Associations of Genetic Factors, Educational Attainment, and Their Interaction With Kidney Function Outcomes

Abstract

Both genetic predisposition and low educational attainment (EA) are associated with higher risk of chronic kidney disease. We examined the interaction of EA and genetic risk in kidney function outcomes. We included 3,597 participants from the Prevention of Renal and Vascular End-Stage Disease Cohort Study, a longitudinal study in a community-based sample from Groningen, the Netherlands (median follow-up, 11 years; 1997–2012). Kidney function was approximated by obtaining estimated glomerular filtration rate (eGFR) from serum creatinine and cystatin C. Individual longitudinal linear eGFR trajectories were derived from linear mixed models. Genotype data on 63 single-nucleotide polymorphisms, with known associations with eGFR, were used to calculate an allele-weighted genetic score (WGS). EA was categorized into high, medium, and low. In ordinary least squares analysis, higher WGS and lower EA showed additive effects on reduced baseline eGFR; the interaction term was nonsignificant. In analysis of eGFR decline, the significant interaction term suggested amplification of genetic risk by low EA. Adjustment for known renal risk factors did not affect our results. This study presents the first evidence of gene-environment interaction between EA and a WGS for eGFR decline and provides population-level insights into the mechanisms underlying socioeconomic disparities in chronic kidney disease.

Abbreviations

BMI

body mass index

CKD

chronic kidney disease

EA

educational attainment

eGFR

estimated glomerular filtration rate

GWAS

genome-wide association study

PREVEND

Prevention of Renal and Vascular End-Stage Disease

SBP

systolic blood pressure

SE

standard error

SNP

single-nucleotide polymorphism

UAE

urinary albumin excretion

WGS

weighted genetic score

Chronic kidney disease (CKD) is a heterogeneous group of disorders characterized by sustained kidney dysfunction and/or signs of kidney damage (1). CKD is associated with cardiovascular morbidity and all-cause mortality (2). It can eventually progress to end-stage kidney disease, necessitating the start of renal replacement therapy. The incidence of CKD is increasing, posing a major global health challenge (3–5).

Over the past 2 decades, evidence has accumulated for a socioeconomic gradient in CKD: Low educational attainment (EA), as an indicator of low socioeconomic status, is associated with reduced kidney function (according to estimated glomerular filtration rate, eGFR) and with higher rates of kidney damage (according to urinary albumin excretion, UAE) (6, 7). Recent data suggest that indicators of socioeconomic status including EA are linked with CKD through poor health behaviors (e.g., smoking, diet, sedentary time), higher prevalence of known clinical risk factors (hypertension, diabetes, hypercholesterolemia, obesity), and poor health-care access (8, 9), each contributing to an environment that is deleterious for kidney health.

In addition to environmental factors, there is strong evidence for a genetic influence on CKD. Familial clustering is observed in CKD (10–13), and heritability of CKD-defining traits has been estimated to be 36%–75%. Further evidence is provided by genome-wide association studies (GWAS) that identified >60 single-nucleotide-polymorphisms (SNPs) associated with creatinine-based eGFR (eGFRcrea) (14). Genetic scores constructed from these SNPs represent a genetic component to kidney function and thus can be interpreted as a proxy of genetic liability to CKD (15–17).

Some evidence exists, albeit conflicting, that higher education counteracts the genetic risk of diabetes (18, 19) and obesity (18, 20, 21), both important determinants of CKD. Therefore, it is possible that higher education also counteracts genetic risk of CKD, or conversely, that low education amplifies the genetic risk of CKD. Uncovering modifying effects of education on genetic risk might facilitate improved risk stratification based on education and genetics. Furthermore, knowledge of modifying effects of education provides support for public health policies (e.g., in managing purported downstream effects of low education to improve kidney outcomes). The joint associations of education and genetic factors have not previously been examined in the context of kidney disease. Thus, our aim was to investigate the interaction between education and genetic predisposition for CKD in the general population. Specifically, we aimed to test the hypothesis that lower EA amplifies genetic risk of reduced kidney function.

METHODS

Study sample and design

We used data from the Prevention of Renal and Vascular End-Stage Disease (PREVEND) Cohort study. PREVEND was initiated to investigate the natural course of increased urinary albumin levels and its association with renal and vascular outcomes. Details have been described elsewhere (22). Briefly, 8,592 individuals, sampled from the general population of Groningen, the Netherlands, underwent an extensive baseline examination between 1997–1998. Four follow-up examinations were completed in 2003, 2006, 2008, and 2012. All subjects gave written informed consent. PREVEND was approved by the medical ethics committee of the University Medical Center Groningen and conducted in accordance with the Helsinki Declaration guidelines. For this study, we used the subset of participants that was genotyped (n = 3,649). Given that participants might receive education into their 20s, we excluded those aged <30 years (n = 52) from the analyses, resulting in n = 3,597.

Measurements

Kidney function.

Kidney function was approximated by eGFR from creatinine and cystatin C. Measurement of serum creatinine was performed by an enzymatic method on a Roche Modular analyzer using reagents and calibrators from Roche (Roche Diagnostics, Mannheim, Germany), traceable to isotope dilution mass spectrometry, with intra- and interassay coefficients of variation of 0.9% and 2.9%, respectively. Serum cystatin C concentration was measured by a Gentian cystatin C Immunoassay (Gentian AS, Moss, Norway) on a Modular analyzer (Roche Diagnostics). Cystatin C was calibrated directly using the standard supplied by the manufacturer (traceable to the International Federation of Clinical Chemistry Working Group for Standardization of Serum Cystatin C) (23). The intra- and interassay coefficients of variation were <4.1% and <3.3%, respectively. Serum creatinine and serum cystatin C were determined in a single run to avoid laboratory day-to-day variation. We calculated eGFR from both serum creatinine and serum cystatin C, using the corresponding Chronic Kidney Disease–Epidemiology collaboration (CKD-EPI) equation (24). Outliers exceeding 4 standard deviations from the mean were excluded.

Genotyping and genetic risk score calculation.

Genotyping details for PREVEND were described previously (17). Briefly, genotyping was performed on the Illumina CytoSNP-12 v2 chip (Illumina, San Diego, California). Samples with call rate <95%, duplicates, and sex discrepancies were excluded. Markers with call rate >95%, Hardy-Weinberg equilibrium P ≥ 10⁻⁵, and minor allele frequency ≥1% were included. Variants were imputed to 1000 Genomes (https://www.internationalgenome.org/home), Phase 1 version 3, using Minimac software (25). To account for population stratification, principal component analysis was performed (26); the resulting principal components represent possible population substructures in PREVEND. In order to remove ethnic outliers, samples with a z score of >3 for any of the first 5 principal components with the highest eigen values were excluded. From the resulting GWAS data, we extracted genotypes of 63 known eGFR SNPs identified in a meta-analysis of GWAS on eGFRcrea in European populations (14). We constructed a weighted genetic score (WGS) comprising these SNPs. Per individual, effect alleles were weighted for their published effect sizes and summed. We then standardized the scores by subtracting the population mean score and dividing by the population standard deviation. Effect alleles were those reported to associate with lower eGFR, thus a higher WGS reflects genetic predisposition toward lower kidney function.

Educational attainment.

Educational attainment (EA) was assessed with self-report questionnaires. EA levels specific to the Netherlands were mapped to the International Standard Classification of Education (27). We then categorized EA into low (no, primary, basic vocational, and secondary education, corresponding to International Standard Classification of Education levels 0–2), medium (senior secondary vocational and general senior secondary education, International Standard Classification of Education levels 3–4), and high (higher professional and higher academic education, International Standard Classification of Education levels 5–6). International Standard Classification of Education levels were imputed to US years of schooling. High EA was the reference category in all analyses.

Covariates.

We adjusted for age, age², and sex. To minimize potential confounding by population stratification, we additionally adjusted for the first 10 genetic principal components. In longitudinal analyses, we additionally adjusted for baseline eGFR. Furthermore, we explored models that include the renal risk factors, body mass index (BMI, calculated as weight (kg)/height (m)²), systolic blood pressure (SBP), glucose, total cholesterol, and smoking status (never-smoker, former smoker, current smoker), each measured at baseline. Furthermore, we adjusted for natural log-transformed urinary albumin excretion (lnUAE), an indicator of kidney damage, measured in two 24-hour urine collections at baseline. In sensitivity analyses, we adjusted for hypertension and diabetes rather than SBP and glucose. For continuous variables, outliers exceeding 4 standard deviations from the mean were excluded.

Statistical analyses

All analyses were performed using R, version 3.5.1 (R Foundation for Statistical Computing, Vienna, Austria) (28).

To assess the explained variance of eGFR by the WGS, conditional on age, age², sex, and the first 10 principal components, ΔR²_adjusted was computed from nested ordinary least squares regression models using the lm() function from the stats R package. We tested associations between the WGS and EA using 1-way analysis of variance implemented in the aov() function from the stats R package.

Cross-sectional analyses, with baseline eGFR as outcome, were performed using ordinary least squares regression analysis using the lm() function implemented in the stats R package. For longitudinal analyses, we performed a 2-step procedure. First, we modeled linear trajectories of eGFR using linear mixed models implemented in the lme4 R package (29), with a random intercept and a random slope for time. Individual trajectories of eGFR change were then extracted and used as outcome variables (i.e., annual eGFR change) in ordinary least squares regression analysis. For both cross-sectional analyses and longitudinal analyses, 10 models were constructed with the main effects of the WGS and EA (models 1–10), in addition to their interaction term, and varying degrees of covariate adjustment (see Web Table 1, available at https://doi.org/10.1093/aje/kwaa237, for model details). Contribution of the WGS × EA interaction term was assessed using model coefficients for separate EA levels (low EA, medium EA, and the interaction of each with the WGS, with high EA as reference category), and computing the difference in adjusted explained variance (ΔR²_adjusted) between 2 nested models (with and without interaction term). To assess significance of the overall interaction term, we used an F test using the anova() function from the stats R package, through which we compared model fit between 2 nested models. We used linear regression models; hence, interaction was assessed on the additive scale. A significant P value for the interaction term indicates departure from additivity. Finally, EA-stratified models, with varying degrees of covariate adjustment, were constructed (models 11–15). For all models, we performed complete-case analysis. We applied a 2-sided significance threshold of α = 0.05 unless otherwise specified.

RESULTS

Baseline characteristics

Baseline characteristics of participants, according to categories of EA, are presented in Table 1. Lower EA was generally associated with a less-favorable renal risk profile (lower eGFR, higher BMI, higher SBP, higher glucose, higher cholesterol, and higher prevalence of smoking).

Table 1.

Baseline Characteristics Overall and According to Educational Attainment, Prevention of Renal and Vascular End-Stage Disease, the Netherlands, 1997–2012

			Educational Attainment
Characteristic	Total (n = 3,597)		Low (n = 1,673)		Medium (n = 889)		High (n = 1,035)
	Mean (SD)	%	Mean (SD)	%	Mean (SD)	%	Mean (SD)	%
Age, years	50 (40–60)a		55 (46–65)a		46 (37–56)a		44 (37–51)a
Male sex		52		49		56		53
eGFR, mL/minute/1.73 m2	94.7 (17.0)		90.5 (17.3)		97.1 (17.0)		99.3 (14.8)
US years of schooling	12.9 (5.0)		8.5 (1.5)		13 (0)		20 (0)
WGS	0 (1.0)		0.02 (1.0)		−0.02 (1.0)		−0.01 (1.0)
Number of effect alleles	62.3 (4.9)		62.3 (4.9)		62.3 (5.1)		62.3 (4.8)
SBP, mm Hg	129 (19.7)		133 (20)		128 (20)		124 (18)
Hypertension		35		46		31		21
Glucose, mmol/L	4.8 (0.8)		5.0 (0.8)		4.7 (0.7)		4.6 (0.6)
Type 2 diabetes		4.0		5.7		3.6		1.7
Body mass indexb	26 (4.1)		27 (4.2)		26 (4.0)		25 (3.5)
Total cholesterol, mmol/L	5.7 (1.1)		5.9 (1.1)		5.6 (1.1)		5.4 (1.0)
Never smoker		27		23		26		36
Former smoker		37		37		38		37
Current smoker		35		40		36		27
Follow-up time, years	11.0 (4.6–11.9)a		9.9 (4.2–11.6)a		11.1 (4.8–12.2)a		11.2 (6.2–12.4)a

Abbreviations: eGFR, estimated glomerular filtration rate; SBP, systolic blood pressure; SD, standard deviation; WGS, weighted genetic risk score.

^a Values are expressed as median (interquartile range).

^b Weight (kg)/height (m)².

We regressed baseline eGFR on the WGS to obtain a crude association. The association of the WGS with baseline eGFR was modest but highly significant (B = −1.68 (standard error (SE), 0.29), R²_adjusted = 0.010; P = 8.6 × 10⁻⁹).

In Web Figure 1, we show a plot of WGS distribution by categories of EA. The WGS was normally and equally distributed in each EA category. The mean WGS did not significantly differ between EA categories (F_{(2, 3,594)} = 0.455; P = 0.635).

Interaction analyses

Cross-sectional analysis.

A plot of baseline eGFR by the WGS and strata of EA is presented in Figure 1. On visual inspection of this data, the association of the WGS with eGFR appeared to be consistent across strata of EA; hence, we anticipated that the term for interaction between the WGS and EA in our models would not be significant. In unadjusted models (models 1–2), both the WGS and EA were independently associated with eGFR (Table 2). A 1-standard-deviation increase in the WGS was associated with 1.61-mL/minute/1.73 m² lower eGFR (model 1, B = −1.61 (SE, 0.28); P = 1.5 × 10⁻⁸), while those with low EA were observed to have the lowest mean eGFR (model 1, low vs. high EA, B = −8.74 (SE, 0.67); P = 5.9 × 10⁻³⁸ (Table 2)). Addition of an interaction term (WGS × EA) did not contribute to the model (model 2 vs. model 1, P = 0.512 (Table 3)). Adjustment for covariates (models 3–4; age, age², sex, and the first 10 principal components) did not affect the association of the WGS with baseline eGFR. However, the association between EA and baseline eGFR disappeared due to strong confounding by age. Inclusion of additional covariates (models 5–8) did not change our conclusions, although, counterintuitively, low EA was significantly associated with higher eGFR in these models. The association of the WGS with baseline eGFR appeared smaller in the low-EA stratum (Figure 2), but the interaction was nonsignificant for all models.

Figure 1.

Plots of estimated glomerular filtration rate (eGFR) versus a weighted genetic score for reduced eGFR, according to educational attainment, Prevention of Renal and Vascular End-Stage Disease, the Netherlands, 1997–2012. Upper panels show plots of cross-sectional eGFR (mL/minute/1.73 m²) versus a weighted genetic score, stratified by levels of educational attainment: high (A), medium (B), and low (C). Lower panels show plots of annual change in eGFR (mL/minute/1.73 m² per year) versus a weighted genetic score, stratified by levels of educational attainment: high (D), medium (E), and low (F). Regression lines with 95% confidence interval are derived from unadjusted ordinary linear regression.

Table 2.

Results of Interaction Analysis From Ordinary Least Squares Regression Analysis, Prevention of Renal and Vascular End-Stage Disease, the Netherlands, 1997–2012^a

Parameter	Model 1			Model 2			Model 3			Model 4
	B	SE	P Value	B	SE	P Value	B	SE	P Value	B	SE	P Value
eGFR, mL/minute/1.73 m2
Intercept	99.27	0.52	0	99.27	0.52	0	91.39	0.56	0	91.39	0.56	0
WGS (per SD)	−1.61	0.28	1.5 × 10−8	−2.04	0.55	1.9 × 10−4	−1.76	0.22	1.4 × 10−15	−2.12	0.42	5.3 × 10−7
Educational attainment
Low	−8.74	0.67	5.9 × 10−38	−8.74	0.67	5.7 × 10−38	0.24	0.56	0.674	0.23	0.56	0.677
Medium	−2.18	0.77	4.9 × 10−3	−2.18	0.77	5.0 × 10−3	0.06	0.60	0.914	0.07	0.60	0.91
High	0	Referent		0	Referent		0	Referent		0	Referent
WGS × educational attainment
WGS × low				0.77	0.69	0.265				0.60	0.53	0.256
WGS × medium				0.29	0.77	0.711				0.29	0.60	0.628
WGS × high				0	Referent					0	Referent
Annual eGFR change, mL/minute/1.73 m2 per yearb
Intercept	−1.089	0.041	0	−1.090	0.041	0	−0.695	0.048	0	−0.697	0.048	0
WGS (per SD)	−0.016	0.007	0.014	0.004	0.013	0.746	−0.027	0.006	2.3 × 10−5	−0.008	0.012	0.52
Educational attainment
Low	−0.125	0.016	3.3 × 10−15	−0.124	0.016	3.7 × 10−15	−0.054	0.016	7.9 × 10−4	−0.054	0.016	8.1 × 10−4
Medium	−0.042	0.018	0.018	−0.042	0.018	0.018	−0.026	0.017	0.131	−0.026	0.017	0.13
High	0	Referent		0	Referent		0	Referent		0	Referent
WGS × Educational attainment
WGS × low				−0.037	0.016	0.018				−0.036	0.015	0.017
WGS × medium				−0.011	0.018	0.537				−0.009	0.017	0.588
WGS × high				0	Referent					0	Referent

Abbreviations: eGFR, estimated glomerular filtration rate; SD, standard deviation; SE, standard error; WGS, weighted genetic score.

^a Models with additional covariate adjustment (models 5–10) are presented in Web Table 1.

^b For longitudinal analysis, baseline eGFR was added to each model. Model 1: WGS + EA; model 2: model 1 + WGS × EA; model 3: WGS + EA + age + age² + sex + genetic principal components 1–10; model 4: model 3 + WGS × EA

Table 3.

Comparison of Nested Ordinary Least Squares Regression Models With and Without a Term for the Interaction of Weighted Genetic Score and Educational Attainment, Prevention of Renal and Vascular End-Stage Disease, the Netherlands, 1997–2012

Model	R 2	R 2 adj	Residual df	RSS	Δdf	F Statistic	Pr(>F) a
eGFR
Model 1b	0.064	0.063	3,362	908,742
Model 2c	0.064	0.063	3,360	908,379	2	0.670	0.512
Model 3d	0.448	0.446	3,349	535,519
Model 4e	0.448	0.445	3,347	535,307	2	0.634	0.515
Model 5f	0.463	0.459	3,230	498,513
Model 6g	0.463	0.459	3,228	498,367	2	0.475	0.622
Model 7h	0.462	0.458	3,195	485,857
Model 8i	0.462	0.458	3,193	485,720	2	0.448	0.634
Model 9j,l	0.462	0.458	3,086	471,547
Model 10k,m	0.462	0.458	3,084	471,432	2	0.377	0.686
Annual eGFR changec
Model 1b	0.041	0.040	3,342	473.93
Model 2c	0.043	0.041	3,340	473.03	2	3.177	0.042
Model 3d	0.112	0.108	3,329	438.62
Model 4e	0.114	0.109	3,327	437.74	2	3.319	0.036
Model 5f	0.130	0.124	3,213	411.88
Model 6g	0.132	0.125	3,211	411.17	2	2.777	0.062
Model 7h	0.132	0.126	3,178	405.28
Model 8i	0.133	0.126	3,176	404.67	2	2.407	0.090
Model 9j,l	0.132	0.125	3,067	403.74
Model 10k,m	0.134	0.126	3,065	402.94	2	3.026	0.049

Abbreviations: df, degrees of freedom; EA, educational attainment; eGFR, estimated glomerular filtration rate; PC, principal components; R², model-explained variance; RSS, residual sum of squares; UAE, urinary albumin excretion; WGS, weighted genetic score.

^a P values Pr(>F) derived from F test using analysis of variance between 2 nested models.

^b Model 1: (WGS + EA)

^c Model 2: model 1 + (WGS × EA)

^d Model 3: WGS + EA + age + age² + sex + genetic PC 1–10

^e Model 4: model 3 + WGS × EA

^f Model 5: WGS + EA + age + age² + sex + PC 1–10 + BMI + SBP + glucose + total cholesterol + smoking

^g Model 6: model 5 + WGS × EA

^h Model 7: WGS + EA + age + age² + sex + PC 1–10 + BMI + SBP + glucose + total cholesterol + smoking + lnUAE

ⁱ Model 8: model 7 + WGS × EA

^j Model 9: WGS + EA + age + age² + sex + PC 1–10 + BMI + hypertension + diabetes + total cholesterol + smoking + lnUAE

^k Model 10: model 9 + WGS × EA

^l Diabetes (fasting glucose >7 mmol/L or non-fasting glucose >11 mmol/L or pharmacy-reported antidiabetic medication or self-reported diabetes) and hypertension (SBP >140 or DBP >90 or pharmacy-reported antihypertensive medication or self-reported hypertension)

^m For longitudinal analysis, baseline eGFR was included in each model.

Figure 2.

Multivariable-adjusted associations of the weighted genetic score with estimated glomerular filtration rate (eGFR) in strata of educational attainment, Prevention of Renal and Vascular End-Stage Disease, the Netherlands, 1997–2012. Estimates of the associations, presented as regression coefficients with 95% confidence intervals, of the weighted genetic score (per standard deviation) with cross-sectional eGFR (mL/minute/1.73 m²) (A) and annual eGFR change (mL/minute/1.73 m² per year) (B), derived from ordinary least squares regression analysis in the entire study population and in strata of educational attainment (high, medium, low). The solid lines represent an estimate of the interaction effect for the unadjusted model (model 11), while the dashed lines represent the interaction effect in the full model (model 15) if it were linear. Model 11: weighted genetic score. Model 12: model 11 + age + age² + sex + genetic principal components 1–10. Model 13: model 12 + BMI + SBP + glucose + total cholesterol + smoking. Model 14: model 13 + lnUAE. Model 15: model 11 + BMI + hypertension + diabetes + total cholesterol + smoking + lnUAE. For longitudinal analysis, baseline eGFR was included in each model.

Longitudinal analysis.

Median follow-up duration was 11 years (interquartile range, 4.6–11.9 years). In the total population, the average change in eGFR was −0.927 (standard deviation, 0.385) mL/minute/1.73 m² per year. A plot of eGFR change by the WGS and strata of EA is presented in Figure 1. In this figure, the WGS is shown to have its strongest association with eGFR change in those with low EA (Figure 1C). In those with medium or high EA (Figure 1A–1B), the WGS had no apparent association with eGFR change. A trend in mean eGFR change was observed across EA levels, with those who had lower EA having faster rates of decline on average.

In unadjusted models (models 1–2), a 1-standard-deviation increase in the WGS was associated with 0.016-mL/minute/1.73 m² per year faster eGFR decline (model 1, B = −0.016 (SE, 0.007); P = 0.014, Table 2), and EA (model 1, low vs. high EA, B = −0.125 (SE, 0.016); P = 3.3 × 10⁻¹⁵) was also independently associated with rate of kidney function decline. Adjustment for covariates (models 3–4; age, age², sex, and the first 10 genetic principal components) increased the association of the WGS with eGFR change (model 3, B = −0.027 (SE, 0.006); P = 2.3 × 10⁻⁵), while attenuating the association of EA with eGFR change (model 3, low vs. high EA, B = −0.054 (SE, 0.016); P = 7.9 × 10⁻⁴). A WGS × EA interaction term was in the expected direction (model 4, low vs. high EA, B = −0.036 (SE, 0.015); P = 0.017), suggesting that the joint association of the WGS and EA is greater than the sum of their main associations. The contribution of the overall term for interaction between the WGS and EA was modest but significant (model 4 vs. model 3, P = 0.036 (Table 3)).

The influence of potential mediators (i.e., BMI, SBP, glucose, total cholesterol, and smoking status) on the interaction was assessed in our final models (models 5–6). Addition of these risk factors did not affect the association between the WGS and eGFR change (model 5, B = −0.026 (SE, 0.006); P = 3.7 × 10⁻⁵), whereas the association of EA was slightly attenuated (model 5, low vs. high EA, B = −0.056 (SE, 0.017); P = 8.5 × 10⁻⁴), suggesting potential mediation by these risk factors (Web Table 1). Potential mediation was further supported by the finding that the overall interaction effect was only borderline significant after addition of these risk factors (model 6 versus model 5, P = 0.062 (Table 3)), although the interaction effect of the WGS with low versus high EA was not attenuated and remained nominally significant (model 6, B = −0.034 (SE, 0.015); P = 0.027) (Web Table 1). Adjustment for lnUAE did not affect our results (models 7–8). The WGS most strongly associated with annual eGFR change in the low-EA stratum (Figure 2).

Sensitivity analysis.

The WGS did not show significantly different distributions between categories of EA. However, Figure 1 and Web Figure 1 are suggestive of slight overrepresentation of a higher WGS in those with lower EA and a lower WGS in those with higher EA. To minimize bias due to potentially influential observations, we excluded 8 observations that exceeded a more stringent cutoff of 3 standard deviations from the mean. These sensitivity analyses yielded essentially the same results as our main analyses, although significance decreased slightly due to reduced statistical power (data not shown).

Furthermore, we repeated all analyses for eGFR estimated from serum creatinine only (eGFRcrea), and from serum cystatin C only (eGFRcysc). Results were generally consistent with our main analysis, with EA being more strongly associated with eGFRcysc than with eGFRcrea. Similarly, interaction effects between the WGS and EA were more pronounced for eGFRcysc than for eGFRcrea (data not shown).

We repeated the interaction analyses using a linear mixed model only. Here, despite some minor discrepancy with longitudinal estimates from ordinary least squares regression analysis, effect estimates were generally and directionally consistent with the ordinary least squares analysis (Web Table 2), and a 3-way interaction term to assess the modifying effect of EA on WGS in eGFR change (WGS × EA × time) was again significant (Web Table 3).

Adjustment for hypertension and diabetes, rather than SBP and glucose, did not affect our results (model 9–10, Table 3, Web Tables 1–3).

DISCUSSION

In the present study, we investigated the associations of genetic factors (summarized by a WGS) and EA, as well as the interaction between the WGS and EA, with kidney function outcomes. We observed additive effects of the WGS and EA for baseline eGFR in cross-sectional analyses, although these were not robust to covariate adjustment. In longitudinal analyses, low EA interacted with high WGS, resulting in faster eGFR decline. This interaction suggests an amplifying effect of low EA on genetic risk, and it could not be explained by a less-favorable renal risk factor profile in those with low EA (i.e., higher BMI, higher SBP, higher glucose, higher cholesterol, and higher prevalence of smoking).

In the present study, participants with low EA had similar genetic risk of lower eGFR compared with those with higher EA, given that the WGS was equally distributed to each stratum of EA. However, the impact of genetic risk on annual eGFR decline was observed to be larger in those with low EA, resulting in a disproportionally fast eGFR decline among the most vulnerable in terms of EA and genetic predisposition. Low EA is unlikely to directly amplify genetic risk of reduced eGFR. Rather, it might act through a range of interrelated purported downstream effects of low EA, such as lower income, poor health behavior, poor health-care access, and higher prevalence of traditional renal risk factors (8, 9). In our analyses, the interaction effect was not explained by traditional renal risk factors. Therefore, other factors likely exist that explain the interaction between EA and a WGS. These might include factors with socioeconomic gradients such as health literacy (30), occupational exposures, and infections (31), whose influence might not be captured by traditional risk factors.

In cross-sectional analyses, low EA was significantly associated with higher eGFR (models 5–10), suggesting a paradoxical protective effect of low EA on kidney function. However, this is likely the result of overadjustment bias (32) in these models, given that many of the covariates (e.g., hypertension and diabetes) are purported mediators in the relationship of EA and eGFR (8, 9).

Each of the 63 SNPs that were identified in previous GWAS on eGFRcrea (14) have small effect sizes. The WGS aggregates these SNPs, thereby greatly increasing statistical power compared with using single SNP associations. Therefore, the WGS is a practical summary score of genetic risk for reduced kidney function. However, some limitations with regards to the WGS must be addressed. The WGS explained only a small fraction of between-individual variation in eGFR in PREVEND. Sample sizes and thus power for GWAS on eGFR have recently greatly increased, facilitating the detection of over 200 additional genetic variants (33). Using a more comprehensive WGS that includes these variants likely increases power to detect interactions. In addition, participants with an equal WGS might have different underlying risk variants. Furthermore, by using a WGS in interaction analysis, it is implicitly assumed that all genetic variants included in the WGS have directionally consistent interaction effects with EA. Another implicit assumption is that the same set of genetic variants affect eGFR in each category of EA. To check these assumptions, single SNP interaction effects would need to be assessed, but this requires infeasibly large sample sizes and is therefore beyond the scope of the present study. Future research could include genome-wide interaction studies to identify the specific genetic variants whose associations with kidney function are modified by EA. Similar studies were performed for blood pressure, BMI, and lipids to identify genetic variants whose effect was modified by smoking, alcohol use, and physical activity (34–37).

For the longitudinal analyses, we reported results from a 2-step method in which we used individual eGFR trajectories, extracted from a linear mixed model, as the outcome variable in ordinary least squares regression analysis. This allows for straightforward estimation of model R² and intuitive interpretation of the WGS × EA 2-way interaction term. The 2-step approach potentially comes at the cost of introducing false precision in eGFR trajectories given that random variation in eGFR measurements during follow-up is ignored to an extent. This might explain the result in previous study in PREVEND, of a WGS comprising 63 SNPs showing similar associations with eGFR change compared with the present study but not reaching statistical significance in linear mixed model analysis (17). Alternatively, the associations of the WGS, EA, and the WGS × EA interaction term on eGFR change can also be modeled in a single linear mixed model, taking into account the random variation and correlation between eGFR measurements. However, R² estimation is not straightforward in linear mixed models, and estimation of the interaction effect on eGFR change requires modeling a 3-way interaction term (WGS × EA × time), the interpretation of which is less intuitive compared with that of a 2-way interaction term. We performed sensitivity analyses using a linear mixed model only. Notwithstanding some discrepancies with the ordinary least squares analysis regarding effect size and statistical significance, the results from linear mixed model were directionally consistent with ordinary least squares analysis, and therefore our conclusions remain unchanged.

Our study adds to the literature on socioeconomic disparities in CKD in that it is, to our knowledge, the first to present evidence of gene-environment interaction between a WGS (based on SNPs associated with eGFR) and EA. Major strengths of this study include the availability of multiple eGFR estimates per individual, based on both serum creatinine and cystatin C values, which were measured in 1 run, allowing precise estimation of glomerular filtration rate, and the considerable follow-up duration. Several limitations, other than those already discussed, need to be addressed. First, the present study population consists exclusively of participants of European ancestry, sampled from a relatively high-income population (i.e., the population of Groningen, the Netherlands). Therefore, the generalizability of these findings to non-European, lower-income populations might be limited. Second, the interaction effects of genetic risk and EA on rate of kidney function decline that we found are modest; replicability and generalizability of these results to other populations is uncertain and therefore require validation in independent samples. Under similar parameters, the interaction effect could be replicated with a sample size of approximately 5,000 (with 80% power at α = 0.05) (Web Figure 2). Third, the observational nature of this study precludes causal conclusions. Fourth, larger samples are needed to examine whether the interaction between a WGS and low EA results in increased rates of CKD. Finally, a higher attrition rate was observed in those with low education. This might have resulted in bias toward the null, or underestimation of effect sizes, due to reduced power and precision of kidney decline outcomes in this group.

Knowledge of the interaction that we found in our longitudinal analyses is unlikely to be useful for risk stratification for preventive medicine, due to the rather modest effect sizes. Furthermore, given the population-based sample, our findings might not translate into the clinic (i.e., in predicting disease progression in CKD patients). However, our results might inform public health policy given that they provide insights into the mechanisms that underlie socioeconomic disparities in CKD. For example, it is possible that downstream effects of low EA contribute to an environment that activates genetic pathways that are detrimental for kidney health. Conversely, deleterious genetic effects are suggested to be completely mitigated by high EA and its downstream effects, at least with regard to kidney function decline. Future study is needed to identify which factors are responsible for this modifying effect, given that these factors are potential targets for intervention to reduce socioeconomic disparities in CKD.

In conclusion, our findings provide population-level insights on the mechanisms underlying socioeconomic disparities in CKD. We observed that a WGS, as a summary of genetic risk, and EA have independent associations with the rate of kidney function decline. Furthermore, our results suggest a subtle amplifying effect of low EA on genetic risk of reduced eGFR. Traditional kidney risk factors that are purported downstream effects of low EA (i.e., higher BMI, higher SBP, higher glucose, higher cholesterol, and higher prevalence of smoking) did not explain the amplifying effect on the WGS, warranting further investigation.