Does uptake of specialty care affect HRQoL development in COPD patients beneficially? A difference-in-difference analysis linking claims and survey data

Chronic obstructive pulmonary disease (COPD) is a serious health condition elevating mortality and morbidity, which claims more than 3 million lives per year [1]. It is the third leading cause of death worldwide [1], and the burden of COPD is even expected to increase further on account of an aging population and continued exposure to COPD risk factors (e.g., smoking) [2]. This inflammatory disease is characterized by respiratory symptoms such as cough, breathlessness, and sputum production [3]. As COPD is a progressive disease with commonly irreversible airflow obstructions, cure cannot be achieved. Therefore, the treatment focus relies on enhancing COPD management to relieve the symptoms, lengthen the lifespan and, above all, maintain or even improve (health-related) quality of life. COPD management can be implemented by, among other things, structured programs such as the German disease management programs (DMPs).

The patients’ health-related quality of life (HRQoL) is one of the most important patient-reported outcomes (PROMs) and also a key outcome in the DMP [4]. It describes how patients perceive the severity of their symptoms and the subjective impact of the disease on their daily lives [5]. Regarding treatment, health care systems and specialty care in Europe vary, for example in the number of specialists and primary care physicians but also in the role of general practitioners (GPs). In many western countries, COPD patients are principally managed by GPs who serve as gate-keepers referring their patients to specialists, whereas patients in central or eastern Europe are rather managed by pulmonologists [6, 7]. In Germany’s Statutory Health Care system, patients have free choice of specialists without a mandatory GP referral. The utilization of specialty care is covered within the scope of the statutory health insurance (SHI) and does not incur any extra costs on behalf of the patients. This means that COPD patients in Germany can freely choose whether they receive their treatment from their primary care physician only or whether a pulmonologist is incorporated in disease management. Thus, the inclusion of pulmonologists in the treatment process, if at all, does not necessarily happen in a coordinated way. Nevertheless, the commonly used care path in Germany usually starts by a referral from the general practitioner to a pulmonologist driven by criteria of medical need [8]. Therefore, the choice of specialty care in COPD treatment on a regular basis or especially for certain severity groups could be a key management aspect that is not yet or rather casually discussed in COPD disease guidelines [9‐11].

Therefore, this real-world data-based study within the COPD DMP aims to analyze the 1 year longitudinal impact of newly initiated pulmonologist care on generic and disease-specific HRQoL of COPD patients to identify eventual subgroups of patients in the DMP for whom management of treatment paths and, thus, program outcome could be improved.

We used a linked real-world data set, combining survey data and pseudonymized health insurance claims data from AOK Bayern, a large SHI fund in the district of Bavaria in southern Germany. With almost 4.5 million insurants, AOK Bayern had a 40.5% share of the district’s SHI market during our study in 2017 [12]. SHI is accessible to everyone in Germany independent of age, comorbidities, or income and covers a broad range of in- and outpatient services at almost no copayment based on the principle of medical need. SHI contributions are calculated risk-independently based on the insurants’ income [13]. Nonetheless, in Germany about 10% of the population are privately insured outside the SHI system, including mainly persons with a high income level, self-employed or civil servants [14, 15].

Our initial data set contained all insurants who participated in the structured COPD DMP of AOK Bayern. This DMP is a voluntary program with comprehensive legal requirements, to which all COPD patients, who meet the inclusion criteria [16], can get enrolled. An overview of DMP inclusion criteria can be found in the supplement (see Supplementary Information (SI) Appendix, Fig. S1). The claims data included, among others, information on sociodemographic characteristics, exacerbations, hospitalizations, comorbidities, outpatient hospital care, outpatient physician visits, and medications. Additionally, DMP-specific documentation was linked to routine claims data, which enabled the inclusion of clinical factors, such as lung function measured in forced expiratory volume in 1 s (FEV₁) and body mass index (BMI) [17, 18].

In addition, we conducted a postal survey in two waves, addressing all 49,662 COPD DMP participants. The survey addressed generic and disease-specific HRQoL, breathlessness, determined by the modified British Medical Research Council (mMRC) [19], and further sociodemographic data (see Fig. S2 for more information). A total of 14,754 (29.7%) responders participated in the first wave in November 2017 and, of those, 9232 (62.6%) participated in the second wave in November 2018.

We then linked the survey data with the DMP and claims data. A graphical overview of data sources can be found in the supplement (Fig. S3).

We excluded all patients from our analysis, who did not respond to both survey waves and who had missing or implausible values in outcomes or covariates. For example, we excluded extreme values in the lung function FEV₁%pred. < 10 or > 150, as we considered values above/below these thresholds as not reliable, compared to a standard lung function: the FEV₁ percent predicted median values for individuals without pulmonary diseases are about 90, and the COPD diseases classification by GOLD starts with the lowest group “I” with a FEV₁%pred. ≤ 80 and ends in highest group “IV” with values < 30. Indeed, there were 27 participants with values below the minimum of 10 and only 37 participants had to be excluded because of values > 150. Additionally, we disregarded patients who had received specialty care in the year before the first survey wave, as we aimed to analyze the impact of newly started specialty care treatment.

The study was approved by the ethics committee of the Ludwig-Maximilians-Universität, Munich, Germany (vote no. 17- 358) and complies with the Declaration of Helsinki.

We aimed to analyze the HRQoL of COPD patients receiving specialty care from pulmonologists compared with patients not treated by a pulmonologist. The main outcome, HRQoL, was measured by the disease-specific COPD Assessment Test (CAT) [20] and the generic 5-level Euro-Qol 5D (EQ-5D-5L) [21, 22], as it is recommended to use both a disease-specific and a generic measure to assess the effect on HRQoL in COPD patients [23].

The EQ-5D-5L is a validated measure of generic HRQoL in COPD [22]. It consists of a valuation section with five dimensions and five problem levels each, as well as a visual analog scale (VAS). For our analysis, we used the VAS as the generic HRQoL outcome, which (in contrast to valuations by population preferences) is a fully patient-reported outcome and thus, regarding HRQoL, also patient relevant [24]. Further, it has also been demonstrated to differentiate more precisely between COPD grades than the valuation section of the EQ-5D [24]. The VAS is a scale from 0 (worst state) to 100 (best state), on which patients can indicate their current general health perception.

The CAT measures the disease-specific HRQoL with a retrograde scale from 40 (worst state) to 0 (best state) [20]. It comprises eight dimensions (such as cough, sputum, energy, and sleep) with six answer levels each (0–5). Reasons for choosing CAT over other disease-specific HRQoL instruments can be found elsewhere [17, 25].

A clinically relevant change in HRQoL is expressed by a minimum important difference (MID), which was found to be a 6.9-point change in VAS for the EQ-5D-5L [22] and a 2.0-point change in CAT [26, 27].

To examine the impact of specialty care on HRQoL, we primarily separated patients into a treatment and a control group. Specialty care in this context is defined as any care given by a pulmonologist, which we categorized using practitioner identification codes available in the claims data.

The treatment group is defined as patients who had no pulmonologist contact in the baseline year before the first survey (t0) and at least one contact with a pulmonologist in the follow-up year between surveys one and two (t1). The control group, on the other hand, comprises all patients who did not receive pulmonologist care during this observation period (see Fig. 1). In both cases, the patients may still have visited their GPs, which is why we assume that the treatment patients received combined care.

We included age, sex, exacerbation history, smoking status, lung function, mMRC, BMI group, Charlson comorbidity index, and education level as covariates (see SI Appendix, Fig. S1 for data sources). Exacerbations are adverse health events with acute worsening in respiratory symptoms [11]. The history of exacerbations (ICD-10 code J441) was based on moderate exacerbations (worsening of COPD symptoms and a doctor’s visit was required) and severe exacerbations (worsening of symptoms and a hospital stay was required) in the last 12 months before the survey [28]. Lung function was measured as FEV₁% predicted (FEV₁%pred.) and was condensed to an annual average wherever several measurements per year and per patient were available. The mMRC describes breathlessness on an ascending scale from 4 to 0 [19]. To calculate the weighted Charlson index, a condition was only considered present if the patient received either one inpatient diagnosis or two confirmed outpatient diagnoses within two different quarters during the year prior to the baseline assessment. This so-called M2Q-criterion is well established in claims data-based research [29]. We implemented the Charlson index using the ICD-10-based coding algorithm of Quan et at. [30] without adjustment for COPD. Further information about the weighted Charlson index conditions and secureness of the diagnosis can be found in a prior description in our research project [28]. Education levels were classified as none, basic (9 years), secondary (10 years), and higher education (12–13 years) or university degree. Reference categories were female and basic education.

We further assessed the disease severity by Global Initiative for Chronic Obstructive Lung Disease (GOLD) ABCD groups [3], based on survey data, which cannot be obtained from claims or DMP data. Here, A indicates the least severe group and D the most severe group. This classification uses a combination of COPD symptoms and exacerbation history. The symptoms differentiate between groups AC versus BD and are classified by breathlessness, where mMRC ≥ 2 identifies the more severe groups B and D. On the other hand, the exacerbation history differentiates between groups AB and CD, with an occurrence of ≥ 2 moderate or at least one severe exacerbation classifying the more severe groups C and D [3].

As we found imbalance in baseline characteristics between the treatment and control group, we calculated propensity scores to balance the baseline characteristics between both groups and thereby reduce possible bias [31, 32]. Especially when the treatment and control group differ in their baseline covariates, matching is recommended to ensure accurate point estimates and better inference [33]. The matching identifies untreated (i.e., control) patients who have similar baseline characteristics to the treated participants. Based on the propensity scores, we used a nearest neighbor matching (NN PSM) with a 1:3 ratio without replacement, in which one treated patient is matched with three untreated patients. We included the previously listed covariates in the matching. The covariates for the matching were retrieved from the baseline year before treatment (t0) to ensure they were not affected by the treatment [32, 34]. Further, we applied restricted cubic splines for lung function to increase the matching quality. We then assessed the performance of the NN PSM graphically and by examining the standardized mean differences (SMDs) for all covariates. For SMDs below 0.1, we considered the groups to be well matched [31, 35]. We used the MatchIt R-Package for matching, which relies on the suggestions of Ho et al. [36].

Before matching, we compared the baseline characteristics of the treatment and the control group. Therefore, we calculated counts and percentages for categorical variables as well as means and standard deviations (SD) for continuous variables. We applied independent parametric t-tests for continuous variables and Chi-square-tests for categorical variables. We additionally evaluated the linear absolute and relative change in HRQoL over 1 year based on paired t-tests.

To evaluate the impact of pulmonologist care on 1 year changes in HRQoL, we applied linear difference-in-difference (DID) models with correction for time-varying variables and cluster-robust standard errors on the propensity score matched sample. By combining matching and DID, we can use the advantages of both methods and achieve a strong quasi-experimental study design with even more robust inferences [35]. For example, the combined approach relies on weaker and more credible assumptions, and it was shown to remove time-invariant systematic differences between the intervention and control group [37, 38]. Further, it can also reduce confounding bias and increase the accuracy of the results [33, 39, 40].

After the matching, we obtained a comparable set of treated and control patients in the pre-treatment period, which is also more balanced in terms of potentially confounding baseline variables. The DID analysis enabled us to estimate the causal effect of an intervention (= pulmonologist care) on the outcome (= HRQoL) by calculating the differences in outcomes between the treatment and control groups, before and after treatment [33].

To account for dependence between observations within the matched pairs, our DID models rely on generalized linear models, and we used the coeftest function from the lmtest R-package [41] and the vcovCL function from the sandwich R-package [42, 43] to estimate the coefficients and cluster-robust standard errors [44, 45].

Our linear DID regression equation can be expressed aswhere Y_it is the outcome (HRQoL) for a patient i at time t. Whether the patient received pulmonologist care is represented by the “Group” dummy variable. The β1 coefficient accounts for the baseline differences regarding our two HRQoL outcomes (VAS and CAT) between the treatment and control group. “Time” is a dummy variable with “0” indicating the year before the first survey and “1” indicating the year between the first and second surveys. The β3 coefficient of the interaction term measures the DID estimate and thus the effect of pulmonologist care. It can be interpreted as the average treatment effect on the treated (ATT) and represents the expected HRQoL effect of pulmonologist care for COPD patients [46]. All included covariates are represented by the X_it term, while ε_it describes the error term. We included the same set of covariates in the DID model as we used in the matching before, as we assumed that they could cause variation in the DID outcome or influence selection into the treatment group [32].

The DID analysis faces two important assumptions: the “parallel trends” assumption and the “common shocks” assumption. The “parallel trends” assumption implies that the outcome for both, the treatment and the control group would change at the same rate, if no treatment had taken place [33, 46]. As we do not have multiple pre-intervention measurements of HRQoL, we cannot test the parallel trends assumption, but it was shown that prior matching can reduce possible bias [33, 35]. The “common shocks” assumption implies that all other occurrences happening within the treatment time span would affect both the treatment and the control group equally which is generally not testable [33]. A further requirement for the DID is that the composition of our treatment and control group does not change during the conduct of our study, and treatment does not “spill-over” from the treatment to the control group [33]. As the composition of the groups in our study cannot change over time by definition, a possible “selection bias across time” [35] is ruled out. A potential “selection bias across groups” [35], in which the groups themselves differ (i.e., those patients who consult a pulmonologist would differ from those who do not) is minimized by the prior matching.

To perceive differential treatment effects of specialty care on different disease severity groups, we did subgroup analysis in which we stratified the main model by GOLD groups AB and CD (because of small patient numbers in single ABCD groups).

All statistical tests were two-sided, and the significance level for p-values was ≤ 0.05. All analyses were performed using R version 4 [47].

We performed a sensitivity analysis for all models to check the robustness of our results. Instead of propensity score matching, the sensitivity analysis used genetic matching, which is distance-based and aims to maximize the balance between the treatment and control group [48, 49]. It is a non-parametric and multivariate matching method and relies on a Mahalanobis distance matching [39, 50, 51], although it does not depend on estimating a propensity score [52]. We used a 1:3 ratio again and the same data set as in the NN PSM for the sensitivity analysis. As in the main analysis, we included all matching variables in the DID model, which should account for possible imbalance, as described by Nguyen et al. [53]. The population size argument identifying the number of random trails and the individuals used for solving the optimization problem was set to 1000.

For this study, we included 2968 COPD patients. A summary of exclusion criteria is given in Fig. 2. An overview of baseline characteristics of the initial data sample is presented in the supplement (SI Appendix, Table S1).

Table 1 illustrates the baseline characteristics of the matched control and treatment group at t0. The treatment group comprised 442 patients, and the control group 2526 unmatched and 1326 matched patients respectively. On average, the treatment group had 1.6 pulmonologist visits (SD = 1.0; min = 1; max = 12). Before the matching, both groups differed significantly from each other in their baseline characteristics (see SI Appendix, Table S2). This reveals a special risk group: patients treated by a pulmonologist in the observation period tended to have not only a lower HRQoL at baseline but also a higher disease burden, as they were significantly younger with a higher number of moderate exacerbations, a lower FEV₁%pred., and a higher mMRC compared with the unmatched control group.

The NN PSM showed a good performance with all SMDs being below 0.1 (see Fig. 3) and thus balanced the baseline characteristics between the treatment and the control group sufficiently well. In the matched treatment group, 58.4% were male, the mean age was 68.5 years, and the mean FEV₁%pred. was 56.7. At baseline, the generic HRQoL/VAS was 56.6 (SD ± 19.9), while the disease-specific HRQoL/CAT was 19.9 (± 7.6). HRQoL baseline values for GOLD groups AB/CD can be found in SI Appendix Table S3. With regard to the ABCD groups, the baseline distribution was 40.3% (A), 41.9% (B), 5.7% (C), and 12.2% (D). Our subgroup analysis AB yielded 363 treated patients while 79 were in groups CD.

The unadjusted 1 year change in HRQoL in the unmatched data is depicted in Table 2 and SI Appendix Table S4. For the control group, we found a statistically significant overall deterioration in generic (VAS –1.3 points) and disease-specific HRQoL (CAT + 0.5 units) from t0 to t1 (see Additional file 1: Appendix, Table S4). For the treatment group, we found a similar absolute and relative change in disease-specific HRQoL (CAT + 0.5 points), which was, however, not statistically significant at the 5% level (see Table 2). On the other hand, generic HRQoL remained stable (VAS − 0.0 points).

Further, we observed an overall downward trend in the unadjusted HRQoL (see Fig. 4). The mean HRQoL baseline values for patients in the treatment group were already lower at baseline (t0) than the values for patients in the control group after the 1 year follow-up (t1).

The results of the DID analysis are summarized in Table 3 and SI Appendix Table S6. The estimator (β3) for the interaction term of the dummy variables “Group” and “Time” is presented in the second column for both HRQoL outcome analyses. We observed an overall positive and significant effect of pulmonologist care on the generic HRQoL (VAS + 2.9) as well as on the disease-specific HRQoL (CAT − 0.8) for the total sample. However, this effect does not reach the MID for either VAS or CAT and is thus not clinically relevant.

Within the subgroup analysis, stratified by GOLD AB/CD groups, the specialty treatment only had a significant impact on the AB groups (VAS + 3.1 and CAT − 0.7) without reaching clinical relevance.