Serum Bicarbonate and Reduced Kidney Function in U.S. Adults: Temporal Replication Across NHANES 2007–2016 and 2017–March 2020 Running title: Serum Bicarbonate and Reduced Kidney Function Authors: 芃莱滨¹* Affiliations: 1. 泰智韬图公团 Correspondence: Omitted in this display version. Word count: ~4,200 Tables: 7 Figures: 3 Supplementary tables: 5 Supplementary figures: 7 Structured Abstract Background: Low serum bicarbonate is associated with kidney disease progression in clinical populations, but the evidence in the general U.S. population has been inconsistent, and few studies have assessed temporal reproducibility. Methods: We conducted a cross-sectional temporal replication analysis using NHANES 2007–2016 (discovery, n = 26,864) and NHANES 2017–March 2020 (validation, n = 7,995). Serum bicarbonate was measured in the BIOPRO subsample. Reduced kidney function was defined as eGFR < 60 mL/min/1.73 m² using the CKD-EPI 2021 race-free creatinine equation. Survey-weighted logistic regression was used with progressive adjustment (Model 4: age, sex, race/ethnicity, BMI, diabetes, hypertension). Restricted cubic splines, quartile analyses, clinical category analyses, subgroup interactions (BH-FDR corrected), E- values, and multiple sensitivity analyses were performed. Results: Each 1 mmol/L higher serum bicarbonate was associated with lower odds of reduced kidney function in both periods (discovery OR 0.883, 95% CI 0.855– 0.912; validation OR 0.888, 95% CI 0.848–0.930). Compared with the 22–29 mmol/L reference, bicarbonate < 22 mmol/L was associated with approximately 2.8-fold higher odds (discovery OR 2.891, 95% CI 2.173–3.846; validation OR 2.741, 95% CI 1.650–4.554). A nonlinear association was detected in the discovery period (nonlinearity P = 5.2 × 10⁻⁵) with directionally consistent but attenuated evidence in validation (P = 0.071). No subgroup interactions survived FDR correction. All 18 logistic sensitivity analyses supported the primary finding. E-
values were moderate (discovery point 1.52, CI 1.42; validation point 1.50, CI 1.36). Conclusions: Lower serum bicarbonate was consistently associated with higher odds of reduced kidney function across two independent NHANES periods, with robust sensitivity analyses and moderate robustness to unmeasured confounding. These findings support serum bicarbonate as a clinically relevant marker associated with kidney function in the general U.S. adult population. Keywords: serum bicarbonate; metabolic acidosis; kidney function; eGFR; NHANES; temporal replication Introduction The kidneys play a central role in systemic acid-base homeostasis through bicarbonate reabsorption, regeneration, and excretion of non-volatile acids [R09]. When kidney function declines, acid retention and metabolic acidosis become increasingly prevalent [R14]. Conversely, metabolic acidosis itself may contribute to kidney injury progression through complement activation, endothelin-1, angiotensin II, and aldosterone-mediated tubulointerstitial fibrosis [R08; R09]. The relationship between serum bicarbonate and kidney outcomes has been examined across diverse populations with variable results. In the Multi-Ethnic Study of Atherosclerosis (MESA), lower serum bicarbonate was associated with a 12% higher odds of rapid kidney function decline (95% CI 6–20%) among community-living adults with baseline eGFR > 60 mL/min/1.73 m² [R01]. Meta- analytic evidence from 14 clinical trials of alkali therapy in CKD found that treatment increased serum bicarbonate by a mean of 3.33 mEq/L and slowed eGFR decline by 3.28 mL/min/1.73 m² [R07]. In the KNOW-CKD study of Korean adults with CKD, low bicarbonate (< 22 mmol/L) was associated with a 27% higher hazard of renal events (HR 1.27, 95% CI 1.01–1.60) [R16]. However, the evidence is not uniformly consistent. In a post-hoc analysis of the RENAAL and IDNT trials among adults with type 2 diabetes and nephropathy, the inverse association between bicarbonate and ESRD was not retained after adjustment for baseline eGFR [R05]. In pediatric nonglomerular CKD, the association was not statistically significant (HR 1.28, 95% CI 0.84–1.94) [R17]. The serum bicarbonate–kidney function association was classified as INCONSISTENT in the literature screening for this project, reflecting variation in study design, population, adjustment, and outcome definitions. Most prior studies have been single-period analyses, limiting assessment of temporal stability. Whether the association between serum bicarbonate and reduced kidney function is consistently observable across independent survey cycles within the same national sampling framework has not been systematically evaluated.
The objective of this study was to evaluate the cross-sectional association between serum bicarbonate and reduced kidney function (eGFR < 60 mL/min/1.73 m²) in the U.S. general adult population using NHANES 2007–2016 as a discovery dataset and to assess temporal replication in NHANES 2017–March 2020 pre-pandemic. We hypothesized that lower serum bicarbonate would be consistently associated with higher odds of reduced kidney function across both periods, and that the association would be robust to progressive covariate adjustment, alternative exposure and outcome definitions, and subgroup stratification.
Methods Study Design This was a cross-sectional, survey-weighted epidemiologic study using publicly available NHANES data. The analysis was conducted in two prespecified periods: a discovery period (NHANES 2007–2016, five 2-year cycles) and a temporal validation period (NHANES 2017–March 2020 pre-pandemic). All analyses used survey weights appropriate to each period and accounted for the complex multistage probability sampling design. Data Source and Participants NHANES is a nationally representative, cross-sectional survey of the civilian non- institutionalized U.S. population, conducted by the National Center for Health Statistics (NCHS). Participants were selected through a complex multistage probability sampling design. The NCHS Ethics Review Board approved the NHANES protocol; all participants provided written informed consent. As this analysis used de-identified publicly available data, no additional institutional review board approval was required. Eligibility criteria were age ≥ 18 years, not pregnant, non-missing serum bicarbonate, non-missing serum creatinine with computable eGFR, non-missing covariates, and non-missing MEC examination weight. The analytic sample comprised 26,864 participants in the discovery period and 7,995 in the validation period. Exposure: Serum Bicarbonate Serum bicarbonate was measured in the NHANES BIOPRO subsample (variable LBXSC3SI, mmol/L). The primary exposure was serum bicarbonate as a continuous variable (per 1 mmol/L increase). Secondary exposure specifications included quartiles (derived from survey-weighted quantiles in the discovery period: Q1 ≤ 24, Q2 > 24 to ≤ 25, Q3 > 25 to ≤ 27, Q4 > 27 mmol/L) and clinical categories (< 22, 22–29 [reference], ≥ 30 mmol/L) based on clinically established thresholds [R15; R24]. Outcome: Reduced Kidney Function Reduced kidney function was defined as eGFR < 60 mL/min/1.73 m², calculated using the CKD-EPI 2021 race-free creatinine equation [R10]. Input variables were serum creatinine (LBXSCR, mg/dL), age (RIDAGEYR), and sex (RIAGENDR). In a sensitivity analysis (S6), eGFR was modeled as a continuous outcome using linear regression. Covariates Covariates were selected based on established risk factors for reduced kidney function and prior literature. Model 1 was unadjusted. Model 2 added age
(continuous, years), sex (male/female), and race/ethnicity (categorical: Non- Hispanic White, Non-Hispanic Black, Mexican American, Other Hispanic, Non- Hispanic Asian, Other/Multi). Model 3 added body mass index (BMI, continuous, kg/m²). Model 4 (fully adjusted) added diabetes (self-reported, based on DIQ010 = 1) and hypertension (self-reported, based on BPQ020 = 1 or BPQ050A = 1). Survey Design and Weight Construction All analyses used NHANES complex survey design variables: primary sampling units (SDMVPSU), strata (SDMVSTRA), and nest = TRUE. For the discovery period (2007–2016, 5 cycles), the analysis weight was constructed as WTMEC2YR / 5. For the validation period (2017–March 2020), WTMECPRP was used as provided. Discovery and validation weights were never combined into a single weight. Statistical Analysis Primary models (A1) Survey-weighted logistic regression models with a quasibinomial link function were fit using the svyglm function in R package survey (version 4.5). The model progression was M1 (unadjusted) → M2 (demographics) → M3 (+BMI) → M4 (fully adjusted). Quartile analyses used the same M4 adjustment. Nonlinear analysis (A2) A restricted cubic spline (RCS) with 4 knots at 21, 24, 26, and 28 mmol/L (5th, 35th, 65th, and 95th survey-weighted percentiles of bicarbonate in the discovery period) was used to model the dose-response relationship. The reference value was 24 mmol/L. Nonlinearity was tested using a Wald test comparing the full RCS model (3 degrees of freedom) to a linear model (1 degree of freedom). Clinical category analyses used the same M4 adjustment with bicarbonate as a three-level categorical variable. An exploratory threshold analysis used a two-piece linear logistic model at candidate cutpoints (22, 24, 26, 28 mmol/L), with a Wald test for slope change. Subgroup and interaction analysis (A3) Stratified models were fit using the M4 formula within each level of six subgroups: sex, age group (< 60, ≥ 60 years), diabetes, hypertension, BMI category (< 25, 25– 29.9, ≥ 30 kg/m²), and race/ethnicity. Formal Wald interaction tests were performed for each subgroup for the continuous bicarbonate exposure. Benjamini- Hochberg false discovery rate (BH-FDR) correction was applied to interaction P- values within each period. Sensitivity analyses (A4) Eight prespecified sensitivity analyses were conducted in both periods: • S1: Primary reproduction (identical to A1 Model 4) • S2: Trimmed bicarbonate (exclude < 1st or > 99th percentile) • S3: Exclude extreme bicarbonate (< 15 or > 35 mmol/L)
• S4: Exclude eGFR < 15 mL/min/1.73 m² • S5: Alternative outcome (eGFR < 45 mL/min/1.73 m²) • S6: Continuous eGFR outcome (linear regression) • S7: Age ≥ 20 years only • S8: Alternative low-bicarbonate thresholds (< 21, < 22, < 23 vs. 22–29 mmol/L) Leave-one-cycle-out (LOCO) analysis and cycle-specific analyses were performed in the discovery period. E-value analysis (B2) E-values [R12] were calculated for the primary M4 continuous bicarbonate odds ratio in both periods. The E-value quantifies the minimum strength of association an unmeasured confounder would need to have with both the exposure and the outcome to fully explain away the observed association. Missing data (B3) A descriptive assessment of missing covariate data was performed within the BIOPRO subsample. Complete-case analysis was the primary approach; full multiple imputation (MICE) was not performed due to computational cost and the low proportion of missing covariate data. Software All analyses were conducted in R version 4.3.3. The R survey package (version 4.5) was used for survey-weighted regression. Figures were generated with ggplot2 (version 3.5.1). Data manipulation used dplyr (version 1.1.4), tidyr (version 1.3.1), and readr (version 2.1.5). JSON configuration files were processed with jsonlite (version 1.8.8).
Results Participant Characteristics A total of 26,864 participants in the discovery period (2007–2016) and 7,995 in the validation period (2017–March 2020) met all eligibility criteria (Table 1). In both periods, participants with low serum bicarbonate (< 22 mmol/L) had higher mean BMI, a higher proportion of female sex, and higher prevalence of diabetes, hypertension, and reduced kidney function compared with those in the normal range (22–29 mmol/L). In the discovery period, mean bicarbonate was 20.35 mmol/L in the low group versus 25.21 mmol/L in the normal group; the prevalence of reduced kidney function was 9.3% versus 5.2%, respectively. Similar patterns were observed in the validation period. Primary Continuous Models In the unadjusted model (M1), the association between serum bicarbonate and reduced kidney function was not statistically significant in the discovery period (OR 0.983, 95% CI 0.951–1.015, P = 0.297) but was significant in the validation period (OR 0.939, 95% CI 0.897–0.982, P = 0.011). After adjustment for demographic covariates (M2), the association became highly significant in both periods (discovery OR 0.870, 95% CI 0.841–0.899, P = 5.3 × 10⁻¹²; validation OR 0.884, 95% CI 0.845–0.924, P = 2.0 × 10⁻⁵). Further adjustment for BMI (M3) and diabetes and hypertension (M4) yielded nearly identical estimates. In the fully adjusted model (M4), each 1 mmol/L higher serum bicarbonate was associated with approximately 12% lower odds of reduced kidney function in both periods: • Discovery: OR 0.883, 95% CI 0.855–0.912, P = 6.6 × 10⁻¹¹ • Validation: OR 0.888, 95% CI 0.848–0.930, P = 8.8 × 10⁻⁵ The temporal replication met all four prespecified criteria: same direction (OR < 1 in both periods), validation P < 0.05, confidence interval overlap (discovery 0.855– 0.912 versus validation 0.848–0.930), and effect-size ratio within 50% (ratio = 1.006). (Table 2) Quartile Results Using discovery-period survey-weighted quartile cutpoints, the inverse association with reduced kidney function was monotonic across quartiles in both periods. Compared with the lowest quartile (Q1, ≤ 24 mmol/L), higher quartiles showed progressively lower odds of reduced kidney function: Discovery: - Q2 (> 24–25 mmol/L): OR 0.790, 95% CI 0.660–0.945 - Q3 (> 25–27 mmol/L): OR 0.602, 95% CI 0.508–0.714 - Q4 (> 27 mmol/L): OR 0.561, 95% CI 0.455–0.691
Validation: - Q2 (> 24–25 mmol/L): OR 0.617, 95% CI 0.435–0.875 - Q3 (> 25–27 mmol/L): OR 0.497, 95% CI 0.379–0.654 - Q4 (> 27 mmol/L): OR 0.550, 95% CI 0.374–0.810 All reported Model 4 quartile contrasts were statistically significant (P < 0.05), including Q2 in validation (P = 0.015). Clinical Category Results Compared with the 22–29 mmol/L reference category, participants with bicarbonate < 22 mmol/L had approximately 2.8-fold higher odds of reduced kidney function in both periods: • Discovery: OR 2.891, 95% CI 2.173–3.846, P = 2.0 × 10⁻¹⁰ • Validation: OR 2.741, 95% CI 1.650–4.554, P = 6.1 × 10⁻⁴ The ≥ 30 mmol/L category was not significantly associated with reduced kidney function (discovery n = 542, OR 0.900, 95% CI 0.638–1.271, P = 0.545; validation n = 352, OR 0.663, 95% CI 0.351–1.253, P = 0.191). (Table 4) Nonlinear Analysis The restricted cubic spline analysis revealed a significant overall association in both periods (discovery P = 1.3 × 10⁻¹¹; validation P = 7.1 × 10⁻⁶). The nonlinearity test was significant in the discovery period (P = 5.2 × 10⁻⁵) but not in the validation period (P = 0.071), although the shape was directionally consistent. The dose-response relationship showed a steeper inverse association at lower bicarbonate concentrations that attenuated at higher values. RCS position estimates (reference 24 mmol/L) included: • 21 mmol/L: discovery OR 1.788 (95% CI 1.544–2.071); validation OR 1.841 (95% CI 1.539–2.201) • 22 mmol/L: discovery OR 1.470 (95% CI 1.333–1.620); validation OR 1.502 (95% CI 1.333–1.692) • 26 mmol/L: discovery OR 0.724 (95% CI 0.674–0.778); validation OR 0.724 (95% CI 0.667–0.786) • 30 mmol/L: discovery OR 0.694 (95% CI 0.546–0.884); validation OR 0.680 (95% CI 0.439–1.054) (Figure 1; Table 3) An exploratory threshold analysis identified 22 mmol/L as a candidate inflection point. The slope-change Wald test was significant in the discovery period (P = 0.00028) but not in the validation period (P = 0.288), although the direction was consistent. This threshold analysis is exploratory and not confirmatory. Subgroup and Interaction Results All 68 subgroup-stratified models converged and showed an inverse association direction (OR < 1) for continuous bicarbonate across all subgroup levels. Among 24 formal Wald interaction tests (6 subgroups × 2 periods × 2 exposures), none
remained significant after BH-FDR correction (minimum FDR P = 0.30). The age- group interaction was nominally significant in both periods for the continuous exposure (raw P = 0.030 in discovery; raw P = 0.042 in validation) but did not survive FDR correction. (Table 5; Figure 2) For the low-bicarbonate binary exposure (< 22 versus 22–29 mmol/L), all subgroup estimates were directionally consistent (OR > 1), although some subgroups had wide confidence intervals due to smaller sample sizes. Sensitivity Analyses All 18 logistic sensitivity analyses (S1–S5, S7, S8a–S8c) supported the main finding in both periods. The continuous eGFR outcome (S6) showed a positive beta coefficient in both periods (discovery β = 0.123, 95% CI −0.006 to 0.253, P = 0.062; validation β = 0.200, 95% CI −0.008–0.409, P = 0.058), consistent with the primary direction but not reaching statistical significance. Alternative low-bicarbonate thresholds (S8) showed graded associations: < 21 mmol/L had the strongest OR (discovery 3.462, 95% CI 2.143–5.592; validation 2.713, 95% CI 1.508–4.882), with diminishing ORs at higher thresholds, consistent with the nonlinear dose-response pattern. (Table 6; Figure 3) Leave-one-cycle-out analysis within the discovery period showed consistent results (OR range 0.866–0.896), with no single cycle driving the overall finding. Cycle- specific analyses showed an inverse association in all five discovery cycles, with moderate heterogeneity (Cochrane Q = 5.63, df = 4, P = 0.229; I² = 29%). (Tables S1–S2) E-Value Analysis The E-values indicated moderate robustness to unmeasured confounding: • Discovery: point estimate E-value = 1.52; CI limit E-value = 1.42 • Validation: point estimate E-value = 1.50; CI limit E-value = 1.36 An unmeasured confounder would need to be associated with both lower serum bicarbonate and reduced kidney function by a risk ratio of approximately 1.5 (on both exposure–confounder and confounder–outcome links) to fully explain away the observed association. Given that the primary model already adjusted for major known confounders (diabetes, hypertension, age, BMI), residual confounding of this magnitude is plausible but not trivial. (Table 7) Missing Data Assessment Within the BIOPRO subsample (n = 35,408 eligible participants), 34,859 (98.4%) had complete covariate data and were included in Model 4. Only 549 participants (1.6%) were excluded due to missing covariates, with the highest missingness in BMI (1.4%). Excluded participants were older and had higher rates of diabetes, hypertension, and reduced kidney function compared with complete-case participants, but the small proportion excluded limits the potential for substantial selection bias. (Table S5)
Discovery–Validation Comparison The primary continuous OR was nearly identical across periods (discovery 0.883, validation 0.888). All four replication criteria were met. The clinical category and quartile results showed similar magnitudes across periods. The nonlinearity finding was statistically significant only in the discovery period but directionally consistent. Sensitivity analyses and E-values were comparable between periods. Together, these results demonstrate robust temporal replication of the serum bicarbonate–reduced kidney function association within NHANES.
Discussion Principal Findings In this cross-sectional temporal replication study of U.S. adults from two independent NHANES periods spanning 2007 to March 2020, each 1 mmol/L higher serum bicarbonate was consistently associated with approximately 12% lower odds of reduced kidney function. The association was robust across progressive adjustment, multiple exposure specifications (continuous, quartile, clinical category), all 18 logistic sensitivity analyses, and both periods. The E- values indicated moderate robustness to unmeasured confounding, and no subgroup interactions survived FDR correction, suggesting the association is general across demographic and clinical subgroups. Comparison with Prior Evidence The magnitude and direction of the association are consistent with prior community-based and clinical studies. Driver et al. (MESA) reported a 12% higher odds of rapid kidney function decline per 1-SD lower bicarbonate [R01], and Raphael et al. (NHANES III) found that low bicarbonate (< 22 mmol/L) was associated with a hazard ratio of 2.56 for mortality in the CKD subgroup [R02]. The present study extends these findings by demonstrating temporal replication within NHANES, using the contemporary CKD-EPI 2021 race-free equation, and by providing a comprehensive sensitivity analysis. The meta-analytic evidence from Navaneethan et al. supports a potential causal role for metabolic acidosis in CKD progression [R07], and clinical trials of alkali therapy have demonstrated slowing of eGFR decline [R06, R07, R20]. Our cross- sectional results are compatible with these intervention findings in direction, though the observational design precludes causal interpretation. Conflicting and Qualifying Evidence The association was not independent of eGFR in diabetic nephropathy [R05], and was null in pediatric nonglomerular CKD [R17], suggesting that the bicarbonate– kidney function relationship may differ by CKD etiology and diabetes status. However, these populations differ substantially from the general U.S. adult population represented in our study, and our subgroup analyses found no evidence of effect modification by diabetes status (interaction P = 0.854 discovery, 0.890 validation). Nonlinear Findings The nonlinear dose-response relationship, with a steeper association at lower bicarbonate values and attenuation at higher values, is broadly consistent with prior evidence of U-shaped associations between serum bicarbonate and mortality in CKD populations [R04, R22, R23]. However, the nonlinearity was statistically significant only in the discovery period (P = 5.2 × 10⁻⁵) and not in the validation period (P = 0.071), although the shape was directionally consistent. The
exploratory threshold analysis identified 22 mmol/L as a candidate inflection point, consistent with clinical guidelines that recommend alkali therapy when bicarbonate falls below this level [R15; R24]. This threshold was directionally supported but not confirmatorily replicated, and further studies with larger validation samples are needed. Biological Plausibility Acid retention, even in the absence of overt metabolic acidosis, promotes tubulointerstitial fibrosis through complement activation and upregulation of the endothelin-1, angiotensin II, and aldosterone pathways [R08]. Impaired ammoniagenesis, a key early feature of CKD, reduces the kidney’s capacity to excrete acid loads, creating a feed-forward cycle of acid retention and progressive tubular injury [R09]. Dietary acid load from modern Western diets may further contribute to low-grade metabolic acidosis in the general population [R19, R25]. The biologic rationale is therefore plausible, though direct mechanistic evidence from human studies remains limited. Clinical and Epidemiologic Implications Our findings suggest that serum bicarbonate, a routinely available and inexpensive laboratory measurement, may serve as a clinically relevant marker of reduced kidney function in the general adult population. The consistent temporal replication across two independent NHANES periods strengthens confidence in the robustness of this association. The approximately 2.8-fold higher odds associated with bicarbonate < 22 mmol/L (the threshold recommended for alkali therapy initiation) highlights the potential clinical relevance of this cutoff, even though the present cross-sectional study cannot directly inform treatment decisions. Strengths This study has several strengths: (1) the use of two independent NHANES periods with prespecified discovery and validation design; (2) comprehensive sensitivity analyses including alternative exposure and outcome definitions, trimming, and leave-one-cycle-out analysis; (3) formal replication assessment using four criteria; (4) E-value analysis for unmeasured confounding; (5) BH-FDR correction for multiple testing in subgroup analyses; (6) use of the contemporary CKD-EPI 2021 race-free eGFR equation; and (7) a large, nationally representative sample with rigorous survey weighting. Limitations Several limitations must be acknowledged. First, the cross-sectional design precludes causal or temporal inference; bicarbonate and kidney function were measured simultaneously, and the direction of the association cannot be determined. Second, diabetes and hypertension were based on self-report, which may introduce misclassification; HbA1c was not available in the KDM-02 data manifest. Third, serum bicarbonate was measured once in the BIOPRO subsample, with no assessment of within-person variability or temporal stability. Fourth, approximately 35–39% of NHANES participants lacked bicarbonate data due to the
BIOPRO subsample design, and the missingness is not at random. Fifth, race/ethnicity coding changed from RIDRETH1 (2007–2010) to RIDRETH3 (2011– 2020), requiring harmonization. Sixth, urine albumin-to-creatinine ratio was not included as a complementary kidney damage marker. Seventh, full multiple imputation was not performed for missing covariate data; however, complete-case analysis included 98.4% of eligible BIOPRO participants, and covariate missingness was negligible (≤ 1.4% per variable). Excluded participants were older and had higher comorbidity burden, which may introduce limited selection bias. Eighth, the ≥ 30 mmol/L bicarbonate group had small sample sizes, limiting interpretation. Ninth, the threshold analysis is exploratory and not confirmatory. Tenth, temporal replication occurred within NHANES, which shares the same sampling framework, and must not be interpreted as independent external replication. Conclusions In this nationally representative temporal replication study, lower serum bicarbonate was consistently associated with higher odds of reduced kidney function across two independent NHANES periods. The association was robust to progressive adjustment, comprehensive sensitivity analyses, and multiple exposure specifications, with moderate E-values indicating robustness to unmeasured confounding. No subgroup interactions survived FDR correction, and the nonlinear dose-response relationship was directionally consistent. These findings support serum bicarbonate as a robust and clinically accessible marker associated with kidney function in the general U.S. adult population and provide a foundation for further investigation into the role of acid-base status in kidney health.
Declarations Ethics approval and consent to participate: NHANES was approved by the NCHS Ethics Review Board. All participants provided written informed consent. This analysis used de-identified, publicly available data and did not require additional institutional review board approval. Data availability: NHANES data are publicly available from the CDC National Center for Health Statistics at https://www.cdc.gov/nchs/nhanes/. Funding: Not reported in this display version. Conflicts of interest: Not reported in this display version. Author contributions: Not reported in this display version. Acknowledgments: Not reported in this display version.
References See supplementary file KDM02_REFERENCE_LIST_FINAL.csv for complete bibliographic details. 1. Driver TH, et al. Low serum bicarbonate and kidney function decline: the Multi-Ethnic Study of Atherosclerosis (MESA). American Journal of Kidney Diseases. 2014;64(4):534–541. doi:10.1053/j.ajkd.2014.05.008. 2. Raphael KL, et al. Serum bicarbonate and mortality in adults in NHANES III. Nephrology Dialysis Transplantation. 2013;28(5):1207–1213. doi:10.1093/ndt/gfs609. 3. Abramowitz MK, et al. Lower serum bicarbonate and a higher anion gap are associated with lower cardiorespiratory fitness in young adults. Kidney International. 2012;81(10):1033–1040. doi:10.1038/ki.2011.479. 4. Kovesdy CP, et al. Association of serum bicarbonate levels with mortality in patients with non-dialysis-dependent CKD. Nephrology Dialysis Transplantation. 2009;24(4):1232–1237. doi:10.1093/ndt/gfn633. 5. Schutte E, et al. Serum bicarbonate and kidney disease progression and cardiovascular outcome in patients with diabetic nephropathy. American Journal of Kidney Diseases. 2015;66(3):450–458. doi:10.1053/j.ajkd.2015.03.032. 6. Goraya N, et al. Clinical evidence that treatment of metabolic acidosis slows the progression of chronic kidney disease. Current Opinion in Nephrology and Hypertension. 2019;28(3):273–280. doi:10.1097/MNH.0000000000000491. 7. Navaneethan SD, et al. Effects of treatment of metabolic acidosis in CKD: a systematic review and meta-analysis. Clinical Journal of the American Society of Nephrology. 2019;14(7):1011–1020. doi:10.2215/CJN.13091118. 8. Ravikumar NPG, et al. Acid-mediated kidney injury across the spectrum of metabolic acidosis. Advances in Chronic Kidney Disease. 2022;29(4):341– 349. doi:10.1053/j.ackd.2022.04.009. 9. Nagami GT, et al. Regulation of acid-base balance in chronic kidney disease. Advances in Chronic Kidney Disease. 2017;24(5):280–285. doi:10.1053/j.ackd.2017.07.004. 10. Inker LA, et al. New creatinine- and cystatin C-based equations to estimate GFR without race. New England Journal of Medicine. 2021;385(19):1737– 1749. doi:10.1056/NEJMoa2102953. 11. Delgado C, et al. A unifying approach for GFR estimation: recommendations of the NKF-ASN Task Force. American Journal of Kidney Diseases. 2022;79(1):128–135. doi:10.1053/j.ajkd.2021.08.003. 12. VanderWeele TJ, et al. Sensitivity analysis in observational research: introducing the E-value. Annals of Internal Medicine. 2017;167(4):268–274. doi:10.7326/M16-2607. 13. Banerjee T, et al. Elevated serum anion gap in adults with moderate chronic kidney disease increases risk for progression to end-stage renal disease.
American Journal of Physiology — Renal Physiology. 2019;316(6):F1155– F1163. doi:10.1152/ajprenal.00496.2018. 14. Clase CM, et al. Relationship between glomerular filtration rate and the prevalence of metabolic abnormalities: results from NHANES III. Nephron Clinical Practice. 2007;106(4):c170–c177. doi:10.1159/000100489. 15. Kraut JA, et al. Metabolic acidosis of CKD: an update. American Journal of Kidney Diseases. 2016;67(2):307–317. doi:10.1053/j.ajkd.2015.08.028. 16. Kim HJ, et al. Metabolic acidosis is an independent risk factor of renal progression in Korean CKD patients: the KNOW-CKD study results. Frontiers in Medicine. 2021;8:707588. doi:10.3389/fmed.2021.707588. 17. Brown DD, et al. Low serum bicarbonate and CKD progression in children. Clinical Journal of the American Society of Nephrology. 2020;15(7):945–955. doi:10.2215/CJN.07060619. 18. Hamidianshirazi M, et al. Dietary acid load and chronic kidney disease. Saudi Journal of Kidney Diseases and Transplantation. 2021;32(2):325–331. doi:10.4103/1319-2442.352409. 19. Goraya N, et al. Pathophysiology of diet-induced acid stress. International Journal of Molecular Sciences. 2024;25(4):2336. doi:10.3390/ijms25042336. 20. Goraya N, et al. Fruit and vegetable treatment of CKD-related metabolic acidosis reduces cardiovascular risk better than sodium bicarbonate. American Journal of Nephrology. 2019;49(5):376–384. doi:10.1159/000500042. 21. Lambert DC, et al. Associations of metabolic syndrome and abdominal obesity with anion gap metabolic acidosis among US adults. Kidney360. 2022;3(2):299–309. doi:10.34067/KID.0002402022. 22. Lombardi G, et al. Longitudinal serum bicarbonate and mortality risk in older patients with advanced CKD: analyses from the EQUAL cohort. Clinical Kidney Journal. 2024;17(12):sfae254. doi:10.1093/ckj/sfae254. 23. Gaggl M, et al. Effect of oral alkali supplementation on progression of chronic kidney disease. Current Hypertension Reviews. 2014;10(2):93–102. doi:10.2174/1573402111666141231123314. 24. Łoniewski I, et al. Bicarbonate therapy for prevention of chronic kidney disease progression. Kidney International. 2014;85(3):512–519. doi:10.1038/ki.2013.401. 25. Adeva MM, et al. Diet-induced metabolic acidosis. Clinical Nutrition. 2011;30(4):416–421. doi:10.1016/j.clnu.2011.03.008.
Tables Table 1. Survey-Weighted Participant Characteristics by Bicarbonate Status Discovery period (NHANES 2007–2016) Characteristic Low (<22) Normal (22–29) Unweighted N 1,480 24,842 Age, years 43.99 ± 0.70 46.66 ± 0.26 BMI, kg/m² 30.92 ± 0.23 28.70 ± 0.08 Bicarbonate, mmol/L 20.35 ± 0.03 25.21 ± 0.04 Female, % 62.4 50.1 Diabetes, % 13.5 8.8 Hypertension, % 34.1 30.6 Reduced kidney function, % 9.3 5.2 Validation period (NHANES 2017–March 2020) Characteristic Low (<22) Normal (22–29) Unweighted N 387 7,256 Age, years 46.25 ± 1.19 47.52 ± 0.60 BMI, kg/m² 32.15 ± 0.40 29.70 ± 0.19 Bicarbonate, mmol/L 20.36 ± 0.06 25.57 ± 0.08 Female, % 59.3 51.5 Diabetes, % 14.9 11.1 Hypertension, % 34.5 31.5 Reduced kidney function, % 10.3 5.4 Values are mean ± SE or percentage, survey-weighted. Participants with bicarbonate ≥ 30 mmol/L excluded from this table (discovery: n = 542; validation: n = 352). Table 2. Primary Survey-Weighted Logistic Regression Results (Model 1–4) Period Model OR 95% CI P-value Discovery M1 (unadjusted) 0.983 0.951–1.015 0.297 Discovery M2 (+demographics) 0.870 0.841–0.899 5.3 × 10⁻¹² Discovery M3 (+BMI) 0.878 0.850–0.907 3.4 × 10⁻¹¹ Discovery M4 (fully adjusted) 0.883 0.855–0.912 6.6 × 10⁻¹¹ Validation M1 (unadjusted) 0.939 0.897–0.982 0.011 Validation M2 (+demographics) 0.884 0.845–0.924 2.0 × 10⁻⁵ Validation M3 (+BMI) 0.891 0.851–0.933 7.8 × 10⁻⁵ Validation M4 (fully adjusted) 0.888 0.848–0.930 8.8 × 10⁻⁵ OR per 1 mmol/L higher serum bicarbonate. Model 4 covariates: age, sex, race/ethnicity, BMI, diabetes, hypertension. Discovery weight: WTMEC2YR/5. Validation weight: WTMECPRP.
Replication assessment: Direction same (both OR < 1), validation P < 0.05, CI overlap, effect ratio 0.888/0.883 = 1.006 (within 50%). Status: REPLICATED. Table 3. Restricted Cubic Spline Position Estimates (Reference: 24 mmol/L) Position Discovery OR Discovery 95% CI Validation OR Validation 95% CI 21 mmol/L 1.788 1.544–2.071 1.841 1.539–2.201 22 mmol/L 1.470 1.333–1.620 1.502 1.333–1.692 26 mmol/L 0.724 0.674–0.778 0.724 0.667–0.786 28 mmol/L 0.672 0.580–0.778 0.717 0.556–0.924 30 mmol/L 0.694 0.546–0.884 0.680 0.439–1.054 Overall P: discovery 1.3 × 10⁻¹¹; validation 7.1 × 10⁻⁶. Nonlinearity P: discovery 5.2 × 10⁻⁵; validation 0.071. Four knots at 21, 24, 26, 28 mmol/L. Model 4 adjustment. Table 4. Clinical Category Analysis Discovery period (NHANES 2007–2016) Category OR 95% CI P value <22 mmol/L 2.891 2.173–3.846 2.0 × 10⁻¹⁰ 22–29 mmol/L 1.000 Reference Reference ≥30 mmol/L 0.900 0.638–1.271 0.545 Validation period (NHANES 2017–March 2020) Category OR 95% CI P value <22 mmol/L 2.741 1.650–4.554 6.1 × 10⁻⁴ 22–29 mmol/L 1.000 Reference Reference ≥30 mmol/L 0.663 0.351–1.253 0.191 Model 4 adjustment. Unweighted N: discovery < 22: 1,480; 22–29: 24,842; ≥ 30: 542. Validation < 22: 387; 22–29: 7,256; ≥ 30: 352. Table 5. Subgroup Interaction Tests (Continuous Exposure, Raw and FDR- Corrected P-values) Discovery period Subgroup Raw P FDR P Sex 0.205 0.547 Age group 0.030 0.300 Diabetes 0.854 0.925 Hypertension 0.327 0.604 BMI category 0.057 0.300 Race/ethnicity 0.108 0.372 Validation period Subgroup Raw P FDR P Sex 0.622 0.878
Subgroup Raw P FDR P Age group 0.042 0.300 Diabetes 0.890 0.925 Hypertension 0.505 0.860 BMI category 0.781 0.894 Race/ethnicity 0.276 0.570 BH-FDR correction applied within each period separately. No interactions significant after FDR correction (minimum FDR P = 0.30). Table 6. Sensitivity Analysis Summary Analysis Description Discovery OR Validation OR Supports Main? S1 Primary reproduction 0.883 0.888 Yes S2 Trim 1st–99th percentile 0.887 0.889 Yes S3 Exclude extreme bicarbonate 0.883 0.889 Yes S4 Exclude eGFR < 15 0.877 0.877 Yes S5 Outcome eGFR < 45 0.853 0.885 Yes S6 Continuous eGFR (β) 0.123 0.200 N/A (linear) S7 Age ≥ 20 only 0.883 0.888 Yes S8a < 21 vs 22–29 mmol/L 3.462 2.713 Yes S8b < 22 vs 22–29 mmol/L 2.925 2.783 Yes S8c < 23 vs 22–29 mmol/L 2.001 2.084 Yes All S1–S5, S7, S8a–S8c logistic models support main finding. S6 is linear regression (beta coefficient, not OR). Table 7. E-Value Summary Period E-value (Point Estimate) E-value (CI Limit) Discovery 1.52 1.42 Validation 1.50 1.36 An unmeasured confounder would need to be associated with both lower serum bicarbonate and reduced kidney function by a risk ratio of at least 1.52 (discovery point) to fully explain away the observed association.
Figure 1. Restricted cubic spline comparison of the association between serum bicarbonate and reduced kidney function in the discovery (2007–2016) and validation (2017–March 2020) periods. The solid line represents the OR; shaded area represents the 95% CI. The dashed vertical line marks the reference value (24 mmol/L). The model is adjusted for age, sex, race/ethnicity, BMI, diabetes, and hypertension (Model 4). Overall association P < 0.001 in both periods. Nonlinearity P = 5.2 × 10⁻⁵ (discovery) and 0.071 (validation).
Figure 2. Subgroup forest plot comparing discovery and validation ORs for the continuous bicarbonate exposure (per 1 mmol/L). Each point represents the OR (Model 4) within each subgroup level. Error bars represent 95% CIs. No interaction survived BH-FDR correction.
Figure 3. Sensitivity forest plot: logistic sensitivity analyses (S1–S5, S7, S8a–S8c) with ORs and 95% CIs for both periods. S6 (continuous eGFR outcome, linear regression) is reported separately as a beta coefficient and is not included in the logistic forest plot. Diamond markers represent discovery; square markers represent validation. All logistic sensitivity analyses support the primary finding direction. Supplement Supplementary Materials: Serum Bicarbonate and Reduced Kidney Function in U.S. Adults Temporal Replication Across NHANES 2007–2016 and 2017–March 2020
Supplementary Methods Survey Weight Construction For the discovery period (NHANES 2007–2016, five 2-year cycles), the combined analysis weight was constructed as: analysis_weight = WTMEC2YR / 5 Division by the number of pooled cycles accounts for the unequal probability of selection across multiple cycles. For the validation period (NHANES 2017–March 2020), WTMECPRP was used as provided by NCHS. Weights were not truncated or trimmed. Variable Harmonization Race/ethnicity was harmonized across cycles using RIDRETH1 (2007–2010) and RIDRETH3 (2011–2020). Categories 5 (Other Hispanic — including Hispanic) and 7 (Non-Hispanic Multiracial) from RIDRETH3 were merged as “OtherMulti” in subgroup analyses due to small cell counts. The primary race/ethnicity variable was coded as Non-Hispanic White, Non-Hispanic Black, Mexican American, Other Hispanic, Non-Hispanic Asian, and Other/Multi. Diabetes was defined by self-report (DIQ010 = 1, “doctor told you have diabetes”). Participants who responded “no” (DIQ010 = 2) or “borderline” (DIQ010 = 3) were classified as non-diabetic. Participants with responses of 7 (refused), 9 (don’t know), or missing were excluded. Hypertension was defined as self-reported ever being told by a doctor that they had hypertension (BPQ020 = 1) or having taken antihypertensive medication (BPQ050A = 1). This approach avoids the auscultatory-to-oscillometric blood pressure measurement method change across NHANES cycles. Restricted Cubic Spline Construction Manual restricted cubic spline basis variables were constructed following Harrell (2015): H1 = (x - k1)³₊ - (x - k3)³₊ × (k4 - k1)/(k4 - k3) + (x - k4)³₊ × (k3 - k1)/(k4 - k3) H2 = (x - k2)³₊ - (x - k3)³₊ × (k4 - k2)/(k4 - k3) + (x - k4)³₊ × (k3 - k2)/(k4 - k3) Knots were placed at the 5th, 35th, 65th, and 95th survey-weighted percentiles of serum bicarbonate in the discovery period: 21, 24, 26, and 28 mmol/L. The reference value for OR calculations was 24 mmol/L. FDR Correction Benjamini-Hochberg false discovery rate (BH-FDR) correction was applied separately within each period. All 24 interaction test P-values (6 subgroups × 2
periods × 2 exposure definitions) were corrected. The FDR q-value threshold was set at 0.05. Replication Criteria A finding was considered replicated if all four criteria were met: 1. Effect direction consistent (OR < 1 or OR > 1 in both periods) 2. Validation period P < 0.05 3. Validation CI overlaps with discovery CI 4. Effect-size ratio within 50% (validation OR / discovery OR between 0.5 and 1.5) E-Value Calculation E-values were calculated for the primary Model 4 continuous bicarbonate OR using the formula: For estimates below 1, the inverse estimate was first used (RR* = 1 / OR). The E-value was then calculated as E-value = RR* + sqrt[RR* × (RR* - 1)]. The same transformation was applied to the confidence limit nearest the null. for the point estimate, with analogous calculation using the CI limit nearest the null. E-values were computed on the risk ratio scale, approximating the OR as the risk ratio given the low outcome prevalence (approximately 5–8% in the reference category). Supplementary Tables Table S1. Leave-One-Cycle-Out Analysis (Discovery Period) Excluded Cycle OR 95% CI P-value Conclusion Change? 2007–2008 0.884 0.851–0.918 3.1 × 10⁻⁸ No 2009–2010 0.889 0.858–0.921 2.1 × 10⁻⁸ No 2011–2012 0.896 0.866–0.928 7.2 × 10⁻⁸ No 2013–2014 0.875 0.843–0.909 3.1 × 10⁻⁹ No 2015–2016 0.866 0.839–0.894 2.8 × 10⁻¹² No All cycles 0.883 0.855–0.912 6.6 × 10⁻¹¹ Reference Model 4 adjustment. No single cycle drives the overall finding. OR range: 0.866– 0.896. Table S2. Cycle-Specific Analysis and Heterogeneity (Discovery Period) Cycle N OR 95% CI P-value 2007–2008 5,395 0.876 0.830–0.923 0.00083 2009–2010 5,793 0.832 0.774–0.895 0.00083 2011–2012 4,990 0.833 0.779–0.891 0.00032 2013–2014 5,455 0.895 0.846–0.947 0.00489 2015–2016 5,231 0.919 0.843–1.001 0.0895
Heterogeneity: Cochrane Q = 5.63, df = 4, P = 0.229; I² = 29% (moderate heterogeneity). All estimates OR < 1. Table S3. Continuous eGFR Outcome (S6) — Linear Regression Period β (per 1 mmol/L bicarbonate) 95% CI P-value Discovery 0.123 −0.006 to 0.253 0.062 Validation 0.200 −0.008 to 0.409 0.058 Positive β indicates higher bicarbonate associated with higher eGFR (consistent with primary logistic finding). Trend-level significance (P < 0.10 but > 0.05) in both periods. Table S4. Exploratory Threshold Analysis Cutpoint Period Below-knot OR Above-knot OR Slope-change P 22 Discovery 0.661 0.919 0.00028 22 Validation 0.765 0.904 0.288 24 Discovery 0.812 0.891 0.093 24 Validation 0.783 0.915 0.087 26 Discovery 0.879 0.865 0.795 26 Validation 0.882 0.856 0.744 28 Discovery 0.879 0.882 0.969 28 Validation 0.888 0.892 0.974 Two-piece linear logistic regression. The 22 mmol/L cutpoint was the preferred exploratory threshold based on slope-change significance in discovery and directionally consistent pattern in validation. Exploratory only; not confirmatory. Table S5. Missing Data Descriptive Assessment Metric Value BIOPRO subsample (eligible) 35,408 Complete-case (Model 4) 34,859 (98.4%) Excluded (missing covariates) 549 (1.6%) BMI missing 1.4% Diabetes missing 0.1% Hypertension missing 0.1% Complete-case vs. excluded comparison: Characteristic Complete-case (n = 34,859) Excluded (n = 549) Age, mean (SD) 48.5 (18.4) 55.1 (20.8) Female, % 50.6% 48.3% BMI, mean (SD) 29.2 (7.1) 29.3 (5.9) Diabetes, % 12.8% 23.9% Hypertension, % 35.2% 50.1% Reduced kidney function, % 7.6% 17.7%
Excluded participants were older and had higher comorbidity burden, but the small proportion excluded (1.6%) limits the potential for substantial selection bias. Full multiple imputation (MICE) was not performed due to computational cost and the low covariate missingness rate. Figure S1. Restricted cubic spline dose-response curve for the discovery period (NHANES 2007–2016).
Figure S2. Restricted cubic spline dose-response curve for the validation period (NHANES 2017–March 2020).
Figure S3. Subgroup forest plot for the discovery period (continuous bicarbonate, per 1 mmol/L).
Figure S4. Subgroup forest plot for the validation period (continuous bicarbonate, per 1 mmol/L).
Figure S5. Leave-one-cycle-out OR and CI plot (discovery period). Figure S6. Subgroup forest plot for the discovery period (low bicarbonate < 22 vs. 22–29 mmol/L).
Figure S7. Subgroup forest plot for the validation period (low bicarbonate < 22 vs. 22–29 mmol/L).