Introduction

Asthma is a chronic disease characterized by inflammation of the lower airways. It expresses itself by recurring exacerbations and symptoms like bronchospasms, dyspnea, mucus hypersecretion, cough, wheezing, and/or tightness of the chest.1,2 Comorbidities can aggravate asthma, influencing treatment needed to control symptoms or leading to under- or overtreatment when mimicking asthma-like symptoms.3 For example, features similar to asthma can also occur due to certain medications (eg, aspirin or beta blockers) or due to dysfunctional breathing caused by, eg, cardiovascular disease (CVD) or anxiety, and upper airway inflammation can be influenced by other systemic inflammatory disorders, eg, rheumatoid arthritis, atopic dermatitis, or Graves’ disease.2 The Global Initiative for Asthma (GINA) acknowledges the increased risk of multimorbidity in worsening asthma control, even if only a few asthma symptoms are present, and has developed a stepwise approach for the management of asthma.4 However, little is known about how the presence of comorbidities affects asthma treatment trajectories in clinical practice. Information on subgroup differences in treatment trajectories, and more specifically, switching between medications, can inform personalized treatment and help to reduce treatment delays by identifying those patients who could benefit from earlier intervention and/or closer monitoring. Moreover, further understanding of the impact of certain comorbidities could guide revisions to management guidelines, with a possibility for stratified recommendations for patients with co-existing health conditions. In the long run, optimizing the management of asthma has the potential to reduce healthcare costs by minimizing complications, emergency department visits, and hospitalizations associated with poor asthma control.5–7 Although some previous studies looked into switches between asthma treatment steps,8,9 to the best of our knowledge, no other study has looked into switching behavior among comorbidity subgroups of adults initiating general asthma treatment.

Treatment of asthma is a dynamic process, entailing multiple treatment steps with the possibility of switching up and down between these steps based on (lack of) disease control. Adequate statistical machinery is needed to model such dynamic regimes over time.10 Although the g-formula was developed for counterfactual causal inference in settings of dynamic treatment regimes, its potential goes beyond causal estimation.10–12 It can produce high-dimensional sets of predictions of an outcome of interest given a set of predictors. In contrast, static regression techniques, such as Cox regression, predict a single outcome over a certain period or at a specific timepoint. The g-formula can utilize regression to predict high-dimensional, longitudinal, treatment trajectories.13,14 The g-formula allows for effects to cascade over time and correct for time-varying-confounding, by projecting previous predictions onto the next one.15,16 Other models, like multi-state or decision tree models, may rely on the Markovian assumption, assume that the probability of moving from one state to another is constant over time, or have to deal with exponential growth when the number of discrete states or time points increases.17 The parametric g-formula is not bound by these assumptions and can more easily deal with high-dimensionality of data, thus offers more flexibility in these regards. Although these assumptions can likely be relaxed using machine learning approaches, such techniques have yet been insufficiently explored by previous research on this topic. Furthermore, the g-formula can provide marginal effect estimates that can be generalized to the population, producing easy-to-interpret results answering “what if?” questions, even in the presence of complex interactions.11–13 Overall, the parametric g-formula can be used to model dynamic treatment regimes and hence might better approximate real-world asthma treatment trajectories as compared to traditional methods.

This study aims to investigate whether the parametric g-formula could be used to approximate empirical asthma treatment trajectories over time and aims to predict subgroup differences in switching behavior within asthma treatment trajectories.

Materials and Methods Study Design and Setting

This observational, longitudinal cohort study used dispensing data from community pharmacies in the Netherlands, between January 1st 1994 and December 31st 2021. Follow-up consisted of 720 days from initiation of inhaled asthma medication.

Study Population

Individuals aged between 16 and 45 years old initiating any of the following medications were included:

– short- or long-acting selective β2-adrenoreceptor agonists (SABA/LABA): salbutamol, terbutaline, salmeterol, or formoterol – inhaled corticosteroids (ICS): beclomethasone, budesonide, fluticasone, or ciclesonide – fixed-dose combination of ICS and LABA (ICS-LABA): salmeterol-fluticasone, formoterol-budesonide, formoterol-beclomethasone, vilanterol-fluticasone, formoterol-fluticasone, or salmeterol-budesonide.

Individuals had to be present in the dataset at least one year prior and one year posterior to initiation of these drugs and should have received at least one more prescription of the same drug within one year. Individuals who initiated treatment with salmeterol, formoterol, or tiotropium bromide only, or whoever received roflumilast, were excluded (due to high probability of chronic obstructive pulmonary disease (COPD)).

Data Source

This study used the University of Groningen IADB.nl dispensing database.18 The IADB.nl database includes anonymized patient-level dispensing data from approximately 120 community pharmacies with a total of 1.3 million patients in the north of the Netherlands from January 1st 1994 onwards. Data include sex, date of birth, and dispensing records with dispensing date, Anatomical Therapeutic Chemical (ATC) code, quantity, and prescribed dose. Inclusion in the database is regardless of health insurance status. Due to the high patient-pharmacy commitment in the Netherlands, the dispensing records for each individual in the IADB.nl database are virtually complete, except for medication dispensed during hospitalization and over-the-counter drugs.18 Of note, over-the-counter sales of asthma inhalation medication were not allowed in the Netherlands during the studied period.

Missingness exclusively concerned the daily dose of ICS and ICS-LABA prescriptions (7% missing) and the number of days for which asthma medication was supplied (10% missing). The daily doses were imputed using the most frequent value, stratified by ATC-code and dose per inhalation. Similarly, missing supply duration was imputed based on the most frequent value, stratified by medication number or, alternatively, by ATC-code, inhaler type (aerosol/powder), and daily inhalations. These imputations were performed before applying a time-dependent proportion of days covered (tPDC)-based algorithm, which proved quite robust in our previous study,9 and before categorizing daily dosage into low, medium, or high doses, minimizing potential misclassification. Consequently, the impact of missing data on study results is likely negligible.

Outcome Variables

The main outcomes were switches within two years of follow-up, the duration until first switch, and number of switches. We were also interested in whether g-formula simulations for this were in line with empirical observations.

We defined asthma treatment steps (0 to 4) as follows, in accordance with the guidelines for treatment of adult asthma from the Dutch College of General Practitioners (NHG), which follow the Global Initiative for Asthma (GINA) report recommendations:4,19

Step 1: SABA or low-dose ICS + formoterol Step 2: Low dose ICS without LABA or long-acting muscarinic antagonist (LAMA) Step 3: Medium dose ICS, or low dose ICS plus LABA or LAMA Step 4 (and 5): High dose ICS, or medium to high dose ICS plus LABA, triple therapy (ICS plus LABA plus LAMA), or biologics Step 0: No current prescriptions meeting any of the criteria of treatment steps 1 to 4

As montelukast is only advised for step 3 or higher, if montelukast was prescribed next to a lower step, this step was updated to step 3 instead. ICS dose was classified into low, medium, or high, based on the total daily dose. The cut-off values were drug and inhaler-specific (Supplement 1). We intended to include step 5 separately, consisting of triple therapy or biologics. However, due to the limited occurrence of these treatments within our study population (<1%), these patients were reclassified into step 4.

A time-dependent proportion of days covered (tPDC)-based algorithm was used to transform the dispensing records into continuous data as described in an earlier study.9 To limit computation time and memory usage, we only registered the mode of each 30 days, resulting in 24 time periods, with a total follow-up duration of 720 days.

Baseline and Time-Varying Predictors

The included baseline predictors were: year, age, sex, and treatment step at initiation, and presence vs absence of chronic drug treatment for diabetes, CVD, mental health problems, arthritis, thyroid disease, atopic diseases, peptic ulcer and gastroesophageal reflux disease (GERD), and use of immunocompromising medication. Timepoint of follow-up, treatment step at previous time point, and an interaction between these last two were added as time-varying predictors.

Chronic drug treatment for comorbidities was defined as having received at least two dispenses of related medication (for ATC-codes, see Supplement 2) within one year before initiation of asthma treatment.20,21

G-Formula

To predict the individual treatment trajectories over time, we utilized parametric micro-simulations with the g-formula.13,22,23 Stable estimates and confidence intervals were obtained by performing of 25 Monte Carlo simulations within 300 bootstrap iterations.

Model Specification

A set of logistic regression models was fitted within each bootstrap loop of the g-formula,24 each with a treatment step at time t+1 as outcome variable, and included all aforementioned predictors (see Supplement 3 for more information on the models’ functional form). Once these models were fitted, we used predictor values observed at the baseline to simulate observations at time point t+1, then used the simulated values at t+1 as input to simulate observations at time point t+2, and so onward until the end of follow-up.

Natural Course Validation

To validate the model, we simulated the natural course in which each patient had their baseline variables set to match the observed data.13 For this validation, an additional logistic regression model was included in the g-formula to predict censoring. This natural course simulation was compared to the empirical, observed data, guarding against gross model misspecification. The focus of this comparison lies on clinical relevance, as the large sample size of this study may lead to statistically significant results (big data paradox) that might not necessarily be clinically meaningful.

Subgroup Predictions

The subgroups of interest were based on age (16 vs 45 years old), sex and chronic drug treatment for diabetes, CVD, mental health problems, arthritis, thyroid disease, atopic diseases, peptic ulcer and GERD, and immunosuppressants (Supplement 2). To approximate subgroup treatment trajectories, the multinomial logistic regression models in the simulation included interaction terms for the subgroup of interest with timepoint of follow-up, with treatment step at previous timepoint and their interaction. Once fitted, the original data at baseline were used as input in our prediction algorithm, except that first all, and then none of the individuals were set to have the comorbidity (or sex or age) of interest, so each individual appeared in all subgroups. The predicted treatment trajectories were reported and compared with each other to determine the significance of subgroup differences. The individual treatment trajectories were visualized using lasagna plots (Supplement 4).

All analyses were performed using R 4.2.1. The source code is available on GitHub at: https://github.com/IreneMommers/R_gformula_public_2024/.

Results Patient Selection and Baseline Characteristics

Of the 24,506 included individuals, the majority were female (65%). Chronic drug treatment was most frequently prescribed for atopic disease (22%; includes allergies, allergic rhinitis, conjunctivitis, and dermatitis), followed by mental health problems (14%; includes anxiety, depression, dementia, and sleep disorders), and arthritis (11%). Treatment was mainly initiated in step 1 (51%). The incidence of asthma treatment increased over time, until it peaked in the period 2011–2015 (Table 1). The treatment steps at initiation for each subgroup can be found in Supplement 5.

Table 1 Patient and Treatment Characteristics

Observed Data Vs Natural Course Simulation

Observed data showed that many individuals stopped asthma treatment within two years of initiation. After two years, those who continued treatment often found themselves in the same treatment step as at initiation (37%), mainly due to never switching (23%). On average, individuals switched between treatment steps twice, and the first switch occurred after eight months on average. The observed treatment trajectories (Figure 1A) were closely approximated by the natural course simulation (Figure 1B and Table 2).

Table 2 Observed Data Versus Natural Course Simulation

Table 3 Subgroup Differences in Switching Patterns

Figure 1 Distributions of individuals across treatment steps over time according to the observed data (A) versus the natural course simulations (B).

Predicted Relation Between Exposures and Treatment Trajectories

Differences in switching behavior were found between subgroups. Individuals aged 45 years were less likely to switch treatment steps compared to those aged 16 years (74.3% vs 78.4%, p < 0.001). Those who did switch at older age, switched earlier (8.1 months vs 8.6 months, p < 0.001) and more often (2.4 vs 2.3 times, p < 0.001) compared to if they had been younger. Females showed a slightly higher switching percentage than males (76.6% vs 75.5%, p = 0.01), with no significant difference in time until the first switch. Chronic drug treatment for atopic diseases, CVD, mental health problems, and peptic ulcer and GERD, was predictive for a decreased proportion of switching (p < 0.001, p < 0.001, p = 0.003, and p < 0.001, respectively), longer duration until the first switch (p = 0.03, p = 0.04, p = 0.003, and p = 0.04, respectively) and sometimes less switches (CVD and mental health problems: both p = 0.02). In contrast, chronic drug treatment for arthritis, diabetes, immunosuppressants, and thyroid disease had no significant association with switching behavior (Table 3). Visualizations of the individual treatment trajectories within the subgroups can be found in Supplement 4.

Within the subgroups atopic diseases, CVD, mental health problems or peptic ulcer and GERD, it stands out that individuals are less likely to switch down to a lower treatment step or quit their medication (−4.2%, −3.4%, −1.1%, and −3.5%, as compared to the subgroups without respective chronic drug treatments) and are more likely to end up in the treatment step in which they started (+2.6%,+2.2%, +1.6%, +2.2%, respectively) or step up (+1.7%, +1.2%, −0.4%, +1.4%, respectively). For the subgroup with chronic drug treatment for atopic diseases, the overall number of individuals switching was 74.9% as compared to 76.5%, a difference of 1.6%. Thus, besides less switching, those who did switch were also more likely (~1%) to ultimately switch back to the same treatment step they started in. For the subgroup with chronic drug treatment for CVD compared to without, the opposite seems to be true. Although fewer individuals switched their medication (−3.9%; 72.7% vs 76.4%), only +2.2% were found in the same step both at the end and beginning of follow-up, indicating that a relatively lower amount switched back to their previous step after first switching away, with as compared to without chronic medication for CVD. For the subgroups with chronic medication for mental health problems or for GERD, the differences in the proportion of individuals switching (74.8% with, versus 76.4% without chronic medication for mental health problems; 76.4% with versus 74.1% without GERD) are in line with the difference in individuals found in the same treatment step at the beginning and end of follow-up (Table 3 and Supplement 6).

Discussion

In this real-world drug utilization study, we found that natural course simulations using the parametric g-formula were successful in approximating empirical asthma treatment trajectories. We identified differences in switching behavior across subgroups defined by age, sex, and chronic drug treatments. Individuals with chronic drug treatment for atopic diseases, CVD, peptic-ulcer and GERD, or mental health problems switched asthma medication less and/or later.

Some questions might arise due to the choice for dynamic models over traditional static regression models. The longitudinal parametric g-formula allowed us to keep the distribution of all variables at baseline equal to those in the target population, except for the ones of interest. By changing the value of a predictor at baseline and modelling the effect of all variables at time t onto time t+1, a simple change at baseline can cascade across all time points. This allows for easy comparison of full treatment trajectories resulting from different baseline characteristics and enables more detailed investigation through the use of, eg, transition matrices or lasagna plots. Although these methods are in line with those used for causal inference, the focus of this study was solely to predict. Therefore, the results should be interpreted predictively rather than causally, especially since we were not able to include all relevant confounders. Nevertheless, by projecting counterfactual scenarios on the entire study population, these methods do allow us to eliminate the influence of differences in the distributions of baseline variables that characterize each subgroup, and thereby to gain more insight into the associations between individual predictors and outcomes of interest along the entire treatment trajectory. This methodology shows potential for application in policy- and decision-making, by predicting the impact of new guidelines, interventions or treatments on the overall population or within specific patient subgroups, or even for informing resource allocation.

The main advantages of the g-formula include its flexibility and dynamic nature. Unlike multi-state models with a limited number of states, the g-formula allows incorporation of as many states as naturally arise from the parameterization (including continuous predictors). Furthermore, employing probabilistic simulations results in less information loss as compared to traditional regression analyses, as full distributions of outcomes are obtained, rather than just the single most likely ones. Additionally, the g-formula overcomes the noncollapsibility problem that is found in many regression analyses.25,26 Lastly, the dynamic g-formula allows multiple interactions while producing output that remains easy to interpret. Nevertheless, it is important to note that the g-formula is highly computationally intensive, making it most useful when an entire trajectory is of interest, rather than just a single outcome.

Our research results provide insight into how certain comorbidities influence the switching behavior of asthma medications. Individuals with chronic drug treatment for atopic diseases, CVD, mental health problems, and GERD switched their asthma medications less frequently, and later compared to those without chronic drug treatment for said comorbidities. Possibly, treating physicians may be more hesitant to decrease asthma medication in patients with multiple conditions, even if asthma symptoms would allow for it, or individuals with chronic drug treatment for these comorbidities may present with more difficult-to-treat asthma symptoms, requiring them to remain on high-dose ICS medication for longer periods. The connection between these comorbidities and asthma is well established in literature: atopic diseases like atopic dermatitis and allergic rhinitis are related to the same IgE-mediated immune response as allergic asthma. These conditions often co-exist and influence each other’s development.27,28 CVD is associated with increased levels of mast cells, eosinophils, inflammatory cytokines, and IgE, suggesting, suggesting a shared pathogenic pathway with asthma.29 For asthma patients with GERD, there are increased odds of suffering from exacerbations, which may be due to pulmonary inflammation or due to bronchoconstriction caused by vagal nerve stimulation, both induced by reflux.30,31 In summary, the presence of comorbidities, such as atopic diseases, CVD, mental health problems, and GERD, can impact switching behavior within asthma treatment, likely due to the complex interplay between these conditions and asthma pathophysiology. Further research should focus on understanding the origin of these differences and exploring how they can inform personalized treatment strategies, improve adherence, and influence long-term outcomes. The results also highlight the complex relationship between mental health problems and asthma medication-switching behavior. Individuals with chronic drug treatment for mental health problems were less likely to either step up or step down their asthma treatment and thus were more likely to remain on the same medication step over time. This switching behavior was not only deviating from those without chronic drug treatment for mental health problems but also from other comorbidities. Previous studies have found that depression and anxiety can result in more severe asthma symptoms and poorer control,32–34 non-adherence to asthma medication,35,36 increased health service usage,37 and a decreased quality of life.34 Perception of worse asthma control in those with anxiety, as well as hyperventilation, can result in misinterpretation of symptoms as asthma or even induce and worsen asthma attacks, preventing individuals from stepping down their asthma treatment.32,38 On the other hand, proper treatment of depressive symptoms may improve asthma outcomes, which could prevent the stepping up of asthma treatment.39,40

Besides its methodology, the strength of this study lies in the use of a large, real-life dispensing database, the IADB. The individual-level dispensing data is proven to be virtually complete, except for over-the-counter drugs and hospitalizations.18 The dispensing rates from this database have been found to be representative of the entire Dutch population.41 However, several limitations should be noted. Due to the nature of the data, no diagnostic information was available, meaning that we had to use dispensed medication as a proxy for disease, which may result in possible misclassification. In addition, the analyses did not include step 5 separately (represented by triple therapy and/or biologics), because of the frequency of this step (n=17 at start and reached by less than 1% of patients ever during the study period) seemed too low to accurately predict the switching behavior to and from step 5. However, with larger sample sizes, incorporation of step 5 could potentially lead to stronger results, as these individuals with severe, uncontrolled asthma might show even more distinct switching behavior. Furthermore, baseline variables were kept constant to prevent confounding, as this study focused only on switching behavior after treatment initiation. Thus, even if no difference was found between subgroups, their real-world baseline characteristics may differ (eg, the treatment steps at baseline differed; Supplement 5). In particular, potential differences in treatment steps at baseline make it challenging to define more precise clinical implications or personalized treatment strategies for patients with specific comorbidities, highlighting a need for further research. However, the presence of differences in switching behavior itself do indicate that earlier intervention and closer monitoring might be of added value. In addition, there can be large differences between scenarios within individuals, even if they were not found for the general population on average. As the focus of this study was to apply the g-formula to predict treatment trajectories at the population level, further research would be needed if this method were to be applied for making predictions on the individual level. Although some relative differences found in this study might appear small, they do represent large numbers of individuals, as asthma is a common chronic disease among young adults.

Conclusion

The parametric g-formula closely approximated the empirical distribution of treatment trajectories in general and within subgroups. The dynamic g-formula provides a good alternative to regression analyses when analyzing processes of a dynamic nature or when interested in full trajectories and shows potential for application in policy- and decision-making, by predicting the impact of new guidelines, interventions or treatments in the full population or within subgroups. This study revealed that higher age, male sex, and chronic drug treatment for atopic disease, CVD, mental health problems, and peptic ulcer and gastroesophageal reflux disease decreased overall switching behavior, indicating that these patients might benefit from earlier intervention or closer monitoring to reduce delays in treatment progression.

Abbreviations

ATC, Anatomical Therapeutic Chemical; ATD, atopic disease(s); COPD, chronic obstructive pulmonary disease; CVD, cardiovascular disease(s); GERD, gastroesophageal reflux disease; GINA, Global Initiative for Asthma; ICS, inhaled corticosteroids; LABA, long-acting selective β2-adrenoreceptor agonist; LAMA, long-acting muscarinic antagonist; MHP, mental health problems; NHG, Dutch College of General Practitioners; SABA, short-acting selective β2-adrenoreceptor agonist; tPDC, time-varying proportion of days covered.

Data Sharing Statement

The dataset used for this study was provided by the University of Groningen IADB.nl dispensing database. The data will be accessible via the IADB (email: [email protected]) upon reasonable request. The source code will be made available through GitHub.

Ethics Approval and Informed Consent

The IADB.nl data is processed in accordance with the Dutch and European guidelines on privacy requirements (GDPR) for handling human data.42 Approval of the medical ethics committee is not needed nor required for research using anonymous medical records, as individuals are not subjected to any actions or rules of conduct.

Consent for Publication

No images or recordings were included in this manuscript that require consent for publication.

Acknowledgments

We would like to acknowledge Marten Koetsier for his invaluable assistance in the execution of the data analysis.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Funding

The authors declare that no funds, grants, or other support were received from any funding agency in the public, commercial, or not-for-profit while preparing this manuscript.

Disclosure

Dr Job FM van Boven reports grants and/or personal fees from ALK, AstraZeneca, Chiesi, GSK, Novartis, Pill Connect, Pfizer, Teva, Trudell Medical and Vertex, outside of this work. The authors report no other conflicts of interest in this work.

References

1. Maslan J, Mims JW. What is asthma? Pathophysiology, demographics, and health care costs. Otolaryngol Clin North Am. 2014;47(1):13–22. doi:10.1016/j.otc.2013.09.010

2. Porsbjerg C, Melén E, Lehtimäki L, Shaw D. Asthma. Lancet. 2023;401(10379):858–873. doi:10.1016/s0140-6736(22)02125-0

3. Kaplan A, Szefler SJ, Halpin DMG. Impact of comorbid conditions on asthmatic adults and children. NPJ Prim Care Respir Med. 2020;30(1):36. doi:10.1038/s41533-020-00194-9

4. Global Initiative for Asthma. Global Strategy for Asthma Management and Prevention, 2023. Available from: www.ginasthma.org. Accessed March1, 2025.

5. van Boven JFM, Lavorini F, Agh T, Sadatsafavi M, Patino O, Muresan B. Cost-effectiveness and impact on health care utilization of interventions to improve medication adherence and outcomes in asthma and chronic obstructive pulmonary disease: a systematic literature review. J Allergy Clin Immunol Pract. 2024;12(5):1228–1243. doi:10.1016/j.jaip.2023.12.049

6. Domínguez-Ortega J, Phillips-Anglés E, Barranco P, Quirce S. Cost-effectiveness of asthma therapy: a comprehensive review. J Asthma. 2015;52(6):529–537. doi:10.3109/02770903.2014.999283

7. Chen W, Lynd LD, FitzGerald JM, et al. Excess medical costs in patients with asthma and the role of comorbidity. Eur Respir J. 2016;48(6):1584–1592. doi:10.1183/13993003.01141-2016

8. Gayle A, Tebboth A, Pang M, et al. Real-life prescribing of asthmatic treatments in UK general practice over time using 2014 BTS/SIGN steps. NPJ Prim Care Respir Med. 2019;29(1):25. doi:10.1038/s41533-019-0137-7

9. Mommers I, van Boven JFM, Schuiling-Veninga CCM, et al. Real-world dispensing patterns of inhalation medication in young adult Asthma: an inception cohort study. Clin Epidemiol. 2023;15:721–732. doi:10.2147/clep.S410036

10. van Geloven N, Swanson SA, Ramspek CL, et al. Prediction meets causal inference: the role of treatment in clinical prediction models. Eur J Epidemiol. 2020;35(7):619–630. doi:10.1007/s10654-020-00636-1

11. Hernán MA, Robins JM. Causal Inference: What If. Boca Raton: Chapman & Hall/CRC; 2020.

12. Naimi AI, Cole SR, Kennedy EH. An introduction to g methods. Int J Epidemiol. 2016;46(2):756–762. doi:10.1093/ije/dyw323

13. Keil AP, Edwards JK, Richardson DB, Naimi AI, Cole SR. The parametric g-formula for time-to-event data: intuition and a worked example. Epidemiology. 2014;25(6):889–897. doi:10.1097/ede.0000000000000160

14. Bijlsma MJ, Tarkiainen L, Myrskylä M, Martikainen P. Unemployment and subsequent depression: a mediation analysis using the parametric G-formula. Soc Sci Med. 2017;194:142–150. doi:10.1016/j.socscimed.2017.10.011

15. Tsiatis AA, Davidian M, Holloway ST, L EB. Dynamic Treatment Regimes: Statistical Methods for Precision Medicine. 1st Edition ed. Chapman and Hall/CRC; 2019.

16. Young JG, Cain LE, Robins JM, O’Reilly EJ, Hernán MA. Comparative effectiveness of dynamic treatment regimes: an application of the parametric g-formula. Stat Biosci. 2011;3(1):119–143. doi:10.1007/s12561-011-9040-7

17. Briggs A, Sculpher M. An introduction to Markov modelling for economic evaluation. PharmacoEconomics. 1998;13(4):397–409. doi:10.2165/00019053-199813040-00003

18. Visser ST, Schuiling-Veninga CC, Bos JH, de Jong-van den Berg LT, Postma MJ. The population-based prescription database IADB.nl: its development, usefulness in outcomes research and challenges. Expert Rev Pharmacoecon Outcomes Res. 2013;13(3):285–292. doi:10.1586/erp.13.20

19. Bottema JW, Broekhuizen L, Chavannes NH, et al. NHG-standaard Asthma bij volwassenen (M27). Nederlands Huisartsen Genootschap 2022.Available from: https://richtlijnen.nhg.org/standaarden/astma-bij-volwassenen.

20. van Boven JF, van Raaij JJ, van der Galiën R, et al. Impact of multiple-dose versus single-dose inhaler devices on COPD patients’ persistence with long-acting β₂-agonists: a dispensing database analysis. NPJ Prim Care Respir Med. 2014;24(1):14069. doi:10.1038/npjpcrm.2014.69

21. Wang Y, Bos JH, Boezen HM, et al. Influence of age on real-life effects of doxycycline for acute exacerbations among COPD outpatients: a population-based cohort study. BMJ Open Respir Res. 2020;7(1). doi:10.1136/bmjresp-2019-000535

22. Diallo AO, Ali MK, Geldsetzer P, et al. Systolic blood pressure and six-year mortality in South Africa: evidence from a country-wide population-based cohort study. Lancet Healthy Longev. 2021;2(2):e78–e86. doi:10.1016/s2666-7568(20)30050-7

23. Jain P, Danaei G, Robins JM, Manson JE, Hernán MA. Smoking cessation and long-term weight gain in the Framingham heart study: an application of the parametric g-formula for a continuous outcome. Eur J Epidemiol. 2016;31(12):1223–1229. doi:10.1007/s10654-016-0200-4

24. Bijlsma MJ, Wilson B. Modelling the socio-economic determinants of fertility: a mediation analysis using the parametric g-formula. J R Stat Soc Ser a Stat Soc. 2019;183(2):493–513. doi:10.1111/rssa.12520

25. Schuster NA, Twisk JWR, ter Riet G, Heymans MW, Rijnhart JJM. Noncollapsibility and its role in quantifying confounding bias in logistic regression. BMC Med Res Method. 2021;21(1):136. doi:10.1186/s12874-021-01316-8

26. Daniel R, Zhang J, Farewell D. Making apples from oranges: comparing noncollapsible effect estimators and their standard errors after adjustment for different covariate sets. Biom J. 2021;63(3):528–557. doi:10.1002/bimj.201900297

27. Yaneva M, Darlenski R. The link between atopic dermatitis and asthma- immunological imbalance and beyond. Asthma Res Pract. 2021;7(1):16. doi:10.1186/s40733-021-00082-0

28. Bousquet J, Anto JM, Bachert C, et al. Allergic rhinitis. Nat Rev Dis Prim. 2020;6(1):95. doi:10.1038/s41572-020-00227-0

29. Cazzola M, Hanania NA, Rogliani P, Matera MG. Cardiovascular disease in asthma patients: from mechanisms to therapeutic implications. Kardiol Pol. 2023;81(3):232–241. doi:10.33963/KP.a2023.0038

30. Mallah N, Turner JM, González-Barcala FJ, Takkouche B. Gastroesophageal reflux disease and asthma exacerbation: a systematic review and meta-analysis. Pediatr Allergy Immunol. 2022;33(1):e13655. doi:10.1111/pai.13655

31. Grandes XA, Talanki Manjunatha R, Habib S, Sangaraju SL, Yepez D. Gastroesophageal reflux disease and asthma: a narrative review. Cureus. 2022;14(5):e24917. doi:10.7759/cureus.24917

32. Ciprandi G, Schiavetti I, Rindone E, Ricciardolo FL. The impact of anxiety and depression on outpatients with asthma. Ann Allergy Asthma Immunol. 2015;115(5):408–414. doi:10.1016/j.anai.2015.08.007

33. Janson C, Björnsson E, Hetta J, Boman G. Anxiety and depression in relation to respiratory symptoms and asthma. Am J Respir Crit Care Med. 1994;149(4 Pt 1):930–934. doi:10.1164/ajrccm.149.4.8143058

34. Lavoie KL, Cartier A, Labrecque M, et al. Are psychiatric disorders associated with worse asthma control and quality of life in asthma patients? Respir Med. 2005;99(10):1249–1257. doi:10.1016/j.rmed.2005.03.003

35. Grenard JL, Munjas BA, Adams JL, et al. Depression and medication adherence in the treatment of chronic diseases in the United States: a meta-analysis. J Gen Intern Med. 2011;26(10):1175–1182. doi:10.1007/s11606-011-1704-y

36. Bender BG. Risk taking, depression, adherence, and symptom control in adolescents and young adults with asthma. Am J Respir Crit Care Med. 2006;173(9):953–957. doi:10.1164/rccm.200511-1706PP

37. ten Brinke A, Ouwerkerk ME, Zwinderman AH, Spinhoven P, Bel EH. Psychopathology in patients with severe asthma is associated with increased health care utilization. Am J Respir Crit Care Med. 2001;163(5):1093–1096. doi:10.1164/ajrccm.163.5.2004020

38. Deshmukh VM, Toelle BG, Usherwood T, O’Grady B, Jenkins CR. Anxiety, panic and adult asthma: a cognitive-behavioral perspective. Respir Med. 2007;101(2):194–202. doi:10.1016/j.rmed.2006.05.005

39. Brown ES, Vigil L, Khan DA, Liggin JD, Carmody TJ, Rush AJ. A randomized trial of citalopram versus placebo in outpatients with asthma and major depressive disorder: a proof of concept study. Biol Psychiatry. 2005;58(11):865–870. doi:10.1016/j.biopsych.2005.04.030

40. Brown ES, Sayed N, Van Enkevort E, et al. A randomized, double-blind, placebo-controlled trial of escitalopram in patients with Asthma and major depressive disorder. J Allergy Clin Immunol Pract. 2018;6(5):1604–1612. doi:10.1016/j.jaip.2018.01.010

41. Sediq R, van der Schans J, Dotinga A, et al. Concordance assessment of self-reported medication use in the Netherlands three-generation Lifelines Cohort study with the pharmacy database iaDB.nl: the PharmLines initiative. Clin Epidemiol. 2018;10:981–989. doi:10.2147/clep.S163037

42. European Union. Regulation (EU) 2016/679 of the European parliament and of the council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (general data protection regulation). Off J Eur Union. 2016;L119:1–88.