Table 1.

Sensitivity analyses

Sensitivity analysis	Justification
The activity analysis was repeated, restricted to patients with at least five years of follow-up	To explore any potential bias caused by patients with atopic eczema with short follow-up periods being more likely to have either none or all of their follow-up with active atopic eczema
The primary analysis was repeated on an incident atopic eczema cohort (exposed patients defined as those joining the cohort when they first fulfil our diagnostic criteria and after the start of the study period)	Covariates measured at entry precede atopic eczema onset so will not be on the causal pathway between atopic eczema and cardiovascular outcomes
The primary analysis was repeated on patients with at least one consultation with their doctor in the year before cohort entry	To exclude practice non-attenders
The primary analysis was repeated on a redefined cohort, where the pool of unexposed patients also included patients with an atopic eczema diagnosis but without two further treatments for the entire duration of their follow-up and patients in the exposed cohort (with an atopic eczema diagnosis and two further treatments) were included as unexposed up until their cohort entry (ie, the latest of their atopic eczema diagnosis and their two further treatments)	To explore the sensitivity of the results to the definition of the exposure
The primary analysis was repeated on a second redefined cohort where exposed patients were those with an atopic eczema diagnosis only (ie, without requiring two atopic eczema treatments), and these patients were eligible for the unexposed cohort up until their atopic eczema diagnosis (some patients may have childhood atopic eczema, but may not have treatment codes recorded if they registered at the practice during adulthood, and therefore may be erroneously excluded from the exposed cohort in the primary analysis)	To explore the sensitivity of the results to the definition of the exposure
The primary analysis was repeated on a subset of patients registered from 2007 onwards	Data on covariates, particularly body mass index and smoking, would be expected to be more complete, thus reducing any selection bias owing to missing data
The primary analysis was repeated on a subset of patients registered from 2007 onwards, additionally adjusting for ethnic group (white, South Asian, black, other, or mixed, identified from Clinical Practice Research Datalink and Hospital Episode Statistics data using a previously developed algorithm)18	To examine whether the omission of this covariate in the primary analysis may have introduced bias
The primary analysis (mediation model only) was repeated with additional adjustment for time updated use of high dose corticosteroids	To examine whether the omission of this covariate in the primary analysis may have introduced bias