The impact of smoking on disease measures in rheumatoid arthritis: the need for appropriate adjustment of time-varying confounding

Abstract

In a recent publication, Quintana-Dunque et al. studied patients with early onset rheumatoid arthritis (RA) and showed that baseline smoking status was inversely associated with disease activity and disability at 36 months. The authors conclude that smoking may not be as deleterious as previously considered in RA disease course. However, the authors fail to highlight several limitations of study design and analysis, including time-varying confounding, which may have a direct impact on results and corresponding conclusions.

In the publication, “The impact of smoking on disease activity disability, and radiographic damage in rheumatoid arthritis: is cigarette protective?” [1], Quintana-Dunque et al. studied 129 patients with early onset rheumatoid arthritis (RA) and showed that baseline smoking status (ever vs. never; current vs. never) was inversely associated with disease activity and disability at 36 months. There was no association between smoking status and erosive disease, radiographic progression, or Sharp scores. The authors conclude that smoking may not be as deleterious as previously considered in RA disease course.

However, the authors fail to highlight several limitations of study design and analysis, which have a direct impact on results and corresponding conclusions. We outline a few of those limitations below.

First, by capturing smoking status only at baseline and not accounting for changes over time, the authors make an assumption that smoking status did not change during the course of three years. This may result in exposure misclassification, which potentially may bias findings toward or away from the null.

Second, there may be other factors that also vary with time. For example, therapy (medication) is likely to change over time, and we are only presented with baseline information. Additionally, factors such as therapy, depression, and obesity are likely strong predictors of disease activity and disability, but are not adjusted for in the analyses.

Improper adjustment for time-varying confounding variables (i.e. changes in exposure status or covariates over time) can lead to biased results, specifically when a covariate serves as both a confounder and an intermediate variable for an exposure of interest [2]. Variables such as medication use, depression, and obesity, serve as both confounders and intermediate variables in the relationship between smoking and disease activity or disability (Figure 1): adjusting for these variables may attenuate the estimated association between exposure and outcome because they lie in the causal pathway as intermediate variables; but not adjusting for them can lead to biased results as they still represent confounders. Standard regression methods fail to appropriately account for this type of bias, but other methods, including marginal structural models (MSMs), through g-computation or inverse probability weighting, or even further, models that do not make assumptions on model form, such as targeted maximum likelihood estimation, are available [3–6].

Figure 1.

Directional acyclic graph demonstrating the longitudinal relationship between current smoking and disease activity over time

Differences in results between standard analytic methods and MSMs, which account for time-varying confounding, have been conducted in previous studies in a variety of fields [7, 8]. A review found substantial differences between MSMs and standard analyses in studies where time-varying confounding was suspected [8]. MSM estimates differed from estimates derived through standard analyses by at least 20% in approximately 40% of exposure-outcome associations in which the direction of the effect was the same, and approximately 11% of papers showed opposite interpretations between the two types of analyses [8].

Given that large differences in conclusions may be found depending on the type of statistical analysis conducted, especially in the context of longitudinal studies and time-varying confounding, it is important for researchers to be mindful of the limitations of certain methods and the biases they may produce.