Lead Investigator: Sara Tedeschi, Brigham and Women’s Hospital
Title of Proposal Research: Causal inference cohort re-analysis of the ‘CARES’ trial to better understand the roles of colchicine, allopurinol, and febuxostat on cardiovascular events among patients with gout
Vivli Data Request: 7142
Funding Source: None
Potential Conflicts of Interest: Dr. Hyon Choi reports research support from Ironwood and Horizon and consulting fees from Ironwood, Selecta, Horizon, Takeda, Kowa, and Vaxart for unrelated projects. This proposed research project is not funded by any of these commercial entities.
Summary of the Proposed Research:
Gout is one of the most common inflammatory arthritis, which reportedly affects about 4% of adults in the US. Gout is caused by the deposition of urate crystals. Thus, urate-lowering medications such as febuxostat and allopurinol are used to prevent future gout attacks. Concerns about the cardiovascular safety of urate-lowering medications resulted in two major clinical trials among gout patients: Cardiovascular Safety of Febuxostat and Allopurinol in Patients with Gout and Cardiovascular Morbidities (CARES) and Febuxostat versus Allopurinol Streamlined Trial (FAST). However, these studies gave somewhat contradictory results.
CARES showed an increase in deaths among febuxostat users in one of the analyses. However, many patients did not complete all planned study visits in this trial, making the interpretation difficult. On the other hand, some analyses in FAST favored febuxostat over allopurinol. However, there were substantial differences in the proportion of patients stopping urate-lowering therapy and the use of colchicine in FAST. These idiosyncrasies have led to lingering questions which call for more advanced analytical approaches.
1. Did the use of colchicine (cardio-protective in other trials) affect the CARES trial results?
2. Did the higher loss to follow-up in the allopurinol group in CARES bias the results?
In this proposed study on CARES, we will use advanced causal inference methods to address these questions. Causal inference is an epidemiological approach to measure the accurate effects of treatment. Although clinical trials are the gold standard of medical research, more advanced analytical methods are needed to overcome imperfections of clinical trials and accurately estimate the impact of medications.
In Aim 1, we will first compare the effect of colchicine prophylaxis on heart attacks compared to no prophylaxis or non-steroidal anti-inflammatory drug prophylaxis using the data from CARES. We will then conduct causal mediation analysis, an approach quantifying the extent to which an intermediate factor (colchicine use during the study in this case) impacted the trial results. In Aim 2, We will estimate the “per-protocol effect” of febuxostat compared to allopurinol on heart attacks. This “per-protocol effect” approach estimates the effect of these medications statistically correcting for the influence from patients who did not complete all planned study visits. Our methodologically innovative project is expected to help reconcile the discrepant results from CARES and FAST, providing reassurance to practitioners. We will also demonstrate the usefulness of causal inference methods in clinical trial analysis for the rheumatology research community at large. We will further disseminate these advanced methods via online and offline methodology tutorials.
Statistical Analysis Plan:
Aim 1. Examine the impact of colchicine on expert-adjudicated cardiovascular (CV) events among patients with gout.
Analysis (1a): In Cardiovascular Safety of Febuxostat and Allopurinol in Patients with Gout and Cardiovascular Morbidities (CARES), the gout flare prophylaxis medications were not randomized and at the discretion of treating physicians. Thus, we will conduct an observational comparative effectiveness study. To control for the potential confounding by indication, we will use propensity score matching weights approach. We previously demonstrated that the matching weights approach could control confounding better than other propensity score approaches, even with a moderate to high level of patient background differences across exposure groups. A multinomial logistic regression model will be fit using the prophylaxis group as the outcome variable and baseline CV risk factors as predictors. The probability of each prophylaxis regime (colchicine, non-steroidal anti-inflammatory drugs [NSAIDs], and none) is predicted based on each individual’s CV risk factor status. Each individual’s matching weight is defined as the minimum of the three predicted probabilities of prophylaxis regimes over the predicted probability of the assigned prophylaxis (one that was received). Once the weights are constructed, we will examine the confounding control by covariate balance via standardized mean difference (< 0.1 is typically considered sufficient confounding control). The subsequent Cox regression will incorporate the calculated matching weights. We will estimate the major adverse cardiovascular events (MACE) outcome hazard ratios by comparing the colchicine users to each one of NSAIDs users and no gout prophylaxis users. The 95% confidence interval will use the robust variance estimator.
Analysis (1b): We will quantify the mediating role of colchicine use via the regression-based causal mediation analysis using a publicly available R package that we previously developed (https://CRAN.R-project.org/package=regmedint). This approach requires simultaneous modeling of the exposure-mediator relationship and the exposure-outcome relationship. For the exposure-mediator regression model, we will specify a logistic regression model to quantify the influence of the febuxostat group assignment on the subsequent colchicine use during follow-up. For the exposure-outcome regression model, we will use a Cox regression model for the MACE outcome. The Cox model will include febuxostat vs. allopurinol group assignment (exposure), colchicine use during the follow-up (mediator), and their interaction. Both the exposure-mediator regression and exposure-outcome regression will adjust for the baseline covariates. These two regression models will be combined by the R package to produce the estimates of the following quantities:
Total Effect: febuxostat’s total impact on MACE irrespective of colchicine,
Natural Direct Effect: febuxostat’s direct impact on MACE, not by promoting increased colchicine use, and
Natural Indirect Effect: febuxostat’s indirect impact on MACE mediated by promoting increased colchicine use during follow-up.
We will summarize the extent of mediation as the proportion mediated, the proportion of the natural indirect effect within the total effect.
Aim 2. Assess febuxostat’s CV safety compared to allopurinol adjusting for selection bias in follow-up.
The steps involved are the following: longitudinal data construction, inverse probability weight estimation, outcome model estimation, and cumulative incidence prediction.
Longitudinal data construction: The main published analyses of CARES did not incorporate the time-varying covariates, although such variables were collected every 2 months. We will, thus, first construct longitudinal analysis datasets with fixed 2-month CARES intervals. When multiple values are available, a time-varying continuous variable will be averaged. For dichotomous variables, such as drug, presence at any visits during the interval will be deemed as present during the interval. Any MACE during an interval will be considered as a presence of the outcome of interest during that interval. For missing values, we will use multi-level multiple imputation via chained equations.
Inverse probability weight estimation: We will construct prediction models for the loss to follow-up and study drug discontinuation during follow-up separately. In each, the “numerator model” will use baseline variables (treatment and covariates) only, whereas the “denominator model” will fully incorporate the time-varying variables. The predicted probability of loss to follow-up (or drug discontinuation) during each interval will be estimated. The final so-called stabilized inverse probability weights will be the cumulative product of the predicted probability of the observed loss to follow-up (or drug discontinuation) status from the “numerator model” over that from the “denominator model” during each interval. We will censor an individual at the interval the assigned urate-lowering therapy (ULT) (febuxostat or allopurinol) is discontinued.
Outcome model estimation: The subsequent outcome analyses will be conducted with a weighted pooled logistic regression model that has the exposure term (febuxostat or allopurinol), spline basis terms for study time, and time-fixed baseline covariates fitted with the generalized estimating equation. The exposure term coefficient will approximate the “per-protocol effect” hazard ratio (HR).
Cumulative incidence prediction: To provide a better understanding of absolute risk difference by the assigned ULT, we will use the fitted model to predict the cumulative incidence curve for each outcome. In addition to the graphical representation, we will estimate the restricted mean time difference as a summary measure.
A Multicenter, Randomized, Active-Control, Phase 3B Study to Evaluate the Cardiovascular Safety of Febuxostat and Allopurinol in Subjects With Gout and Cardiovascular Comorbidities
Data Contributor: Takeda
Study ID: NCT01101035
Sponsor ID: TMX-67_301
Tedeschi S, Hayashi K, Zhang Y, Choi H, Solomon D. Identifying Optimal Serum Urate Levels to Reduce Gout Flares in Patients Taking Urate Lowering Therapy: A Post-hoc Cohort Analysis of CARES with Consideration of Drop-out [abstract]. Arthritis Rheumatol. 2023; 75 (suppl 9).