This study retrospectively collected data from patients and was conducted in accordance with the ethical standards of the institutional review board (vote #21-323) and with the 1964 Helsinki Declaration and its later amendments. Only data from patients who provided general consent for retrospective research were included in the study.

Generative AI tools (ChatGPT (OpenAI), SciSpace, and DeepL) were used to support R coding, data analysis design, literature searches, language translation, and native-language refinement. In addition, large language models were employed for a preliminary internal review of the manuscript to identify inconsistencies, improve clarity, and ensure terminological accuracy. All AI-assisted outputs were critically reviewed and verified by the authors to ensure accuracy, methodological rigor, and adherence to ethical standards.

This study included patients with allergy-related diseases who underwent serum eosinophilic cationic protein (ECP) testing at the University of Lübeck’s allergy outpatient clinic between January 2014 and November 2025. For each ECP test, data were extracted on patient age, sex, test date, eosinophil granulocyte counts (if available), and ICD-10 diagnosis codes from the same quarter.

The primary diagnostic codes of interest were D84.1 (hereditary angioedema, HAE) and T78.3 (angioedema), alongside adjustment variables including J30 (rhinitis), J32 and J33 (sinusitis and nasal polyposis), L20 (atopic dermatitis), L50 (urticaria), D47 (including mastocytosis), and J45 (allergic asthma).

Initial documentation occasionally listed D84.1 as a tentative diagnosis. Therefore, all patients with recorded D84.1 were systematically reviewed. If HAE could not be confirmed based on predefined criteria, the diagnosis was marked as “not present.”

Diagnosis of HAE was based on clinical practice guideline recommendations (1) and internationally consented criteria as described by Buttgereit et al. and Zuraw et al. (35, 36).

For HAE-C1INH (types 1/2), confirmation required repeated laboratory testing demonstrating reduced concentration or function of C1 inhibitor (C1INH). Acquired C1-INH deficiency was excluded based on normal C1q levels and absence of autoimmune/lymphoproliferative disease; hereditary disease was supported by childhood onset, family history and/or pathogenic SERPING1 variant.

HAE (all types) was considered present if the following conditions were fulfilled:

Prophylactic treatments were not systematically documented across all records; however, a substantial proportion of ECP measurements in the HAE cohort were obtained during ongoing long-term prophylaxis with berotralstat or lanadelumab as part of routine follow-up care. In contrast, most control patients were sampled at first presentation and were typically not receiving long-term anti-allergic medication at the time of blood collection. Systemic corticosteroids were rarely used in either cohort and were generally reserved for emergency situations; given the elective outpatient setting, relevant corticosteroid exposure at the time of sampling was uncommon. The ECP dataset contained no classical missing data: all extracted ECP measurements had complete accompanying information on age, sex, and test date. Eosinophil counts were simply not ordered at some visits and were therefore absent by design rather than missing. Accordingly, we analyzed (i) the full ECP cohort and (ii) a filtered eosinophil sub-cohort comprising only entries with available eosinophil counts. No imputation was required.

Blood samples were obtained as part of routine outpatient care during regular clinic hours. For serum preparation, blood was collected using serum tubes with clot activator beads (S-Monovette^®, 7.5 mL; Sarstedt AG & Co. KG, Nümbrecht, Germany; article no. 01.1601). For complete blood counts including absolute eosinophil numbers, blood was collected using EDTA tubes (S-Monovette^® EDTA K3E, 2.7 mL; Sarstedt AG & Co. KG, Nümbrecht, Germany; order no. 04.1917). Serum and EDTA samples were obtained simultaneously during the same venipuncture procedure.

All samples were collected during symptom-free intervals; no blood samples were obtained during acute angioedema attacks.

After blood collection, serum samples were kept at room temperature for approximately 1.5 hours to allow clot formation. Serum was then separated from blood cells by centrifugation according to standardized procedures established in the automated routine clinical laboratory. No deviations from normal laboratory protocols occurred.

Serum ECP levels were measured using the ImmunoCAP^® ECP assay (Thermo Fisher Scientific, Waltham, MA, USA), a CE-marked in-vitro diagnostic test routinely used in clinical practice. Absolute eosinophil counts were obtained from peripheral blood using standard certified automated hematology analyzers as part of routine complete blood counts.

As primary outcome, we determined whether ECP levels measured in a real-world setting were specifically increased in cases with confirmed HAE diagnosis. A secondary outcome was the specific increase of absolute eosinophil counts in HAE patients. As exploratory outcomes, we analyzed ECP levels and absolute eosinophil in HAE subgroups (i.e. HAE-C1INH/HAE type 1 and 2, and HAE-nC1INH/HAE type 3) and acquired mast-cell mediated angioedema (denoted as AAE.URT according to the current DANCE classification). Furthermore, we performed sex-stratified analyses on all HAE types and, for balanced comparison, restricted to HAE-C1INH only.

All analyses were conducted in R (version 4.4.3, 2025-02–28 ucrt; R Foundation for Statistical Computing, Vienna, Austria; (URL: http://www.r-project.org/, last accessed 29 Sep 2025). The workflow relied on a modern tidyverse-based data structure and visualization ecosystem, and on specialized causal-inference and Bayesian tools. We used tibble (modern data frames), dplyr (data manipulation), tidyr (data reshaping), stringr (string handling), and forcats (factor handling). For visualization we used ggplot2 as the core plotting engine, ggpubr for multi-panel figure assembly, ggnewscale for multiple independent color/fill scales, ggbeeswarm for compact raw-data overlays, ggtext for rich-text axis labels, and scales for axis transformations. Bayesian visualization and posterior summarization used bayesplot, bayestestR, and ggdist. Covariate balance diagnostics were performed using cobalt. Propensity scores were estimated via penalized Firth logistic regression using logistf, ensuring stable estimation under data sparsity. Inverse-probability weighting and post-stratification incorporated Dirichlet draws from MCMCpack. Outcome modeling was performed with INLA (Integrated Nested Laplace Approximation; URL: http://www.r-inla.org/, last accessed 29 Sep 2025), using spline bases from splines for nonlinear age and seasonality effects. E-values for unmeasured-confounding sensitivity analyses were computed using the EValue package. Word-export and table generation used officer and flextable. File naming was handled using filenamer.

The interaction between data variables in the study was visualized and analyzed using DAG as implemented with dagitty (URL: https://www.dagitty.net/dags.html, last accessed 29. Sep.2025). The dagitty code for our DAG is provided in the Supplementary Materials.

Power calculations for the Student’s t-test were performed using the pwr package’s pwr.t2n.test function. The assurance metric, representing the Bayesian probability of study success given the prior uncertainty, was calculated using the package bayesassurance with the function assurance_nd_na (37, 38). A power of 0.8 is generally considered adequate for study planning, whereas no universally accepted target value exists for the assurance metric. The R script for power and assurance calculations (Commented_Bayesian power calculation.R) is provided in the Supplementary Materials.

For data analysis, we used a multilevel regression and poststratification (MRP) framework (39). To obtain unbiased and efficient estimates of HAE’s effect on ECP levels and eosinophil counts, we implemented a doubly robust causal-inference strategy that incorporated adjustment variables both in the propensity score model and in the outcome regression (40, 41). For regression analyses, especially those requiring hierarchical model formulations, we used Integrated Nested Laplace Approximation (INLA), a Bayesian inference framework developed by Rue, Martino, and Chopin (42).

In this framework, the variable traditionally referred to as treatment in causal inference Ti, represents HAE status (1 = HAE case, 0 = non-HAE control). Although HAE status is not a treatment in the clinical sense, we retain the standard causal-inference terminology –specifically the ATT (Average Treatment Effect on the Treated) and ATO (Average Treatment Effect for the Overlap Population) estimands – because these definitions refer to the target population of the estimand rather than to an actual therapeutic intervention. This convention is widely used in observational causal inference for exposures rather than treatments.

Propensity scores were estimated using penalized Firth logistic regression, predicting each observation’s probability of belonging to the HAE cohort given the adjustment variables (41, 43). Firth regularization was chosen instead of standard logistic regression to prevent bias and instability arising from data sparsity or separation (44).

Inverse-probability treatment weights (IPTW) were derived for the ATT estimand and, in sensitivity analyses, for the ATO estimand. Let pi=P(Ti=1∣Xi) denote the propensity score of patient with PatientID i. We defined the weights as:

ATT weights: wiATT={1, Ti=1pi1−pi , Ti=0

ATO weights: wiATO={1−pi,Ti=1pi, Ti=0

The resulting effective sample size (ESS) was calculated to quantify the amount of information retained after inverse-probability weighting. For a set of weights wi, the ESS is defined as

which corresponds to the size of an equivalent sample with equal weights that would yield the same level of precision (45). ESS was computed separately for the HAE group and the control group for both the ATT and ATO weighting schemes.

The IPTW approach is well suited for repeated-measures data because weights operate at the level of individual observations, allowing multiple measurements per patient to be incorporated without compromising the causal interpretation. In INLA, inverse-probability weights were supplied through the weights argument, which reweights each observation’s contribution to the likelihood and thus implements standard IPTW in a Bayesian framework. To account for within-subject correlation, the outcome model included a patient-specific random intercept using f(PatientID, model=“iid”), capturing individual-level heterogeneity across repeated measurements.

To mitigate selection bias due to non-random measurement times, we incorporated the weekday of blood collection as a bias-breaking variable, following the approach of Geneletti et al. within our multilevel MRP framework (46, 47). In detail, the Geneletti method applies a bias breaking variable B that fulfils the following basic independence assumption:

Using DAG analysis, we determined the weekday of blood collection as an optimal bias breaking variable, because it does not influence the ECP value, but has a considerable impact on the selection of patients into the cohort.

The Geneletti method is described for odds ratios, while our log ECP value y is a continuous variable. Therefore, we need to adapt the calculation. For each weekday, we determine the difference in log ECP values using our linear regression model described above.

with the stratum (weekday) specific HAE effect Δb being

We are interested in an debiased effect estimator Δ, which can be calculated given weekday probabilities p(B=b), i.e. the probability that a patient was included on a specific weekday. We can calculate this by marginalizing over all K weekdays

However, while the effect estimates for Δb are the posterior distributions inferred from the INLA regression, the weekday probabilities p(B=b) are in fact unknown in our case. We addressed this by resampling both the weekday-specific effect estimates and the weekday probabilities. The latter were drawn from a Dirichlet distribution with varying dispersion parameter κ, representing different assumptions about the degree of uncertainty in weekday selection:

We set κ=0.5 (minimally informative Dirichlet prior/Jeffrey’s prior) for high dispersion, κ = 10 (medium dispersion) and κ=20 (low dispersion).

By aggregating samples values Δ(s) obtained above, we approximated the posterior distribution of the overall effect estimator  Δ:

where δΔ(s)(·) denotes the Dirac measure, i.e. a unit mass centered at Δ(s), representing one posterior draw.

Using INLA, we discerned posterior distributions of model parameters, with the bayestestR R package aiding in the extraction of the maximum a posteriori estimator, 95% highest posterior density interval, and Bayesian p-value. This Bayesian p-value aligns closely with its frequentist equivalent, although it differs conceptually, as it reflects the probability that the parameter lies above or below zero given the data (48). We therefore considered a Bayesian p-value ≤ 0.05 as statistically significant.

To give an estimate of the robustness of our estimates, we further calculated the Evalues by VanderWeele et al. (49).

Due to the rare-disease context, the small cohort size, and the longitudinal structure of the data, patient-level raw data cannot be fully anonymized without loss of essential information and are therefore not made publicly available in accordance with GDPR requirements.

Fully commented R scripts, model specifications, directed acyclic graph (DAG) code, and aggregated summary data supporting the findings of this study are provided in the Supplementary Materials. De-identified patient-level data may be made available from the corresponding author upon reasonable request and subject to appropriate ethical and data protection approvals.

The final analysis included 1,928 patients who visited the allergy outpatient clinic of our department and had serum ECP levels measured (ECP measurement cohort, Table 1). Absolute eosinophil counts were available for 1,119 of these patients (eosinophil measurement cohort). The cohort comprised approximately twice as many female as male patients, with a median age of 45–47 years. Among all patients, 48 had a confirmed diagnosis of hereditary angioedema (HAE): 32 with HAE-C1-INH type I, none with type II, and 16 with HAE with normal C1-INH (HAE-nC1INH). Patients contributed variable numbers of repeated measurements for both ECP and eosinophil counts. On average, HAE patients had 2.71 ECP measurements compared to 1.25 in non-HAE controls, with similar patterns in eosinophil count measurements.

For transparency, cohorts were additionally stratified by sex, and sex-specific demographic characteristics and measurement frequencies are provided in Supplementary Table 1. Sex-stratified analyses were performed to assess the robustness of the main findings across demographic strata, rather than to formally test sex as an a priori effect modifier. Importantly, the cohort of patients with HAE-nC1INH consisted almost exclusively of females (N = 15), with only a single male patient (N = 1). In contrast, patients with HAE-C1INH showed a balanced sex distribution, comprising 16 males and 16 females.

To estimate the strength and representativity of our study, we used both the classical power calculation and the Bayesian assurance value, i.e. the probability of success under varying preconditions (Supplementary Figure 1). These calculations were based upon N_HAE=48 HAE patients and a three times higher number of controls (N_controls=144), because we intended to increase the power by leveraging the large number of controls. For a one-sided hypothesis (increased ECP values in HAE patients) we found that an increase of about 1.35, equivalent to an effect size (Cohen’s d) of 0.5 was sufficient to reach a power of 0.8, and, depending on the a-priori assumed uncertainty in the effect estimator (denoted as τ_Design) a Bayesian assurance metric of about 0.6 – 0.9, both values indicating a reasonable study design. It has to be noted here that Bayesian analyses do not depend on power calculations, such as frequentist analyses. The assurance value is generally lower than the power value, because it integrates more sources of uncertainty in pre-study assumptions.

For all patients treated in our department, treatment diagnoses were recorded using the ICD10 code for medical billing purposes. From our database, we identified diagnoses known to be associated with type 2 inflammation as potential confounders. These included allergic rhinitis, sinusitis, atopic dermatitis, bronchial asthma, chronic spontaneous mast cell-mediated angioedema, and chronic spontaneous urticaria. In cases where patients had multiple ECP measurements and potentially different diagnosis codes associated with each measurement, we combined all recorded diagnosis codes. It is important to note that all the mentioned diagnoses are chronic conditions. We assumed that differences in the recorded diagnosis codes were more likely due to documentation errors rather than the presence or absence of the respective disease.

Considering the substantial number of patients with allergic rhinitis in our database, we included the month of blood collection as a seasonal factor. Particularly, for pollen allergies, we anticipated an increase in eosinophilic inflammation during spring and summer.

Our retrospective study design is inherently prone to selection bias. Although our outpatient clinic is integrated into a general allergy service, it has a particular focus on hereditary angioedema, and patients with the full spectrum of allergic, inflammatory, and allergy-like diseases are seen during the same consultation hours. Patient recruitment and sample collection varied across weekdays, as different physicians were responsible for each clinic day. Consequently, inclusion into the analytic cohort is moderately dependent on the weekday. In contrast, the clinical laboratory follows a standardized and largely automated workflow throughout the week, ensuring that ECP measurements and complete blood counts can be considered independent of weekday-related factors.

We used dagitty to visualize and analyze the assumptions about the interactions of data variables in a directed acyclic graph (DAG) (Figure 1). The HAE status, which was additionally curated in the database, is the exposure, and we are interested in its causal effect on the ECP level and eosinophil counts (absolute) as outcomes. The existence of the ECP value is the precondition for inclusion into the cohort (selection variable). Age, sex, season of the year, (recorded) sampling time and comorbidities that have a potential influence were used as adjustment variables, while comorbidities such as type IV sensitization against contact allergens without present disease are considered as having no influence on the ECP value, but lead to presentation in the outpatient clinic with ECP value determination and therefore selection into the cohort.

Using dagitty, causal and biasing paths can be identified by specifying the attributes of the nodes. In our DAG, “inclusion into the cohort” acts as a collider, while the biasing path runs through the weekday variable. Accordingly, poststratification can block this biasing path and thus enable estimation of an unbiased causal effect of HAE status on ECP levels and eosinophil counts. Age, sex, season, and eosinophil-associated conditions were treated as confounders and controlled for through doubly robust regression, further reducing potential bias. Although not depicted in the DAG, unmeasured confounding could still introduce bias; to assess the robustness of our findings, we therefore calculated E-values according to VanderWeele et al. (49). These considerations justify the use of weekday as a bias-breaking variable.

The original Geneletti approach for bias-breaking variables relies on external reference data for poststratification. Because such external weekday distributions were unavailable, we instead implemented Bayesian bootstrapping, assuming a variable distribution of weekday frequencies modeled by a Dirichlet distribution with equal base parameters but varying dispersion (low, moderate, and high). This dispersion propagates into the final posterior estimate of the HAE effect as an additional layer of uncertainty, mathematically attenuating statistical significance and thereby yielding highly conservative effect estimates.

The primary outcome was the serum ECP level. We tested whether patients with hereditary angioedema (HAE) exhibited higher ECP concentrations compared to non-HAE controls (Figure 2). All available ECP measurements from the 48 HAE patients and all controls were included. Propensity scores were estimated using Firth logistic regression based on covariates identified in the directed acyclic graph (sex, age, comorbidities, and month of sample collection). These covariates were used both for constructing inverse-probability-of-treatment weights (ATT estimand) and, in the doubly robust models, as covariates in the outcome regression. The weekday of blood collection was incorporated as a bias-breaking variable based on the Geneletti criterion.

Among 2,349 control measurements, ATT weighting yielded an ESS of 576.87 and a weighted median ECP level of 20 µg/L. Among 130 HAE measurements, the ESS equaled the observed sample size (130), with a weighted median of 32 µg/L (Figure 2). Because repeated measurements inflate simple descriptive statistics, direct group comparisons are not appropriate; therefore, we used hierarchical linear models with patient-specific random intercepts to accommodate multiple measurements per patient.

In contrast to conventional matching approaches, we used inverse-probability-of-treatment weighting (IPTW) to retain all observations and avoid unnecessary data loss. The resulting weights achieved satisfactory covariate balance and adequate overlap between HAE and control patients. Propensity score distributions after weighting are shown in Figure 3, demonstrating substantial common support for estimating the ATT. Covariate-wise standardized mean differences (SMDs) are presented in the corresponding Love plot (Figure 3), confirming that all covariates reached acceptable balance after weighting, i.e. a standardized mean difference of below 0.1.

Rather than reporting a single adjusted estimate, we constructed a set of complementary regression specifications, each targeting the same causal contrast (the effect of HAE on ECP levels) under different statistical assumptions (Figure 3). These included (1): a naïve unweighted regression (2); IPTW-only models (3); doubly robust models (IPTW with covariate adjustment) (4); Bayesian poststratification using low-, medium-, and high-dispersion Dirichlet priors to reflect increasing uncertainty in the weekday distribution; and (5) an analysis using overlap (ATO) instead of ATT weights. Weekday-specific HAE effects were additionally summarized in the result table. Together, these models constitute a structured series of sensitivity analyses that assess the stability of the estimated effect across a wide analytical space.

Among these specifications, the post stratified model with high Dirichlet dispersion – which assumes the greatest uncertainty regarding weekday sampling – yields the most conservative estimate and was therefore designated as the primary effect measure. Even under this highly conservative scenario, HAE was associated with a 1.52-fold increase in ECP levels (95% credibility interval: 1.24–1.90; Bayesian p = 0.00088). This corresponds to a moderate effect size consistent with biologically plausible differences in eosinophil activation. The VanderWeele E-value was 2.72 (1.91 under the most conservative assumptions), indicating substantial robustness to unmeasured confounding. Of note, the ATT estimator yields a larger effect size than the unweighted estimator because it directly targets the counterfactual question of how ECP levels would differ within the subgroup of HAE patients if these patients did not have HAE. By reweighting control observations to resemble the covariate distribution of HAE patients, the ATT focuses on the clinically relevant comparison for this population.

Because our cohort included a relatively large number of patients with HAE-nC1INH (or HAE type 3), we performed additional analyses stratified by HAE subtype: HAE-C1INH (HAE type 1/2, Supplementary Figure 2), HAE-nC1INH (HAE type 3, Supplementary Figure 3), and mast-cell-mediated angioedema (AAE.URT; Supplementary Figure 4). The raw data for each subgroup is shown in Supplementary Figure 5. The pattern observed in the overall cohort was reproduced in all subgroups. ECP levels were significantly elevated in HAE-C1INH (1.37-fold increase, p=0.01684) and even more pronounced in HAE-nC1INH (1.7-fold increase, p=0.04268). A small but detectable increase was also observed in mast-cell-mediated angioedema (1.15-fold increase, p=0.01712). However, covariate balance in the subgroup analyses was somewhat reduced compared with the main analysis, with standardized mean differences of up to approximately 0.2 for several covariates in each subgroup. This reflects the smaller sample sizes of the individual strata and should be taken into account when interpreting subgroup results. These findings support the internal validity of our results: HAE-nC1INH patients display a similar and even stronger pattern of eosinophil activation than HAE-C1INH patients, whereas the effect in mast-cell–mediated angioedema is minimal, consistent with the pathophysiology of that condition. Notably, the observation in HAE-C1INH – where the diagnosis can be established with high certainty based on C1-INH functional assays – further supports that the ECP elevation is genuinely related to bradykinin-mediated angioedema rather than misclassification with mast-cell-mediated disorders.

We additionally explored potential sex-specific differences in the association between HAE and ECP levels. Owing to the limited number of male HAE patients and controls, these analyses were exploratory. An elevated ECP level associated with HAE was observed in females (1.6-fold increase, p=0.003, Supplementary Figure 6), whereas estimates in males were less precise and did not reach statistical significance (1.28-fold increase, p=0.17308, Supplementary Figure 7). Inspection of the raw data suggested slightly higher baseline ECP levels in male compared with female controls, while ECP levels were similar between males and females among HAE patients (Supplementary Figures 8A, B).

Notably, inverse probability weighting achieved less satisfactory covariate balance in males, as reflected by standardized mean differences exceeding 0.2, which limits the interpretability of male-specific estimates. Because HAE-nC1INH occurred almost exclusively in females, we performed additional sex-stratified analyses restricted to patients with HAE-C1INH only. In this more balanced subgroup, the association between HAE and elevated ECP levels was directionally consistent in both females (1.44-fold increase, p=0.01324, Supplementary Figure 9) and males (1.19-fold increase, p=0.31128, Supplementary Figure 10). Corresponding raw data are shown in Supplementary Figure 11A, B. Taken together, these findings demonstrate that elevated ECP levels in HAE patients are reproducible across all modeling frameworks and remain robust to increasingly conservative assumptions. The consistency of results across multiple analytic perspectives supports both the stability and the credibility of the estimated HAE effect.

Although ECP levels reflect a different biological aspect of eosinophil activity than absolute eosinophil counts, we examined whether eosinophil counts showed a similar increase in HAE patients (Figure 2). The database was primarily designed to extract ECP values from the clinical information system and to include eosinophil counts only when available. Consequently, several eosinophil count values were missing, and one HAE patient lacked eosinophil data (N = 47, with 124 measurements, ESS = 124 with ATT weighting). The number of controls was also smaller (N = 1072, with 1386 measurements, ESS = 347.67 with ATT weighting). The weighted median of absolute eosinophil counts was 0.12/nL in controls and 0.15/nL in HAE cases.

We performed a regression analysis with the same pattern as with the ECP values before (Figure 4). The Love plot demonstrated good balance between HAE case and control measurements, as indicated by a standard mean difference below 0.1 and a satisfactory overlap of propensity score distributions after weighting.

The regression analysis itself indicated that HAE had no measurable effect on eosinophil counts across all model specifications. The slightly increased HAE effect in the Monday stratum further illustrates the utility of the post-stratification approach: it highlights that weekdays affect patient selection, whereas eosinophil counts themselves remain unaffected. A genuine weekday-related effect on eosinophil counts would imply a systematic weekday bias in laboratory measurement quality, which is unlikely given the standardized and automated laboratory procedures.

We performed similar subgroup analyses as for the ECP level here: for HAE-C1INH (HAE type 1/2, Supplementary Figure 12), HAE-nC1INH (HAE type 3, Supplementary Figure 13), and mast-cell-mediated angioedema (AAE-URT, Supplementary Figure 14). The raw data comparison (Supplementary Figure 15) showed a slight increase in absolute eosinophil counts in HAE-nC1INH cases, compared to corresponding controls. However, this is not reflected in the regression analyses, which confirmed no effect of HAE-C1INH, HAE-nC1INH, or mast-cell mediated angioedema on absolute eosinophil counts. It has to be noted that the weighting was less optimal in subgroup analyses, compared to the main analysis.

Similar results were obtained in the sex-stratified analyses, which were performed equivalently to the ECP analyses. Neither females with HAE (Supplementary Figure 16), nor males with HAE (Supplementary Figure 17) showed an increase in absolute eosinophil counts. The corresponding raw data is shown in Supplementary Figures 8C, D. Restriction to patients with HAE type 1 and 2 did not reveal any effect of the disease on absolute eosinophil counts in females (Supplementary Figure 18) or males (Supplementary Figure 19). Likewise, the corresponding raw data is shown in Supplementary Figures 11C, D.