Smoking Status and Cause-Specific Discontinuation of Tumour Necrosis Factor Inhibitors in Axial Spondyloarthritis

Sizheng Steven Zhao; Kazuki Yoshida; Gareth T. Jones; David M. Hughes; Stephen J. Duffield; Sara K. Tedeschi; Houchen Lyu; Robert J. Moots; Daniel H. Solomon; Nicola J. Goodson

Disclosures

Arthritis Res Ther. 2019;21(177) 

In This Article

Methods

Study Design and Population

We used data from the British Society for Rheumatology Biologics Register for Ankylosing Spondylitis (BSRBR-AS), a UK-wide prospective cohort study of patients fulfilling the Assessment of SpondyloArthritis international Society (ASAS) criteria for axSpA.[8] The current analysis focused on new users of TNFi (Humira, Enbrel/Benepali, Cimzia and Simponi) from December 2012 to June 2017. Patients were assessed at baseline, 3, 6 and 12 months and annually thereafter, with additional follow-ups permitted in the interim. TNFi start and stop dates were ascertained from their medical notes. Participants who completed a valid baseline questionnaire (within 1 year before and 7 days after starting TNFi) and recorded smoking status were eligible for analysis.

Exposure and Outcomes

Self-reported smoking status at baseline (current, ex- or never) was used to define the exposure. The outcomes of interest were time to all-cause and cause-specific TNFi discontinuation. Reasons for stopping were categorised in the BSRBR-AS as adverse events, inefficacy, symptom in remission, or other, with accompanying free-text to further elaborate. Each entry was manually reviewed and, where necessary, re-categorised. To allow comparison with existing studies, and so that each specific cause has sufficient events with which to perform the analyses, we grouped causes as (1) infection-related adverse events, (2) all other adverse events, and (3) inefficacy or other reasons (e.g. patient choice and social circumstances). The censoring date was defined as the last study visit or, if no data existed past the baseline visit, as 3 months (the minimum per protocol follow-up interval).

Covariates

The following covariates were recorded at baseline and chosen a priori for their possible associations with TNFi discontinuation based on prior literature:[5] age, gender, symptom duration, education, elevated baseline CRP (above upper normal limit), classification as ankylosing spondylitis (modified New York criteria), social deprivation (quintiles),[9] alcohol use (current, previous, never), comorbidity (categorised as 0, 1 or ≥ 2 from a list of 13 conditions[8]), body mass index (BMI), TNFi agent and year of TNFi initiation. Patient-reported variables were measured at each follow-up; those known to be associated with TNFi discontinuation include BASDAI, spinal pain, functional impairment (BASFI) and fatigue (Chalder Fatigue Scale). Since these variables capture more than just disease activity, we refer to them collectively as "disease severity" throughout the text.

Statistical Analysis

Baseline participant characteristics were summarised by smoking status. Unadjusted comparison of time to all-cause discontinuation according to smoking status was made using Kaplan-Meier estimators.

Prior studies considered each cause of TNFi discontinuation as censoring.[10] In traditional survival analysis, observation time for each subject is censored when they leave the study or at the end of follow-up. Furthermore, a key assumption of Cox models is that censored patients should be representative of remaining individuals at that time point (i.e., censoring is random). Patients who discontinue treatment are not representative of those who continue; therefore, Cox models cannot be used.

Marginal structural models use inverse-probability weights to account for confounders rather than adjusting for them in conventional outcome regression.[11] This approach is more flexible, allowing us to account for baseline imbalance in characteristics between smoking status, time-varying disease activity, and competing risk events within one model. For all following analyses, differences in baseline covariates between smoking categories were balanced using inverse probability of "treatment" weights (IPTW)[11] to allow unconfounded descriptive comparisons.[12] Causal inference is not straightforward regarding the effect of baseline smoking status on TNFi discontinuation, since we cannot randomly assign an individual to "having smoked for 20 years" at the onset of a hypothetical trial. IPTW provides an approach to estimate causal effects under theoretical ideal conditions and may be advantageous over regression adjustment (see Additional file 1). We derived IPTW using predicted values from a multinomial logistic regression model, using baseline smoking status as the dependent variable and all baseline covariates specified above as independent variables.

We applied baseline time-invariant inverse probability of censoring weights (IPCW) to account for participants excluded from the sample eligible for analysis, such that baseline characteristics of the analysis set resembled the eligible TNFi exposed cohort. These IPCWs were constructed in the same manner as IPTWs, with inclusion/exclusion status as the dependent variable.

We used marginal structural Cox proportional hazards models to estimate hazard ratios of TNFi discontinuation according to baseline smoking status.[11,13] Parameters were estimated using weighted pooled logistic models that are equivalent to Cox models when using discrete time[14] and allow the use of subject-specific time-varying weights. Time was split into integer months.[13]

Censoring becomes non-random when competing risk events are specified as censoring. We therefore calculated time-varying IPCWs, such that censoring becomes random at each time point with respect to baseline characteristics and history of disease activity.[15] For analysis of each specific cause of TNFi discontinuation (e.g. infection), the other competing risk events (e.g. other reasons) were treated as censoring events.

Weighted pooled logistic models included only the outcome, exposure (smoking status) and time, which was included in all models as restricted cubic splines.[13] All weights were "stabilised" to have a mean of 1, allowing the overall sample size to remain unchanged.[16] Further details of the methods are given in.[13] Multiple imputation was used for missing covariates. Analyses were conducted using Stata version 13.

processing....