Real-world comparison of the effects of etanercept and adalimumab on well-being in non-systemic juvenile idiopathic arthritis: a propensity score matched cohort study

The “Pharmacovigilance in JIA patients treated with biologic agents and/or MTX” (Pharmachild) registry started in 2011 and is currently ongoing. Its primary objective is to assess safety and efficacy of DMARD therapies in patients with JIA. Inclusion criteria are children with JIA as per ILAR classification criteria that are receiving NSAIDs, glucocorticoids, cs-DMARDs or b-DMARDs per physician decision. Currently, patients are enrolled from 85 centers that are part of the Pediatric Rheumatology International Trials Organization (PRINTO) from 31 countries worldwide [19]. Pharmachild consists of patients for whom only retrospective data have been collected at enrolment and patients for whom also prospective data is collected. In brief, Pharmachild collects demographic, clinical and laboratory data, information on drug exposure and adverse events and the cross-culturally adapted version of the Juvenile Arthritis Multidimensional Assessment Report (JAMAR) [20]. The JAMAR assesses patient-reported outcomes in JIA, including functional status, pain, disease activity, health-related quality of life, well-being and satisfaction with disease status [21]. It has been translated into 54 languages and both a parent and child version exist. JAMAR questionnaires in Pharmachild are only available for patients with prospective data. Further details of the Pharmachild registry are available elsewhere [15].

Data of patients with prospective data were extracted on 12 November 2020. For inclusion into the current study, patients or their parents should have completed a “baseline” JAMAR on the day of starting ETN or ADA therapy or at maximum 1 month earlier, provided they had not received any b-DMARD previously. In case both a parent and child JAMAR was completed for the same visit, the child version was selected. In this way, patient-reported outcomes were prioritized over parent-reported outcomes, without excluding information of visits for which only a parent or child JAMAR was available. Other exclusion criteria were systemic JIA, and a history of uveitis. Systemic JIA patients were excluded since this form of JIA is distinct from other subtypes with different clinical features and therapy options [2]. Furthermore, a “follow-up” JAMAR should have been completed 3–12 months after having started ETN or ADA. In case two or more follow-up JAMARs were completed by/for one patient, the JAMAR closest to 6 months after start of ETN or ADA was selected.

We compared study outcomes between patients who started ETN versus patients who started ADA. The primary outcome in this study was the improvement in JAMAR visual analogue scale (VAS) well-being score compared to baseline at the follow-up time-point closest to 6 months, with a minimum of 3 and maximum of 12 months. This 21-point VAS score reflects the answer to the following question: “considering all the ways the illness affects you/your child, please evaluate how you/he/she feels at the moment “, and ranges from 0 (very well) to 10 (very poorly). Secondary outcomes were the decrease in active joint count from baseline to follow-up, the number of adverse events reported by the patient or their parent(s) at follow-up and the number of uveitis events that occurred during follow-up.

Other covariates measured at baseline were patient/parent-reported pain, patient/parent-reported evaluation of disease activity, the physician global assessment of disease activity (all measured on a 21-point VAS), the physical and psychosocial domains of the pediatric rheumatology quality of life scale (composite scores of 5 items measured on a 4-point Likert scale), the juvenile arthritis functional score (a composite score of 15 items measured on a 4-point Likert scale), the patient acceptable symptom state (satisfied or not satisfied with current condition) and the Juvenile Arthritis Disease Activity Score (a composite measure consisting of the physician global assessment, VAS well-being, erythrocyte sedimentation rate and the active joint count) [22].

It is difficult to ascertain causal relationships from observational studies due to the lack of randomization typical of clinical trials, which often leads to confounding by indication. This latter term means that certain patients are more likely to receive a treatment of interest than others and therefore run a different risk for the outcome of interest. We addressed this problem by propensity score matching: ETN and ADA starters were matched at baseline on the probability of being prescribed ADA instead of ETN. The following variables at baseline that could play a role in the decision between ETN or ADA therapy [10] were used in a logistic regression model to predict the propensity score: ILAR category of JIA, sex, age, country of medical center, VAS pain, adverse events while on methotrexate therapy and VAS well-being. Before matching the patients, a distribution of propensity scores for ETN and ADA starters was made and patients outside the range of propensity scores that was common for both groups were excluded. This was done in order to eliminate violation of the positivity assumption, which requires that there are no subjects in one treatment group that are not comparable to subjects in the other treatment group based on propensity score [23]. Subsequently, patients were matched 1 to 1 without replacement based on the logit propensity score. For this matching, we used an acceptable distance (i.e. caliper) of 0.2 times the standard deviation of the logit propensity score, as recommended in the literature [24]. Patients with propensity scores outside of the caliper remained unmatched and were excluded for further analysis. After matching, balance in covariates at baseline was assessed by comparing descriptive statistics and by means of the area under the receiver operating characteristic curve (AUC) of the propensity model fitted in the balanced cohort. Several examples of propensity score matching studies exist within the field of rheumatic diseases [25‐29], and the authors believe that innovative statistical methods like these are of additive value for evidence-based practice in (pediatric) rheumatology.

Covariates at baseline were compared between ETN and ADA starters using the Mann–Whitney U test, Chi-squared test or Fischer’s exact test. In addition, VAS well-being scores at follow-up, time from baseline measurements to start of the b-DMARD, and time from start of the b-DMARD to follow-up measurements were compared between ETN and ADA starters using the Mann–Whitney U test. Missing outcomes at follow-up were handled by multiple imputation using chained equations. All analyses were run for 20 imputed datasets and the different estimates were combined using the theory of Rubin’s rules, which takes into account both uncertainty from one imputed dataset (within-imputation variability) and uncertainty due to the missing information (between-imputation variability) [30]. Outcomes were analyzed using linear and logistic mixed effects models with a random intercept per treatment center to correct for dependence of observations. We performed an intention-to-treat analysis, that is, patients who started ETN or ADA were analyzed in their respective groups regardless if they stopped or changed initial therapy. The analyses of improvement in VAS well-being and decrease in active joint count (quantitative variables) were adjusted for baseline VAS well-being and baseline active joint count respectively in order to increase statistical power and address the problem of regression to the mean [31]. As a sensitivity analysis, all analyses were repeated for the unmatched cohort of patients meeting the positivity assumption while adjusting for the propensity score (instead of matching). For this analysis, we transformed the propensity score using restricted cubic splines with 4 knots in order to correctly model the relation between this numerical variable and the outcomes of interest [32]. For all analyses, statistical significance was set at P < 0.05. All analyses were performed with R version 4.0.0 and the packages rms, mice, lme4, pROC and Matching [33].

As of 12 November 2020, a total of 2,907 non-systemic JIA patients without a history of uveitis were enrolled in the prospective cohort of Pharmachild. Out of these, 158 patients completed a JAMAR at start of ETN/ADA and 3–12 months thereafter (Fig. 1). After calculating propensity scores, another 24 patients who had started ETN had to be excluded because of violation of the positivity assumption. The distribution of propensity scores is provided in an additional figure [see Additional file 1]. Clinical characteristics were similar between included and excluded patients. These are summarized in an additional table [see Additional file 2]. 45/60 ETN starters and 45/74 ADA starters were subsequently matched on propensity score, for whom characteristics used in the propensity score model were similar (Table 1). Further characteristics of the matched patients are summarized in an additional table [see Additional file 3]. The AUC of the propensity score model fitted in the matched baseline cohort was low (0.56, 95% CI: 0.32 – 0.56), indicating a good balance of confounders between ETN and ADA starters. The percentage of patients with a child version JAMAR was comparable for ETN (33.3%) and ADA starters (37.8%). Moreover, the median year of starting ETN (2015, interquartile range (IQR): 2015 – 2016) was close to the median year of starting ADA (2016, IQR: 2015 – 2016). Patients who started ETN had a longer disease duration than patients who started ADA (median 2.9 years versus median 1.5 years, P = 0.31). The median VAS pain score in the overall matched cohort was 4.0 (IQR: 1.0 – 6.5), median VAS well-being score was 4.0 (IQR: 1.5 – 6.0) and median active joint count was 3.0 (IQR: 1.0 – 5.8). The median duration from completing a JAMAR to starting a b-DMARD was similar (P = 0.15) for ETN (0 days, IQR: 0 – 1) and ADA starters (0 days, IQR: 0 – 7).

The median duration from starting a b-DMARD to completing a follow-up JAMAR was not significantly different (P = 0.51) for ETN (183 days, IQR: 168 – 199) and ADA (176 days, IQR: 168 – 195) starters. The distribution of days from starting a b-DMARD to completing a baseline and follow-up JAMAR is provided in an additional figure [see Additional file 4]. At follow up, 42/45 (93%) ETN starters still used ETN and 36/45 (80%) ADA starters still used ADA (P = 0.12). VAS well-being scores at follow-up were better (P = 0.01) for ETN starters (median 0.0, IQR: 0.0 – 1.0) than ADA starters (median 1.0, IQR: 0.0 – 3.5) (Fig. 2). Nevertheless, a median improvement in VAS well-being of 2 was observed for both ETN (IQR: 0.0 – 5.0) and ADA (IQR: 0.0 – 4.0). The estimated mean difference in VAS well-being improvement for ETN versus ADA starters was 0.89 (95% CI: -0.01 – 1.78) (Table 2). For both groups, 3 patients reported considerable worsening of well-being (VAS well-being increase of ≥ 2). Median active joint count at follow-up was 0 for both ETN and ADA starters (Fig. 3). The estimated mean difference in active joint count decrease for ETN versus ADA starters was -0.36 (95% CI: -1.02 – 0.30). At follow-up, 11 (24.4%) ETN starters and 15 (34.9%) ADA starters reported adverse events. The estimated odds ratio for adverse events between the two groups was 0.48 (95% CI: 0.16 – 1.44). MTX co-medication at follow-up was common for both ETN (60%) and ADA (67%) starters. Patients who started ETN reported more gastric complaints than patients who started ADA, whereas the latter group reported more mood swings and sleep disturbances (Table 3). During follow-up, one event of uveitis occurred in the ETN group.

When analyzing all follow-up measurements of the full unmatched cohort of patients meeting the positivity assumption (n = 134), median VAS well-being was 0.5 (IQR: 0.0 – 2.0), median active joint count was 0 (IQR: 0 – 0), 36/132 patients (27.3%) reported adverse events and no additional events of uveitis were reported. Median improvement of VAS well-being and decrease in active joint count from baseline was 2.0 (IQR: 0.0 – 4.3) and 3.0 (IQR: 1.0 – 6.5) for ETN starters and 1.8 (IQR: 0.0 – 4.0) and 2.0 (IQR: 1.0 – 4.0) for ADA starters, respectively. 15 ETN starters (25.0%) and 21 ADA starters (29.2%) reported adverse events. While adjusting for propensity score, the estimated mean difference in VAS well-being improvement for ETN versus ADA starters was 0.70 (95% CI: -0.05 – 1.45). The estimated mean difference in active joint count decrease for ETN versus ADA starters, adjusted for propensity score, was -0.37 (95% CI: -1.27 – 0.52). Finally, the adjusted odds ratio for adverse events between the two groups was 0.45 (95% CI: 0.17 – 1.19). The results from follow-up measurements for the unmatched cohort are summarized in an additional table [see Additional file 5].