Regional variation in non-traumatic major lower limb amputation in England: observational study of linked primary and secondary care data

Abstract

Background

Peripheral artery disease and diabetes are the main primary risk factors for non-traumatic major lower limb amputation. Regional variation in incidence of major lower limb amputation has yet to be fully described in terms of these risk factors and explained. The aim of this study was to estimate yearly incidence of major lower limb amputation over a 10-year interval (2010–2019) across England, by related condition and by region and, additionally, to investigate reasons for regional variation.

Methods

This observational study utilized primary care (Clinical Practice Research Datalink Aurum), secondary care (Hospital Episode Statistics), death and demographic data in England. Adults registered with a practice using Clinical Practice Research Datalink Aurum and with Hospital Episode Statistics linkage were included. Patients with a record of major lower limb amputation during the interval 1 January 2010 to 31 December 2019 were identified and yearly incidence rates of major lower limb amputation were calculated. Co-morbidities analysed were cardiovascular disease (including coronary artery disease, peripheral artery disease and cerebrovascular disease), diabetes (of any type) and cancer. Demographic and socioeconomic covariates analysed were age, sex, ethnicity, deprivation level, region and urban/rural categorization.

Results

The study included 18 397 483 individuals, 8584 of which had a record of major lower limb amputation. The age-standardized yearly incidence rate of major lower limb amputation in England decreased by 30% from 11.2 per 100 000 person-years in 2010 to 7.8 in 2019. The incidence rate in those with diabetes fell by 30% over the 10-year interval, rose by 20% for those with both diabetes and cardiovascular disease, and changed little in those with cardiovascular disease. In 2019, the age-standardized incidence rate was highest in the North East (14.8 per 100 000 person-years) and lowest in the East of England (4.5 per 100 000 person-years). Between 2010 and 1019, incidence rates decreased across all regions, the largest decrease of 56% in the East Midlands and the smallest of 8% in the North East. Statistically significant regional variation remained after full adjustment for demographic, socioeconomic data and related conditions.

Conclusion

Whilst the incidence of major lower limb amputation is decreasing overall, significant regional variation in major lower limb amputation exists and is unexplained by demographic, socioeconomic and health data. Regional differences in service provision and accessibility should be investigated to provide further explanation.

This study estimated the yearly incidence of major lower limb amputation in England over 10 years, overall, by conditions and by regions, while exploring reasons for regional variations. Overall incidence fell by 30%, with similar declines for those with diabetes, but a 20% increase was observed for individuals with both diabetes and cardiovascular disease. Significant regional variation was found, being highest in the North East and lowest in the East of England, remaining after adjustment for demographic, socioeconomic data and related conditions, thus indicating the need to investigate differences in service provision and accessibility.

Introduction

Peripheral artery disease (PAD) and diabetes are the predominant causes of non-traumatic major lower limb amputation (MLLA)¹. While improvements in the management of chronic limb-threatening ischaemia (CLTI; the end-stage of PAD) have improved outcomes for those presenting to vascular surgery with potentially salvageable legs, reducing PAD and diabetes-related major amputations requires whole-system improvements in healthcare and close collaboration of primary and secondary care². The epidemiology of MLLA is therefore an important public health measure and regional differences in incidence of amputation need to be understood to achieve ongoing improvements and reduce inequalities of healthcare. However, incidence trends of MLLA in England are widely debated^3–11. Studies agree that there is regional variation in MLLA incidence, with incidence higher in the north^3,6,9–11. The reasons for this variation are unexplained by deprivation, age or sex differences. These studies suggest that regional differences in both ethnicity and prevalence of related conditions may be driving regional variation¹¹. A systematic review identified that studies either reported for those with PAD or diabetes; no studies reported for or compared MLLA incidence between both¹¹.

PAD is a manifestation of systemic cardiovascular disease, and coprevalence of PAD with other cardiovascular diseases, such as coronary artery disease and cerebrovascular disease, is common^12,13. Accuracy in the recording of PAD diagnoses within general practice (GP) data is debated and PAD is also widely underdiagnosed^14,15. This renders the process of accurately identifying those with PAD in electronic health record studies difficult. Therefore, in order to capture as many individuals at risk of MLLA due to PAD in the population as possible, cardiovascular disease (CVD) can be used as a proxy for PAD.

Studies to date have used the secondary care database Hospital Episode Statistics (HES) to ascertain MLLA data and separate aggregate data sources, such as national health survey and census data, as population data^3,9,16,17. As it is event based, HES does not contain general population data. Using aggregate population data, which differs in source from the case data, results in differing inclusion criteria between the case and population data, introducing bias into incidence calculations and leading to inaccurate results. Furthermore, these aggregate data are only stratified by a minimal number of subgroups, often lacking health-related data. A more accurate way to assess if the regional morbidity rate differences affect MLLA incidence is to ascertain patient-level case and population morbidity rate information from primary care health records.

This study aimed to describe time trends in non-traumatic MLLA across England between 2010 and 2019, by morbidity rate (CVD and/or diabetes) and by region using linked secondary and primary care data, and also to investigate reasons for regional variation.

Methods

A population-based observational study was performed using the population contained within the primary care database of Clinical Practice Research Datalink (CPRD) Aurum (CPRD registration number 21_000364, approved on 17 March 2021).

Inclusion/exclusion criteria

Adults were included if they were registered within a CPRD Aurum practice that was considered by CPRD to be ‘registered acceptable’ with ‘up-to-standard follow-up’ and had available HES linkage.

MLLA was defined as a lower limb amputation above the ankle. Individuals were defined as having an incident MLLA if they had an Office of Population Censuses and Surveys Classification of Surgical Operations and Procedures (4th revision) (OPCS-4) code for MLLA recorded in the HES linkage within the study interval of 1 January 2010 to 31 December 2019 inclusive. The recording of MLLA in primary care data may be inaccurate¹⁸. MLLA were excluded where a code for traumatic injury was present within the same hospital episode. Should an individual be recorded as having more than one MLLA, either on the same or a different limb, then only the amputation at the highest level was included in analysis to reflect greatest severity. Where bilateral amputations at the same level were recorded, only the first recorded was included to avoid double counting.

The denominator data set was created to match the criteria for that of the case data set minus criteria related to MLLA.

CPRD limitations on health data access meant that, whilst it was possible to use CPRD demographic data for the whole-study population, access to health data was only granted for individuals with a case of MLLA and for a large random sample of the CPRD population. Where analysis involved health data, the sample data were used and upweighted to represent the population. Where health data were not used in the analysis, the above defined population data were used.

Covariates

Demographic and socioeconomic covariates analysed were age, sex, ethnicity, deprivation level, region and urban/rural categorization, chosen based on availability of data and potential to explain regional variation suggested by a systematic review¹¹. Age was grouped as: below 50, 50–59, 60–69, 70–79 and 80 years and over. Ethnicity categories were defined by the Office of National Statistics (ONS) including a category for missing ethnicity¹⁹. Ethnicity data was taken from HES records then CPRD where missing in HES²⁰. Patient-level Index of Multiple Deprivation 2015 quintiles were used to categorize deprivation levels²¹. Region was defined as patient-level Strategic Health Authority (SHA) regions. Urban/rural 2011 classification data were obtained at practice level due to restrictions by CPRD. A ‘missing’ category was created for all categorical variables where missingness was found; individuals were excluded who had missing region data, however, missingness of region was negligible.

As accuracy in the recording of PAD diagnoses within GP data is debated and PAD underdiagnosed; accurately identifying those with PAD in electronic health record studies is difficult^14,15. CVD was used as a proxy in order to capture as many patients at risk as accurately as possible. CVD included coronary artery disease, PAD and cerebrovascular disease codes. Subsequently, co-morbidities analysed were CVD, diabetes (of any type) and cancer. Individuals were included as having these co-morbidities on the first instance of a relevant code. Limb amputation due to cancer is rare, consequently, the inclusion of cancer-related amputations is unlikely to impact results. However, cancer and CVD share many risk factors and a history of cancer can increase an individual’s risk of CVD²². Amputations of those with a history of cancer were therefore not excluded; cancer was included in the analysis as a variable that may increase risk/explain regional variation²³.

Condition diagnoses are defined using the medical code lists that are available online at: https://github.com/AnnaMeffen/MLEA_incidence_thesis_supplements.

Analysis methods

Covariates were summarized by data set, case, sample population and whole CPRD population (where data were available) and presented as mean(s.d.) for continuous variables and count (%) for categorical variables.

Crude and age-standardized incidence rates were calculated yearly for the whole CPRD population and by region. As diagnoses data were not available for the whole CPRD population, a sample and upweighting method was used to calculate the yearly incidence rate stratified by CVD and diabetes (Supplementary methods).

Person-years at risk were defined as the difference between the entry and exit date divided by 365.25. Entry date was defined as the latest of 1 January of the year the person turned 18 years old, study start date and registration date. Exit date was defined as the earliest of study end date, date of death and registration end date. Study start and end dates were defined calendar yearly for the whole 10-year study interval.

Death date was taken from ONS and then from CPRD where ONS death data was not recorded. CPRD provides death data, however, ONS is believed to be more accurate²⁴.

The incidence rate was defined as the total number of new events out of the total person-years at risk and presented per 100 000 person-years. Age-standardized incidence rates were calculated using the age group cut-points defined above with rates standardized to the age distribution of the whole CPRD denominator data set in 2010. Upweighted and age-standardized incidence rates were stratified by morbidity rate: those with diabetes, those with CVD, and individuals with both diabetes and CVD (these groups are not mutually exclusive).

Factors influencing regional variation were investigated by applying negative binomial models to the upweighted data sets for the whole 10-year study interval then varying the included covariates. Initially, an unadjusted model containing region as the only covariate, based on the crude regional incidence estimates, was applied to provide comparison and assess crude levels of variation. As upweighting was performed using region, sex and age group for each study year, these covariates were added to form the adjusted base model. The remaining covariates were added individually to the adjusted base model and then together as a fully adjusted model. The incidence rate ratios (IRRs) of these models were compared by region, with London (the most populous region) as the base comparator.

Analyses were performed using Stata v. 18²⁵, R 4.3.1²⁶ and RStudio 2023.03.0²⁷.

Results

Case data set demographic

There were 8584 patients who underwent at least one MLLA within the time interval identified (Table 1). The mean age at event was 70 years with over two-thirds being male (67%), the majority were of white ethnicity (89%) and most lived in an urban setting (87%). CVD was more prevalent than diabetes (58% versus 53%) and about a fifth of patients had a diagnosis of cancer. Deprivation quintiles were not evenly distributed, with 90% more MLLA in the most deprived quintile than in the least deprived quintile. The highest number of MLLA was seen in the North West of England with the lowest in the East Midlands.

Table 1.

Patient demographics by data set

Demographic	MLLA patients	Sample population	CPRD population
Age (years), mean(s.d.)	70(20)	46(29)	46(29)
Age at death (years), mean(s.d.)	76(17)	82(17)	82(17)
Sex
Male	5768 (67.19)	441 489 (48.47)	8 915 486 (48.46)
Female	2816 (32.81)	469 373 (51.53)	9 481 652 (51.54)
Undetermined	0 (0.00)	24 (<0.01)	345 (<0.01)
Ethnicity
White	7679 (89.46)	571 053 (62.69)	–
Black	236 (2.75)	35 559 (3.90)	–
Asian	153 (1.78)	69 361 (7.61)	–
Other	55 (0.64)	21 101 (2.32)	–
Mixed	38 (0.44)	18 622 (2.04)	–
Missing	423 (4.93)	195 190 (21.43)	–
Region
North West	1925 (22.43)	146 008 (16.03)	2 946 618 (16.02)
West Midlands	1470 (17.12)	135 221 (14.84)	2 737 123 (14.88)
South West	1193 (13.90)	113 263 (12.43)	2 277 777 (12.38)
London	954 (11.11)	199 671 (21.92)	4 040 209 (21.96)
South central	932 (10.86)	112 604 (12.36)	2 258 663 (12.28)
South East coast	710 (8.27)	74 243 (8.15)	1 504 187 (8.18)
North East	461 (5.37)	28 326 (3.11)	576 047 (3.13)
Yorkshire	379 (4.42)	35 522 (3.90)	719 842 (3.91)
East of England	352 (4.10)	38 381 (4.21)	777 045 (4.22)
East Midlands	208 (2.40)	27 647 (3.04)	559 972 (3.04)
Deprivation level
1 (least deprived)	1227 (14.29)	175 281 (19.24)	–
2	1451 (16.90)	179 641 (19.72)	–
3	1658 (19.32)	175 420 (19.26)	–
4	1896 (22.09)	199 013 (21.85)	–
5 (most deprived)	2340 (27.26)	179 259 (19.68)	–
Missing	12 (0.14)	2272 (0.25)	–
Urban/rural classification
Urban	7476 (87.09)	810 251 (88.95)	–
Rural	1108 (12.91)	100 635 (11.05)	–
Diabetes type 1	815 (9.49)	5579 (0.61)	–
Diabetes type 2	3665 (42.70)	57 295 (6.29)	–
Diabetes—other	31 (0.36)	4889 (0.54)	–
CVD	4940 (57.55)	66 520 (7.30)	–
Cancer	1725 (20.10)	74 199 (8.15)	–
Total	8584	910 886	18 397 483

Values are n (%) unless otherwise specified. Age is at event for MLLA patients and at mid-study for denominator populations. MLLA, major lower limb amputation; CPRD, Clinical Practice Research Datalink; CVD, cardiovascular disease.

Denominator data sets demographic

In terms of age, sex and region, the sample data set was comparable to that of the whole CPRD data set. Whilst it is not possible to compare other demographics between the sample and whole CPRD populations the proportions of those with diabetes and CVD are similar to those cited in the UK population^28,29.

Incidence rate of MLLA in England

The crude yearly incidence rate of MLLA in England decreased by 28% over the 10-year study interval whilst the age-standardized yearly incidence rate decreased by 30%. The crude incidence rate fell from 11.3 per 100 000 person-years in 2010 (95% c.i. 10.6 to 12.0) to 8.1 per 100 000 person-years in 2019 (95% c.i. 7.5 to 8.6) and the age-standardized yearly incidence rate fell from 11.2 per 100 000 person-years in 2010 (95% c.i. 10.5 to 11.9) to 7.8 per 100 000 person-years in 2019 (95% c.i. 7.2 to 8.3). The effect of age-standardization was a slightly lower incidence rate overall. The precision of incidence rate estimates remained similar (Fig. 1).

Fig. 1.

Crude and age-standardized yearly incidence rates of MLLA in England per 100 000 person-years 2010–2019

Age-standardized to the age distribution of the CPRD study population in mid-2010 categorized as aged less than 50, 50–59, 60–69, 70–79 and 80 years and over. CPRD, Clinical Practice Research Datalink.

Incidence rate of MLLA by morbidity rate

The age-standardized incidence rate of MLLA in those with diabetes fell by 30% over the 10-year interval (55.5 to 39 per 100 000 person-years) despite an increase in person-years at risk within the study population over the study interval (Fig. 2). The incidence rate rose slightly by 3% (71.9 to 74.3 per 100 000 person-years) in those with CVD despite a slight rise in person-years at risk within the study interval. There was a moderate overall increase in the incidence rate of those with both diabetes and CVD over the 10-year study interval of 20% (195.8 to 235.6 per 100 000 person-years), however, there was less stability in incidence rates over time for this group with clear fluctuations.

Fig. 2.

a Yearly incidence rate of MLLA and b number of MLLA and number of person-years at risk by morbidity rate

Uses sample data upweighted to the CPRD study population in mid-2010 in terms of age (categorized as aged less than 50, 50–59, 60–69, 70–79 and 80 years and over), sex and region. Age-standardized to the age distribution of the CPRD study population in mid-2010 categorized as aged less than 50, 50–59, 60–69, 70–79 and 80 years and over. CVD, cardiovascular disease; MLLA, major lower limb amputation; py, person years; CPRD, Clinical Practice Research Datalink.

Incidence rates in those with both CVD and diabetes were highest and were up to 30 times higher than that of the whole population (2019), up to six times higher than those with diabetes and up to three times higher than those with CVD (2019).

Regional incidence of MLLA

At the start of the study interval, crude and age-standardized incidence rates were highest in the North East (crude: 16.9 per 100 000 person-years, age-standardized: 16.3 per 100 000 person-years with 95% c.i. (11.9 to 20.7)); crude incidence was lowest in London (7.3 per 100 000 person-years) and age-standardized incidence was lowest in the East of England (8.3 per 100 000 person years 2010 with 95% c.i. (5.7 to 11.0)). All regions experienced a decrease in age-standardized incidence rates from 2010 to 2019. The largest decrease of 57% was seen in the East Midlands and the smallest of 8% in the North East (Fig. 3). Despite the overall decrease in incidence over the study interval, there was instability in incidence rates with peaks in incidence rates above the initial 2010 rate in some areas.

Fig. 3.

a Crude versus b age-standardized incidence rate by region and year

Age-standardized to the age distribution of the CPRD study population in mid-2010 categorized as aged less than 50, 50–59, 60–69, 70–79 and 80 years and over. Regions are grouped into north, Midlands and south by colour. CPRD, Clinical Practice Research Datalink.

The effect of age-standardization was a reduction in the magnitude of regional variation (Fig. 3), suggesting that differences in age between regions plays a small part in regional variation.

Looking at regions grouped into north, Midlands and south (Fig. 3), the incidence of MLLA was higher in northern regions, with little difference between the Midlands and southern regions.

An interactive R shiny app has been created to visualize crude and age-standardized incidence rates of MLLA in map format and is available at: https://annameffen.shinyapps.io/MLLA_by_SHA/.

Reasons for regional variation

Over the 10-year interval, crude incidence was significantly higher than London in all regions and highest in the North East (IRR 2.66, 95% c.i. 2.37 to 2.98) (Fig. 4). Age and sex (model B) explained some of the variation in MLLA incidence compared with London, particularly in the East Midlands and the East of England where the incidence rate was not significantly different from London after age and sex adjustment. Variation in MLLA incidence rates between London and other regions was most often explained by differences in ethnicity compared with London (model C). After adjustment for age, sex and ethnicity, incidence rates were not significantly different from London in Yorkshire, south central and the South East; incidence rates were significantly lower than London in the East Midlands (IRR 0.81, 95% c.i. 0.68 to 0.97) and the East of England (IRR 0.79, 95% c.i. 0.68 to 0.93). Differences in proportions of deprivation (model E) and proportions of those with diabetes (model F) or cancer (model I) played little part in variation according to these models. Regional variation was not explained by any of the models in the North East, North West and South West, where incidence rates were highest. When fitting the fully adjusted model (model I), MLLA incidence remained significantly different from London in all regions.

Fig. 4.

a Incidence rate ratios and confidence interval of each model applied by region compared with London over the whole 10-year study interval b Model definitions

Model A uses the full study data and is based on crude incidence rates. All remaining models (models B through to I) use sample data upweighted to the CPRD study population in each mid-year in terms of age (categorized as aged less than 50, 50–59, 60–69, 70–79 and 80 years and over), sex and region. Model B, including age, sex and region, acts as a base model given the data was upweighted using these variables. CPRD, Clinical Practice Research Datalink; CVD, cardiovascular disease.

Discussion

While the overall incidence of MLLA in England is declining, there are significant regional variations that are partially influenced by age and ethnicity. However, differences in deprivation, diabetes or cancer prevalence did not account for this. This suggests that variation in design and delivery of healthcare services may be contributing to the differing MLLA incidence rates across regions.

Comparison to previous studies is challenging due to methodological differences¹¹. While many studies focused on diabetes-related amputations^{7,8,10,11,30–38}, only two examined CVD-related amputations (PAD specific) each with varying data sets and criteria^5,9,11. None compared amputation incidence between CVD and diabetes or reported rates for the general population in a similar interval. One diabetes-focused study reported an incidence rate but their narrower criteria (type 2 diabetes only) and the lack of standardization may explain their higher rate of 96 per 100 000 person-years.

Despite methodological differences, similar conclusions were reached in studies by Holman¹⁰, Ahmad³, Moxey⁶ and Maheswaran⁹, showing higher amputation incidence in the north and lower rates in regions that are more ethnically diverse, as seen here in London and the East Midlands, though not in the West Midlands. While previous studies suggested that regional variation might reflect varying prevalence of related conditions, this study found no evidence to support that. Additionally, these studies suggest that regional variation may stem from regional differences in healthcare accessibility and service provision.

It is encouraging that MLLA incidence rates steadily declined for those with diabetes despite diabetes prevalence increasing over the same interval³⁹. This has coincided with national efforts to improve and standardize diabetic footcare. The England and Wales National Diabetic Foot Audit and Quality Improvement Collaborative has demonstrated regional variation in terms of service delivery and outcomes⁴⁰. They recommend increasing accessibility of, and time to assessment by specialist, multidisciplinary foot services⁴⁰. However, it is notable that, in this study, the majority of MLLAs were performed in those with CVD and no such decrease in MLLA was observed in this cohort. PAD is often recognized late and potential missed opportunities in timely specialist referrals have been identified, with delays to specialist vascular assessment likely increasing the risk of MLLA^41,42. Rapid-access specialist limb-salvage clinics have demonstrated reduction in MLLA incidence^43–45. UK guidelines now specify time targets from referral to revascularization of those with CLTI and additional National Health Service (NHS) funding to achieve them, which may reduce some regional variation⁴⁶. However, much of the regional variation may exist upstream of specialist vascular clinics^47,48. It is notable that few PAD guidelines (including guidelines on referral) included primary care clinicians and it is vital future guidelines include all key stakeholders in their author group⁴⁹.

This was both the first study to compare incidence of MLLA in England by morbidity rate over time, and the first to assess time trends and reasons for regional variation using primary care data. More specifically, this study is the first to explore regional differences in morbidity rate as a reason for regional variation. Importantly, the numerator and denominator data were derived from the same population data source, minimizing potential bias within the results. This study includes detailed descriptions of methodology, precisely defined statistical terminology and adhered to recommended reporting guidelines (Table S1).

The primary limitation of this study is that CPRD does not cover the entire population of England, meaning this study is more akin to a cohort study. The study population size is further restricted by CPRD data size access limitations leading to the necessity of the upweighting process. Population coverage limitations are balanced by the accuracy gained from capturing both numerator and denominator data from the same population, as discussed earlier. Research has shown CPRD data to be representative in terms of age, sex and ethnicity for the whole population, however, this is not confirmed regionally^50,51. Clinical representativeness in terms of morbidity rate prevalence is not described in available literature. CPRD data is also likely to be biased towards a less healthy population as healthy people may not register at a GP surgery and be represented. Additionally, these data may contain duplicate records where individuals have changed GP practices, thus contributing to the CPRD database under multiple patient identifiers⁵².

On stratification, there was little MLLA patient data for non-white ethnicities, which could be reflective of the population and the nature of the causes of MLLA (which is most likely), but there remains a small risk that it could be reflective of bias due to unreliable data collection methods; it is not possible to confirm with the available data^53,54. As any missing data in MLLA patients with non-white ethnicity could potentially make a small, but likely insignificant, difference to results, caution is advised to not overstate the role of ethnicity. Further limitations may lie in the accuracy of diagnostic coding. Whilst care has been taken, there are risks of inaccuracies in both the code lists and in recording of all collected data, though the risks are likely to be small with little impact on results.

Significant regional variation in MLLA incidence in England exists after adjustment for patient demographic, socioeconomic factors and related conditions. Routinely collected electronic health records are unlikely to provide further insight into the remaining variation; regional differences in service provision and accessibility should be investigated to provide further explanation and determine where healthcare system redesign may reduce regional disparity in MLLA incidence.

Funding

A.M.’s PhD studentship that led to this work was funded by the George Davies Charitable Trust donation, the funder had no role in the this study.

Disclosure

All authors have completed the ICMJE uniform disclosure form at http://www.icmje.org/disclosure-of-interest/ and declare: M.J.R. and L.J.G. are supported by National Institute for Health and Care Research (NIHR) Applied Research Collaboration East Midlands (ARC EM), A.M., M.J.R., L.J.G. and N.B. are supported by Leicester NIHR Biomedical Research Centre (BRC). J.S.M.H. is supported by an NIHR Academic Clinical Lectureship. J.S.M.H. and A.M. have received support from the George Davies Charitable Trust donation. R.D.S. is the National Chair of the Vascular Clinical Reference Group and a National Specialty Advisor for vascular services, both roles for NHS England. The funders had no role in the study.

Supplementary material

Supplementary material is available at BJS Open online.

Data availability

Data are not publicly available. Data access via CPRD is subject to protocol approval via CPRD’s Research Data Governance (RDG) Process.

Author contributions

Anna Meffen (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing—original draft, Writing—review & editing), Mark Rutherford (Methodology, Writing—review & editing), Rob D. Sayers (Conceptualization, Methodology, Supervision, Writing—review & editing), John Houghton (Methodology, Writing—review & editing), Naomi Bradbury (Visualization, Writing—review & editing) and Laura Gray (Conceptualization, Methodology, Supervision, Writing—review & editing).