Study setting and data source
The ATTRA registry, established in 2001, is a non-interventional, prospective, national, observational cohort study. Its primary purpose is to evaluate the safety and effectiveness of bDMARDs/tsDMARDs in patients with chronic inflammatory rheumatic diseases. Patients with RA (and ankylosing spondylitis, psoriatic arthritis, juvenile idiopathic arthritis and systemic lupus erythematosus) starting bDMARDs or tsDMARDs are recruited from fifty-six practices sites (private or academic), and the registry captures more than 95% of patients with RA treated with bDMARDs/tsDMARDs in the Czech Republic (CZ).
At the start of therapy, baseline data are collected including demographics (gender, age at diagnosis, age at the start of 1st line treatment, height, weight, presence of comorbidities), disease characteristics (disease duration, presence of rheumatoid factor (RF) and anti-citrullinated protein antibodies (ACPA), presence of joint erosions on X-ray), disease activity (swollen and tender joint count (0–28), patient global assessment (PtGA) of disease activity and physician global assessment of disease activity (MDGA) on a 100-mm visual analogue scale (VAS; 0 – best, 100 – worst), erythrocyte sedimentation rate (ESR, mg/h) and C-reactive protein (CRP, mg/L)) and 28-joint disease activity score index (DAS28; 0–10) (11), Health Assessment Questionnaire (HAQ) for patient function with values from 0 to 3 (the higher, the worse disability) (12), EuroQol EQ-5D questionnaire for quality of life with values from − 0.59 to 1 (the higher, the better quality of life) (13), and current or previous anti-rheumatic therapies and therapy with glucocorticoids (GCs). Follow-up data on disease activity, disease function and antirheumatic therapies are collected after three and six months, and then every six months for three years, with disease activity and anti-rheumatic therapy data collected annually after that.
for ATTRA was granted by the Czech Multicentre Research Ethics Committee (no. 201611 S300) and Institutional Ethics Committee of Institute of Rheumatology, Prague, Czech Republic (no. 10113/2016). No additional ethical approval was required for the current analysis. All subjects provided their written consent for collecting and storing data before participation. All procedures were performed following the Declaration of Helsinki.
In this study, we used two separate datasets for analyses to validate our results – primary dataset (older cohort) and validation dataset (newer cohort). The primary dataset included all bio-naive adult patients diagnosed with RA starting TNFi therapy within a period from the registry data collection start (2001) until 31/12/2017. The validation dataset consisted of all bio-naive adult patients with RA diagnosis starting TNFi therapy between 01/01/2018 and 01/01/2020. Patients without filled SF36 questionnaire at baseline and without at least oneyear follow-up with available 6-month and 12month visits were excluded from the analysis (see flow charts Fig. 1).
We divided patients meeting the inclusion criteria according to their response (definitely/mostly yes, definitely/mostly no, do not know) to Q11A 'I seem to get sick a little easier than other people', and Q11C 'I expect my health to get worse' at baseline. We further analysed only patients who answered definitely/mostly yes/no, because we focused only on decisive patients. Therefore, patients who responded 'definitely yes' and 'mostly yes' were analysed together (as well as patients responding 'definitely no' and 'mostly no'). Patients' subgroups based on their responses are shown in pie charts Supplementary Fig. 1. We used two separate cohorts (primary and validation datasets) to validate our results. As part of a sensitivity analysis, we performed the whole analysis on the propensity-score matched datasets as well.
Objectives and Outcome measures
Our goal was to investigate whether the two selected SF-36 questions Q11A 'I seem to get sick a little easier than other people' and Q11C 'I expect my health to get worse', could predict therapeutic response in patients starting their first TNFi therapy. The therapeutic response was evaluated through remission achievements throughout the first year of TNFi therapy and drug retention.
Our primary outcome was remission (REM) achievement at 6 and 12 months since TNFi treatment initiation. Remission was defined through the disease activity index as DAS28-ESR < 2.6. Besides remission rates, odds ratios (ORs) of remission with 'no' group as a reference were calculated. Our secondary outcome was drug retention, computed as the time from the first-line TNFi initiation until the date of drug discontinuation (for any reason) or the last update of patients in the registry. Primary and secondary outcomes were evaluated across studied subgroups ('definitely/mostly yes' vs 'definitely/mostly no') in both datasets (primary and validation) and propensity-score matched datasets afterwards.
A descriptive summary of patients' demographic and treatment characteristics and disease activity measurements was performed for patients answering 'definitely/mostly yes' and 'definitely/mostly no' to Q11A and Q11C. For continuous variables, we calculated the median with interquartile range (IQR, 25th–75th percentiles). For a description of categorical variables, we used absolute and relative frequencies (i.e., percentages). We performed the non-parametric Mann-Whitney U test for continuous variables (after normality checks) and Pearson's chi-squared test for categorical variables to test differences between two patients' groups. In case the assumption of Pearson's chi-squared test was violated, Fisher's exact test was used instead. For all tests, P values < 0.05 were considered to be statistically significant.
We computed univariable logistic regression models to obtain odds ratios of remission achievement after 6/12 months of treatment for patients answering 'yes' vs 'no' to studied questions. Next, we performed multivariable logistic regression models with baseline HAQ and DAS28-ESR to obtain odds ratios adjusted for potential confounders.
Drug retention was computed through the Kaplan-Meier survival method. Drug survival probabilities were displayed through Kaplan-Meier curves and supplemented by numbers of patients at risk beneath the graphs. We also present numbers of discontinuations, one-year and two-year survival rates and median survival time with corresponding confidence intervals. The probabilities of drug discontinuations were compared across the studied groups through the Log-rank test. If the curves were crossing, we also computed the Breslow and Tarone-Ware tests. Finally, we employed Cox regression models to estimate hazard ratios (HRs) for treatment discontinuation for patients answering 'yes' vs 'no' to studied questions. Besides crude hazard ratios, we obtained adjusted versions with baseline HAQ and DAS28-ESR as confounders.
For the sensitivity analysis, we created balanced datasets for both subgroups (answering 'yes' and 'no'). We used propensity score matching to match patients answering 'yes' to patients responding 'no' within each studied question. We performed logistic regression with the outcome variable 'yes' (= 1) vs 'no' (= 0) and selected baseline covariates for matching. The covariates were chosen based on statistically significant differences in baseline characteristics with respect to clinical relevance and multicollinearity. We chose the matching ratio 1:1 and set the caliper to 0.2. The adequacy of the final propensity score model was checked through the balance diagnostics (standardised mean differences should be less than 0.1 to ensure balance in selected covariates). We used matching to make both groups comparable in baseline characteristics and to minimise confounding by other factors in evaluating REM achievements at the 6-/12-month visit and in the evaluation of drug retentions.
We did not impute missing data in this analysis and performed an availablecase analysis instead. We used IBM SPSS Statistics 25.0 to compute all descriptive statistics and comparisons. The propensity score model was performed in R (version 3.5.3).