The National Health and Nutrition Examination Survey (NHANES) is a nationally representative, cross-sectional surveillance program conducted by the Centers for Disease Control and Prevention (CDC) to systematically assess the health and nutritional status of the United States population. As the only national survey that combines comprehensive health interviews, standardized physical examinations, and detailed laboratory testing across all age groups, the NHANES database offers a unique and robust data source for population health surveillance. The survey collects multidimensional data, including sociodemographic characteristics, self-reported health behaviors and conditions, clinically validated physical measurements, and objective biomarkers from biological specimens, thereby supporting rigorous epidemiological research and evidence-based policy development.

The present study utilized data from 10 consecutive National Health and NHANES cycles, covering the period from 1999 to 2018. To maintain consistency in the availability of key exposure variables, data from the 2011–2014 cycles were excluded, as these did not include the specific exposure measures required for our analysis. The remaining 8 cycles (1999–2010 and 2015–2018) provided a robust, nationally representative sample for our investigation, allowing for a comprehensive evaluation of trends and associations over nearly two decades.

The National Health and NHANES, administered by the National Center for Health Statistics (NCHS) of the CDC, is a nationally representative survey of the United States civilian, non-institutionalized population. Approved by the NCHS Research Ethics Review Board, NHANES collects data with informed written consent from all participants. The datasets used in this study are publicly available on the NHANES website¹.

In this study, rigorous exclusion criteria were applied to ensure the integrity and validity of the analysis. Participants were excluded based on the following criteria: (i) age under 18 years; (ii) pregnancy; (iii) incomplete or missing data on arthritis status; (iv) missing data on CRP or ALB levels; (v) missing data on osteoarthritis (OA), physical activity (PA), or other critical variables; and (vi) missing data on relevant covariates (Figure 1). After applying these criteria, a final analytical cohort of 27,033 participants was established.

To account for the complex survey design and ensure nationally representative estimates, Mobile Examination Center (MEC) weights were applied in accordance with NHANES analytical protocols. For the 1999–2000 and 2001–2002 cycles, weights were calculated using the formula 2/8 × wtmec4yr, while for the 2003–2010 and 2015–2018 cycles, the formula 1/8 × wtmec2yr was used. This methodological approach is consistent with established NHANES guidelines and enhances the generalizability and reliability of the study findings.

Herein, the CAR was calculated as the ratio of CRP concentration (mg/L) to ALB level (g/L). CRP and ALB measurements were obtained from blood samples collected at the MEC and subsequently analyzed in a central laboratory. Prior to collection, blood specimens were screened using predefined exclusion criteria to ensure data quality. NHANES also implemented rigorous quality control procedures throughout the collection and laboratory testing processes to ensure the accuracy and reliability of results. This standardized approach enhances the validity of CAR as a biomarker in our analysis. For regression analysis, the CAR variable underwent natural logarithmic (Ln) transformation to address the right-skewed distribution of inflammatory indices. This transformation improved data normality and strengthened the robustness of the statistical models.

Arthritis diagnosis was assessed using a validated self-report instrument (MCQ160a), which employed a binary response format (yes/no) to determine whether participants had received a clinician-confirmed arthritis diagnosis. For those who answered affirmatively, subsequent items identified the disease subtype through categorical options: “RA,” “Osteoarthritis,” “Psoriatic arthritis,” “Other,” “Refused,” and “Do not know.” This approach is consistent with previous validation studies demonstrating 85% concordance between self-reported arthritis diagnoses and clinical records, supporting its methodological validity for epidemiological surveillance in this cohort (19).

For this study, we extracted comprehensive data on various demographic, socioeconomic, lifestyle, and clinical factors. Demographic and clinical covariates included self-reported sex, age, race/ethnicity (categorized as Mexican American, Other Hispanic, Non-Hispanic Black, Non-Hispanic White, or Other Race), and education level. Anthropometric measures included body mass index (BMI), stratified using ≤ 25.0 kg/m², 25.0–30.0 kg/m², and > 30.0 kg/m². Socioeconomic status was assessed using the poverty-to-income ratio (PIR), categorized as <1.3, 1.3–3.5, and > 3.5. Lifestyle factors consisted of self-reported smoking history (defined as having smoked ≥100 cigarettes in a lifetime) and alcohol consumption in the past 12 months. Comorbidities included hypertension and diabetes. Laboratory parameters comprised complete blood count indices, focusing on CRP and ALB levels. These variables were selected to enable a multidimensional assessment of participant characteristics and health status, ensuring a robust analysis of the relationships under investigation.

All analyses were conducted using R statistical software (version 4.2.2; R Foundation for Statistical Computing). To account for the complex, multistage probability sampling design of the NHANES, we applied appropriate examination weights (wtmec4yr for combined cycles 1999–2002 and wtmec2yr for individual cycles) in all analyses. Continuous variables are presented as weighted means with standard deviations, while categorical variables are expressed as weighted percentages. Between-group comparisons were performed using design-adjusted χ² tests for categorical variables and weighted Student’s t-tests for continuous variables.

Given the right-skewed distribution of the CAR values, we performed a logarithmic transformation prior to analysis. Weighted multivariable logistic regression models were constructed to examine the association between RA and CAR across three progressively adjusted models: Crude Model: unadjusted analysis; Model 1: adjusted for demographic factors (age, sex, race/ethnicity); Model 2: further adjusted for socioeconomic status (PIR, education level) and BMI; and Model 3: additional adjustment for clinical covariates (diabetes mellitus, hypertension status) and behavioral factors (alcohol consumption, smoking status). To ensure the robustness of our findings, we conducted sensitivity analyses using complete-case analysis and unweighted multivariable logistic regression to evaluate potential sampling design effects. All analyses accounted for the NHANES complex survey design and incorporated appropriate sampling weights.

To evaluate the non-linear association between Ln(CAR) and RA risk, we categorized Ln(CAR) into quartile-based strata and applied restricted cubic splines (RCS) for flexible modeling of dose-response relationships. Advanced analytical techniques, including generalized additive models (GAMs) with smooth curve fitting, were used to investigate potential threshold effects and identify the inflection point where the risk association pattern changes. To assess potential effect modification, we performed interaction testing and stratified analyses across key sociodemographic and clinical factors, including age, PIR, BMI, race, education, diabetes status, hypertension, alcohol consumption, and smoking behavior. Statistical significance was defined as p < 0.05, with rigorous adjustment for multiple comparisons to control for Type I error. This comprehensive analytical approach allowed for the examination of heterogeneity in Ln(CAR)-RA associations across diverse population subgroups.

Variable selection was performed using least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation to determine the optimal penalty parameter (λ). Model discrimination was evaluated through receiver operating characteristic (ROC) curve analysis, with performance quantified by the area under the curve (AUC). We further assessed clinical utility using decision curve analysis (DCA) to estimate the net benefit across a range of probability thresholds.

The final prediction model was presented as a nomogram, which visually displays the weighted contribution of each predictor variable and enables calculation of individualized RA risk probabilities. Points are assigned for each risk factor based on its regression coefficient, with the sum of points corresponding to a predicted probability of RA development.

From Table 1: this study included 27,033 participants, consisting of 14,017 males (51.39%) and 12,926 females (48.61%). Of these, 18,674 individuals (79.93%) were under 60 years old, while 8,359 individuals (20.07%) were over 60 years old. The racial composition was as follows: 12,294 White People (69.60%), 540 Black People (10.59%), 5,351 Mexican Americans (8.33%), and 3,984 individuals from other races (11.48%). The study focused on four groups based on logarithm-transformed C-reactive protein-to-albumin ratio (LnCAR), with RA present in 239 individuals (2.69%) in Q1, 3,729 individuals (4.10%) in Q2, 488 individuals (5.78%) in Q3, and 645 individuals (7.78%) in Q4.

The results of the weighted multivariate logistic regression analysis demonstrate a strong association between an elevated CAR and an increased risk of developing RA (Table 2). This association remained consistent across sequentially adjusted models. In the unadjusted model, each unit increase in Ln(CAR) was associated with a 40.1% higher odds of RA (odds ratio [OR] 1.41; 95% CI, 1.34–1.48; P < 0.001). After adjusting for sex, age, and race in model 1, the effect estimate remained significant (OR 1.34; 95% CI, 1.27–1.41; P < 0.001). Model 2, which incorporated additional covariates such as PIR, BMI, and education level, yielded a modestly attenuated but still significant association (OR 1.21; 95% CI, 1.14–1.28; P < 0.001).

To evaluate the sensitivity of these findings and explore potential non-linear relationships, Ln(CAR) was analyzed as a categorical variable stratified by quartiles. Compared with the lowest quartile (Q1), individuals in the highest quartile (Q4) exhibited a 213% increase in the odds of RA in the unadjusted analysis (odds ratio [OR] 3.13; 95% CI, 2.51–3.91; P < 0.001). This elevated risk remained significant in Model 1 (OR 2.54; 95% CI, 2.02–3.18; P < 0.001) and Model 2 (OR 2.01; 95% CI, 1.59–2.56; P < 0.001), with a gradual attenuation of effect size as covariates were added. In the fully adjusted Model 3, the highest quartile of Ln(CAR) still conferred a 77% higher odds of RA compared with the lowest quartile (OR 1.77; 95% CI, 1.39–2.25; P < 0.001), highlighting the independent prognostic value of CAR in RA risk stratification.

To further assess the consistency of the CAR-RA association and the potential impact of weighting methods, we conducted sensitivity analyses using an unweighted approach (Table 3). These additional analyses confirmed the primary findings, showing a consistent positive relationship between elevated CAR and RA risk. In all models, from unadjusted to fully adjusted, the direction and statistical significance of this association remained consistent, with effect sizes similar to those in the weighted analysis. These results reinforce the robustness of our primary findings, establishing CAR as a reliable biomarker for RA risk stratification, regardless of the analytical weighting method used. The consistency across different methodological approaches further supports the biological plausibility and clinical relevance of the CAR-RA relationship.

Figure 2 illustrates the adjusted association between the natural LnCAR, modeled as a continuous variable, and the risk of incident RA. After comprehensive adjustment for demographic, socioeconomic, lifestyle, and clinical confounders, a clear linear dose-response relationship was observed, with higher LnCAR values corresponding to progressively increased RA risk. Specifically, each one-unit increase in LnCAR was associated with a significant and graded rise in the adjusted odds ratio (OR) for RA, indicating a strong and biologically plausible risk gradient. These findings remained consistent across sensitivity analyses, further supporting the independent role of systemic inflammation, as reflected by CAR, in RA pathogenesis.

Multivariable logistic regression models were constructed with adjustments for key confounders, including sex, age, race/ethnicity, BMI, educational attainment, diabetes status, hypertension, alcohol consumption, and smoking history. The solid red line represents the adjusted OR for RA per one-unit increase in natural LnCAR, while the shaded red area indicates the 95% CI. The overall association was highly significant (P < 0.001), and no evidence of non-linearity was observed, as indicated by a non-significant P-value (>0.05) from the likelihood ratio test comparing nested linear and non-linear models. These findings support a strong linear dose-response relationship between elevated systemic inflammation, as reflected by CAR, and increased RA risk.

The results of the interaction test (Figure 3) indicate that age, sex, race, PIR, BMI, hypertension, diabetes, alcohol consumption, and smoking history did not significantly modify the association between CAR and RA risk, as all interaction P-values were greater than 0.05.

Prior to model development, we performed LASSO regression to identify predictive variables. Through 10-fold cross-validation, the optimal regularization parameter (λ = 0.006) was selected to minimize prediction error. Variables with an odds ratio (OR) of 1 (indicating no association) were excluded.

The final prediction model incorporated eight clinically relevant variables: age, sex, CAR, hypertension, smoking status, platelet-to-lymphocyte ratio (PLR), BMI, and diabetes mellitus. Results of the variable selection process (LASSO coefficient trajectories and cross-validation) are presented in Figure 4, while the final model predictors are detailed in Figure 5.

To assess the predictive accuracy of the CAR and the RA prediction model, ROC curves were analyzed. The CAR model demonstrated modest discrimination, with an AUC of 0.615 (95% CI, 0.602–0.628). In contrast, the comprehensive prediction model showed significantly improved performance, achieving an AUC of 0.749 (95% CI, 0.738–0.760), with a sensitivity of 72.2% and specificity of 66.3% (Figure 6).

To facilitate clinical implementation, we developed a nomogram (Figure 5) that integrates all predictor variables into an intuitive scoring system. Each variable is weighted according to its contribution to RA risk, and the total score corresponds to a predicted probability of RA development.

Decision curve analysis further validated the model’s clinical value, demonstrating superior net benefit compared to the “treat all” and “treat none” strategies across a clinically relevant range of threshold probabilities (Figure 6). This supports its utility for individualized risk assessment and decision-making in clinical practice.