Glucagon-like Peptide 1 Receptor Agonists in Asthma Exacerbations: an Application of High-dimensional Iterative Causal Forest to Identify Subgroups

Glucagon-like Peptide 1 Receptor Agonists in Asthma Exacerbations: an Application of High-dimensional Iterative Causal Forest to Identify Subgroups | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL Pharmacoepidemiology and Drug Safety This is a preprint and has not been peer reviewed. Data may be preliminary. 10 January 2025 V1 Latest version Share on Glucagon-like Peptide 1 Receptor Agonists in Asthma Exacerbations: an Application of High-dimensional Iterative Causal Forest to Identify Subgroups Authors : Tiansheng Wang 0000-0002-0980-8896 [email protected] , Jeanny Wang H , Alan C. Kinlaw , Richard Wyss , Virginia Pate , Zhuoyue Gou , John B. Buse , Corinne Keet A , Michael Kosorok R , and Til Stürmer 0000-0002-9204-7177 Authors Info & Affiliations https://doi.org/10.22541/au.173650335.52689097/v1 Published Pharmacoepidemiology and Drug Safety Version of record Peer review timeline 370 views 260 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract BACKGROUND : Glucagon-like Peptide 1 Receptor Agonists (GLP1RA) may reduce asthma exacerbation (AE) risk, but it is unclear which populations benefit most. Recent pharmacoepidemiologic studies have employed iterative causal forest (iCF), a machine learning (ML) algorithm to identify subgroups with heterogeneous treatment effects. While iCF do not rely on prior knowledge of treatment-variable interactions, it may be constrained by missing or misdefined variables in pharmacepidemiology studies. METHODS: We applied the high-dimensional iterative causal forest (hdiCF), a causal ML algorithm that does not reply on predefined variables, to MarketScan 2016-2020 claims data to identify populations with asthma that might benefit most from GLP1RA in reducing AE risk. We built a GLP1RA vs sulfonylurea new-user cohort with ≥ 1 inpatient or 2 outpatient asthma encounters, excluding patients with non-asthma indications for systemic steroids. Using 599 high-dimensional features from inpatient/outpatient services and pharmacy claims, patients were followed for 6 months from their second prescription. The outcome was acute AE (hospital admission or emergency department visit for asthma). RESULTS: In the overall population, GLP1RA decreased AE risk relative to sulfonylurea: aRD -1.4% (-2.0%, -0.8%). hdiCF identified 3 subgroups based on systemic steroid prescription fills (0, 1, and ≥2): patients with ≥2 systemic steroid prescriptions (GLP1RA: 34 events/1367 individuals; sulfonylurea: 53/1013) benefited most from GLP1RA: aRD -3.8% (-5.3%, -2.2%). CONCLUSIONS: This study demonstrates how automated feature identification can pinpoint clinically relevant subgroups with varying treatment effects. Systemic steroid use, as a proxy for severe asthma, may guide personalized predictions of GLP1RA’s short-term benefits on acute AE. Glucagon-like Peptide 1 Receptor Agonists in Asthma Exacerbations: an Application of High-dimensional Iterative Causal Forest to Identify Subgroups running head: Causal machine learning for subgroup identification Tiansheng Wang, [email protected] Jeanny H Wang, [email protected] Alan C Kinlaw, [email protected] Richard Wyss, [email protected] Virginia Pate, [email protected] Zhuoyue Gou, [email protected] John B Buse, [email protected] Corinne A Keet, corinne_keet@@med.unc.edu Michael R Kosorok, [email protected] Til Stürmer, [email protected] 1 Department of Epidemiology, University of North Carolina Gillings School of Global Public Health, Chapel Hill, NC, USA 2 Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital, Boston, MA, USA 3 Department of Pharmaceutical Outcomes and Policy, University of North Carolina Eshelman School of Pharmacy, Chapel Hill, NC, USA 4 Department of Biostatistics and Epidemiology at the Rutgers School of Public Health 5 Department of Medicine, University of North Carolina School of Medicine, Chapel Hill, NC, USA 6 Department of Pediatrics, University of North Carolina School of Medicine, Chapel Hill, NC, USA 7 Department of Biostatistics, University of North Carolina Gillings School of Global Public Health, Chapel Hill, NC, USA maximum words: 3000 Current word count: 2962 Abstract (maximum 250 words, current word count: 250) BACKGROUND : Glucagon-like Peptide 1 Receptor Agonists (GLP1RA) may reduce asthma exacerbation (AE) risk, but it is unclear which populations benefit most. Recent pharmacoepidemiologic studies have employed iterative causal forest (iCF), a machine learning (ML) algorithm to identify subgroups with heterogeneous treatment effects. While iCF do not rely on prior knowledge of treatment-variable interactions, it may be constrained by missing or misdefined variables in pharmacepidemiology studies. METHODS: We applied the high-dimensional iterative causal forest (hdiCF), a causal ML algorithm that does not reply on predefined variables, to MarketScan 2016-2020 claims data to identify populations with asthma that might benefit most from GLP1RA in reducing AE risk. We built a GLP1RA vs sulfonylurea new-user cohort with ≥ 1 inpatient or 2 outpatient asthma encounters, excluding patients with non-asthma indications for systemic steroids. Using 599 high-dimensional features from inpatient/outpatient services and pharmacy claims, patients were followed for 6 months from their second prescription. The outcome was acute AE (hospital admission or emergency department visit for asthma). RESULTS: In the overall population, GLP1RA decreased AE risk relative to sulfonylurea: aRD -1.4% (-2.0%, -0.8%). hdiCF identified 3 subgroups based on systemic steroid prescription fills (0, 1, and ≥2): patients with ≥2 systemic steroid prescriptions (GLP1RA: 34 events/1367 individuals; sulfonylurea: 53/1013) benefited most from GLP1RA: aRD -3.8% (-5.3%, -2.2%). CONCLUSIONS: This study demonstrates how automated feature identification can pinpoint clinically relevant subgroups with varying treatment effects. Systemic steroid use, as a proxy for severe asthma, may guide personalized predictions of GLP1RA’s short-term benefits on acute AE. Keywords: high-dimensional iterative causal forest, GLP-1 receptor agonist, asthma exacerbation, real-world data, heterogeneous treatment effect, machine learning, precision medicine SUMMARY (up to 5 bullet points, of around 100 words only) • This study applies the high-dimensional iterative causal forest (hdiCF) algorithm to identify subpopulations may benefit more from GLP1 receptor agonists ( GLP1RA) relative to sulfonylureas on asthma exacerbations MarketScan population with type 2 diabetes and asthma. • We observed an overall protective effect of GLP1RAs compared to sulfonylureas and patients with ≥2 systemic steroid prescriptions (reflecting more severe asthma) benefit the most. • By using diagnosis, procedure, and prescription codes as high-dimensional variables, hdiCF uncovers treatment effect modifiers that might be overlooked by predefined variables. • These findings support personalized therapy strategies and highlight the need for confirmatory research in broader populations. PLAIN LANGUAGE SUMMARY (maximum 200 words, current word count: 150) People with type 2 diabetes often also have asthma, which can lead to severe flare-ups or “exacerbations.” Certain diabetes drugs known as GLP1 receptor agonists (GLP1RAs) may help reduce these attacks, but it is unclear which patients benefit the most. In this study, we examined a large U.S. health insurance database (MarketScan) from 2016 to 2020. We compared 8,684 new users of GLP1RAs or sulfonylureas (another diabetes drug) and followed them for six months, tracking severe asthma flare-ups that led to hospital stays or emergency department visits. Overall, patients taking GLP1RAs had a 1.4% lower risk of a serious asthma attack than those taking sulfonylureas. Using a new machine learning method (high-dimensional iterative causal forest), we identified that the greatest benefit appeared in patients who had two or more prescriptions for oral steroids, a sign of more severe asthma—these individuals had a 3.8% lower risk of an asthma attack when on GLP1RAs. By pinpointing who benefits most, our findings may guide doctors to choose GLP1RAs for patients with both diabetes and difficult-to-control asthma, and highlight the potential of personalized treatment approaches. INTRODUCTION Intended and unintended treatment effects of antihyperglycemic drug classes are likely to vary across subpopulations of patients with type 2 diabetes due to heterogeneity of individuals. 1,2 While recent studies suggests that Glucagon-like Peptide 1 Receptor Agonists (GLP1RAs, approved by FDA for treating diabetes or obestiy 1 ) may reduce the risk of asthma exacerbations, it remains unclear which subpopulations derive the greatest benefit from GLP1RA treatment. 3,4 Identifying these patients can improve patient responses, guide randomized studies targeting specific sub-populations, and enable personalized predictions of the short-term benefits of GLP1RAs in reducing asthma exacerbation risk. Machine learning algorithms applied to real-world data have the potential to uncover previously unknown heterogeneity and suggest personalized therapies. In earlier work, we developed the iterative causal forest (iCF), a causal forest 5-7 –based subgrouping algorithm 8 and applied it to claims data and identified patients with ≥2 emergency department (ED) visits—a proxy for severe asthma—as the subgroup with the largest conditional average treatment effect (CATE) for GLP1RA relative to sulfonylurea. While iCF provided valuable insights into previously unknown subpopulations, its reliance on pre-specified covariates may overlook important features and lead to inaccurate subgrouping. To address this limitation, we developed high-dimensional iterative causal forest (hdiCF) algorithm, a causal forest-based, semi-automated subgrouping method that does not require prior knowledge of treatment effect modifiers or predefined variables. 11 This approach may reveal additional plausible heterogeneous treatment effect (HTE) in real-world data and provide novel insights into GLP1RA’s potential impact on asthma exacerbations. In this study, we applied hdiCF to assess the potential HTE of GLP1RA on asthma exacerbations and identify which subpopulations of patients with diabetes and asthma may benefit most from GLP1RA treatment. RESEARCH DESIGN AND METHODS: In this study, we used recently published methods 9 to construct a new-user cohort of GLP1RA versus sulfonylureas in the Merative MarketScan® Research Databases from 2015 to 2020, focusing on healthcare encounters after October 2015 (when ICD-10-CM became standard). We provide a brief overview of these methods here and direct readers to our previous paper 9 for additional details. Data Source We used the MarketScan® Research Databases, a proprietary U.S. claims database containing longitudinal data from younger adults across outpatient, and pharmacy claims, and enrollment files. Study Population We identified MarketScan enrollees aged 18–65 years from October 1, 2015, to December 30, 2020, and used an active-comparator, new-user design to compare new users of GLP1RAs with those of sulfonylureas 12,13 . Sulfonylureas were chosen as the comparator because their mechanism (increased insulin secretion) may affect asthma risk 14 , and they have previously been used as a comparator showing a protective effect of GLP1RAs on asthma exacerbation 2,9 . Restricting the comparison to these two drug classes helps ensure similar disease severity and indication for second-line glucose-lowering therapy. To construct the new-user cohort, we started identifying new use of GLP1RAs or sulfonylureas from October 1, 2016, allowing a one-year baseline assessment period under ICD-10-CM. New use was defined by a first prescription after a 12-month washout with no use of either drug class 12,13 . We required a second prescription within 90 days after the first prescription’s estimated days’ supply ( Figure S1 ). We identified patients with asthma using at least one inpatient or two outpatient claims for asthma during the baseline year 3 . Patients with vocal cord dysfunction, chronic congestive heart failure, COPD or other respiratory diseases, or conditions requiring systemic steroid treatment were excluded ( Tables S1–S3 ). 3,4 Outcome Asthma exacerbation was defined as an inpatient admission or an ED visit for asthma. 3,8 Hospital admissions required a primary asthma diagnosis or a secondary asthma diagnosis plus a primary diagnosis of respiratory symptoms ( Tables S1 ). 3,9,15 Risk Period In the intention-to-treat analysis, patients were followed from cohort entry (second prescription date) until the end of MarketScan enrollment or end of study, for up to 0.5 years, based on GLP1RA adherence patterns in real-world. 16,17 We required cohort entry no later than July 1, 2020, to allow 0.5-year follow-up ( Figure S1 ). Covariates During the 12-month baseline, we obtained high-dimensional, ordinal variables from ICD-10-CM diagnoses, procedures (HCPCS/CPT), and prescriptions (ATC codes), as well as predefined variables (demographics, comorbidities, co-medications, and healthcare utilization). 11 High-dimensional variables plus age and sex were used in hdiCF for subgroup decisions, while predefined variables were used to estimate propensity scores (PS) controlling for treatment decision factors. 18 Obesity, a key confounder, 19 was defined using a validated ICD-10-CM algorithm indicating “overweight, obese, or severely obese” (BMI ≥ 25), which has a positive predictive value of 97.2% and comparable sensitivity/specificity compared to alternative definitions ( Table S4 ). 20 Statistical Analysis We applied the hdiCF algorithm to our original cohort to classify patients into subgroups at baseline. 11 Briefly, hdiCF first identifies up to n (e.g., 200) of the most prevalent codes across various data dimensions, applying a prevalence cutoff (e.g., based on the frequency of these codes over the baseline period ( Figure S2 ). Next, it fits a PS model with these high-dimensional variables, applies PS trimming, and merges low-frequency categories until each level has at least m (e.g., 20) observations. Finally, it executes an iterative causal forest (iCF) to identify stable subgroups. 8 In our study, we used the 200 most prevalent codes in each dimension, requiring a prevalence above 1% (cutoff c =0.01, and below 99%). For ICD-10-CM, we used 3-digit codes; for CPT, 5-digit codes; and for ATC codes, the 3rd-level pharmacological subgroup. We then trained a causal forest model and selected variables with the top 5% variable-importance scores to build each forest, employing 5-fold cross-validation and using 1,000 trees and 100 iterations at each depth. 7,8 Because the focus of hdiCF is on detecting subgroups with heterogeneous treatment effects rather than confounding control, we used our predefined variables (e.g., demographics, comorbidities, co-medications) to estimate the PS by logistic regression within each identified subgroup. We then constructed inverse probability treatment weights (IPTWs), where each individual was weighted by the inverse of the probability of receiving the treatment they actually received (PS for the GLP1RA group; 1 − PS for the sulfonylurea group). These were stabilized by the marginal prevalence of the actual treatment. We assessed the CATE as the IPTW-adjusted risk difference (aRD) ( Figure 1 ). 21 To further refine confounding control, we performed asymmetric PS trimming, excluding patients with PSs below the 0.1th percentile of the GLP1RA group or above the 99.9th percentile of the sulfonylurea group, avoiding extreme treatment allocations contrary to prediction. 22 We evaluated covariate balance using absolute standardized mean differences (ASMDs). 13 In parallel, we estimated the average treatment effect for the entire population as an IPTW-adjusted risk difference. We then conducted a bias analysis to evaluate how much unmeasured confounding would be required to attenuate the observed 0.5-year RD to the null through bias analysis. 23-25 Specifically, we considered two scenarios: First, An unmeasured confounder with an inverse association with the outcome (negative RD) and higher prevalence in the GLP1RA group (positive prevalence difference [PD]). Second, An unmeasured confounder with a positive RD and lower prevalence in the GLP1RA group (negative PD). We estimated the crude RD for the confounder–outcome association in the comparator group and its PD between GLP1RA and sulfonylurea groups, then took their product. We examined the top three confounders with the most negative product, plus known key confounders such as ED visit for asthma 26 and overweight/obesity. 19 We ran four sensitivity analyses based on primary analysis, modifying one parameter at a time: 1) Replacing “overweight/obese/severely obese” with two separate categories, “overweight” and “obese/severely obese.” 20 2) Omitting asymmetric PS trimming; 3) adjusting PS trimming to exclude PSs below the 1st percentile of the GLP1RA group or above the 99th percentile of the comparator group; 4) truncating IPTWs at 10 while omitting PS trimming. 27 All hdiCF subgroup analyses were performed in R (version 4.1), with the algorithm available at https://github.com/tianshengwang/hdiCF. Data management and estimation of CATEs were conducted in SAS (version 9.4). Average treatment effect We identified 8,684 eligible initiators—4,878 GLP1RA initiators and 3,806 sulfonylurea initiators ( Figure S3 ). Table 1 and Table S4 provide the distribution of key predefined variables (identified from previous work 9 ) and full baseline characteristics, respectively. Compared to sulfonylurea users, GLP1RA initiators tended to be younger, female, and have more diabetes-related complications, overweight/obesity diagnoses (64% vs 49%), and more frequent A1C and lipid tests. After PS weighting, covariate balance was adequate (median ASMD: 0.011; Table S5 ). In the PS-trimmed cohort, asthma exacerbation risks were 1.5% for 4,843 GLP1RA initiators versus 2.6% for 3,748 sulfonylurea initiators (aRD: −1.4%; 95% CI: −2.0%, −0.8%). Figure 2 shows how a single unmeasured confounder would have driven the 0.5-year RD observed in our study towards or beyond the null. Such a confounder would need to surpass the confounding bias of known and measured confounders (e.g., baseline ED visit for asthma, overweight/obese). Measured overweight/obese had a RD of 0.2% and a PD of 14.8%, and ED visit for asthma has the highest RD of 8.8% and a PD of -2.3% ( Table S6 ). Conditional average treatment effect Using the n=200 most prevalent codes in each dimension (prevalence trimming (N=7,683) and merging low counts (n<20) into adjacent frequency categories, systemic steroid use was the most important high-dimensional variable ( Figure S4 ). We used minimum-leaf sizes of N/25, N/40, N/50, and N/70 to iteratively grow forests ( Figure S5 ). Models built from depth-3 causal forests had the smallest prediction error (0.025) and were chosen as the final hdiCF solution ( Appendixes 1–7 ). The hdiCF approach identified three subgroups based on systemic steroid prescriptions (≥2, =1, =0). After PS trimming (reducing sample size by sulfonylurea groups showed good overlap ( Figure S6 ), and covariates are well balanced (median ASMD: 0.016–0.022; Tables S4 and S5 ). Subgroup 1 (≥2 systemic steroid prescriptions) contained 2,380 patients, with a higher prevalence of overweight/obesity, bronchitis, smoking, asthma medication use, and ED visits. GLP1RA initiators had 2.5% asthma exacerbation risk vs. 5.2% for sulfonylurea (aRD: −3.8%; 95% CI: −5.3%, −2.2%). Subgroups 2 (1 systemic steroid prescription) and 3 (no systemic steroid prescription) showed no significant benefit (aRDs of 0.1% [95% CI: −1.0%, 1.2%] and −0.3% [95% CI: −0.9%, 0.3%], respectively) ( Figure 3 , Table 2 ). Sensitivity analyses Across four sensitivity analyses, both the average and subgroup-specific treatment effects were generally consistent with the primary findings: 1) splitting “overweight/obese/severely obese” into “overweight” and “obese/severely obese” yielded similar results but residual confounding arose in subgroup 2 (max weighted ASMD=0.175; Table S8 and S9 ). The largest CATE remained −3.9% (95% CI: −5.5%, −2.3%) ( Table S10 ); 2) without PS trimming, subgroups 1 and 2 showed larger ASMDs (0.166 and 0.165), with a reduced largest CATE of −3.4% (95% CI: −4.9%, −1.8%) ( Table S11 ); 3) PS Trimming at 1st vs. 99th Percentiles removed ~8% of patients and reduced residual confounding (max ASMD <0.1), and reduced the largest CATE in subgroup 1 from −3.8% to −2.7% (95% CI: −4.2%, −1.2%) ( Table S12 ); 4) truncating IPTW at 10 decreased the largest CATE in subgroup 1 to −3.3% (95% CI: −4.9%, −1.8%) ( Table S13 ). Clinical characteristics of the PS-trimmed and IPTW-truncated cohorts were similar to the original cohort both overall ( Table S14 ) and by subgroup ( Tables S15–S17 ). Minimum PS values for GLP1RA initiators and sulfonylurea initiators were approximately 0 and 0.07, respectively. Trimming at the 0.1th vs 99.9th and 1st vs 99th percentiles resulted in truncating the left-tail of PS distribution at Compared to trimming at 1st vs 99th, the 0.1th vs 99.9th percentile cutoffs yielded a distribution more similar to the original cohort. DISCUSSION We aimed to use ICD-10 era claims data and the hdiCF machine learning subgrouping algorithm to identify patients with type 2 diabetes and asthma who might derive greater benefit from GLP1RA in preventing asthma exacerbations. Our finding of an overall beneficial effect supports prior results from Foer et al. comparing GLP1RAs and sulfonylureas, as well as our own earlier work with ICD-9/ICD-10 era data. 3,9 The hdiCF algorithm found that patients with ≥2 systemic steroid prescriptions showed the largest CATE. Because ED visits often correlate with systemic steroid use, 28,29 this aligns with previous iCF analyses that flagged patients with ≥2 ED visits as having the highest benefit. 9 To our knowledge, few studies employ high-dimensional features (rather than predefined variables) for HTE detection, 11 making this semi-automated approach particularly useful. Compared to our prior iCF analysis of a 6,084 GLP1RA vs. 11,004 sulfonylurea cohort (2007–2019), 9 the new subgroup defined by ≥2 steroid prescriptions is a better proxy for treatment-effect modifiers than the previously identified ≥2 ED visits. Not only does it show a larger CATE (−3.8% vs. −2.8%), but it also has more precise estimates despite a smaller sample. These subgroups may guide clinicians in identifying patients most likely to benefit and help researchers design future trials focusing on high-benefit groups. 2 Our bias analysis assessed that a single unmeasured confounder would need to be stronger than any of our individual measured confounders (and independent of them) in order to negate the beneficial effect, which strengthens the credibility of our findings of treatment heterogeneity. The magnitude of residual confounding identified in our bias analysis for prior ED visits for asthma, indicated that this confounder may conceal an even greater beneficial effect for GLP1RA and asthma exacerbations. Bias analysis for confounding by overweight/obese indicated that from the direction of measured confounding for overweight/obese, the residual confounding from actual overweight/obese may not reduce the observed beneficial effect (RD*PD>0). Obesity is associated with a higher risk of asthma and asthma exacerbation. 19 The observed low RD (0.002) of overweight/obese in GLPRA group vs. sulfonylurea group could be partly due to measurement error, as claims codes may underreport this condition. 20 We saw large ASMDs in subgroup 2 when splitting overweight/obesity into two categories, suggesting residual confounding or measurement error 20 . Theoretically, the lack of detailed specification of overweight/obese in the binary variable could hinder adequate control of confounding, 30 and our observation may be partly due to measurement error. Notably, PS trimming and IPTW truncation effectively reduced confounding (e.g., subgroups 1 and 2 with large ASMDs in the untrimmed cohort showed small ASMDs after trimming). 22 The similar distribution of clinical features of the PS-trimmed or IPTW-truncated cohorts and the original cohort further shows that the trimmed/truncated cohorts closely resemble the original population. However, heavier PS trimming decreased the largest CATE in the highest-risk subgroup, possibly because the tails of the PS distribution hold individuals with more extreme baseline risks. Future studies are needed to evaluate the impact of PS trimming on identifying and assessing HTE. Systemic steroid use serves as a proxy for asthma severity and control, with a previous exacerbation being the strongest predictor for future events. 19, 26 It is highly correlated with baseline ED visits, with 5% of patients with ≥2 systemic steroid prescriptions having baseline ED visits for asthma, compared to less than 0.5% in those with <2 prescriptions. Since asthma severity could not be directly measured in the available claims data, and decision trees tend to divide populations based on proxies (indirect measures) of certain features, 31 the identified subgroups should be interpreted cautiously. The largest CATE was observed in the ≥2 systemic steroid subgroup, likely because this group is at the highest risk for asthma exacerbations, 3,4,9 and GLP1RA has been shown to have a beneficial effect on asthma outcomes. 32-35 Notably, these findings align with Foer et al.’s study, which showed that patients with moderate/severe asthma (receiving inhaled corticosteroids, either alone or with other controllers or biologics) experienced greater benefit from GLP1RA compared to sulfonylurea (sample size: 211 vs. 1,052, incidence rate ratio 0.45 [0.27-0.75]) than the overall population during the 0.5-year follow-up period. 3 This consistency supports the reliability of the hdiCF. CATEs differ across subgroups defined by the number of systemic steroid prescriptions, 28 demonstrating clinically meaningful differences in risk reduction. 36 Third, sulfonylureas might not perfectly match GLP1RAs in baseline characteristics, especially if patients on costlier GLP1RAs receive more intensive care. Finally, our younger, commercially insured population limits generalizability to older adults or lower socioeconomic groups. This study has several limitations. First, residual confounding that varies by subgroup could create spurious heterogeneity. For example, we were unable to account for Forced expiratory volume in one second (FEV1), a key measure of lung function, as this data is often unavailable in claims. 4, 37 We cannot rule out the possibility that hdiCF may identify subgroups where the magnitude of effect is confounded by strong bias. Further research is needed to address the challenge of distinguishing true heterogeneity from hidden bias when estimating CATE. Second, while using hdiCF with high-dimensional variables allows us to capture treatment effect modifiers that predefined variables might miss, we relied on predefined variables to control for confounding within each subgroup for simplicity, assuming that all confounders could be prespecified. However, misidentification or misdefinition of these variables could still bias CATE estimates. Although sensitivity and bias analyses provide insights into unmeasured confounding, they cannot fully resolve these concerns. Future research should explore the validity of using high-dimensional variables throughout the entire process of identifying and assessing HTE. Third, we used sulfonylurea as the comparator, but residual confounding may still exist, particularly if patients prescribed the more expensive GLP1RAs also receive higher-quality care. Lastly, our younger, commercially insured population limits generalizability to older adults or lower socioeconomic groups. Our study demonstrates that an semi-automated machine learning subgrouping algorithm, hdiCF, can identify clinically relevant subgroups with varying treatment effects. Consistent with previous findings, we observed that patients with at least two systemic steroid prescriptions (indicating more severe asthma) benefit most from GLP1RA, as identified by hdiCF. Clinicians managing patients with both type 2 diabetes and asthma may consider the number of systemic steroid prescriptions to tailor treatment and maximize the benefit of GLP1RA in reducing asthma exacerbation risk. Further validation of hdiCF is needed, and future investigations should explore methods for estimating CATE entirely with high-dimensional variables. Table 1. Distributions of selected important variables in overall population and subgroup 1 before and after PS weighting in patients with baseline asthma and without COPD in primary analysis*. GLP1RA N=4,843 SU N=3,748 ASMD§ Weighted GLP1RA N=3,720 Weighted SU N=5,016 Weighted ASMD§ GLP1RA N=1,367 SU N=1,013 ASMD¶ Weighted GLP1RA N=1,006 Weighted SU N=1,411 Weighted ASMD# Demographic characteristics Age group 1: 18≤age≤30 153 (3.2) 185 (4.9) 0.09 146 (3.9) 210 (4.2) 0.013 39 (2.9) 38 (3.8) 0.05 33 (3.2) 51 (3.6) 0.022 2: 30<age≤40 630 (13.0) 590 (15.7) 0.078 501 (13.5) 662 (13.2) 0.008 200 (14.6) 150 (14.8) 0.005 141 (14.0) 187 (13.2) 0.023 3: 40<age≤50 1,457 (30.1) 902 (24.1) 0.136 1,030 (27.7) 1,394 (27.8) 0.002 432 (31.6) 241 (23.8) 0.175 292 (29.0) 392 (27.7) 0.028 4: 50<age≤60 1,942 (40.1) 1,500 (40.0) 0.002 1,495 (40.2) 2,014 (40.1) 0.001 528 (38.6) 443 (43.7) 0.104 400 (39.8) 596 (42.3) 0.05 5: 60<age≤65 661 (13.6) 571 (15.2) 0.045 547 (14.7) 736 (14.7) 0.001 168 (12.3) 141 (13.9) 0.048 140 (13.9) 185 (13.1) 0.024 Sex, Males††† 1,300 (26.8) 1,267 (33.8) 0.152 1,111 (29.9) 1,438 (28.7) 0.027 292 (21.4) 285 (28.1) 0.157 246 (24.4) 336 (23.8) 0.014 Cardiovascular disorders Arrhythmia disorders 259 (5.3) 191 (5.1) 0.011 195 (5.2) 237 (4.7) 0.023 83 (6.1) 59 (5.8) 0.01 62 (6.2) 93 (6.6) 0.015 Overweight, obese or severely obese ††† 3,100 (64.0) 1,843 (49.2) 0.303 2,153 (57.9) 2,949 (58.8) 0.019 936 (68.5) 558 (55.1) 0.278 636 (63.2) 870 (61.7) 0.033 Respiratory disorders Bronchitis 487 (10.1) 348 (9.3) 0.026 360 (9.7) 456 (9.1) 0.02 236 (17.3) 159 (15.7) 0.042 166 (16.4) 217 (15.4) 0.029 Smoking and smoking cessation 278 (5.7) 224 (6.0) 0.01 218 (5.9) 270 (5.4) 0.021 97 (7.1) 79 (7.8) 0.027 71 (7.0) 88 (6.2) 0.033 Other comorbidity Electrolyte disorder 425 (8.8) 360 (9.6) 0.029 348 (9.4) 459 (9.2) 0.007 147 (10.8) 129 (12.7) 0.062 120 (11.9) 166 (11.7) 0.006 Iron deficiency anemia 632 (13.0) 418 (11.2) 0.058 445 (12.0) 608 (12.1) 0.005 197 (14.4) 127 (12.5) 0.055 127 (12.7) 162 (11.5) 0.036 Depression 1,017 (21.0) 631 (16.8) 0.106 709 (19.1) 995 (19.8) 0.02 331 (24.2) 210 (20.7) 0.084 228 (22.7) 341 (24.2) 0.035 Obstructive sleep apnea 1,472 (30.4) 835 (22.3) 0.185 1,019 (27.4) 1,417 (28.3) 0.019 485 (35.5) 279 (27.5) 0.171 330 (32.8) 463 (32.8) 0.001 Hypothyroidism 1,152 (23.8) 588 (15.7) 0.205 764 (20.6) 1,100 (21.9) 0.034 370 (27.1) 166 (16.4) 0.261 231 (23.0) 375 (26.5) 0.082 Medications for Asthma or COPD No. of prescriptions for ICS + LABA product‡ 0 2,864 (59.1) 2,464 (65.7) 0.137 2,294 (61.7) 3,095 (61.7) 0.001 673 (49.2) 529 (52.2) 0.06 508 (50.5) 730 (51.7) 0.024 1~5 744 (15.4) 455 (12.1) 0.094 525 (14.1) 711 (14.2) 0.001 273 (20.0) 165 (16.3) 0.096 186 (18.5) 251 (17.8) 0.018 6~10 659 (13.6) 397 (10.6) 0.093 464 (12.5) 626 (12.5) 0 222 (16.2) 139 (13.7) 0.071 153 (15.2) 205 (14.5) 0.02 ≥ 11 576 (11.9) 432 (11.5) 0.011 436 (11.7) 584 (11.6) 0.002 199 (14.6) 180 (17.8) 0.087 159 (15.8) 226 (16.0) 0.005 Inhaled steroid 0 4,171 (86.1) 3,200 (85.4) 0.021 3,197 (86.0) 4,357 (86.9) 0.026 1,112 (81.3) 829 (81.8) 0.013 825 (81.9) 1,179 (83.5) 0.042 1~5 466 (9.6) 356 (9.5) 0.004 343 (9.2) 418 (8.3) 0.031 183 (13.4) 118 (11.6) 0.053 123 (12.2) 151 (10.7) 0.048 ≥ 6 206 (4.3) 192 (5.1) 0.041 179 (4.8) 241 (4.8) 0.001 72 (5.3) 66 (6.5) 0.053 59 (5.8) 82 (5.8) 0.002 Ipratropium 398 (8.2) 246 (6.6) 0.063 291 (7.8) 392 (7.8) 0 240 (17.6) 141 (13.9) 0.1 164 (16.3) 201 (14.2) 0.057 No. of prescriptions for SABA 0 1,315 (27.2) 974 (26.0) 0.026 984 (26.5) 1,341 (26.7) 0.006 248 (18.1) 174 (17.2) 0.025 186 (18.5) 250 (17.7) 0.021 1~5 1,644 (33.9) 1,173 (31.3) 0.057 1,206 (32.4) 1,609 (32.1) 0.008 408 (29.8) 246 (24.3) 0.125 277 (27.5) 397 (28.2) 0.014 6~10 975 (20.1) 825 (22.0) 0.046 790 (21.2) 1,106 (22.1) 0.02 318 (23.3) 257 (25.4) 0.049 255 (25.3) 365 (25.9) 0.013 ≥ 11 909 (18.8) 776 (20.7) 0.049 740 (19.9) 960 (19.1) 0.019 393 (28.7) 336 (33.2) 0.096 288 (28.6) 399 (28.3) 0.008 No. of prescriptions Albuterol-Ipratropium 328 (6.8) 199 (5.3) 0.061 240 (6.5) 327 (6.5) 0.003 187 (13.7) 114 (11.3) 0.073 133 (13.2) 168 (11.9) 0.039 Systemic steroid** 0 3,648 (75.3) 2,982 (79.6) 0.102 2,857 (76.8) 3,821 (76.2) 0.014 608 (44.5) 543 (53.6) 0.183 476 (47.3) 681 (48.2) 0.019 1 80 (1.7) 36 (1.0) 0.061 50 (1.3) 67 (1.3) 0.001 36 (2.6) 17 (1.7) 0.066 22 (2.1) 33 (2.3) 0.012 ≥ 2 1,115 (23.0) 730 (19.5) 0.087 813 (21.9) 1,127 (22.5) 0.015 723 (52.9) 453 (44.7) 0.164 509 (50.6) 698 (49.4) 0.023 History of medications use Insulin 1,359 (28.1) 325 (8.7) 0.517 743 (20.0) 1,139 (22.7) 0.067 324 (23.7) 107 (10.6) 0.354 186 (18.5) 289 (20.5) 0.05 Aspirin 293 (6.0) 197 (5.3) 0.034 214 (5.8) 300 (6.0) 0.009 93 (6.8) 65 (6.4) 0.016 68 (6.8) 100 (7.1) 0.013 Estrogen 451 (9.3) 303 (8.1) 0.044 336 (9.0) 449 (9.0) 0.003 147 (10.8) 94 (9.3) 0.049 110 (10.9) 171 (12.1) 0.037 Measures of healthcare utilization N of hospital admissions 0 4,245 (87.7) 3,220 (85.9) 0.051 3,233 (86.9) 4,338 (86.5) 0.013 1,152 (84.3) 851 (84.0) 0.007 847 (84.2) 1,201 (85.1) 0.026 1 470 (9.7) 414 (11.0) 0.044 390 (10.5) 534 (10.7) 0.005 168 (12.3) 140 (13.8) 0.045 130 (12.9) 183 (12.9) 0 ≥ 2 128 (2.6) 114 (3.0) 0.024 97 (2.6) 144 (2.9) 0.016 47 (3.4) 22 (2.2) 0.077 29 (2.9) 28 (2.0) 0.062 N of emergency department visits 0 2,887 (59.6) 2,169 (57.9) 0.035 2,192 (58.9) 3,032 (60.5) 0.031 657 (48.1) 439 (43.3) 0.095 468 (46.5) 688 (48.7) 0.045 1 945 (19.5) 705 (18.8) 0.018 704 (18.9) 909 (18.1) 0.021 302 (22.1) 225 (22.2) 0.003 221 (22.0) 296 (20.9) 0.025 ≥ 2 1,011 (20.9) 874 (23.3) 0.059 823 (22.1) 1,075 (21.4) 0.017 408 (29.8) 349 (34.5) 0.099 317 (31.5) 428 (30.3) 0.026 N of emergency department visits for asthma 0 4,591 (94.8) 3,468 (92.5) 0.093 3,503 (94.2) 4,730 (94.3) 0.005 1,214 (88.8) 841 (83.0) 0.167 880 (87.5) 1,243 (88.1) 0.019 1 168 (3.5) 192 (5.1) 0.082 153 (4.1) 212 (4.2) 0.006 96 (7.0) 115 (11.4) 0.15 85 (8.4) 119 (8.4) 0.001 ≥ 2 84 (1.7) 88 (2.3) 0.043 64 (1.7) 74 (1.5) 0.019 57 (4.2) 57 (5.6) 0.068 41 (4.1) 49 (3.5) 0.032 *Data are given as number (percentage), unless otherwise indicated. All variables (except for the number of days between first and second prescriptions) are measured 12 months before the first prescription date. Only important predefined variables (with a variable importance value greater than the mean) identified in our previous study 9 is shown in this table. ASMD, absolute standardized mean differences; ACEI indicates angiotensin-converting enzyme inhibitor; ARB, angiotensin receptor blocker; BB, β blocker; CCB, calcium channel blocker; CKD, chronic kidney disease; DBP, diastolic blood pressure; DPP-4i, dipeptidyl peptidase-4 inhibitor; GLP1RA, glucagon-like peptide-1 receptor agonists; A1C, hemoglobin A1C; MI, myocardial infarction; SGLT2i, sodium-glucose cotransporter 2 inhibitors; ICS = inhaled steroid; LABA=long-acting b-agonist; SABA=short-acting b2-agonist.NA, not applicable. ††Selected important variables by raw causal forest for the GLP1RA vs SU cohort using 2007 to 2019 in our previous study; †††sex and overweight or obesity are not selected as important variables. ‡including ICS/LABA combo product or using both ICS and LABA simultaneously. A full list of variables are shown in Table S5 & S6 , for all variables after PS weighting: § (0, 0.517), median 0.048, mean 0.072; || range: (0, 0.067), median 0.011, mean 0.014; ¶ range: (0.001, 0.354), median 0.065, mean 0.077; # range: (0, 0.093), median 0.022, mean 0.024. The defined variable for systemic steroid** prescriptions includes hydrocortisone (H02AB09), prednisone (H02AB06), methylprednisolone (H02AB04), methylprednisolone-combo (H02BX01), prednisolone (H02AB06), and dexamethasone (H02AB02). Please note that this definition includes methylprednisolone-combo (H02BX01) in addition to other steroids coded under (H02A). This contributes to the small number of patients, which may not align with the subgroup definition. For example, in Subgroup 1 (patients with ≥2 systemic steroid prescriptions), 36 GLP1RA initiators and 17 sulfonylurea initiators had only 1 systemic steroid prescription. Table 2. Crude and adjusted risk differences for asthma exacerbation associated with use of GLP1RA in population by initial treatment analysis in a maximum of 6-month follow-up by subgroup decisions from hdiCF algorithm in primary analysis*. Overall population GLP1RA 4,843 2,239 72 3.2 (2.4 to 4.4) 1.5 12.0 -1.1 (-1.7 to -0.5) -1.4 (-2.0 to -0.8) SU 3,748 1,699 98 5.8 (4.2 to 7.8) 2.6 13.6 Subgroup 1 : ≥ 2 systemic steroid prescriptions GLP1RA 1,367 624 34 5.4 (3.7 to 8.0) 2.5 12.5 -2.7 (-4.3 to -1.1) -3.8 (-5.3 to -2.2) SU 1,013 455 53 11.6 (7.9 to 17.2) 5.2 11.6 Subgroup 2 : 1 systemic steroid prescription GLP1RA 1,239 573 16 2.8 (1.5 to 5.1) 1.3 12.4 -0.5 (-1.6 to 0.6) 0.1 (-1.0 to 1.2) SU 882 400 16 4.0 (1.7 to 9.5) 1.8 15.5 Subgroup 3 : No systemic steroid prescriptions GLP1RA 2,164 1,010 18 1.8 (0.9 to 3.5) 0.8 11.6 -0.3 (-0.9 to 0.3) -0.3 (-0.9 to 0.3) SU 1,767 810 20 2.5 (1.6 to 3.9) 1.1 13.5 N, number; IPTW, inverse probability treatment weight. *All patients were required to enter the cohort no later than July 1, 2020 so that we could potentially follow patients for 0.5-year (study ends on Dec 31, 2020). Patients were censored for insurance disenrollment. PS trimming was applied in each identified subgroup. FIGURE LEGEND Figure 1. Applying hdiCF algorithm to identify subgroups and assess heterogeneous treatment effects. RD, risk difference. CATE, conditional average treatment effect (subgroup-specific treatment effect). The hdiCF algorithm utilizes high-dimensional variables to identify subgroups with heterogeneous treatment effects. For simplicity, in each hdiCF identified subgroup, predefined variables are used to predict PS by logistic regression and assess CATE by inverse probability treatment weight adjusted RD (aRD). Figure 2. Bias analyses for unmeasured confounding. RD, risk difference; PD, prevalence; ED, emergency department; Index year, the year initiated GLP1RA or comparator drug. The shaded area indicates the strength of confounding implied if the true RD is null or harmful. The dotted line indicates the strength of confounding implied if true RD is -1.0% in GLP1RA vs sulfonylureas cohort. Negative RD suggests a beneficial effect in the untreated cohort. Negative PD suggests a higher frequency in the GLP1RA group. The label of circle shape (●) indicates the observed RD and PD of a confounder displayed on the absolute scale X-axis and Y-axis. Note, only the following two scenarios involving a confounder can nullify our observed beneficial effect (negative RD): 1) the confounder has a negative RD (beneficial effect) and a positive PD (higher frequency in the GLP1RA group); 2) the confounder has a positive RD (harmful effect) and a negative PD (lower frequency in the GLP1RA group). Figure 3 . Subgroup decisions from hdiCF algorithm based on baseline high-dimensional features for GLP1RA vs sulfonylureas cohort in primary analysis. SU, sulfonylureas; aRD (%), adjusted risk difference (%) by inverse probability treatment weight; Negative values indicate decreased risk of exacerbation (benefit from GLP1RA), whereas positive values indicate increased risk of exacerbation (harm from GLP1RA). The sample size, event, and aRD are for PS trimmed population within each subgroup. Treatment effect is additionally detailed for the subpopulation before splitting in the dotted box. REFERENCE: 1. American Diabetes Association Professional Practice Committee; 9. Pharmacologic Approaches to Glycemic Treatment: Standards of Care in Diabetes—2024. Diabetes Care 2024; 47 (Supplement_1): S158–S178. https://doi.org/10.2337/dc24-S009 2. Chung WK, Erion K, Florez JC, Hattersley AT, Hivert MF, Lee CG, McCarthy MI, Nolan JJ, Norris JM, Pearson ER, Philipson L. Precision medicine in diabetes: a consensus report from the American Diabetes Association (ADA) and the European Association for the Study of Diabetes (EASD). Diabetes care 2020;43(7):1617-35. 3. Foer D, Beeler PE, Cui J, Karlson EW, Bates DW, Cahill KN. Asthma Exacerbations in Patients with Type 2 Diabetes and Asthma on Glucagon-like Peptide-1 Receptor Agonists. Am J Respir Crit Care Med 2021;203(7):831-840. 4. Albogami Y, Cusi K, Daniels MJ, Wei YJ, Winterstein AG. Glucagon-Like Peptide 1 Receptor Agonists and Chronic Lower Respiratory Disease Exacerbations Among Patients With Type 2 Diabetes. Diabetes Care 2021; 44(6):1344-1352. 5. Wager S, Athey S. Estimation and inference of Heterogeneous Treatment Effects using random forests. J AM STAT ASSOC 2018 . 113 (523), 1228-1242 6. Athey S, Imbens G. Recursive partitioning for heterogeneous causal effects. PNAS 2016; 113; 7353-60 7. Athey S, Wager S. Estimating treatment effects with causal forests: An application. Observational studies 2019;5(2):37-51. 8. Wang T, Keil AP, Kim S, Wyss R, Htoo PT, Funk MJ, Buse JB, Kosorok MR, Stürmer T. Iterative Causal Forest: A Novel Algorithm for Subgroup Identification. Am J Epidemiol 2024;193(5):764-776. doi: 10.1093/aje/kwad219. 9. Wang T, Keil AP, Buse JB, Keet C, Kim S, Wyss R, Pate V, Jonsson-Funk M, Pratley RE, Kvist K, Kosorok MR, Stürmer T. Glucagon-like Peptide 1 Receptor Agonists and Asthma Exacerbations: Which Patients Benefit Most? Ann Am Thorac Soc 2024;. doi: 10.1513/AnnalsATS.202309-836OC. Epub ahead of print. PMID: 39012183. 10. Schneeweiss S, Rassen JA, Glynn RJ, Avorn J, Mogun H, Brookhart MA. High-dimensional propensity score adjustment in studies of treatment effects using health care claims data. Epidemiology 2009; Jul;20(4):512-22. doi: 10.1097/EDE.0b013e3181a663cc. Erratum in: Epidemiology 2018; 29(6):e63-e64. doi: 10.1097/EDE.0000000000000886. 11. Wang T, Pate V, Wyss R, Buse JB, Kosorok MR, Stürmer T. High-dimensional Iterative Causal Forest (hdiCF): a Novel Algorithm for Subgroup Identification in Claims Data. Am J Epidemiol 2024; https://doi.org/10.1093/aje/kwae322. 12. Lund JL, Richardson DB, Stürmer T. The active comparator, new user study design in pharmacoepidemiology: historical foundations and contemporary application. Curr Epidemiol Rep 2015;2(4):221-228. 13. Stürmer T, Wang T, Golightly YM, Keil A, Lund JL, Jonsson Funk M. Methodological considerations when analysing and interpreting real-world data. Rheumatology 2020;59(1):14-25. 14. Peters MC, Schiebler ML, Cardet JC, et al. The Impact of Insulin Resistance on Loss of Lung Function and Response to Treatment in Asthma. Am J Respir Crit Care Med 2022;206(9):1096-1106. 15. Altawalbeh SM, Thorpe CT, Zgibor JC, Kane-Gill S, Kang Y, Thorpe JM. Antileukotriene Agents Versus Long-Acting Beta-Agonists in Older Adults with Persistent Asthma: A Comparison of Add-On Therapies. J Am Geriatr Soc 2016 Aug;64(8):1592-600. 16. Divino V, DeKoven M, Hallinan S, Varol N, Wirta SB, Lee WC, et al. Glucagon-like peptide-1 receptor agonist treatment patterns among type 2 diabetes patients in six European countries. Diabetes Ther 2014;5:499–520. 17. Johnston SS, Nguyen H, Felber E, Cappell K, Nelson JK, Chu BC, et al. Retrospective study of adherence to glucagon-like peptide-1 receptor agonist therapy in patients with type 2 diabetes mellitus in the United States. Adv Ther 2014;31:1119–1133. 18. Webster-Clark M, Stürmer T, Wang T, Man K, Marinac-Dabic D, Rothman KJ, Ellis AR, Gokhale M, Lunt M, Girman C, Glynn RJ. Using PSs to estimate effects of treatment initiation decisions: State of the science. Stat Med 2021 Mar 30;40(7):1718-1735. 19. Miethe S, Karsonova A, Karaulov A, Renz H. Obesity and asthma. J Allergy Clin Immunol 2020;146:685–693 20. Suissa K, Schneeweiss S, Lin KJ,Brill G, Kim SC, Patorno E. Validation of obesity-related diagnosis codes in claims data. Diabetes Obes Metab 2021;23(12):2623-2631 21. Murray EJ, Caniglia EC, Swanson SA, Hernández-Díaz S, Hernán MA. Patients and investigators prefer measures of absolute risk in subgroups for pragmatic randomized trials. J Clin Epidemiol 2018;103:10-21 22. Stürmer T, Rothman KJ, Avorn J, Glynn RJ. Treatment effects in the presence of unmeasured confounding: dealing with observations in the tails of the PS distribution – a simulation study. Am J Epidemiol 2010;172:843-54. 23. Miller M, Swanson SA, Azrael D, Pate V, Stürmer T. Antidepressant dose, age, and the risk of deliberate self-harm. JAMA internal medicine 2014; 174(6):899-909. 24. Schneeweiss S. Sensitivity analysis and external adjustment for unmeasured confounders in epidemiologic database studies of therapeutics. Pharmacoepidemiology and drug safety 2006; 15(5):291-303. 25. VanderWeele TJ, Arah OA. Bias formulas for sensitivity analysis of unmeasured confounding for general outcomes, treatments, and confounders. Epidemiology 2011; 22(1):42-52. 26. Miller MK, Lee JH, Miller DP, Wenzel SE, TENOR Study Group. Recent asthma exacerbations: a key predictor of future exacerbations. Respiratory medicine 2007; 101(3):481-9. 27. Stürmer T, Webster-Clark M, Lund JL, Wyss R, Ellis AR, Lunt M, Rothman KJ, Glynn RJ. Propensity Score Weighting and Trimming Strategies for Reducing Variance and Bias of Treatment Effect Estimates: A Simulation Study. Am J Epidemiol 2021; 190(8):1659-1670. 28. Chipps BE, Murphy KR, Oppenheimer J. 2020 NAEPP guidelines update and GINA 2021—asthma care differences, overlap, and challenges . J Allergy Clin Immunol Pract 2022; Jan 1;10(1):S19-30. 29. Lefebvre P, Duh MS, Lafeuille MH, Gozalo L, Desai U, Robitaille MN, Albers F, Yancey S, Ortega H, Forshag M, Lin X, Dalal AA. Acute and chronic systemic corticosteroid-related complications in patients with severe asthma. J Allergy Clin Immunol 2015; 136(6):1488-1495. doi: 10.1016/j.jaci.2015.07.046. Epub 2015 Sep 26. PMID: 26414880. 30. Lash, T. L., VanderWeele, T. J., Haneuse, S., & Rothman, K. J. (2021). Measurement and measurement error. In Modern epidemiology (4th ed., pp. 287–314). Wolters Kluwer. 31. Hastie T, Tibshirani R, Friedman JH, Friedman JH. The elements of statistical learning: data mining, inference, and prediction. New York: springer; 2009 Aug. 32. Ramsahai JM, Hansbro PM, Wark PA. Mechanisms and management of asthma exacerbations. Am J Respir Crit Care Med 2019; 199(4):423-32. 33. Viby NE, Isidor MS, Buggeskov KB, Poulsen SS, Hansen JB, Kissow H. Glucagon-like peptide-1 (GLP-1) reduces mortality and improves lung function in a model of experimental obstructive lung disease in female mice. Endocrinology 2013;154:4503–4511. 34. Toki S, Goleniewska K, Reiss S, Zhang J, Bloodworth MH, Stier MT, et al. Glucagon-like peptide 1 signaling inhibits allergen-induced lung IL-33 release and reduces group 2 innate lymphoid cell cytokine production in vivo. J Allergy Clin Immunol 2018;142:1515–1528.e8. 35. Hur J, Kang JY, Kim YK, Lee SY, Lee HY. Glucagon-like peptide 1 receptor (GLP-1R) agonist relieved asthmatic airway inflammation via suppression of NLRP3 inflammasome activation in obese asthma mice model. Pulm Pharmacol Ther 2021;67:102003. 36. Bonini M, Di Paolo M, Bagnasco D, Baiardini I, Braido F, Caminati M, Carpagnano E, Contoli M, Corsico A, Del Giacco S, Heffler E. Minimal clinically important difference for asthma endpoints: an expert consensus report. European Respiratory Review 2020; 29(156):190137. 37. Foer D, Cahill KN; Comment on Albogami et al. Glucagon-Like Peptide 1 Receptor Agonists and Chronic Lower Respiratory Disease Exacerbations Among Patients With Type 2 Diabetes. Diabetes Care 2021;44:1344–1352. Diabetes Care 2021; 44 (8): e165–e166 Information & Authors Information Version history V1 Version 1 10 January 2025 Peer review timeline Published Pharmacoepidemiology and Drug Safety Version of Record 27 Jul 2025 Published Copyright This work is licensed under a Non Exclusive No Reuse License. Collection Pharmacoepidemiology and Drug Safety Keywords asthma exacerbation glp-1 receptor agonist heterogeneous treatment effect high-dimensional iterative causal forest machine learning precision medicine real-world data Authors Affiliations Tiansheng Wang 0000-0002-0980-8896 [email protected] The University of North Carolina at Chapel Hill Gillings School of Global Public Health View all articles by this author Jeanny Wang H The University of North Carolina at Chapel Hill Gillings School of Global Public Health View all articles by this author Alan C. Kinlaw The University of North Carolina at Chapel Hill Division of Pharmaceutical Outcomes and Policy View all articles by this author Richard Wyss Brigham and Women's Hospital Division of Pharmacoepidemiology and Pharmacoeconomics View all articles by this author Virginia Pate The University of North Carolina at Chapel Hill Gillings School of Global Public Health View all articles by this author Zhuoyue Gou Rutgers School of Public Health Department of Biostatistics and Epidemiology View all articles by this author John B. Buse The University of North Carolina at Chapel Hill Department of Medicine View all articles by this author Corinne Keet A The University of North Carolina at Chapel Hill Department of Pediatrics View all articles by this author Michael Kosorok R The University of North Carolina at Chapel Hill Department of Biostatistics View all articles by this author Til Stürmer 0000-0002-9204-7177 The University of North Carolina at Chapel Hill Gillings School of Global Public Health View all articles by this author Metrics & Citations Metrics Article Usage 370 views 260 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Tiansheng Wang, Jeanny Wang H, Alan C. Kinlaw, et al. Glucagon-like Peptide 1 Receptor Agonists in Asthma Exacerbations: an Application of High-dimensional Iterative Causal Forest to Identify Subgroups. Authorea . 10 January 2025. DOI: https://doi.org/10.22541/au.173650335.52689097/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.173650335.52689097/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a00df78a0ba5e2c5',t:'MTc3OTY0MzE4Mw=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();