Association Between Atopic Eczema and Cancer in England and Denmark

Two large matched cohort studies using health care data from England and Denmark investigate cancer risk in people with and without atopic eczema, including variation in risk associated with eczema severity and activity.


eAppendix 2. STUDY POPULATIONS England
In England, the population eligible for inclusion comprised all adults aged 18 years and older registered with CPRD practices eligible for linkage to HES/ONS databases between 2 January 1998 and 31 March 2016 (period of complete linkage to HES and ONS version 14) (Figure 1). All individuals had to be registered with a practice meeting CPRD quality control standards and had to have at least 12 months of registration prior to study entry (to allow adequate time for recording of baseline data in the patient record). We restricted our study population to adults due to limited follow up (mean follow up time = 5 years 8 ) and the rarity of childhood cancer outcomes.
From this population, we identified individuals with atopic eczema (eczema-exposed cohort) using a validated algorithm 3 based on a record of one eczema diagnosis code from primary or secondary care and at least two primary care records (diagnosis code or prescription) for eczema therapy (eAppendix 1). We considered the date on which the individual fulfilled the full eczema algorithm (eczema diagnosis code or second eczema therapy record, whichever occurred first) to be the date of eczema diagnosis. Eczema-exposed individuals entered the cohort (index date) on the latest of the following: one year after date of registration with their primary care practice; date practice met CPRD quality control standards; start of study (2 January 1998); 18th birthday; or 12 months after the date of eczema diagnosis (to limit potential reverse causality).
We identified an age-, sex-, calendar period-and general practice-matched cohort of individuals without eczema. We randomly matched (without replacement) up to five individuals without atopic eczema for every individual with eczema in calendar date order (i.e., individuals in the matched cohort were assigned first to individuals with eczema with earliest cohort entry to avoid time-related bias). Individuals without atopic eczema entered the cohort on the same date as the matched eczema-exposed individual (index date). We allowed a 15-year age difference for matching to maximise the generalisability of our cohort because of the small pool of eligible matched comparators (particularly among women and older people). We accounted for the wide age-matching window in our analysis by finely adjusting for age as the underlying timescale.
We followed individuals from cohort entry until the earliest of the following: no longer registered with practice, practice no longer contributing data to CPRD, death, end of study (31 March 2016), diagnosis of atopic eczema (matched cohort only), or cancer diagnosis (any cancer diagnosis, excluding non-melanoma skin cancer [NMSC]). We used morbidity coding in the CPRD and HES databases and cause-of-death coding in the ONS database to identify all records of cancer.

Denmark
In Denmark, we used data from linked nationwide registries: 1) The Danish National Patient Registry includes data on non-psychiatric hospital admissions from 1977, and all Danish hospital admissions, outpatient clinic appointments and emergency room contacts from 1994 9 ; 2) The Danish Cancer Registry includes data on incident cancers diagnosed since 1943 4 ; 3) The Civil Registration System includes demographic data (e.g., civil status, vital status and address) for the entire population recorded from 1968 10 ; 4) The Danish National Prescription Registry includes data on all prescriptions filled at community pharmacies since 1995 11 ; and 5) A number of socioeconomic sources, including education level and income data, gathered by Statistics Denmark, the central statistical authority in Denmark. 12,13 We used the Danish National Patient Registry to identify all individuals born in Denmark who received a hospital diagnosis of atopic eczema at any age between 1 January 1982 and 30 June 2016 (period of coverage from main data sources). Individuals entered the cohort on the date of their first atopic eczema diagnosis, recorded either during a hospital admission or an outpatient appointment. Individuals had to be living in Denmark on the date of cohort entry. The Danish atopic eczema cohort represents individuals with moderateto-severe atopic eczema, as they were identified through hospital records.
We identified a matched cohort of individuals without atopic eczema by matching up to 10 individuals (with replacement) on sex and exact birth year to each individual with atopic eczema. Individuals entered the comparator cohort on the eczema diagnosis date of the matched eczema-exposed individual. They were required to be born and currently living in Denmark, with no previous atopic eczema diagnosis recorded on or before cohort entry.
We followed cohorts from cohort entry (date of atopic eczema diagnosis both for individuals with eczema and their matched comparators) to the earliest of the following: emigration, death, end of study (30 June 2017), diagnosis of atopic eczema (matched cohort only), or first-ever cancer diagnosis recorded in the Danish Cancer Registry. eAppendix 3. SECONDARY ANALYSES In secondary analyses, we investigated whether the association between atopic eczema and cancer was: (1) more pronounced in individuals with severe or active eczema and (2) modified by age, sex, or asthma. In Denmark, sample sizes were insufficient for analyses of eczema severity and activity. However, as we had no access to Danish primary care data (where mild eczema is managed), it is likely that people with eczema included in the Danish cohort had disease that required more intensive topical therapy and systemic treatment, corresponding to people classified with moderate/severe eczema in the English study population.

Eczema severity
In analyses examining eczema severity, we classified individuals with atopic eczema as having mild, moderate, or severe disease. We then compared their cancer risk to the risk among individuals without atopic eczema, using Cox regression to estimate HRs.
To define atopic eczema disease severity, we considered individuals to have mild disease by default. We classified individuals as having moderate disease from the first of: (1) a second potent topical corticosteroid prescription within one year; or (2) a first prescription for a topical calcineurin inhibitor. We defined severe atopic eczema from the first of: (1) use of phototherapy or systemic treatment for atopic eczema (excluding systemic glucocorticoids, as they may have been prescribed for coexisting asthma); (2) hospital admission for eczema (i.e., when eczema was recorded as the primary diagnosis of an admission); or (3) referral to a dermatologist. Severity was updated over time, i.e., individuals were classified as having mild eczema until the date they first satisfied the requirements of the moderate-eczema definition (unless they developed severe eczema). When (and if) a person satisfied the requirements of the definition for severe eczema, they switched to the severe category and remained in it for the rest of follow-up (similar to established procedures for defining severity in psoriasis studies 14 ). At any given point during follow-up, individuals with atopic eczema therefore belonged to a single severity category.

Eczema activity
In addition to capturing maximum atopic eczema disease severity, we examined eczema activity, which reflects the proportion of life lived with eczema symptoms. It is possible that the association between atopic eczema and cancer differs between people experiencing severe eczema of short duration and those with longstanding mild eczema. We explored the role of atopic eczema disease activity in a post-hoc analysis, as results from our pre-specified analysis were suggestive of time-related bias (see 'Original algorithm' below). In the analysis of eczema activity, we classified individuals with atopic eczema into three categories based on the proportion of the 12 months preceding start of follow-up during which they experienced active disease (no active eczema; active disease for less than 50% of the time; or active disease for at least 50% of the time) and compared their risk of cancer to individuals without eczema using Cox regression. As cohort entry in the English study was delayed until at least 12 months after the eczema diagnosis date, we identified eczema disease activity status in the 12 months before cohort entry. We defined disease activity as lasting for three months (considered a reasonable interval between prescriptions), starting from any medical record suggesting atopic eczema, including an eczema-related primary care consultation, a hospital admission, or a prescription for medication used to treat eczema. If an individual had a second medical record of eczema activity within the first three-month period, we extended the period of active disease by another three months from that second record (and so on for subsequent records). We then calculated disease activity as the percentage of time with active disease in the 12 months before start of follow-up.
We aimed to identify a measure of atopic eczema disease activity that captured long-term disease control (e.g. the proportion of an individual's life with eczema symptoms). In the past, eczema disease activity has been operationalised either as: (1) as a well-controlled week, defined as up to two days of symptoms or treatment in a week; 15 or (2) using flares, 16 defined as treatment intensification (i.e., stepped-up treatment) and identifying those with frequent exacerbations as having active eczema. However, these approaches cannot easily be applied to electronic health record data.

Original algorithm
In our study protocol, we identified active atopic eczema as two or more primary care consultations, primary care prescriptions, or hospital records for eczema (diagnostic codes or records for eczema therapies [including prescribed medications and records for phototherapy]) within a one-year period. 17 Individuals started contributing person-time in the "active eczema" group on the date of the second consultation or prescription and were classified as having active eczema for the next 12 months, unless a new consultation/prescription was recorded, in which case they were classified as having active eczema for a further 12 months (and so on).
Atopic eczema disease activity was then classified into one of three categories based on the percentage of total follow-up time with active eczema: (1) never active; 2) active during less than 50% of follow-up time; or (3) active during 50% or more of follow-up time.
However, we found that our protocol-defined eczema disease activity algorithm was prone to bias. Individuals with less follow up were more likely to have either all or none of their follow-up time defined as active. One reason is that it is necessary to survive to the second record indicating active disease, meaning that individuals with less follow-up time may be more likely to be classified as never having active disease (insufficient time to record the second active disease record). A second reason is that individuals with less follow-up time may also be more likely to have all of their observed time defined as active due to the algorithm classifying active disease in 12-month increments. Thus, if an individual has only 12 months of follow-up and they have two eczema records during this time, much of the time will be defined as active. This could lead to erroneously high rates of cancer in those who either never have active disease or whose eczema is active for more than 50% of follow up time, as our original algorithm may differentially classify individuals with short follow-up time due to cancer in the 'never active group' or in the group whose disease is active during 50% or more of followup time.
Further, our original algorithm determined active disease status at any time point during follow-up based on data from the whole follow-up period. Therefore, the algorithm potentially misclassified activity status at an earlier time point based on what happened during a later time point.

Post hoc algorithm
Given the likelihood of time-related biases, we chose an alternative atopic eczema disease activity definition.
We classified individuals with atopic eczema into three categories based on the proportion of the 12 months prior to the start of follow up spent with active disease (never active eczema, active eczema for less than 50% of the 12 months, or active eczema for at least 50% of the 12 months) and compared their risk of cancer to individuals without eczema using Cox regression. As cohort entry in the English study was delayed until at least 12 months after the eczema diagnosis date, we identified eczema disease activity status during the 12 months before cohort entry. We defined disease activity as lasting for three months (considered a reasonable interval between prescriptions) from any medical record suggesting eczema, including an eczema-related primary care consultation, a hospital admission, or therapy (prescribed medication or record of phototherapy). If an individual had a further record of activity within the three months, we extended the period of active disease by three months from the date of the second record (and so on for subsequent records). We then calculated disease activity as the percentage of time with active disease in the 12 months before start of follow-up.
Our post-hoc algorithm does not capture the totality of eczema disease experience throughout the follow-up period. For example, the algorithm may: (1) misclassify individuals with early active disease that remits during follow up as having highly active disease (leading to underestimation of the effect of active disease); or (2) misclassify individuals with early inactive disease that flares during follow up as having limited disease activity (again leading to underestimation of the effect of active disease). Overall, the effect of potential misclassification introduced by capturing a measure of eczema disease activity at cohort entry rather than throughout follow up (effectively an 'intention-to-treat' analysis) would be to bias our results to the null. Thus, any estimates of the effect of eczema disease activity using our new definition are likely to be conservative. eFigure 1. England: Flowchart illustrating identification of study participants. *NB: numbers of individuals with and without eczema do not necessarily sum to the total number of individuals included, as individuals with eczema could be included in the matched comparison cohort up until the date of their first eczema diagnosis. Abbreviations: CPRD: Clinical Practice Research Datalink; HES: Hospital Episode Statistics; ONS: Office for National Statistics. eFigure 3. England: Association (HR [99% CI]* comparing individuals at each level of eczema disease activity to those without eczema) between atopic eczema disease activity and cancer. *Estimated hazard ratios obtained from Cox regression using current age as the underlying timescale, stratified by matched set (matched on age at cohort entry, sex, general practice, and date of cohort entry). All models fitted to individuals with complete data for all variables included in each model and from valid matched sets, including one eczema-exposed individual and at least one unexposed individual without eczema. All models implicitly adjusted for sex, date of cohort entry, and practice (due to stratification by matched set), and age (due to underlying timescale). Potential mediators: Body mass index, harmful alcohol use, and smoking. Abbreviations: IMD: Index of Multiple Deprivation; CNS: central nervous system. Specific CNS tumours (meningioma, brain neoplasm, and other CNS tumours) were too rare to consider individually. eTable 1. Sensitivity analyses. Sensitivity analysis description Justification England* Denmark* Restricted to individuals with a 3-year cancer-free window following eczema diagnosis (main analysis in English studies required a 12-month window, Danish studies did not have this 12-month window to maximise study power). This analysis was restricted to individuals with at least three years of followup in the CPRD prior to cohort entry, with no record of a cancer diagnosis prior to cohort entry.
We included a 12-month cancer-free window following eczema diagnosis in the English studies to limit reverse causality as an explanation for any observed association (to maximise power, given the smaller sample size, we did not do this for the main analysis for the Danish studies). By increasing this window to 3 years, we aimed to account for any lag between cancer onset and diagnosis, and therefore to further reduce reverse causality.
Yes (E1) Yes (D1) Restricted to individuals with at least one consultation with their GP in the year prior to cohort entry.
To exclude practice non-attenders.

Yes (E2) No
Restricted to individuals with a diagnosis of newly active atopic eczema in the exposed cohort (exposed individuals defined as those who joined the cohort when they first fulfilled the study's diagnostic criteria after the start of the study period) and their matched sets.
Covariates measured at entry precede atopic eczema onset, so will not be on the causal pathway between atopic eczema and cancer outcomes.

Yes (E3) No
The main analysis was repeated using a redefined cohort (different unexposed pool, exposed individuals unchanged, unless unmatched), where: (1) the pool of unexposed people also included individuals with an atopic eczema diagnosis but without two treatments for the entire duration of their followup; and (2) individuals in the exposed cohort (with an atopic eczema diagnosis and two treatments) were classified as unexposed up until cohort entry (i.e., at the time of their atopic eczema diagnosis or their two treatments, whichever occurred later). This cohort was matched separately to the main analysis cohort.
To explore the sensitivity of the results to the definition of the exposure.

Yes (E4) No
The main analysis was repeated using a second redefined cohort, in which exposed individuals had only an atopic eczema diagnosis (i.e., without requiring two atopic eczema treatments) and these individuals were eligible for inclusion in the unexposed cohort up until their atopic eczema diagnosis.
(Some individuals may have had childhood atopic eczema; however, they might not have recorded treatment codes if they registered with their GP during adulthood. These individuals therefore may have been excluded erroneously from the exposed cohort in the primary analysis). This cohort was matched separately to the main analysis cohort.
To explore the sensitivity of the results to the definition of the exposure.

Yes (E5) No
We repeated the analysis after removing the censoring criterion of an atopic eczema diagnosis for members of the matched cohort.
To avoid potential informative censoring by atopic eczema in the matched cohort.

No Yes (D5)
Restricted to a subset of individuals registered from 2006 onwards, and additionally adjusting for ethnicity (White, South Asian, Black, other, or mixed), identified from CPRD and HES data using a previously developed algorithm 18 Records for ethnicity became more complete following the introduction of remuneration for including ethnicity data in the Quality and Outcomes Framework).
To examine whether the omission of ethnicity as a covariate in the main analysis introduced bias.

Yes (E6) No
Restricted to individuals entering the cohorts from 2004 onwards.
To account for changes in diagnostic and coding practices over time -specifically those introduced by the Quality and Outcomes Framework in 2004. This is likely to be more important for cancer outcomes and some covariates. However, as there are no specific dermatology indicators in the Quality and Outcomes Framework, it is unlikely that eczema coding was affected).

Yes (E7) No
Additionally including diabetes mellitus as a covariate. Diabetes mellitus was classified as absent, type 1, type 2, or type unknown, based on primary or secondary care diabetes diagnoses (using definite diagnostic codes only). Onset of diabetes was defined as the date of the earliest recorded diagnostic code. Diabetes mellitus is part of the Quality and Outcomes Framework, which remunerates primary care practices for recording specific management and monitoring practices for some chronic conditions. Therefore, diagnostic coding is likely for individuals with diagnosed diabetes, eliminating the need to supplement the definition with prescriptions for anti-diabetic drugs or possible diabetes codes.
Our rationale for including diabetes mellitus as a covariate and stratifying by diabetes type (rather than simply classifying as present or absent) is as follows: (1) Type 1 diabetes mellitus is immune-mediated (resulting from autoimmune destruction of pancreatic beta-cells); (2) Research suggests that there is a link between diabetes and cancer. 19 (3) We believe that any link between eczema and cancer will be via an immune mechanism. Therefore, it seems important to consider diabetes as a potential confounder in the relationship between eczema and cancer. However, as this is a somewhat tenuous link, we only adjusted for diabetes in this sensitivity analysis.
Yes (E8) Adjusted for diabetes (as a proxy measure of lifestyle factors) in analysis adjusting for potential mediators Repeated main analysis using a cohort that was matched without allowing for a 12-month cancer-free window starting from the eczema diagnosis date (i.e. start of observation for eczema-exposed individuals was the earliest of: diagnosis date + one year, current registration date + one year, up-tostandard date, or 18 th birthday, whichever occurred first).
To be comparable with methods used with Danish data.

Yes (E9) No
Additionally adjusting for immunosuppression. Immunosuppression was defined using coding of immunosuppressive disorders and records of prescriptions for drugs leading to immunosuppression (excluding oral corticosteroid use, as this is likely to be transient).
Immunosuppressive disorders included HIV, haematopoietic stem cell or bone marrow transplant, myeloma, or other unspecified cellular immune deficiencies (e.g., pancytopenia), leukaemia, and lymphoma. Myeloma, leukaemia, and lymphoma were removed from the definition of immunosuppression in the respective analyses in which they were outcomes.
As immunosuppression might mediate the relationship between eczema and specific cancers, we only adjusted for immunosuppression in a sensitivity analysis.

Yes (E10) Yes (D2)
Additionally adjusting for oral glucocorticoids. We adjusted for oral glucocorticoids as a time-updated variable categorized as ever/never prescribed, with status changing at the time of the first prescription.
Oral glucocorticoids may be prescribed more frequently to people with atopic eczema because of coexisting asthma and may also increase the risk of certain cancers.

Yes (E11) Yes (D3)
Additionally adjusting for systemic therapies. We adjusted for systemic therapy (methotrexate, ciclosporin, azathioprine, or mycophenolate), as a time-updated variable categorized as ever/never prescribed, with status changing at the time of the first prescription. Oral glucocorticoids were not included.
Systemic immunosuppressive drugs (cyclosporin, azathioprine, mycophenolate and methotrexate) may be prescribed for severe atopic eczema and may also affect cancer risk.

Yes (E12) Yes (D4)
We repeated the overall cancer analysis after excluding any skin cancer diagnoses from the overall cancer outcome definition.
To explore potential ascertainment bias introduced by investigating skin cancer in people with and without atopic eczema. People with eczema are more likely to be diagnosed with skin cancer as clinicians examine skin more carefully in those with existing skin disease.

Yes (E13) Yes (D5)
Repeated specific cancer analyses without censoring at diagnosis of other types of cancer (i.e., censoring only at diagnosis of the specific cancer under investigation).
In the main analyses we censored at the time of any cancer diagnosis, as some cancers may represent secondary sites of an original cancer (rather than a primary cancer diagnosis). We repeated the analysis to explore the impact of not censoring at diagnosis of other types of cancer.

Yes (E14) No
Repeated the main analysis after restriction to individuals with complete smoking and BMI data in valid matched sets.
We undertook a complete case analysis for analyses additionally adjusting for potential mediators (i.e. smoking status, BMI, and harmful alcohol use). We therefore repeated the main analysis in individuals with complete BMI and smoking data after adjusting for quintile of IMD score and calendar period only (no further adjustment for potential mediators) to ensure that any difference between effect estimates from analyses adjusted for confounders and from analyses additionally adjusted for potential mediators were due to the additional covariables adjusted for and not a result of the restriction of the study sample.

Yes (E15) No
Repeated analysis after restriction to individuals with an index date on or after 1 January 1996.
To limit left-censoring of variables (i.e., incorrect identification of variables due to insufficient data to capture them) by ensuring at least 1 year of registration history for all registries prior to cohort entry (index) date.

No Yes (D6)
Additionally adjusting for socioeconomic status by adding baseline educational level (short, medium, long, missing), partnership status (married/cohabiting, single) and gross personal income (low, intermediate, high, very high) to the minimally adjusted model. We restricted the analysis to those aged 30 years or older on the cohort entry date, as people are more likely to have recorded partnership and income information and to have attained their highest education level by this age.
To explore the potential role of socioeconomic outcomes as confounders of the association between atopic eczema and cancer. Prior to this analysis, we also repeated the main analysis for individuals aged 30 years and older without adjusting for SES to ensure that any differences in effect estimates after adjusting for SES variables were due to SES and not to restriction to people over 30 years old (D7a).

No
Yes ( Mean (SD) duration of follow-up (years) 6.1 (4.9) 5.6 (4.8) 6.5 (5.0) 6.1 (5.0) 4.1 (3.9) 4.2 (4.0) 2.9 (3.2) 3.6 (3.8) Individuals could contribute data as both eczema exposed and unexposed. Follow-up based on censoring at the earliest of: death, no longer registered with practice, practice no longer contributing to CPRD, or any cancer diagnosis (excluding non-melanoma skin cancer) Abbreviations: IQR: Interquartile range; SD: Standard deviation; IMD: Index of multiple deprivation. BMI: Body mass index. a. Follow-up based on censoring at the earliest of: death, no longer registered with practice, practice no longer contributing to CPRD, or any cancer diagnosis (excluding non-melanoma skin cancer). b. Age or calendar period at cohort entry. c. IMD based on individual-level data (from 2007) if available, supplemented with practice-level data (from 2010) if individual-level data not available. d. Smoking and BMI based on records closest to index date. e. Based on records on or before cohort entry. eTable 3. Denmark: Characteristics of the study population at cohort entry, for the overall cohort and for individuals aged 30 years or over (included in sensitivity analysis adjusting for socioeconomic status). Values are n (%) unless stated otherwise. *Estimated hazard ratios from Cox regression with current age as the underlying timescale, stratified by matched set (matched on age at cohort entry, sex, general practice, and date at cohort entry). All models fitted to individuals with complete data for all variables included in each model and from valid matched sets, including one eczema exposed individual and at least one unexposed individual without eczema. All models implicitly adjusted for sex, date at cohort entry and practice due to stratification by matched set, and age due to underlying timescale. Abbreviations: BMI: Body mass index; IMD: Index of multiple deprivation. eTable 5. Denmark: Association (HR [99% CI]) between atopic eczema and cancer outcomes: comparing risk of cancer in those with atopic eczema to those without. Abbreviations: SES, socioeconomic status a. Estimated based on a Cox regression model with time since index date (date of atopic eczema diagnosis in eczema exposed, and index date of matched individual with atopic eczema in the cohort without eczema) as underlying timescale, stratified by matched set to account for matching factors (sex and birth year) and adjusted for any additional variables in the sensitivity analysis. Reference group is cohort without atopic eczema. b. Estimated based on a Cox regression model with time since index date (date of atopic eczema diagnosis in eczema exposed, and index date of matched individual with atopic eczema in the cohort without eczema) as underlying timescale, stratified by matched set to account for matching factors (sex and birth year) and adjusting for time-varying lifestyle-related diseases as well as any additional variables in the sensitivity analysis. Reference group is cohort without atopic eczema. *Note: Full methods and justification of each sensitivity analysis are included in Supplementary eTable 1 (All analysis numbers [e.g. D1, D2 etc.] reference the sensitivity analyses in eTable 1). *Estimated hazard ratios from Cox regression with current age as the underlying timescale, stratified by matched set (matched on age at cohort entry, sex, general practice, and date at cohort entry). All models fitted to individuals with complete data for all variables included in each model and from valid matched sets, including one eczema exposed individual and at least one unexposed individual without eczema. All models implicitly adjusted for sex, date at cohort entry and practice due to stratification by matched set, and age due to underlying timescale. eTable 9. England: Association (HR [99% CI]*) between atopic eczema disease activity and cancer outcomes. *Estimated hazard ratios from Cox regression with current age as the underlying timescale, stratified by matched set (matched on age at cohort entry, sex, general practice, and date at cohort entry). All models fitted to individuals with complete data for all variables included in each model and from valid matched sets, including one eczema exposed individual and at least one unexposed individual without eczema. eTable 10. England: Adjusted hazard ratios (99% CIs) for the association between atopic eczema and cancer, stratified by sex (adjusted for calendar period and IMD). *Estimated hazard ratios from Cox regression with current age as the underlying timescale, stratified by matched set (matched on age at cohort entry, sex, general practice, and date at cohort entry). All models fitted to individuals with complete data for all variables included in each model and from valid matched sets, including one eczema exposed individual and at least one unexposed individual without eczema. All models implicitly adjusted for sex, date at cohort entry and practice due to stratification by matched set, and age due to underlying timescale. P-values from likelihood ratio tests comparing models with and without an interaction term between atopic eczema and sex. eTable 11. Denmark: Adjusted hazard ratios (99% CIs) for the association between atopic eczema and cancer, stratified by sex. a. Estimated based on a Cox regression model with time since index date as underlying timescale, stratified by matched set to account for matching factors (sex, birth year). b. Mediation model: Adjusted additionally for time-varying lifestyle-related diseases. *To preserve patient confidentiality, where numbers of events are less than 5, we have not presented person years at risk. eTable 12. England: Adjusted hazard ratios (99% CIs) for the association between atopic eczema stratified by current age (adjusted for calendar period and IMD). *Estimated hazard ratios from Cox regression with current age as the underlying timescale, stratified by matched set (matched on age at cohort entry, sex, general practice, and date at cohort entry). All models fitted to individuals with complete data for all variables included in each model and from valid matched sets, including one eczema exposed individual and at least one unexposed individual without eczema. All models implicitly adjusted for sex, date at cohort entry and practice due to stratification by matched set, and age due to underlying timescale. P-values from likelihood ratio tests comparing models with and without an interaction term between atopic eczema and age band. eTable 13. Denmark: Adjusted hazard ratios (99% CIs) for the association between atopic eczema and cancer, stratified by age.