II -V, please see online supplement).

Cost estimates
The estimated mean annual cost per AF patient was £3861 (95% CI: £3842-£3880). The estimated total costs and distribution of costs according to sex are shown in Table 3. While there is little difference between the total costs and the distribution of costs for inpatient, outpatient and prescription costs, the difference seems more pronounced when comparing the care home component of costs (5% of total costs among males vs 7% of total costs among females).
The average annual cost per AF patient by age and for each health or care home sector is shown in Figure 1. The contribution of each setting to the total health and care home costs by the number of existing comorbidities is illustrated in Figure 2. While inpatient and total costs vary considerably with the number of comorbidities, outpatient and care home contributions remain fairly constant.
The estimated mean annual cost per AF patient alive at the end of the five-year follow-up period was £3110 (95% CI: £3090-£3131). The average annual cost per AF patient by age and for each health or care home sector is presented in the supplementary online material ( Figure   I, please see online supplement). For these patients, inpatient costs was the main driver across all age groups; a gradient between age and costs indicated increasing costs as the cohort ages.
Similar patterns were observed for care home costs. On the contrary, outpatient and prescribing costs remained constant up to 74 years, but decreased slightly for older patients.
The estimated mean annual cost per AF patient who died during the five-year follow-up period, was £2299 (95% CI: £2279-£2319) ( Figure II, please see online supplement). For these patients, inpatient costs was the main driver across all age groups; a gradient between age and costs indicated decreasing costs as the cohort ages. This was also observed for outpatient and prescribing costs; but care home costs on average increased across age groups.

Discussion
A greater proportion of AF patients were found in areas with the highest index of deprivation.
This, combined with the likelihood for people living in the most deprived quintile having longer inpatient stays due to a lack of support at home, may explain the difference in inpatient care utilisation between patients from the most and the least deprived areas, with associated costs being higher for the former group. As AF is more likely to affect the elderly, AF related costs were expected to increase with age. As health deteriorates with age, older age groups are assumed to make greater use of healthcare services, and therefore incur higher costs than younger age groups. However, age was found to have a modest impact on overall healthcare costs, being fairly consistent across age groups. This finding is in line with existing evidence indicating that healthcare expenditure depends not only on patients' calendar age, but is also significantly associated with remaining lifetime [27].
Any observed correlation between healthcare expenditure and age may therefore be attributable to the fact that the proportion of patients who are at the end of their lives is substantially greater in older rather than younger age groups [27]. On the other hand, comorbidity had a considerable effect on the overall cost, increasing significantly in patients with more than one comorbidity.
However, the decrease in healthcare utilisation by age is more pronounced in patients with more comorbidities than in those with only one comorbidity. Decreasing inpatient and outpatient costs for the oldest patients were offset by increasing care home costs, in particular for women. Indeed, the main cause for higher overall costs incurred by women is attributable to the higher likelihood for elderly women to reside in care homes. Interestingly, care home contribution to the overall costs was noticeably lower for patients with multiple comorbidities than for those with none or one comorbidity. This may suggest that sicker patients are more likely to be in hospital than in a care home. To date, only one single study published in 2004 has estimated the cost of AF in Scotland; the authors estimated the cost of AF in 1995/1996, and projected these to the year 2000 [28].
Previous work has focussed on a 12-months follow-up, which seems limited in order to capture all healthcare resource utilisation for AF patients. Our study offers a longer follow-up and a contemporary estimate of healthcare costs related to AF including all relevant care settings.
Our study offers a distinct advantage over previous work as costs, rather than being based on extrapolated rates using a prevalence-based approach [28], are estimated with an incidencebased method using patient-level morbidity records. Using an incidence based approach to costing and a broad perspective to capture the majority of costs associated with AF, several routinely collected administrative datasets from Scotland were combined, including care home utilisation.
Existing studies, including ours, regardless of econometric model choice and covariates used, show that costs due to inpatient admission are the main contributor to overall AF related healthcare cost. This is a pertinent finding that may well support future policies on opportunistic screening in the population at risk of AF, and in particular in Scotland where 1 in 3 patients with AF are currently undiagnosed [29]. The European AF management guidelines and the Scottish Cross-Party Group 'Heart Disease and Stroke', recently recommended that people who are 65 years or older and at risk of AF and associated comorbidities such as cardiovascular disease, diabetes or respiratory disease should be screened opportunistically in primary care, pharmacies or community settings [29,30]. With rigorous screening and appropriate treatment, hospitalisations could be avoided and costs reduced. Although we have captured most healthcare sectors and related costs, we were not able to obtain national data on primary care consultations, as these data are currently not routinely available for linkage in Scotland. Not capturing these data, may lead to an underestimation of the size of the AF cohort and associated costs.
However, the costs associated with primary care consultations is expected to have a limited impact on the overall total AF related costs. Such underestimation could also result from AF going undiagnosed and clinical miscoding of morbidity records. Nevertheless, by using a cohort of patients hospitalised with AF we were able to capture more severe cases of AF.
Prescribing and care home data were only available respectively from 2009 to 2012, their contribution to overall AF related costs might also be underestimated. Other limitations are inherent to the nature of administrative data, such as missing records or incomplete data.
Further, we acknowledge the issue concerning attributing AF related costs to patients with a structural heart disease, as AF may manifest subsequently because of this. In our analysis, we identified about 14% of AF patients with a structural heart disease; these were patients with systolic dysfunction, valvular heart disease or heart valve replacement. However, from the hospital data it was not possible to establish causation between structural heart disease and AF.
Recognising these limitations, we were nevertheless able to harness high quality patient-level linked data to identify a cohort of AF patients and to estimate AF related costs in Scotland. The inclusion of all available cost components is crucial for establishing overall costs, as these often extend beyond hospitalisation. The study identifies hospitalisation as the main cost driver and suggests that the implementation of AF screening policies could substantially reduce AF related health care costs. Most importantly, the study concludes that patient's age has a limited impact on the overall AF related cost, and therefore may contribute much less to future growth of AF related cost in an ever-ageing Scottish population .
Future work will be able to utilise Scottish Stroke Care Audit (SSCA) records, allowing for the identification of additional AF patients; these are patients hospitalised with a stroke, where AF has been recorded in audit data as an underlying comorbidity. Being able to complement inpatient records with SSCA records will allow us to capture more AF patients in Scotland.
There is a lack of generalisable studies based on large national population datasets that examine the total and the distribution of costs associated with AF [11]. The aim of this study was to quantify the inpatient, outpatient, prescribing and care home costs associated with AF over a five-year period. Using record-linkage of national datasets from Scotland, we also examined the distribution of costs that are attributable to AF.

Methods
Cost analyses or cost of illness studies typically adopt either the prevalence or incidence based approaches [12]. In the context of AF, the prevalence-based approach determines costs attributable to all cases of AF in a given year, while the incidence-based approach determines costs of new cases of AF in a given time period. In the present study, costs were estimated with an incidence-based approach. A further distinction between costing analyses is between the medicalized and the global comprehensive approaches. In the first case, only expenditures directly attributable to a particular disease are used for estimating the overall costs. While the medicalized approach can be used to identify highly specific expenditures, it may also lead to underestimation or overestimation of the economic burden of a given disease; this may happen when cost estimation is not adequately adjusted for confounders highly correlated with the disease of interest. Conversely, the global comprehensive approach, used in this analysis, includes all the expenditures incurred by a population with a particular disease [13]. These expenditures are not necessarily related to the disease of interest; for instance, expenditures related to orthopaedics surgery or cancer treatment incurred by a patient with AF, will count towards the global comprehensive cost of AF. Records 01 (SMR01). These records contain all general acute admissions, categorized as inpatients or day cases, discharged from non-obstetric and non-psychiatric specialties [14].

Data
Incident AF events (ICD10 code I48) were identified using all six diagnostic positions in SMR01, with a look back period of five years to minimise double counting. After checking for data entry errors and removal of duplicate records, the final AF cohort consisting of 278,286 individuals hospitalised with a diagnosis of AF or atrial flutter was identified. doctors and a small number of specialist appliance suppliers [16]. The quality of PIS data is guaranteed by an electronic data capture, and it passes several stages of quality control before and after data are submitted [17]. The care home census combines the former Residential Care Home Census (run by the Scottish Government) and the Private Nursing Homes Census (run by ISD Scotland). Items reported in the care home census include discharge dates to care home residency such as NHS and private nursing homes, as well as an indication on whether nursing care is required [16].
Patients were followed up for five years following incident AF event in terms of their healthcare resource use, care home admissions and mortality. Since AF is often a precursor of stroke and cardiovascular conditions, an estimation of costs for a period of five years post AF event would allow us to fully capture costs associated with an AF patient.

Costing
Inpatient care costs were obtained from the latest (2013/2014) Scottish National Tariff (SNT), a list of standard average prices based on Healthcare Resource Groups (HRGs) [17,18]. The SNT uses HRG4 for grouping clinically similar treatments that use similar levels of healthcare resources. After defining a total cost per episode, the total cost for a continuous inpatient stay (CIS) was calculated. A CIS describes the entire duration of an inpatient stay from the date of admission to the date of discharge and can consist of several episodes in different specialties. Since the SNT is based on spells of care (inpatient stay within the same specialty) rather than individual inpatient episodes or a CIS, a CIS was partitioned into spells when a change in specialty occurred [17].
If within a CIS, two or more episodes were in the same specialty, only the highest incurred cost was taken into account, and the remaining episodes were replaced with a zero cost. Outpatient costs were obtained by assigning outpatient specialty costs, to outpatient attendances [17]. Unit costs were specific to whether the outpatient attendance took place at a consultant or nurse led clinic [15].
The cost of each prescription dispensed per patient was obtained from PIS [19]. Firstly, the price per unit was obtained by dividing the item price by the pack size. Secondly, the total number of items dispensed was obtained by multiplying the number of items dispensed by the number of instalments. Care home costs, obtained from the care home census, were based on length of stay or residency. Care home residency was established from care home census records, reporting admission to a care home like structure [16]. An average of care home charges for long stay residents was calculated using information on whether nursing care was provided or not. The average weekly care home charge was expressed per day, so that only the effective days spent in a care home were costed. The tariffs used for costing account for inflation, therefore further cost adjustment was not needed.

Econometric model
Healthcare expenditure data are typically characterised by: i) a significant proportion of zerocost observations for individuals who have not utilised any healthcare resources in a given time period, and ii) a skewed distribution for positive costs. A two-part model was used [20,21]. In the first part of the model, the probability of using a healthcare service in a given time period In order to account for the skewed nature of cost data, generalised linear models (GLMs) were used. These were compared against ordinary least squares regression (OLS) and log transformed OLS by means of the Akaike Information Criterion (AIC), which measures goodness of fit. When comparing the different models, GLM reported the lowest AIC indicating the best fit for the given set of data. A user-written STATA programme "glmdiagnostic.do" [20], performing four different tests simultaneously, was used to identify the most appropriate distributional family and link function.

Econometric model covariates
The two-part model adjusted for age, sex, year of inpatient admission, socio-economic status, urban-rural classification, health board, comorbidities and mortality. These covariates are considered to be the main confounders that have an effect on costs incurred by an AF population. We controlled for age because AF and associated comorbidities are age-related conditions, and may have an impact on the overall costs. We also assumed costs to vary between males and females, in particular those for care home residency. Variation in healthcare utilisation and associated costs and care home residency by socio-economic status is controlled for using the Scottish Index of Multiple Deprivation (SIMD). The SIMD reflects areas of multiple deprivation ranked from the most to the least deprived and expressed as quintiles where the most and the least deprived areas are represented by 1 and 5 respectively [22]. In Scotland, there are 14 regional health boards responsible for the provision of healthcare [23]. Hence, potential differences in healthcare utilisation and prescribing costs, may reflect variation in clinical practice and prescribing behaviour rather than the ability of patients to access care. Patients living in urban areas may have easier access to care compared to patients living in more remote areas, which is controlled for including the 8-fold classification measuring rurality [24].
Patients with one or more comorbidities are expected to incur significantly higher costs than those with none. We accounted for this by including the Charlson comorbidity index, where 1 indicates the absence of comorbidities, 2 the presence of only a single comorbidity and 3 the presence of more than one comorbidity [25]. Two interaction terms between age and comorbidities, and mortality and SIMD were included in the econometric model. Intuitively, a relationship of direct proportionality between age and comorbidities suggests that the level of comorbidities increases, as patients get older. Similarly, the socio-economic status may significantly influence the rate of socio-economic inequalities in mortality [26].

Sensitivity analyses
In order to ascertain whether mortality had an impact on overall AF related healthcare costs, average annual cost per patient by age and for each health or care home sector, was estimated for patients who were alive and those who were dead at the end of the five-year follow-up period. The two econometric models (Equation IV and V, please see online TABLEent) followed the same structure of the model described in the previous section and used for the main analysis; however, those models were not adjusted for mortality.

Ethics statement
The authors state that no ethical approval was needed.

Patients and public involvement
There was no patients or public involvement

Cohort characteristics
Of the 278,286 AF patients with a mean age of 74 years (SD 12.

Econometric modelling results
Regression results for both modelling parts are presented in Table 2. Overall, an inversely U-  The use of health or social care services and associated costs also increased significantly for patients living in the most deprived areas, when compared with patients living in areas with the lowest level of deprivation. The effect of socio-economic status on healthcare utilisation was also measured for those who are alive at the end of the five-year follow-up period through an interaction term between SIMD and mortality, but no statistically significant effect was found.
Full details of regression results for interaction terms are presented in the supplementary online material (Table I, please see online supplement).
For patients with comorbidities, the probabilities of utilising healthcare services were greater than the probability for those with no comorbidities. Although healthcare utilisation increased with the number of comorbidities, the interaction term between age and comorbidities indicated that as patients get older the use of healthcare services on average is lower for patients with one or more comorbidities than those with none. The decrease in healthcare utilisation by age is more pronounced in patients with more comorbidities than in those with only one comorbidity. The difference in healthcare costs between comorbidity categories indicated that in the presence of one or more comorbidities, on average healthcare costs decrease as patients get older. Full details of regression results for patients who were alive and those who were dead at the end of the five-year follow-up period are presented in Table 3 and Table 4 respectively, while regression results for interaction terms are presented in the supplementary online material (Table II and Table III, please see online supplement).

Cost estimates
The estimated mean annual cost per AF patient was £3785 (95% CI: £3767-£3804). The estimated total costs and distribution of costs according to sex are shown in Table 5. While there is little difference between the total costs and the distribution of costs for inpatient, outpatient and prescription costs, the difference seems more pronounced when comparing the care home component of costs (5% of total costs among males vs 7% of total costs among females).
The average annual cost per AF patient by age and for each health or care home sector is shown in Figure 1. The estimated mean annual cost per AF patient alive at the end of the five-year follow-up period was £3047 (95% CI: £3027-£3067). The average annual cost per AF patient by age and for each health or care home sector is presented in the supplementary online material ( Figure   I, please see online supplement). For these patients, inpatient cost was the main driver across all age groups; a gradient between age and costs indicated increasing costs as the cohort ages.
Similar patterns were observed for care home costs. On the contrary, outpatient and prescribing costs remained constant up to 74 years, but decreased slightly for older patients. The estimated mean annual cost per AF patient who died during the five-year follow-up period, was £2304 (95% CI: £2284-£2324) ( Figure II, please see online supplement). For these patients, inpatient cost was the main driver across all age groups; a gradient between age and costs indicated decreasing costs as the cohort ages. This was also observed for outpatient and prescribing costs; but care home costs on average increased across age groups.

Discussion
A greater proportion of AF patients were found in areas with the highest index of deprivation.
This, combined with the likelihood for people living in the most deprived quintile having longer inpatient stays due to a lack of support at home, may explain the difference in inpatient care utilisation between patients from the most and the least deprived areas, with associated costs being higher for the former group. As AF is more likely to affect the elderly, so that costs were expected to increase with age. As health deteriorates with age, older age groups are assumed to make greater use of healthcare services, and therefore incur higher costs than younger age groups. However, age was found to have a modest impact on overall healthcare costs, being fairly consistent across age groups. This finding is in line with existing evidence indicating that healthcare expenditure depends not only on patients' calendar age, but is also significantly associated with remaining lifetime [27].
Any observed correlation between healthcare expenditure and age may therefore be attributable to the fact that the proportion of patients who are at the end of their lives is substantially greater in older rather than younger age groups [27]. On the other hand, comorbidity had a considerable effect on the overall costs, increasing significantly in patients with more than one comorbidity.
However, the decrease in healthcare utilisation by age is more pronounced in patients with more comorbidities than in those with only one comorbidity. Decreasing inpatient and outpatient costs for the oldest patients were offset by increasing care home costs, in particular for women. Indeed, the main cause for higher overall costs incurred by women is attributable to the higher likelihood for elderly women to reside in care homes.
Interestingly, care home contribution to the overall costs was noticeably lower for patients with multiple comorbidities than for those with none or one comorbidity. This may suggest that sicker patients are more likely to be in hospital than in a care home.
To date, only one single study published in 2004 has estimated the cost of AF in Scotland; the authors estimated the cost of AF in 1995/1996 with the medicalised approach , and projected these to the year 2000 [28]. Previous work has focussed on a 12-months follow-up, which seems limited in order to capture all healthcare resource utilisation for AF patients. Our study offers a longer follow-up and a contemporary estimate of healthcare costs related to AF including all relevant care settings. Our study offers a distinct advantage over previous work as costs, rather than being based on extrapolated rates using a prevalence-based approach [28], are estimated with an incidence-based method using patient-level morbidity records. Using an incidence based approach to costing and a broad perspective to capture the majority of costs associated with AF, several routinely collected administrative datasets from Scotland were combined, including care home utilisation.
Existing studies, including ours, regardless of econometric model choice and covariates used, show that costs due to inpatient admission are the main contributor to overall AF related healthcare cost. This is a pertinent finding that may well support future policies on opportunistic screening in the population at risk of AF, and in particular in Scotland where 1 in 3 patients with AF are currently undiagnosed [29]. Further, we acknowledge the issue concerning attributing AF related costs to patients with a structural heart disease, as AF may manifest subsequently because of this. In our analysis, we identified about 14% of AF patients with a structural heart disease; these were patients with systolic dysfunction, valvular heart disease or heart valve replacement. However, from the hospital data it was not possible to establish causation between structural heart disease and AF. Recognising these limitations, we were nevertheless able to harness high quality patient-level linked data to identify a cohort of AF patients and to estimate their associated healthcare utilisation and costs in Scotland.
The inclusion of all available cost components is crucial for establishing overall costs, as these often extend beyond hospitalisation. The study identifies hospitalisation as the main cost driver and suggests that the implementation of AF screening policies could substantially reduce AF related health care costs. Most importantly, the study concludes that patient's age has a limited impact on the overall AF related cost, and therefore may contribute much less to future growth of AF related cost in an ever-ageing Scottish population.

Article summary
Strengths and limitations of this study  Costs are estimated through an incidence-based approach using patient-level morbidity records.
 Sufficient follow-up time is used to capture all relevant global costs to generate a contemporary estimate of health and care home costs related to AF.
 Scotland offers a robust record linkage system, where administrative patient-level health data are routinely collected.
 Data on primary care consultations were not available for linkage at a national level, however the impact this might have on overall costs is expected to be small.
There is a lack of generalisable studies based on large national population datasets that examine the total and the distribution of costs associated with AF [11]. The aim of this study was to quantify the inpatient, outpatient, prescribing and care home costs associated with AF over a five-year period. Using record-linkage of national datasets from Scotland, we also examined the distribution of costs that are attributable to AF.

Methods
Cost analyses or cost of illness studies typically adopt either the prevalence or incidence based approaches [12]. In the context of AF, the prevalence-based approach determines costs attributable to all cases of AF in a given year, while the incidence-based approach determines costs of new cases of AF in a given time period. In the present study, costs were estimated with an incidence-based approach. A further distinction between costing analyses is between the medicalized and the global comprehensive approaches. In the first case, only expenditures directly attributable to a particular disease are used for estimating the overall costs. While the medicalized approach can be used to identify highly specific expenditures, it may also lead to underestimation or overestimation of the economic burden of a given disease; this may happen when cost estimation is not adequately adjusted for confounders highly correlated with the disease of interest. Conversely, the global comprehensive approach, used in this analysis, includes all the expenditures incurred by a population with a particular disease [13]. These expenditures are not necessarily related to the disease of interest; for instance, expenditures related to orthopaedics surgery or cancer treatment incurred by a patient with AF, will count towards the global comprehensive cost of AF. Records 01 (SMR01). These records contain all general acute admissions, categorized as inpatients or day cases, discharged from non-obstetric and non-psychiatric specialties [14].

Data
Incident AF events (ICD10 code I48) were identified using all six diagnostic positions in SMR01, with a look back period of five years to minimise double counting. After checking for data entry errors and removal of duplicate records, the final AF cohort consisting of 278,286 individuals hospitalised with a diagnosis of AF or atrial flutter was identified.  [15]. The PIS database includes prescribing records for all medicines and their associated costs, which are prescribed and dispensed by community pharmacies, dispensing doctors and a small number of specialist appliance suppliers [16]. The quality of PIS data is guaranteed by an electronic data capture, and it passes several stages of quality control before and after data are submitted [17]. The care home census combines the former Residential Care Home Census (run by the Scottish Government) and the Private Nursing Homes Census (run by ISD Scotland). Items reported in the care home census include discharge dates to care home residency such as NHS and private nursing homes, as well as an indication on whether nursing care is required [16].
Patients were followed up for five years following incident AF event in terms of their healthcare resource use, care home admissions and mortality. Since AF is often a precursor of stroke and cardiovascular conditions, an estimation of costs for a period of five years post AF event would allow us to fully capture costs associated with an AF patient.

Costing
Inpatient care costs were obtained from the latest (2013/2014) Scottish National Tariff (SNT), a list of standard average prices based on Healthcare Resource Groups (HRGs) [17,18]. The SNT uses HRG4 for grouping clinically similar treatments that use similar levels of healthcare resources. After defining a total cost per episode, the total cost for a continuous inpatient stay (CIS) was calculated. A CIS describes the entire duration of an inpatient stay from the date of admission to the date of discharge and can consist of several episodes in different specialties. Since the SNT is based on spells of care (inpatient stay within the same specialty) rather than individual inpatient episodes or a CIS, a CIS was partitioned into spells when a change in specialty occurred [17].
If within a CIS, two or more episodes were in the same specialty, only the highest incurred cost was taken into account, and the remaining episodes were replaced with a zero cost. Outpatient costs were obtained by assigning outpatient specialty costs, to outpatient attendances [17]. Unit costs were specific to whether the outpatient attendance took place at a consultant or nurse led clinic [15].
The cost of each prescription dispensed per patient was obtained from PIS [19]. Firstly, the price per unit was obtained by dividing the item price by the pack size. Secondly, the total number of items dispensed was obtained by multiplying the number of items dispensed by the number of instalments. Care home costs, obtained from the care home census, were based on length of stay or residency. Care home residency was established from care home census records, reporting admission to a care home like structure [16]. An average of care home charges for long stay residents was calculated using information on whether nursing care was provided or not. The average weekly care home charge was expressed per day, so that only the effective days spent in a care home were costed. The tariffs used for costing account for inflation, therefore further cost adjustment was not needed.

Econometric model
Healthcare expenditure data are typically characterised by: i) a significant proportion of zerocost observations for individuals who have not utilised any healthcare resources in a given time period, and ii) a skewed distribution for positive costs. A two-part model was used [20,21]. In the first part of the model, the probability of using a healthcare service in a given time period was estimated using a probit model (Equation I,  indicating the best fit for the given set of data. A user-written STATA programme "glmdiagnostic.do" [20], performing four different tests simultaneously, was used to identify the most appropriate distributional family and link function.

Econometric model covariates
The two-part model adjusted for age, sex, year of inpatient admission, socio-economic status, urban-rural classification, health board, comorbidities and mortality. These covariates are considered to be the main confounders that have an effect on costs incurred by an AF population. We controlled for age because AF and associated comorbidities are age-related conditions, and may have an impact on the overall costs. We also assumed costs to vary between males and females, in particular those for care home residency. Variation in healthcare utilisation and associated costs and care home residency by socio-economic status is controlled for using the Scottish Index of Multiple Deprivation (SIMD). The SIMD reflects areas of multiple deprivation ranked from the most to the least deprived and expressed as quintiles where the most and the least deprived areas are represented by 1 and 5 respectively [22]. In Scotland, there are 14 regional health boards responsible for the provision of healthcare [23]. Hence, potential differences in healthcare utilisation and prescribing costs, may reflect variation in clinical practice and prescribing behaviour rather than the ability of patients to access care. Patients living in urban areas may have easier access to care compared to patients living in more remote areas, which is controlled for including the 8-fold classification measuring rurality [24].
Patients with one or more comorbidities are expected to incur significantly higher costs than those with none. We accounted for this by including the Charlson comorbidity index, where 1 indicates the absence of comorbidities, 2 the presence of only a single comorbidity and 3 the presence of more than one comorbidity [25]. Two interaction terms between age and comorbidities, and mortality and SIMD were included in the econometric model. Intuitively, a relationship of direct proportionality between age and comorbidities suggests that the level of comorbidities increases, as patients get older. Similarly, the socio-economic status may significantly influence the rate of socio-economic inequalities in mortality [26].

Sensitivity analyses
In order to ascertain whether mortality had an impact on overall AF related healthcare costs, average annual cost per patient by age and for each health or care home sector, was estimated for patients who were alive and those who were dead at the end of the five-year follow-up period. The two econometric models (Equation IV and V, please see online TABLEent) followed the same structure of the model described in the previous section and used for the main analysis; however, those models were not adjusted for mortality.

Ethics statement
No ethics approval was sought as this study does not require consenting/contacting patients directly. All data used by researchers is pseudonymised, data reported is aggregated to minimise risk of identification and output clearance is required.

Patients and public involvement
There was no patients or public involvement

Cohort characteristics
Of

Econometric modelling results
Regression results for both modelling parts are presented in Table 2. Overall, an inversely U-  The use of health or social care services and associated costs also increased significantly for patients living in the most deprived areas, when compared with patients living in areas with the lowest level of deprivation. The effect of socio-economic status on healthcare utilisation was also measured for those who are alive at the end of the five-year follow-up period through an interaction term between SIMD and mortality, but no statistically significant effect was found.
Full details of regression results for interaction terms are presented in the supplementary online material (Table I, please see online supplement).
For patients with comorbidities, the probabilities of utilising healthcare services were greater than the probability for those with no comorbidities. Although healthcare utilisation increased with the number of comorbidities, the interaction term between age and comorbidities indicated that as patients get older the use of healthcare services on average is lower for patients with one or more comorbidities than those with none. The decrease in healthcare utilisation by age is more pronounced in patients with more comorbidities than in those with only one comorbidity. The difference in healthcare costs between comorbidity categories indicated that in the presence of one or more comorbidities, on average healthcare costs decrease as patients get older. Full details of regression results for patients who were alive and those who were dead at the end of the five-year follow-up period are presented in Table 3 and Table 4 respectively, while regression results for interaction terms are presented in the supplementary online material (Table II and Table III, please see online supplement). The estimated mean annual cost per AF patient was £3785 (95% CI: £3767-£3804). The estimated total costs and distribution of costs according to sex are shown in Table 5.

Cost estimates
While there is little difference between the total costs and the distribution of costs for inpatient, outpatient and prescription costs, the difference seems more pronounced when comparing the care home component of costs (5% of total costs among males vs 7% of total costs among females).
The average annual cost per AF patient by age and for each health or care home sector is shown in Figure 1. The estimated mean annual cost per AF patient alive at the end of the five-year follow-up period was £3047 (95% CI: £3027-£3067). The average annual cost per AF patient by age and for each health or care home sector is presented in the supplementary online material ( Figure   I, please see online supplement). For these patients, inpatient cost was the main driver across all age groups; a gradient between age and costs indicated increasing costs as the cohort ages. The estimated mean annual cost per AF patient who died during the five-year follow-up period, was £2304 (95% CI: £2284-£2324) ( Figure II, please see online supplement). For these patients, inpatient cost was the main driver across all age groups; a gradient between age and costs indicated decreasing costs as the cohort ages. This was also observed for outpatient and prescribing costs; but care home costs on average increased across age groups.

Discussion
A greater proportion of AF patients were found in areas with the highest index of deprivation.
This, combined with the likelihood for people living in the most deprived quintile having longer inpatient stays due to a lack of support at home, may explain the difference in inpatient care utilisation between patients from the most and the least deprived areas, with associated costs being higher for the former group. As AF is more likely to affect the elderly, so that costs were expected to increase with age. As health deteriorates with age, older age groups are assumed to make greater use of healthcare services, and therefore incur higher costs than younger age groups. However, age was found to have a modest impact on overall healthcare costs, being fairly consistent across age groups. This finding is in line with existing evidence indicating that healthcare expenditure depends not only on patients' calendar age, but is also significantly associated with remaining lifetime [27].
Any observed correlation between healthcare expenditure and age may therefore be attributable to the fact that the proportion of patients who are at the end of their lives is substantially greater in older rather than younger age groups [27]. On the other hand, comorbidity had a considerable  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y   13 effect on the overall costs, increasing significantly in patients with more than one comorbidity.
However, the decrease in healthcare utilisation by age is more pronounced in patients with more comorbidities than in those with only one comorbidity.
Decreasing inpatient and outpatient costs for the oldest patients were offset by increasing care home costs, in particular for women. Indeed, the main cause for higher overall costs incurred by women is attributable to the higher likelihood for elderly women to reside in care homes.
Interestingly, care home contribution to the overall costs was noticeably lower for patients with multiple comorbidities than for those with none or one comorbidity. This may suggest that sicker patients are more likely to be in hospital than in a care home.
To date, only one single study published in 2004 has estimated the cost of AF in Scotland; the authors estimated the cost of AF in 1995/1996 with the medicalised approach , and projected these to the year 2000 [28]. Previous work has focussed on a 12-months follow-up, which seems limited in order to capture all healthcare resource utilisation for AF patients. Our study offers a longer follow-up and a contemporary estimate of healthcare costs related to AF including all relevant care settings. Our study offers a distinct advantage over previous work as costs, rather than being based on extrapolated rates using a prevalence-based approach [28], are estimated with an incidence-based method using patient-level morbidity records. Using an incidence based approach to costing and a broad perspective to capture the majority of costs associated with AF, several routinely collected administrative datasets from Scotland were combined, including care home utilisation.
In addition, this is likely to have a marginal impact on our conclusions, as the global comprehensive approach used in this study include expenditures that are not necessarily related to AF.
We also acknowledge that specifying whether patients had received cardiovascular procedures (e.g. cardioversion, echocardiograms and angiograms) would improve the accuracy of our cost estimation, as it would indicate whether costs should be attributable to AF or other forms of structural heart disease. However, this information is not currently available in our routinely collected data of hospital admissions.
Recognising these limitations, we were nevertheless able to harness high quality patient-level linked data to identify a cohort of AF patients and to estimate their associated healthcare utilisation and costs in Scotland.

Introduction
Background/rationale 2 Explain the scientific background and rationale for the investigation being reported 3 Objectives 3 State specific objectives, including any prespecified hypotheses 3

Study design 4
Present key elements of study design early in the paper 3-4 Setting 5 Describe the setting, locations, and relevant dates, including periods of recruitment, exposure, follow-up, and data (b) For matched studies, give matching criteria and number of exposed and unexposed n/a  Discuss the generalisability (external validity) of the study results 14