Study Design: Cohort study.
Data source: Data from CPRD were used. Incepted in the year 1987, CPRD is a longitudinal anonymised electronic database containing health records of >10 million people in the UK(9). CPRD participants are representative of the UK population in terms of age, sex, and ethnicity(9).
Ethical approval: ISAC of the Medicines and Healthcare products Regulatory Authority (ISAC Reference: 18_227R).
Study population: Age ≥40 years, diagnosed with knee or hip OA between 1st January 1990 and 31st December 2013, at-least 2-year disease and exposure free prior registration in the CPRD before OA diagnosis, and contributing acceptable research quality data in up-to-standard GP practices (Appendix 1).
Exposure: New continuous β-blocker prescription, defined as ≥2 prescriptions of β-blockers within a 60-day period after the first OA diagnosis (new user design).
Unexposed: Participants without a prescription of β-blocker, or with a single β-blocker prescription after OA diagnosis date, matched to exposed participants for age at OA diagnosis (5-year age band), sex, OA location (knee or hip) and propensity score (PS) for β-blocker prescription (Appendix 1).
Start of follow-up (index date): Date of first β-blocker prescription for the exposed. The duration between OA diagnosis date and first β-blocker prescription date in the exposed was added to the OA diagnosis date of the matched unexposed to obtain their start of follow-up date. Thus, the exposed and matched unexposed participants had the same duration of OA prior to start of follow-up. This minimised any potential bias due to unequal disease duration prior to start of follow-up in exposed and unexposed participants as the risk of joint replacement increases with duration of OA(10). If this approach was not taken, exposed participants would have had longer follow-up with OA prior to cohort entry (time taken to develop a comorbidity for which beta-blockers may be prescribed) and consequently be at higher risk of outcome.
Outcomes: [1] Knee or hip TJR (primary outcome), [2] knee TJR, and [3] hip TJR.
Exclusion criteria:
[1] Consultation for the below prior to start of follow-up:
-
Conditions causing chronic pain: autoimmune inflammatory rheumatic diseases (rheumatoid arthritis, psoriatic arthritis, ankylosing spondylitis, lupus, polymyalgia rheumatica); gout; radiculopathy; neuropathy; and fibromyalgia.
-
Contra-indications for β-blockers: chronic obstructive pulmonary disease or asthma; peripheral vascular disease; heart block, aortic stenosis, and hypertrophic obstructive cardiomyopathy.
[2] Knee or hip TJR prior to or within 90 days of start of follow-up. Knee or hip TJR within 90 days after start of follow-up were excluded as we do not expect β-blockers to influence the rate of TJR immediately, and they may have been commenced during a pre-anaesthetic check-up.
End follow-up: Exposed participants were followed up from the index date. Participant follow-up ended at the earliest of date of first outcome, death date, transfer out date, date of last data collection, or study end (31/12/2018).
Ascertainment of exposure, outcomes and covariates: Read codes and product codes were used to ascertain these factors (Appendix 2).
Statistical Analyses: As participants prescribed β-adrenoreceptor blocking drugs are likely to have comorbidities, be older and have a high body mass index (BMI), a PS for β-blocker prescription was calculated using a cumulative logit regression model. Greedy nearest neighbour 1:1 matching, without replacement specifying a maximum calliper width of 0.001, was undertaken to match the exposed to unexposed participants. Missing values for BMI and smoking status were categorised as missing data for the purpose of PS matching as people with healthy lifestyle and normal BMI are more likely to have missing data in consultation-based databases such as CPRD where lifestyle and demographic factors are collected opportunistically(11). Mean, standard deviation (SD), n (%), and standardised mean difference (SMD) were used to examine the covariate balance between matched exposed and unexposed participants. Any variables that were in imbalance after PS matching were included in the model if the SMD was >0.10 as recommended(12).
Hazard ratios (HRs) and 95% confidence interval (CI) were calculated for each incident outcome (first Read code for the event) using a Cox proportional hazards model. Covariates that are likely to influence outcomes, e.g. progression of OA, or reflect health-seeking behaviour were included in the Cox model for additional confounder adjustment. These were:
-
number of GP consultations for knee or hip injury, non-osteoporotic fractures (defined as fractures in any bone except the femur, distal radius, and vertebrae after the age of 18 years but before the age of 50 years in women and 60 years in men) prior to start of follow-up,
-
number of analgesic prescriptions between the first consultation for knee/hip OA and start of follow-up,
-
number of GP consultations, hospital out-patient referrals, hospital admissions in the 12-month period preceding start of follow-up,
-
bisphosphonate or glucosamine/chondroitin sulphate prescription in the 12-month period prior to start of follow-up,
-
new diagnosis of interphalangeal or thumb base OA, neck or back pain or spinal degenerative diseases after start of follow-up.
These analyses were stratified according to class of β-blocker drug used namely, β1 selective, intrinsic sympathomimetic activity, membrane stabilising effect, and lipophilic properties (low versus high). Number needed to treat (NNT) and 95% CI for a 2-year treatment duration were computed using the aHR and survival probability in control group as described previously(13).
Additionally, we performed multiple imputation to handle missing BMI values and smoking status using chained equations as a sensitivity analysis. Demographic factors, relevant diagnoses and prescriptions (Appendix 3), covariates that are likely to influence outcomes or reflect health-seeking behaviour listed above, primary outcome variable, and Nelson-Aalen estimator of baseline cumulative hazard were included in the imputation model as recommended(14). Ten imputed datasets were created to account for random variability(15). PS calculation, matching and Cox regression analyses were undertaken in each imputed dataset. However, we did not find substantial difference between the results with missing values as a dummy category and the imputed values (see Tables S1, S2, and S3 in the supplementary material). Thus, results are only reported with the missing category approach. Data management and analysis were performed in Stata MP v15.