Background/Purpose: Informative censoring in epidemiologic research studies (the loss of data from the risk set that is related to the exposure and/or outcome) can cause bias in the estimation of causal risks. An example of this bias occurs when there is a competing risk of death for an outcome that occurs in older individuals, such as dementia. Gout and hyperuricemia have been associated with a lower risk of dementia in observational studies, leading to questions about the potential neuroprotective role of hyperuricemia. However, these prior findings may result of this type of bias. In this situation, gout may appear to lower the risk of dementia because many patients with gout do not live long enough to develop it. The purpose of this study was to evaluate distinct methods to estimate causal risks of gout on the risk of dementia using data from the Veterans Affairs (VA) health system.

Methods: We studied patients with gout and age- and sex- matched controls from national VA health record data. We defined incident cases of gout based on the presence of an ICD-9 or ICD-10 code from at least two separate encounters at least 30 days apart and matched to controls of similar age, sex, and calendar year. The outcome of dementia was defined using previously validated administrative algorithms. An extensive list of comorbidities, medications, and other covariates were extracted from the medical record and defined using validated algorithms. We tested three approaches to estimating the causal risks of gout: 1) a traditional Cox model, 2) a commonly used competing risks model (Fine & Gray) and 3) a joint frailty model that accounts for informative censoring by considering the propensity for dementia among those that died (Figure 1). Due to large sample sizes and computational requirements, we meta-analyzed up to 30 models for each method run on the same randomly selected samples of 0.1% of the population adjusting for potential confounders identified in the descriptive analysis.

Results: There were significant and expected differences in patient characteristics between the 533,465 patients with gout and 4,935,717 matched controls including higher body mass index (BMI) and higher rates of cardiovascular comorbidity (not shown). In the traditional adjusted Cox model that assumes that censoring is uninformative, the risk of dementia among patients with gout was significantly lower than controls [HR: 0.93 (95% CI: 0.87, 0.98)]. The protective effect was larger in Fine & Gray competing risks models [HR 0.90 (95% CI: 0.85, 0.96)]. In contrast, the joint frailty model accounting for informative censoring suggested no significant reduction in dementia risk in gout compared to controls [HR: 0.97 (95% CI: 0.91, 1.03)] (Figure 2).

Conclusion: Estimates of the risk of dementia in patients with gout varied according to the method used to address the competing risk of death. Analyses performed using a method that more comprehensively accounts for informative censoring suggested no reduction in the risk of dementia among patients with gout. These findings call into question the results of multiple prior studies and should help to reassure clinical providers that urate-lowering to manage gout is not likely to increase the long-term risk of dementia.

Figure 1: Illustration of modeling approaches to consider the competing risk of death.

Figure 2: Hazard Ratios for the risk of dementia for patients with gout based on 3 distinct modeling approaches: 1) Cox proportional hazards model (circle), 2) Fine and Gray competing risks model (triangle), and 3) the joint frailty model (square).

« Back to ACR Convergence 2025

ACR Meeting Abstracts - https://acrabstracts.org/abstract/methods-to-address-survival-bias-and-competing-risks-in-estimating-the-causal-risks-of-gout-on-dementia-risk/