Statistical methods
Age at AD debut (≤/> 1 yr), aeroallergen sensitization
(yes/no) and allergic rhinitis (yes/no) were categorized as binary
variables. Severity of AD by means of SCORAD was included in the models
as a continuous variable including the highest SCORAD level measured in
the child (early-onset from 0-1 year and late-onset from 1-6 years).
Associations between AD (binary age at onset or continuous SCORAD) and
aeroallergen sensitization and allergic rhinitis were analyzed by
logistic regression analyses expressing results as odds ratios (OR) with
95% confidence intervals (CI).
Furthermore,
we used a logistic regression general estimating equations (GEE) model
to compute the overall OR for allergy endpoints using compiled data from
both time points and accounting for repeated measures. All results are
calculated as crude and adjusted for potential confounders including
sex, older siblings, maternal allergic rhinitis, paternal allergic
rhinitis, breastfeeding and smoking during 3rdtrimester.
The analysis of whether the association between AD and later
sensitization/allergic rhinitis depended on the AD-debut age (yes/no)
was analyzed using a logistic regression model where AD status is a
three-level variable (early, late, never). The analysis of whether the
association between AD and later sensitization/allergic rhinitis
depended on the AD-debut age (SCORAD) was based on two models using the
quantitative measure of SCORAD as a predictor; one based on early debut,
and one based on late debut. As all non-AD children had a SCORAD of 0,
these are excluded from the analysis. From these models the slopes are
compared using a Wald’s tests associated with H0: b(early) = b(late)
which hence reveal whether the association between severity and AR/sIgE
is different between early and late AD-debuts.
In the analyses of early-onset AD (debut≤1yr, yes/no) we used all the
other children as a control group, whereas in the analyses of late-onset
AD (debut 1-6 years, yes/no) we excluded the children with an
early-onset AD debut from the control group.
We additionally stratified for FLG mutation. Further, effect
modification by FLG was evaluated by second order interaction
models by adding cross-products to the models.
All statistical analyses were performed with the statistical software
package R version 4.0.2 and RStudio version 1.1.442 (RStudio Inc,
Boston, MA, USA). Missing observations were treated as missing data.
P-values <0.05 were considered statistically significant.