A Comparison of Different Estimates of Albuminuria in Association with Mortality in Epidemiologic Research

Albumin in urine—even in small amounts—is associated with mortality in populations with and without kidney disease (1). Albumin is commonly measured in spot urine samples. To account for variation in spot urine concentration, albuminuria is estimated by calculating the albumin-creatinine ratio. An alternative is the albumin-osmolality ratio. However, age, race/ethnicity, diet, muscle mass, tubular function, and chronic diseases influence urine creatinine and osmolality and are also associated with mortality. This may distort the association of each ratio with mortality (2). Therefore, we explored whether the association between urine albumin and mortality is different when urine albumin is not divided by, but adjusted for, urine creatinine or osmolality as a separate variable in regression models.

We used the National Health and Nutrition Examination Survey in 2009 and 2010, a cohort representative of the general American population (3). Urine albumin, creatinine, and osmolality were measured in random void urine samples using a fluorescent immunoassay, an enzymatic assay, and freezing point depression, respectively. We estimated the GFR (eGFR) using the Chronic Kidney Disease Epidemiology Collaboration equation. Mortality was captured through 2015. Of all 10,537 participants, we excluded 4010 younger than 18 years, 68 pregnant women, 22 on dialysis, 785 without urine measures or serum creatinine, and 17 without follow-up. We converted urine measures into logarithmically transformed normally distributed Z scores and into quintiles to facilitate comparisons. We evaluated differences between participants who died and survived using the Wilcoxon rank-sum test. We evaluated associations with mortality using Cox proportional hazards regression with and without multivariable adjustment for age; sex; race/ethnicity; smoking; histories of hypertension, cardiovascular disease, and diabetes mellitus; waist circumference; body mass index (BMI); systolic BP; and eGFR. We imputed missing values for histories of hypertension, cardiovascular disease, and diabetes mellitus; waist circumference; BMI; and systolic BP (all in <5% of participants) using multivariable normal regression on the basis of all of the other variables. On the basis of each unadjusted regression model, we calculated the Harrell C statistic as a measure of discrimination. We compared hazard ratios by producing a thousand bootstrap samples and testing their absolute differences on a logarithmic scale.

The 5642 included participants had a median (interquartile range) age of 48 (32–63) years, 50% were women, 49% were White, 17% were Black, 71% had a BMI>25 kg/m², 10% had an albumin-creatinine ratio >30 mg/g, and 8% had an eGFR<60 ml/min per 1.73 m². Over a median follow-up of 71 (65–77) months, 377 (7%) died. Compared with those who survived, those who died had lower urine creatinine and osmolality but higher urine albumin and higher albumin-creatinine and albumin-osmolality ratios (P<0.001 for all) (Figure 1A).

Figure 1.

Estimates of albuminuria in relation to mortality in the general American population (n=5642). (A) Medians with interquartile ranges are given for urine measures in participants who survived (n=5265) and participants who died (n=377). Distributions of all urine measures were different between both groups (P<0.001 for all). (B) Hazard ratios with 95% confidence intervals are given for Z scores of urine measures with and without multivariable adjustment. Each urine measure was associated with mortality, also with multivariable adjustment (P<0.001 for all). Asterisks indicate hazard ratios that were different from the hazard ratio of albumin; circles indicate hazard ratios that were different from the hazard ratio of the albumin-creatinine ratio or albumin-osmolality ratio (P<0.001 for unadjusted and P<0.006 for adjusted). (C) Hazard ratios with 95% confidence intervals are given for quintiles of urine measures with and without multivariable adjustment, with each lowest quintile as reference. The lower limits of the second through highest quintiles were 3.0, 5.2, 8.0, and 14.9 mg/L, respectively, for albumin; 3.6, 5.0, 7.0, and 11.8 mg/g, respectively, for the albumin-creatinine ratio; and 6.1, 8.8, 12.8, and 23.0 mosm/kg, respectively, for the albumin-osmolality ratio. Quintiles of each urine measure were associated with mortality, also with multivariable adjustment (P<0.001 for all trends). Asterisks indicate hazard ratios of quintiles that were different from the hazard ratio of the corresponding quintile of albumin (P<0.001 for unadjusted and P<0.04 for adjusted); circles indicate hazard ratios of quintiles that were different from the hazard ratio of the corresponding quintile of the albumin-creatinine or albumin-osmolality ratio (P<0.04 for unadjusted and P<0.05 for adjusted).

Figure 1B compares hazard ratios for Z scores. The hazard ratio for albumin adjusted for creatinine was greater than the hazard ratios for albumin and for the albumin-creatinine ratio. Likewise, the hazard ratio for albumin adjusted for osmolality was greater than the hazard ratios for albumin and for the albumin-osmolality ratio. Hazard ratios were not different between albumin adjusted for creatinine and albumin adjusted for osmolality or between the albumin-creatinine and albumin-osmolality ratios. C statistics were greatest and mutually similar for albumin adjusted for osmolality (0.72; 95% confidence interval [CI], 0.69 to 0.75), albumin adjusted for creatinine (0.71; 95% CI, 0.69 to 0.74), and the albumin-creatinine ratio (0.71; 95% CI, 0.68 to 0.74). The C statistic for albumin adjusted for osmolality was significantly greater than those for the albumin-osmolality ratio (0.70; 95% CI, 0.67 to 0.73) and for albumin (0.66; 95% CI, 0.63 to 0.69), whereas the C statistic for albumin adjusted for creatinine was significantly greater than that for albumin (P<0.001 for all).

Figure 1C compares hazard ratios for quintiles. The hazard ratios for quintiles of albumin adjusted for creatinine or osmolality increased more steeply than the hazard ratios for quintiles of albumin, the albumin-creatinine ratio, and the albumin-osmolality ratio.

We obtained similar results with unadjusted and multivariable-adjusted models when restricting the analyses to men or women; Whites or Blacks; or those with or without a BMI>25 kg/m², an albumin-creatinine ratio >30 mg/g, or an eGFR<60 ml/min per 1.73 m² and when excluding instead of imputing missing values (data not shown).

In conclusion, we found that different estimates of albuminuria discriminated similarly between those who died and survived, but their association with mortality was stronger when urine albumin was adjusted for rather than divided by urine creatinine or osmolality. Although ratios are ubiquitously used in medicine, forcing two associations into one variable in a regression model has methodological disadvantages (4,5). If the numerator and denominator have different associations with the outcome, the association with the ratio may be blunted. This study empirically demonstrates such a bias when using the albumin-creatinine or albumin-osmolality ratio in epidemiologic research. A similar bias may arise when dividing other urine biomarkers by urine creatinine or osmolality. Further studies including measurements of 24-hour albumin excretion, participants with more severe albuminuria, other urine biomarkers, and other clinical outcomes such as kidney failure and cardiovascular disease are needed to confirm and extend this finding.

Disclosures

J.H. Ix has received an investigator-initiated research grant from Baxter International, unrelated to the current study. M.G. Shlipak has received personal fees from Cricket Health and has served as an advisor for Tai Diagnostics, all unrelated to the current study. S.S. Waikar has received personal fees from Bunch and James, Cerus, CVS, GE Health Care, GlaxoSmithKline, Harvard Clinical Research Institute, Janssen, JNJ, Kantum Pharma, Mallinckrodt, Mass Medical International, Pfizer, Public Health Advocacy Institute, Roth Capital Partners, Strataca, Takeda, Venbio, and Wolters Kluwer; has received grants and personal fees from Allena; and has served as an expert witness for litigation related to cisplatin nephrotoxicity, Granuflo, mercury exposure, Omniscan, and statins, all unrelated to the current study. All remaining authors have nothing to disclose.

Funding

J.H. Ix and M.G. Shlipak are supported by National Institutes of Health grant U01DK102730. J.J.E. Koopman was supported with a Niels Stensen Fellowship and a NVLE Fund travel grant. S.S. Waikar is supported by National Institutes of Health grant U01DK0856660.