Accuracy of glomerular filtration rate estimation using creatinine and cystatin C for identifying and monitoring moderate chronic kidney disease: the eGFR-C study

Primary objectives

The study evaluated the comparative performance of GFR-estimating equations, including those incorporating cystatin C, in assessing and monitoring GFR in people with stage 3 CKD. The data were analysed to assess the impact of ethnicity, albuminuria, diabetes and other characteristics on equation performance. The aims of the study were:

1. to estimate and compare the accuracy of GFR-estimating equations at baseline based on the MDRD equation and three CKD-EPI equations using either creatinine or cystatin C or a combination of both in individuals with stage 3 CKD

2. to estimate the accuracy of the GFR-estimating equations according to ethnic group (particularly Caucasian, South Asian and African-Caribbean), and baseline diabetes, albuminuria, age, gender, BMI and mGFR level

3. to evaluate and compare how accurately these GFR-estimating equations track mGFR and detect change in mGFR over 3 years

4. to estimate the biological variability of mGFR and eGFR

5. to establish which GFR-estimating equation, together with urinary ACR, or ACR alone, most accurately predicts mortality and those individuals that have progressive loss of kidney function (CKD progression)

6. to estimate and model disease progression (decline in GFR or increase in ACR) and differences in progression between ethnic groups (Caucasian, South Asian and African-Caribbean), baseline diabetes and albuminuria status and other potential risk factors

7. to explore the comparative cost effectiveness of monitoring strategies for identifying people who have progressive loss of kidney function (CKD progression) utilising different GFR-estimating equations.

Secondary objectives

1. To estimate and compare the accuracy of more recent GFR-estimating equations that have been published while the study has been ongoing including BIS1 and BIS2 equations, CAPA equation, LMR equation, FAS_creatinine equation, FAS_{creatinine-cystatin} equation, EKFC equation and the 2021 revisions of the CKD-EPI equations (Table 1).

2. To evaluate and compare how accurately these newer GFR-estimating equations reflect and detect change in GFR over 3 years.

3. To estimate and compare the performance of the MDRD equation and CKD-EPI equations using the Haycock equation92 for BSA adjustment instead of the Du Bois equation. 93

4. To assess the impact of cystatin C calibration on the performance of the CKD-EPI equations.

5. To assess the impact of creatinine methodology [enzymatic vs. isotope dilution mass spectrometry (ID-MS)] on the performance of the MDRD equation and creatinine-based CKD-EPI equations.

Recruitment

Adults with stage 3 CKD were recruited to the study at six centres, with a target case mix as follows:

1. Birmingham – 50% Caucasian, 25% South Asian, 25% African-Caribbean from secondary care

2. Canterbury – predominantly Caucasian cohort from secondary care

3. Derby – predominantly Caucasian cohort from primary care

4. Leicester – 50% Caucasian, 50% South Asian, from primary and secondary care

5. Salford – predominantly Caucasian cohort from secondary care

6. London – Kings College Hospital – 50% Caucasian, 50% African-Caribbean from secondary care.

Participants were recruited from both primary and secondary/tertiary care. Recruitment in secondary/tertiary care was primarily from CKD clinics. Potential participants were identified by the research nurse/co-investigator at each of the recruiting centres from the individual renal unit databases. Recruitment from primary care utilised the following approaches. In Leicester, letters were sent to general practitioners (GPs) inviting them to participate. Research active practices were approached by the Clinical Research Network and consenting practices were given instructions to help identify eligible people from the CKD register, for example using READ code searches. MIQUEST (Morbidity Information QUery and Export SynTax) software was used to extract an up-to-date data set. The practices sent invitation letters with a stamped envelope to eligible people. The invitation was sent out with the patient information sheet. Potential recruits were given a dedicated study phone line to use to indicate willingness to participate and the research assistant then telephoned willing participants to further discuss the study and schedule hospital attendance. Signed informed consent was obtained upon hospital attendance for the reference GFR test (see below). In Derbyshire, a similar process was followed except that eligible participants were identified from a database of participants from the Renal Risk in Derby study,94 a cohort study of persons with CKD in primary care.

Sampling and data collection

Baseline visit: Participants were asked to attend hospital in the morning after having been advised to consume a light breakfast (no meat or fish). A clinical and drug history was recorded using a standardised questionnaire taken by research nurses on the day of hospital attendance. Vascular disease was defined as the presence of MI (including ST-elevation myocardial infarction and non-ST elevation myocardial infarction), angina, congestive cardiac failure (heart failure) or a requirement for coronary intervention (angioplasty, coronary artery bypass graft or pacemaker), cerebrovascular or peripheral vascular disease. Information on ethnicity was gathered using a modified version of the 2011 UK Census Questionnaire, with ethnicity being mapped to the following codes:

Caucasian 31, 32, 33, 34; South Asian 39, 40, 41; African-Caribbean 44, 45, 46.

Height was measured to the nearest 0.1 cm with a rigid stadiometer. Body weight was measured in light indoor clothing to the nearest 0.1 kg. Waist circumference was recorded to the nearest 0.1 cm at the mid-point between the lower costal margin and the level of the anterior superior iliac crest. Brachial BP was measured as recommended by the British and Irish Hypertension Society [https://bihsoc.org/resources/bp-measurement/measure-blood-pressure/ (accessed 26 July 2023)] three times in the sitting position using standardised Omron M7 digital sphygmomanometers (Omron Healthcare, Milton Keynes, UK). The average of the second and third BP readings was recorded.

Baseline blood was taken for serum creatinine and cystatin C, and a urine sample was collected for ACR. Blood samples were also taken for haemoglobin and glycated haemoglobin (only if known to have diabetes) measurement. Further aliquots of serum and urine were stored for potential analysis of future markers of GFR or disease progression. Blood samples were collected using standard venepuncture and phlebotomy procedures including the use of a tourniquet. Blood was collected in appropriate Greiner Vacuette^TM tubes [www.gbo.com (accessed 26 July 2023)] following the manufacturer’s recommended order of draw. The urine sample was taken into a plain Sterilin pot. Samples were transported to the local laboratory, where plasma/serum was separated within 4–6 hours of venepuncture by centrifuging at 2000 g for 10 minutes. Aliquots of serum/plasma and urine were then stored at −80 °C pending transportation to the central laboratories [St. Thomas’s (iohexol, ID-MS creatinine) or Canterbury (enzymatic creatinine, cystatin C, ACR) depending on analyte] and analysis.

Glomerular filtration rate was measured using an iohexol clearance method. 97 A 5 ml bolus of Omnipaque 240 (518 g/l iohexol corresponding to 240 g/l of iodine, GE Healthcare [www.gehealthcare.co.uk/ (accessed 12 April 2023)] followed by 10 ml of normal saline was injected into the antecubital vein. Blood samples were collected at 5, 120, 180 and 240 minutes after injection. Exact time of the samples in relation to the bolus injection was accurately recorded. Participants were allowed free access to fluids during the collection procedure but were asked to refrain from protein intake (i.e. biscuits/toast would be permitted) and to refrain from excessive exercise. Samples were stored at −80 °C prior to analysis. Iohexol was determined using an ID-MS method (see below) and GFR calculated. 98

Glomerular filtration rate was estimated using published GFR-estimating equations: the simplified ID-MS traceable version of the MDRD equation and the three CKD-EPI equations (CKD-EPI_creatinine, CKD-EPI_cystatin and CKD-EPI_{creatinine-cystatin}) for the primary study objectives. For the secondary study objectives, the BIS1 and BIS2, CAPA, LMR, FAS_creatinine, FAS_{creatinine-cystatin}, EKFC and the 2021 revisions of the CKD-EPI equations were studied (Table 1).

Follow-up: Participants were followed for 3 years. All the above measurements and the clinical history were repeated at 36 months. At each 6-month interval, blood samples of all participants were taken for serum creatinine, according to standard care,40 in addition to cystatin C and a urine sample was collected for ACR: GFR was estimated as above. All measurements were undertaken in accordance with standard operating procedures by trained staff. Nursing staff familiar with these procedures at the Canterbury centre cascade trained other recruiting centres. Clinicians and others involved in patient care were blinded to the reference measured (iohexol) GFR and cystatin C-based eGFR results for the duration of the study. During the course of the study, participants were given advice regarding the management of their CKD according to standard evidence-based practice. 40

Laboratory analyses

Iohexol was measured using electrospray isotope dilution tandem mass spectrometry on an ABSCIEX API6500 Q-trap (ABSCIEX, Warrington, UK) mass spectrometer. 97 Iohexol stock standard, 10 mmol/l, was prepared by diluting Omnipaque 300 solution (647 g/l) in deionised water and stored in 1 ml aliquots at −80 °C. Aqueous iohexol calibrators (0, 10, 100 and 500 µmol/l) were prepared from the stock iohexol standard by dilution and stored in 0.5 ml aliquots at −80 °C. Iohexol stable isotope, d5-iohexol (Toronto Research Chemicals Inc., Toronto, ON, Canada), was obtained from 2BScientific Ltd, Upper Heyford, UK, dissolved in deionised water at circa 10 mmol/l, and stored at −80 °C. Plasma control samples were prepared by spiking a plasma pool with iohexol stock standard at 10, 100 and 400 µmol/l. Calibrators, controls, patient samples and stable isotope stock solutions were thawed from frozen on a roller mixer at room temperature for no more than 60 minutes, and then centrifuged for 4 minutes at 1500 g at 4 °C (Eppendorf 5810R centrifuge, VWR International Ltd, Lutterworth, UK). Working iohexol stable isotope was prepared by diluting the circa 10 mmol/l solution 1 : 200 with deionised water. Calibrators, controls and samples were pipetted (20 µl) into 2 ml microcentrifuge tubes [000-MICR-200, Elkay Laboratory Products (UK) Ltd, Basingstoke, UK] and 50 µl working iohexol stable isotope, followed by 200 µl acetonitrile (Rathburn Chemicals Ltd, Walkerburn, UK) were added to each tube. Samples were capped, vortex-mixed for 5 seconds and centrifuged for 5 minutes at 20,800 g at 4 °C (Eppendorf 5417R centrifuge, VWR International Ltd, Lutterworth, UK). Supernatants (200 µl) were then transferred into a 96 deep well plate and loaded onto the autosampler. The sample (2 µl) was automatically injected into a mobile phase stream of acetonitrile : water (1: 1) with 0.025% formic acid using a Hewlett-Packard 1100 Series autosampler and pump (Applied Biosystems, Warrington, UK) at 250 µl/minute. Chromatography was performed on a Chirobiotic T 100 × 2.1 mm column with a 2 cm × 4.0 mm guard column (Sigma-Aldrich Company Ltd, Poole, UK).

Tandem mass spectrometry was performed in positive ion multiple reaction monitoring mode: iohexol 821.849/602.8, d5-iohexol 826.849/607.8. Data acquisition time was 6 minutes with a pause time of 5.0070 milliseconds between transitions and a scan speed of 10 Da/s. Iohexol concentrations were calculated in Analyst 1.6 (ABSCIEX, Warrington, UK) using the ratio of sample peak area to stable isotope peak area. Between-day imprecision (CV, %) was 1.0%, 0.8% and 1.5% at 10, 100 and 400 μmol/l, respectively. The laboratory participated in an international proficiency testing scheme [EQUALIS, https://equalis.se/en/ (accessed 5 April 2023)] for iohexol measurement with good performance.

Iohexol concentrations were log-transformed (natural log) and plotted as a function of time. GFR was calculated from the slope-intercept method using a single compartment model:99,100

To ensure integrity of the iohexol procedure (dose administration, sample collection, sample labelling, and iohexol analysis), the iohexol data were rigorously reviewed for every mGFR. The 5-minute sample enabled identification of procedures where the iohexol was given subcutaneously in error, or where saline flushing of the infusion line was suboptimal, as demonstrated by low and high iohexol concentrations respectively. In addition, all procedures where the iohexol concentration versus time correlation coefficient (r) was < 0.98 (< 6% of total procedures) were re-analysed to check for any within-assay sample transposition.

Glomerular filtration rate was adjusted for BSA using the Du Bois equation93 and corrected for the fast exponential. 98 In a separate analysis, GFR was adjusted for BSA using an alternative approach (Haycock equation92) to assess the impact of these adjustments on the accuracy of GFR estimation.

Serum creatinine was measured using an enzymatic assay on an Abbott Architect analyser [Abbott Diagnostics Ltd, www.abbott.co.uk/ (accessed 12 April 2023)] standardised to the reference material, NIST SRM 967 and 914. Between-day imprecision (CV, %) was 0.8%, 0.3% and 0.4% at concentrations of 75, 176 and 760 μmol/l, respectively. The laboratory participated in an international proficiency testing scheme [UKNEQAS, https://birminghamquality.org.uk/ (accessed 5 April 2023)] for creatinine measurement and GFR estimation with satisfactory performance. Additionally, serum creatinine was measured using ID-MS on an ABSCIEX API6500 Q-trap mass spectrometer.

Cystatin C was measured by a turbidimetric immunoassay on an Abbott Architect analyser. Between-day imprecision was 2.3% and 1.6% at concentrations of 0.9 and 4.0 mg/l, respectively. The laboratory participated in an international proficiency testing scheme (EQUALIS) for cystatin C measurement and GFR estimation with good performance.

During the course of the study, we became aware of published data describing a significant positive bias of the Abbott cystatin C assay. 101 This was supported by information from the EQUALIS proficiency testing scheme and our own re-analysis of historical stored samples (data not presented here). To investigate this, a recovery study was undertaken in which lyophilised human serum cystatin C ERM-DA471/IFCC [Sigma-Aldrich Chemical Co., www.sigmaaldrich.com (accessed 5 April 2023)] was added to pooled non-uraemic serum to give samples with a range of expected concentrations covering 1.47–3.41 mg/l. These samples were analysed and the mean recovery calculated.

To further explore this bias, in a subset of samples (n = 106) covering a representative range of concentrations cystatin C was also measured by a particle-enhanced nephelometric immunoassay according to the manufacturer’s instructions on a Siemens BN Prospec analyser [www.siemens.com (accessed 12 April 2023)]. Between-batch imprecision (n = 38) for the Siemens assay was 3.5% at 0.87 mg/l and 3.6% at 4.64 mg/l. Both Abbott and Siemens assays were calibrated against the internationally certified reference material ERM-DA471/IFCC for cystatin C. 51

Prior to analysis, samples were thawed at room temperature, mixed by inversion and centrifuged prior to measurement. For the biological variation study, all samples from each individual subject were measured in duplicate in random order in a single assay. Each of the biomarker analyses was undertaken by a single operator blinded to participant data using a single instrument. Creatinine and cystatin C measurements were undertaken in an accredited laboratory by scientists registered with the Health and Care Professions Council.

Main study: prospective longitudinal cohort study

The sample size calculation for the main study focused on the ability to detect differences in accuracy of measurement between the MDRD equation and the CKD-EPI_cystatin equation. We also made secondary comparisons with the other equations. While it is relatively easy to estimate the relationship between each equation and the reference standard in smaller samples, a relatively large sample is required to have adequate statistical power to make the comparison in accuracy. Similarly, while it is possible to show a relationship between eGFR and progression within a cohort, to show that one equation predicts progression better than another is more challenging.

The measure of accuracy we used was the P30, the percentage of eGFR values within 30% of ‘true’ GFR. Describing percentiles of the distribution of the differences between estimated and mGFR was endorsed as a useful measure of accuracy by the National Kidney Foundation – Kidney Disease Outcomes Quality Initiative (NKF-KDOQI) in 2002 and has been widely used subsequently in the GFR field. 102 The approach captures aspects of both imprecision (measurement error) and bias (systematic over- and/or underestimation). NKF-KDOQI suggested that GFR equations should achieve a P30 value in excess of 90%. Having undertaken initial algebraic sample size calculations based on comparing variance estimates of measurement error, we defined our final sample size based on a simulation study to estimate differences in P30 and estimates of rate of change which are not amenable to algebraic solution. Our simulations modelled the full structure of the study including random variability, and computed statistical power through noting the percentage of simulations yielding statistically significant results for analysis of each outcome. Power estimates were based on 1000 simulations.

Values of P30 for the alternative eGFR equations have previously been estimated to range between 73% and 95% across the literature. 41 The original MDRD validation cohort achieved a P30 of 91%,31 whereas in the CKD-EPI cohort, this fell to 81%, compared to 84% for the CKD-EPI equation itself. 32 In the external validation data sets of Inker et al. 28 the CKD-EPI_cystatin and CKD-EPI_{creatinine-cystatin} equations achieved P30 values of 86% and 92%, respectively. In our recent study of GFR-estimating equation performance in older people, the MDRD equation achieved a P30 of 81% compared to 86% for the CKD-EPI_{creatinine-cystatin} equation. 52

We evaluated the ability of the present study to detect a difference of 5% in P30, between 81% and 86%, which is of a magnitude considered clinically important and likely to occur with the expected scale of differences in imprecision between the equations. With 1000 evaluable subjects our simulations showed 87% power for detecting a difference at the 5% level. We thus aimed to recruit 1300 people, which, allowing for 15–20% dropout, would deliver over 90% power. This calculation was conservative in that it only took imprecision into account. In the presence of systematic bias (a reasonable assumption) the power was estimated to be greater than this. In relation to estimating equation performance amongst the ethnic groups, the proposed sample size of these subgroups allowed P30 estimates to be reported with 95% CIs of 10%.

The annual rate of change over 3 years can be estimated by the observed difference between follow-up and baseline measurements or by using regression techniques for estimated and mGFR. Padala et al. proposed two cut-points for estimating accuracy of rates of change comparing estimated to mGFR: > 3 ml/minute/1.73 m²/year error and > 5% error. 84 Our simulations predicted that with the magnitude of measurement error that corresponds with P30 measures of 81% and 86% the study would have over 90% power to detect differences in the proportions with > 3 ml/minute/1.73 m²/year error and over 80% power to detect differences in proportions with > 5% error.

Ideally, we would like to have been able to detect differences in progression to stage 4 CKD between estimating equations. While it is possible to estimate a relationship with progression for each equation, a study would need to have 10,000 participants to be able to have 90% power to detect a difference in predictive abilities to stage 4 CKD between equations. Therefore, our study did not have power to evaluate this as an outcome, but it was considered in the model-based analysis.

Substudy of disease progression

A target sample size of 375 was chosen based on practical considerations to allow investigation of 12 covariates of interest (gender, age, diabetes, duration of diabetes, ethnicity, albuminuria, baseline GFR, BP, BMI, waist circumference, smoking status and presence of vascular disease) in addition to variables for time, drug effects and random effects. Another consideration was to include a reasonable number of subjects in each of the ethnic groups and high- and low-risk subgroups. General sample size formulae103 for multiple regression models indicate that a sample of 300 individuals, with 15–20% attrition, will provide 90% power to detect a change in R² of 0.11 (medium-to-small effect size103) attributed to 20 independent variables, 6 controlled independent variables and a significance level of 0.05.

Study of intraindividual biological variation

Sample sizes in biological variation studies are somewhat dictated by the practical limitations and costs of handling the large numbers of analyses generated while minimising the effects of pre-analytical and analytical variation. However, biological variation studies are known to be robust to the effects of sample size. For example, using 4 samples on 10 subjects, Gowans et al. estimated a CV_I for creatinine of 4.1%,76 whereas Keevil et al. using 10 samples from 12 subjects obtained a CV_I of 4.9%. 77

The sample size was based on the precision of CV_I, which was estimated to be 10%. With 20 participants recruited, tested on 4 occasions and assayed in duplicate and assuming data are log-normally distributed, an approximate 95% CI for CV_I has limits ± 2% (absolute).

Main study: prospective longitudinal cohort study

Data were analysed to address three main questions.

Substudy of disease progression

Analysis was undertaken using Stata/IC version 16.1.

The rate of decline (ml/minute/1.73 m²/year) in mGFR and the difference between mGFRs and eGFRs (bias), assessed every 12 months, were modelled over time using a longitudinal linear random coefficients regression model, to estimate average and variability in disease progression and bias. 84 The random coefficients model included random effects for intercept and slope, allowing a different intercept and slope for each individual within the model. Rates of change in mGFR and bias were estimated from the slopes of the regression model, and reported with 95% CIs. Participants were included in the analysis if they had measurements recorded on more than one occasion during the study. Measured GFR was modelled on the natural and log-transformed scales. Parameters of the model for mGFR and bias were estimated using maximum likelihood. Between- and within-patient variability in the rate of decline of mGFR was also estimated.

Diagnostic plots of residuals, fitted values and marginal predictors for intercept and slope were assessed for normal distribution and constant variance assumptions and goodness of fit. The final models for mGFR and eGFRs on the natural scale indicated that normal distribution assumptions and goodness of fit were acceptable; for ACR, the diagnostic plots of log-transformed data were acceptable, and there was evidence of non-normality on the natural scale.

Covariates explored in the disease progression model for mGFR were as described earlier (see Substudy of disease progression). The time-varying covariates in the model were BMI, waist circumference, systolic BP and diastolic BP, the mean of the second and third BP measurements at each time was used for systolic and diastolic BP.

The effect of covariates on the population average intercept and longitudinal time effect (progression) was assessed. The method of backward elimination was used to remove covariates that were not significant from the model. As this was an exploratory analysis and the sample size was relatively small, parameters were retained in the model if the p-value was < 0.20. This enabled detection of possible associations which may be more significant in a larger sample. Where there was evidence of an interaction with time, indicating a difference in progression between categories defined by the covariate (e.g. between males and females), a factor was included in the model to estimate separate intercepts for categories of the covariate as well as separate slopes. Similarly, for continuous covariates in the model, if there was evidence that progression changed with different values of the covariate, an estimate of the change in intercept for the covariate was also included in the model.

The effect of drug class on the rate of progression was explored for drug classes where at least 10 participants in the substudy were taking medication within that class. Drug classes included in the analysis were thiazide diuretics, loop diuretics, beta-blockers, calcium channel blockers (CCBs), ACE inhibitors, A2RBs, alpha-blockers, statins, xanthine oxidase inhibitors (allopurinol) and antiplatelet agents.

It was originally planned to use the population Fisher information matrix optimal design algorithms (R open source software) to calculate the D-optimal105 sampling times from the disease progression model based on reference GFR for people with diabetes and/or albuminuria, and for those with neither of these conditions. We intended to select optimal monitoring strategies from a set of designs with sampling every 6 months and compare monitoring strategies with a number of sampling points (between two and six). However, the mGFR disease progression model was linear and therefore these methods were not needed, as the optimal design solution is simplified for linear models, intuitively the slope is estimated optimally by the two design points with the greatest spread. 105 More frequent monitoring will increase accuracy of estimation.

Additional disease progression modelling was performed on eGFR (CKD-EPI_creatinine and CKD-EPI_cystatin) and ACR measured every 6 months. The covariates and drug classes explored in the mGFR analysis were also explored in the eGFRs and ACR models. Parameter estimates, 95% CIs and estimates of within- and between-patient variability were compared to those for mGFR.

While our longitudinal cohort did not have adequate power to detect differences in progression, our data on mGFR and eGFR over time (study 1), patterns and determinants of progression (study 2), and intraindividual biological variation (study 3) were combined in a measurement model to evaluate the impact of alternative monitoring strategies on detection of NICE’s combined progression criteria (see Clinical guidelines) and/or progression to stage 4 CKD. True GFR values were modelled over time for representative cohorts of people, and the comparative performance of alternative monitoring strategies in detecting progression (varying in choice of eGFR equation) was simulated utilising estimates of measurement error and accuracy. Outcome variables that were assessed included FP progression rates, and the sensitivity and delays in detecting progression.

Further assessment of covariates was performed by combining the data from the disease progression substudy with the main study and fitting the final covariate models for the substudy to the full data set. The main purpose of this was to check the inferences are consistent across the two data sets, as the full data set had more participants and hence more covariate information, while the substudy has more time points so progression is better defined.

Study of intraindividual biological variation

Data were log-transformed and normality tests were performed using the Shapiro–Wilk test. Outliers between duplicate measurements and of within-subject variance were excluded using Cochran’s test and outliers amongst mean values of subjects were excluded using Reed’s test. 35 Sensitivity analyses were also performed without exclusion of identified outliers. Log transformation was used to simplify calculation and because it improved the normality of the data as assessed by an increase in Shapiro–Wilk W statistic and visual examination of the distributions.

Terminology used was as proposed by Simundic et al. 106 Analytical (CV_A), individual (CV_I) and between-subject (CV_G) components of variation were calculated using standard approaches35 of linear random-effects modelling with restricted maximum likelihood estimation (allowing for the clustering of observations within time points and repeated observations per patient) (Stata version 15). Exact geometric CVs [(exp(S2)−1)×100]107,108 were calculated. CIs for SDs and CVs were estimated as described by Burdick and Graybill. 109 Differences in measures of CV, comparing the eGFR measures to mGFR, were investigated using multilevel models accounting for the clustering of test observations within individuals, using unstructured covariance matrices, in addition to the clustering of test results (multiple results per person, observation points and assessments). The critical difference (RCV) for significant changes in serial results with 95% probability was calculated using the approach for log-normal data giving a negative and positive limit. 110 The derived RCV for the reference GFR was used to test the ability of eGFR equations to detect a true change in GFR. The number of specimens (n) required to produce a precise estimate of the homeostatic set-point with 95% confidence within ± 10% was calculated as:

For each biomarker, the index of individuality (II) was calculated as:

To confirm kidney function was stable across the study period, the iohexol GFR measures were modelled to identify trend with time using a multilevel linear regression model (allowing for clustering of assessments within time points and observations within individuals).

Secondary objectives

Secondary analyses covered a range of more recently published equations and modified versions of the equations evaluated in the primary analysis.

Clinical accuracy and health economic analysis of monitoring with different estimated glomerular filtration rate equations

If cystatin C-based eGFR was found to be an improved approach to monitoring progression of CKD, then this would have cost implications for the healthcare sector. Cystatin C is more expensive than creatinine (£3.80 vs. £0.43); however, prompt identification of worsening kidney disease function may reduce additional treatment costs, offsetting the additional cost of testing. Furthermore, more accurate identification of individuals at high risk using cystatin C may reduce costs due to overdiagnosis and provide better management of low-risk individuals. 25,85

A systematic review was conducted to identify previous studies that have assessed the cost-effectiveness of test-based strategies for CKD using a decision-analytic model. This review was conducted and published at the beginning of the project and then updated in the last year of the project (full details and results can be found in Appendix 2). A key objective was to examine how existing models have captured the clinical impact of monitoring kidney function over time, given that we only have accuracy data available from this study.

To better understand definitions around kidney disease progression and triggers for clinical action (which have been updated since the initiation of this study), we then summarised the latest key clinical guidelines (see Clinical guidelines). We then used the data and results from this study to compare the accuracy of the different estimating equations for predicting CKD progression (using the definitions warranting clinical action based on the latest clinical guidelines) – we describe this as ‘clinical accuracy’ [see Comparative accuracy of estimated glomerular filtration rate equations for predicting accelerated progression (measurement model analysis)]. This analysis is based on a simulated measurement model, which considers a 10-year horizon of annual GFR monitoring, with GFR estimation error simulated onto mGFR values based on the bias profile information derived from the current study.

A cost–utility analysis was planned to estimate the health and economic consequences of implementing cystatin C-based eGFR or a combination of both cystatin C and creatinine-based eGFR for monitoring subjects who are initially stage 3 CKD compared to MDRD (creatinine-based) eGFR alone. In the absence of an accuracy improvement, a comparative cost analysis was reported to demonstrate the economic implications of implementing cystatin C based equations for monitoring those who are initially stage 3 CKD (see Comparative cost of monitoring with different estimated glomerular filtration rate equations).

Main study: prospective longitudinal cohort study

In total, 29,845 people were screened for potential suitability for study inclusion: 15,340 were deemed unsuitable at an early stage from informatics and clinic lists due to not having CKD stage 3. Other identified reasons for unsuitability included that they were too unwell, had recently had AKI, were unable to consent, were known alcohol or drug abusers, had had a previous reaction to iodine, were amputees, were under 18 years of age or were pregnant or breastfeeding. People considered suitable for study inclusion (n = 6209) were approached in person and/or sent participant information sheets. A further 4000 individuals were contacted through their primary care provider. Reasons for declining to participate were recorded in 928 cases. The major reasons for declining were that the participants were not interested in research; that they had too many medical appointments; that the 5-hour appointment time was too long; that too much travel was involved; that they were already in other research studies; and that the study involved too many injections.

A total of 1229 participants were recruited to the main study between April 2014 and January 2017, representing 95% of the target recruitment number. Recruitment was primarily from secondary/tertiary care, with 72 patients being recruited from primary care. Details of cumulative recruitment per month and recruitment by site and withdrawal by site may be found in Appendix 1 (see Appendix 1, Tables 37 and 38). Figure 2 shows the flow of participants through the study, using the format recommended by the Consolidated Standards of Reporting of Trials. 111

FIGURE 2.

Consolidated Standards of Reporting Trials flow diagram illustrating recruitment and follow-up in the prospective longitudinal cohort study.

Of the 1229 participants recruited to the study, 1205 and 1180 respectively had evaluable estimated and mGFRs at baseline, with 1167 (95.0%) having both eGFRs and mGFRs recorded. At 36 months 976 participants remained in the study, of whom 875 (71.2%) had evaluable estimated and mGFRs at both baseline and 36 months. The last patient visit occurred in January 2020. Following laboratory analyses and data queries, the study database was locked on 30 April 2021 (see Appendix 1, Table 36).

Overall, 253 (20.6%) participants withdrew from the study. Consent was withdrawn by 112 (9.1%), 79 (6.4%) were lost to follow-up (LTFU) and there were 62 (5.0%) deaths. Causes of death were cancer (16), cardiovascular disease (14), respiratory disease (11), kidney disease (4), cerebrovascular (1), liver disease (1), other (6) and unknown (9). In addition to death and loss to follow-up, reasons for withdrawing from the study included illness (60), commencement of kidney replacement therapy (12), inability to complete sampling (9), suspected reaction to iohexol (5), loss from area (3), carer commitment (2), other reasons (7) and no reason given (14). A further 25 participants were unable to complete their 36-month mGFR test, increasing the total effective dropout rate to 22.6%. Study dropout appears to escalate at the final time point of the study: this reflects closure of records of participants that the research teams had been trying to retain in the study up to this point (e.g. the participant was no longer contactable).

Substudy of disease progression

Fewer participants were recruited to the substudy than planned. In addition, due to difficulties in recruiting participants from some ethnic groups, more than the planned number of Caucasian participants and fewer South Asian and African-Caribbean participants were recruited. A total of 278 participants were recruited to the substudy of patterns of disease progression between April 2014 and December 2016, representing 93% of the target recruitment number. Of these, 128 (46%) had diabetes and/or albuminuria. There were 196 Caucasian (70%), 47 African-Caribbean (17%) and 35 South Asian participants (13%) in the substudy.

Figure 3 shows the flow of participants through the substudy. Of the 278 participants recruited to the study, 269 and 273 respectively had evaluable mGFRs and at least one eGFR recorded at baseline, with 265 (95%) having both estimated and mGFRs recorded. Attrition was 22.7% for the substudy, with 215 participants remaining in the study at the end of year 3. Of these, 214 and 198 respectively had evaluable estimated and mGFRs, with 197 (71%) having both mGFRs and at least one eGFR recorded. The last patient visit occurred in December 2019.

FIGURE 3.

Consolidated Standards of Reporting Trials flow diagram illustrating recruitment and follow-up in the substudy of disease progression.

Overall, 63 (22.7%) participants dropped out of the study. Consent was withdrawn by 31, 22 were LTFU and there were 10 deaths. Of those who were LTFU, 50% were African-Caribbean. Causes of death were cancer (2), cardiovascular disease (4), respiratory disease (2) and unknown (2). In addition to death and loss to follow-up, reasons for withdrawing from the study included illness (20), suspected reaction to iohexol (1), loss from the area (1), other reasons (1) and no reason given (8). A further 17 participants were unable to complete their 36-month mGFR test, increasing the total effective dropout rate to 28.8%. Seventeen participants crossed over from the substudy to the main study (i.e. withdrawing from additional iohexol-mGFR tests). One patient crossed over to the main study at one time point but subsequently returned to the substudy. Patients have been included in the substudy analysis if data were collected at more than one time point.

The number of participants recruited was sufficient to perform the planned analysis and investigate the 12 covariates of interest. There were only three participants with type I diabetes; therefore type of diabetes was not included in the analysis model. Smoking status changed very little during the study and the number of current smokers was small so only baseline smoking status was used in the analysis. Similarly, most participants with diabetes were diagnosed before the study began and baseline diabetes status was used in the analysis model and the duration of diabetes was calculated at baseline. Vascular disease was defined as described above (see Sampling and data collection). Due to the low numbers of South Asian and African-Caribbean participants, separate variance parameters were not included in the analysis models for these ethnic groups. It was assumed that variability is similar across ethnic groups. Differences between ethnic groups were explored, and estimates and CIs were calculated. Inferences drawn from the substudy analysis results should be interpreted more cautiously due to the limited number of South Asian and African-Caribbean participants.

Study of intraindividual biological variation

Participants (n = 20) were recruited to the study between August 2014 and July 2015. 97 Figure 4 shows the flow of participants through the study. All 20 patients attended all four iohexol-mGFR procedures except one patient who missed their final appointment. Results from five iohexol clearances (five separate patients) were excluded before analysis, as the dose given was not fully administered or it was given subcutaneously. Medications were held constant during the 4 weeks of the study, except that two patients received a 1-week course of amoxicillin (500 mg tds) due to chest infection. Sixteen of the participants subsequently entered the main study, with their first iohexol-mGFR being used as their baseline test.

FIGURE 4.

Consolidated Standards of Reporting Trials flow diagram illustrating recruitment and follow-up in the study of intraindividual biological variation.

Main study: prospective longitudinal cohort study

Characteristics of the study subjects are shown in Table 2. The median age of participants was 67.5 years; 714 (58%) were male and 1066 (87%) were Caucasian. Diabetes was a pre-existing diagnosis in 28.3% of participants. The median mGFR at baseline was 46.8 ml/minute/1.73 m² and 56.8% had albuminuria (ACR ≥ 3 mg/mmol). There were 71 people in the baseline cohort with mGFR below 30 ml/minute/1.73 m² and 211 people with mGFR ≥ 60 ml/minute/1.73 m², with a total range of values from 11.9 to 103.7 ml/minute/1.73 m².

Substudy of disease progression

Characteristics of the study participants are shown in Table 3. The characteristics of those who remained in the study at the end of year 3 were similar to those at baseline. The proportion of participants in the substudy who were current smokers was low. There were very few missing values for the participant characteristics and continuous covariates at baseline. Analysis was performed on observed data and missing values were not estimated.

The evaluable population for disease progression modelling included those participants with paired mGFR and eGFR at two or more time points. We included individuals with at least two times so the data for each individual could contribute to the random intercept and slope terms in the model, and by including those with both mGFR and eGFR at two time points comparisons of mGFR and eGFR models were based on the same samples.

Study of intraindividual biological variation

Characteristics of the study subjects are shown in Table 4. Application of Cochran and Reed’s tests led to the exclusion of between one and three duplicate measurements for mGFR or eGFR and to the exclusion of one outlying within-subject measurement for iohexol clearance. Overall, no patient was completely excluded and all calculations of biological variation for mGFRs and eGFRs were based on a minimum of 3 weeks’ data in all individuals.

Accuracy of Glomerular Filtration Rate-estimating equations including the Modification of Diet in Renal Disease equation and three Chronic Kidney Disease-Epidemiology Consortium equations using either creatinine or cystatin C or a combination of both in people with stage 3 chronic kidney disease: baseline analysis

During the course of the study, we became aware of significant published concerns regarding bias of the Abbott cystatin C assay. Re-analysis of some historical stored samples suggested to us that the Abbott assay had undergone a significant change in standardisation around the time when the present study commenced (data not shown). In a laboratory analytical recovery study, the Abbott cystatin C assay was found to over-recover added international reference preparation cystatin C by an average of 109.1%. Given the potential negative impact of this on cystatin C eGFR, this was further explored by comparing cystatin C values with those obtained with an alternative assay (Siemens). Cystatin C results obtained using the Siemens method were lower than those using the Abbott assay, the relationship between the two methods being described by the linear regression equation Siemens = −0.08 + 0.94(Abbott). CIs for the intercept and slope were −0.12, −0.03 and 0.92, 0.96, respectively (see Appendix 1, Figure 14). Deming regression produced a similar equation (see Appendix 1, Table 39).

Glomerular Filtration rate estimates of cystatin C containing CKD-EPI equations were recalculated using a recalibration of cystatin C values based on this regression equation. The rationale for this recalibration is considered further in the Discussion section of this report. In this section (Accuracy of glomerular filtration rate-estimating equations including the Modification of Diet in Renal Disease equation and three Chronic Kidney Disease-Epidemiology Consortium equations using either creatinine or cystatin C or a combination of both in people with stage 3 chronic kidney disease: baseline analysis), results for both the initial (Abbott) cystatin C equations and the recalibrated (Siemens) cystatin C containing equations are shown to illustrate the impact of calibration. In subsequent sections of the results chapter, only the recalibrated cystatin C-containing equations are shown.

All estimates of GFR relating to the primary study objectives were negatively biased overall compared to mGFR (see Table 5 and Figure 5). As would be predicted, recalibration of the cystatin C-containing equations with Siemens results improved the negative bias. There was no difference in bias overall against mGFR between the MDRD, CKD-EPI_creatinine, CKD-EPI_cystatin and CKD-EPI_{creatinine-cystatin} equations. However, both the MDRD and CKD-EPI_creatinine equations demonstrated positive bias at lower levels of GFR and increasing negative bias at higher levels of GFR: this effect was largely attenuated when the CKD-EPI_cystatin and CKD-EPI_{creatinine-cystatin} equations were used (Figure 6). CKD-EPI equations incorporating recalibrated cystatin C values showed an improvement in performance (P30, 95% CI) compared to the same equations using Abbott cystatin C results (Table 5); for example, CKD-EPI_cystatin 72.5 (69.8 to 75.0) versus 89.5 (87.6 to 91.2) when adjusted values were used. Accuracy (P30) of the recalibrated CKD-EPI_cystatin equation did not differ from that of the MDRD and CKD-EPI_creatinine equations (Table 6), while accuracy of the recalibrated CKD-EPI_{creatinine-cystatin} equation was superior to that of all of these equations (94.9% CI 93.5 to 96.1) (see Table 6 and Figure 5). In general, P15 values followed a similar rank order between equations to P30 values but were significantly lower.

FIGURE 5.

Bias of GFR-estimating equations compared to mGFR shown as box and whisker plots. The box shows the median and the first (Q1) and third quartiles (Q3). The whiskers span all data points within 1.5 IQR of the nearer quartile, with Tukey outliers outside of this range (< Q1–1.5 IQR or > Q3 + 1.5 IQR). The CKD-EPI_cystatin and CKD-EPI_{creatinine-cystatin} equations are shown both before and after recalibration against the Siemens assay.

FIGURE 6.

Bias of GFR-estimating equations compared to mGFR shown as lowess plots. Bias plots for the baseline cohort (n = 1167) are shown with lowess (locally weighted scatterplot smoothing) function (red solid line). The solid black line shows zero bias, and the dashed red line shows mean bias. The CKD-EPI_cystatin and CKD-EPI_{creatinine-cystatin} equations are shown both after (c, d) and before (e, f) recalibration against the Siemens assay. (a) MDRD eGFR vs. mGFR; (b) CKD-EPI_creatinine eGFR vs. mGFR; (c) CKD-EPI_cystatin eGFR vs. mGFR (Siemens); (d) CKD-EPI_{creatinine-cystatin} eGFR vs. mGFR (Siemens); (e) CKD-EPI_cystatin eGFR vs. mGFR (Abbott); and (f) CKD-EPI_{creatinine-cystatin} eGFR vs. mGFR (Abbott).

Accuracy of glomerular filtration rate-estimating equations according to age, gender, diabetes, albuminuria, body mass index, measured glomerular filtration rate level and ethnic group (particularly Caucasian, South Asian and African-Caribbean)

Performance (P30) of GFR-estimating equations was examined by categories of age, gender, diabetes, albuminuria, BMI, level of GFR (Table 7) and ethnic group (particularly Caucasians, South Asian and African-Caribbean) (Table 8). For each characteristic, data for equations incorporating cystatin C utilised cystatin C values obtained following assay recalibration (see Accuracy of glomerular filtration rate-estimating equations including the Modification of Diet in Renal Disease equation and three Chronic Kidney Disease-Epidemiology Consortium equations using either creatinine or cystatin C or a combination of both in people with stage 3 chronic kidney disease: baseline analysis). Although in some cases point estimates suggest differences with certain characteristics (e.g. inferior performance in people with higher BMI), for none of the equations were any of the changes significant (i.e. confidence intervals of the P30 estimates were overlapping across categories).

In the overall cohort, removal of the race adjustment factors had no impact (overlapping CIs) on the median bias or accuracy (P30) of the MDRD, CKD-EPI_creatinine and CKD-EPI_{creatinine-cystatin} equations (see Table 5). There was no evidence (overlapping CIs) of a difference in accuracy of any of the equations across ethnic groups (Table 8). Removal of the adjustment factor for African-Caribbean ethnicity in the MDRD, CKD-EPI_creatinine and CKD-EPI_{creatinine-cystatin} equations resulted in decreased point estimates of P30 for these participants, although this only achieved significance (p < 0.05) for the MDRD equation (see Table 8). The 2021 revisions of the CKD-EPI equations, which were specifically remodelled to address concerns around interethnic performance, were also included here for comparison.

Performance of glomerular filtration rate-estimating equations at detecting changes in measured glomerular filtration rate over 3 years

The median mGFR (ml/minute/1.73 m²) at baseline was 48.1 falling to 43.6 at 3-year follow-up. The equivalent changes for the eGFRs were: 44.6–41.5 (MDRD); 45.7–42.0 (CKD-EPI_creatinine); 43.6–40.4 (CKD-EPI_cystatin); and 43.4–40.6 (CKD-EPI_{creatinine-cystatin}). Of the 875 participants with mGFR and eGFR at both baseline and 3-year follow-up, 7 (0.8%) reached kidney failure (mGFR < 15) at follow-up. Of the 875 participants over the study duration, 268 (30.6%) had a change in GFR > 10; 235 (26.8%) had a decline in GFR > 10; 272 (31.1%) had a GFR decline of > RCV; 156 (17.8%) had a GFR decline of > 25%; and 139 (15.9%) had a GFR decline of > 25% in combination with a change in disease category.

As opposed to assessments of equation accuracy at baseline (see Table 5), in the analyses shown in this section which relate to monitoring GFR over time, the ‘race adjustment factor removed’ versions of equations are not shown. This adjustment involves scaling an element of the equation in African-Caribbean individuals only by a constant factor on each occasion, so monitoring an individual over time will have no impact on the ability of an equation to reflect change over time. This was confirmed empirically, and the data have not been included to aid clarity. In all tables in this section, data for equations incorporating cystatin C utilise cystatin C values obtained following assay recalibration (see Accuracy of glomerular filtration rate-estimating equations including the Modification of Diet in Renal Disease equation and three Chronic Kidney Disease-Epidemiology Consortium equations using either creatinine or cystatin C or a combination of both in people with stage 3 chronic kidney disease: baseline analysis).

To estimate and model disease progression (decline in glomerular filtration rate or increase in albumin-to-creatinine ratio) and differences in progression between ethnic groups (Caucasian, South Asian and African-Caribbean), and baseline diabetes and albuminuria status and other potential risk factors

In the substudy of disease progression, 239 participants with paired mGFR and eGFR on at least 2 occasions were evaluable and included in the disease progression covariate models. There were no obvious changes in the covariates BP, BMI and waist circumference (Table 16) or mGFR or eGFR over time (Table 17), although there was a tendency towards lower GFR and higher ACR values over time (see Appendix 1, Figures 15–17).

The difference between estimated and mGFR (bias) at each time point is summarised in Table 18. The baseline differences (bias) between estimated and mGFR in the substudy were similar to those observed in the main study, and did not change during the course of the study, although the bias at baseline was larger than observed at later time points for all eGFRs except CKD-EPI_creatinine (see also Appendix 1, Figure 17).

The estimated slopes (progression) and variance estimates for mGFRs and eGFRs and bias (eGFR minus mGFR) from the random coefficients regression model without covariates are presented in Table 19.

The rates of decline were steepest for mGFR and CKD-EPI_creatinine eGFR, and slope estimates were similar. The rates of decline were slowest for cystatin C eGFR. The model-predicted slopes for mGFR and CKD-EPI_creatinine and CKD-EPI_cystatin eGFR are shown in Appendix 1, Figure 18.

The variability of slopes was larger for the eGFRs compared to mGFR and the residual variability was lower. This is likely due to the difference in the number of observations for each individual for mGFR and eGFR. In the regression model for eGFRs, more data points are included for each individual (includes 6-monthly data), resulting in less shrinkage towards the population mean slope, more variability between slopes and lower residual variability (variability between observations and individual slope fitted values).

There was little evidence of decline in bias (difference) over time based on the difference between CKD-EPI_creatinine GFR and mGFR (slope estimate 0.101, 95% CI −0.272 to 0.474). Similarly for the difference between MDRD and mGFR, the slope estimate was 0.325 and 95% CI −0.038 to 0.688. In contrast, differences of mGFR from CKD-EPI_cystatin and CKD-EPI_{creatinine-cystatin} showed an incline in slope over time indicated by the positive slope estimates with 95% CI not overlapping zero (e.g. 0.840, CI 0.441 to 1.24 for difference between CKD-EPI_cystatin GFR and mGFR), suggesting an increase in bias (i.e. eGFR minus mGFR) over time. The median bias was negative at all times for CKD-EPI_cystatin and CKD-EPI_{creatinine-cystatin} (i.e. these estimates of GFR were always lower than the mGFR), but the median bias at later time points was closer to zero compared to the median bias at baseline (Table 18).

Prior to conducting the random coefficients regression modelling with covariates, correlations between the covariates to be included in the model were explored to identify collinearity (Table 20). Correlations were calculated using the data from baseline and from all time points for evaluable participants (i.e. those with estimated and mGFR available on at least two time points). The final covariate models for mGFR, eGFR and ACR should be interpreted bearing in mind the observed associations between covariates. Where collinearity exists, some covariates associated with the GFR measure were excluded from the final multiple regression model. A correlation of 0.3 or higher and < 0.5 was considered moderate, and correlations of 0.5 or above high. Table 20 shows the correlation coefficients for each covariate pair. The highest correlation was seen between waist circumference and BMI (0.856). Waist circumference was also moderately correlated with gender (0.337) and diabetes status (0.384). BMI and diabetes status were moderately correlated (0.383), as were systolic and diastolic BP (0.400). All other correlations were < 0.3, although there were a number of correlations between 0.2 and 0.3 which could have impacted covariate selection in the model.

Regression modelling to explore the effect of each covariate individually on the intercept and progression slope was conducted prior to modelling the combined covariates in the full models. For mGFR, the individual unadjusted covariate regression models had coefficients with a p-value of < 0.2 for the intercept term for all covariates with the exception of BMI and smoking status. Similarly, the individual regressions had coefficients with p < 0.2 for the rate of change (progression) for covariates age, diabetes status, ethnicity group, albuminuria, BMI, waist circumference and smoking status (data not shown).

Measured glomerular filtration rate random coefficients final covariate model

Estimates from the random coefficients final covariate model for mGFR for both the substudy and the combined substudy and main study data combined are shown in Table 21. All covariates were controlled for other terms in the model, that is the estimated coefficients show the additional effect of the covariate after all other variables in the model have been accounted for.

TABLE 21 - Measured GFR covariate model

See note in text regarding effect of one individual on impact of African-Caribbean ethnicity slope factor.

Note

For the substudy, terms were retained in the model where p < 0.2 and intercept terms retained where the interaction with slope p < 0.2. Associations observed to be significant in the substudy were then assessed in the combined data set. Units for progression are ml/minute/1.73 m²/year. Values show regression coefficient (95% CI), p-value.

Model term/covariate	Measured GFR (ml/minute/1.73 m2)
	Substudy data only	Main study and substudy data combined
Intercept constant (Caucasian, no albuminuria, no CCB antagonists)	0.202 (−4.518 to 4.922), 0.933	0.042 (−1.857 to 1.943), 0.965
Age (years)	−0.036 (−0.081 to 0.008), 0.111	−0.019 (−0.038 to −0.001), 0.042
Baseline mGFR (ml/minute/1.73 m2)	0.957 (0.916 to 1.000), < 0.001	0.984 (0.968 to 1.000), < 0.001
Ethnicity group
African-Caribbean	0.146 (−1.401 to 1.693), 0.853	0.096 (−0.788 to 0.979), 0.832
South Asian	−0.660 (−2.308 to 0.998), 0.433	−0.270 (−1.12 to 0.585), 0.536
Albuminuria
3–30 mg/mmol	0.527 (−0.708 to 1.761), 0.403	0.123 (−0.353 to 0.600), 0.612
> 30 mg/mmol	−0.069 (−1.533 to 1.395), 0.927	−0.021 (−0.584 to 0.541), 0.941
Systolic BP (mmHg)	0.031 (0.006 to 0.056), 0.016	0.015 (0.004 to 0.025), 0.008
Slope constant (progression rate of change) (ml/minute/1.73 m2/year)	1.051 (−0.800 to 2.902), 0.266	2.121 (1.226 to 3.016), < 0.001
Ethnicity group * slope
African-Caribbean	1.014 (−0.198 to 2.226), 0.101a	1.113 (0.171 to 2.055), 0.021
South Asian	−0.655 (−1.917 to 0.606), 0.309	−0.218 (−1.095 to 0.659) 0.625
Albuminuria * slope
3–30 mg/mmol	−0.935 (−1.758 to −0.112), 0.026	−0.549 (−0.959 to −0.140), 0.009
> 30 mg/mmol	−1.619 (−2.631 to −0.607), 0.002	−1.521 (−2.007 to −1.036), < 0.001
Baseline mGFR * slope	−0.037 (−0.069 to 0.005), 0.025	−0.064 (−0.080 to −0.048), < 0.001

The coefficients for each of the categorical variates in the regression model correspond to the difference from the reference category, for example the albuminuria coefficients for the two higher categories correspond to the difference from those with no albuminuria (< 3 mg/mmol). To estimate the intercept for a particular subgroup, the regression coefficients of the constant term and each of the covariate main effect terms (not interactions) are used. As an example consider the estimated intercept from the mGFR substudy final model for a baseline mGFR of 60 ml/minute/1.73 m², ACR of 10 mg/mmol, age 70 years, South Asian ethnicity, and systolic BP 120 mmHg: the estimate of intercept is 0.202 − (0.036 × 70) + (0.957 × 60) − 0.660 + 0.527 + (0.031 × 120), which is equal to 58.7 ml/minute/1.73 m². Similarly, progression can be estimated using the slope constant term and the coefficients of the interactions with the slope (the terms appearing below the slope constant term in the table). For the example given above, the estimate of progression would be 1.051 – 0.655 – 0.935 – (0.037 × 60), which is equal to a rate of change of −2.76 ml/minute/1.73 m²/year.

In the substudy data, the intercept (including covariates) was the expected mean value of mGFR when the time was equal to zero (baseline). In the final model, this was derived using baseline mGFR, age (increasing age lowers the intercept) and systolic BP (higher systolic BP increases the intercept) (p < 0.2). Of these, baseline-mGFR was the main determinant. Intercept terms were also included in the model for covariates which were associated with slope. In the final model, the rate of change in mGFR over time was associated with baseline GFR, ethnicity group and albuminuria (p < 0.2). Baseline GFR increased the intercept by 0.957 ml/minute/1.73 m² (95% CI 0.916 to 1.000) for each unit of baseline GFR, and decreased the progression slope (steeper decline) by −0.037 ml/minute/1.73 m²/year (95% CI −0.069 to 0.005) for each unit. African-Caribbean ethnicity increased the progression slope (slower decline) by 1.01 ml/minute/1.73 m²/year (95% CI −0.198 to 2.226). However, inspection of the individual data revealed one participant (ID 19877) who had much higher values of mGFR at baseline and at 3 years in the African-Caribbean ethnicity group. Sensitivity analysis excluding this participant (ID 19877) showed no evidence of effects of ethnicity on progression or intercept (p > 0.2).

The rate of decline was steeper for those with higher levels of albuminuria: the slope decreased by −0.935 ml/minute/1.73 m²/year (95% CI −1.758 to −0.112) for those with ACR between 3 and 30 mg/mmol (inclusive) and −1.619 ml/minute/1.73 m²/year (95% CI −2.631 to −0.607) for those with ACR > 30 mg/mmol.

Age and systolic BP were associated with the intercept only. There was a decrease of −0.036 ml/minute/1.73 m² (95% CI −0.081 to −0.008) for each year of age, and an increase of 0.031 ml/minute/1.73 m² (95% CI 0.006 to 0.056) for each unit of systolic BP (mmHg).

Plots of the fitted and observed data by participants for the final covariate model for mGFR are shown for completeness in Report Supplementary Material 1.

The parameter estimates from the model using the full data set were similar to those from the substudy for all covariates (Table 21). The association between African-Caribbean ethnicity and progression (i.e. with African-Caribbean ethnicity being associated with reduced progression) was more significant when the model was fitted to the full data set (p = 0.021). Sensitivity analysis excluding the data for one participant (ID 19877) discussed earlier made very little difference to the statistical significance of this association (p = 0.032) (see Appendix 1, Figure 19). It should be noted that the number of evaluable participants in the full data set was still quite low for African-Caribbean (n = 46) and South Asian (n = 51) ethnicity groups.

Figure 7 illustrates the changes in mGFR progression over time for albuminuria status. The median mGFR of those with albuminuria declines more steeply compared to those who do not have albuminuria. The equivalent figure for the combined main and substudy data may be found in Appendix 1, Figure 20.

FIGURE 7.

Median and IQR mGFR over time by albuminuria status (substudy data).

Creatinine estimated glomerular filtration rate random coefficients final covariate model

The individual covariate regression models for CKD-EPI_creatinine had coefficients with a p-value of < 0.2 for the intercept term for all covariates with the exception of ethnicity group and systolic BP. Similarly, the individual regressions had coefficients with p < 0.2 for the rate of change (progression) for covariates age, diabetes status, ethnicity group, albuminuria, diastolic BP, waist circumference and smoking status (data not shown).

Table 22 shows the estimates from the random coefficients regression final covariate model for CKD-EPI_creatinine for all main study and substudy data compared to the model estimates including the substudy data only.

TABLE 22 - Estimated GFR CKD-EPIcreatinine covariate model

Note

For the substudy, terms were retained in the model where p < 0.2 and intercept terms retained where the interaction with slope p < 0.2. Associations observed to be significant in the substudy were then assessed in the combined data set. Units for progression are ml/minute/1.73 m²/year. Values show regression coefficient (95% CI), p-value.

Model term/covariate	CKD-EPIcreatinine (ml/minute/1.73 m2)
	Substudy data only	Main study and substudy data combined
Intercept constant (Caucasian, no diabetes, no albuminuria, no beta-blockers, no A2RB)	7.971 (3.462 to 12.480), 0.001	5.125 (2.819 to 7.431), < 0.001
Age (years)	−0.038 (−0.080 to −0.003), 0.069	−0.017 (−0.039 to 0.004), 0.109
Baseline CKD-EPIcreatinine (ml/minute/1.73 m2)	0.945 (0.905 to 0.985), < 0.001	0.909 (0.888 to 0.930), < 0.001
Diabetes (yes)	−0.094 (−1.294 to 1.106), 0.878	−0.515 (−1.099 to 0.067), 0.083
Ethnicity group
African-Caribbean	0.786 (−0.604 to 2.177), 0.268	0.952 (−0.144 to 2.048), 0.089
South Asian	−0.621 (−2.150 to 0.908), 0.426	0.174 (−0.876 to 1.224), 0.745
Albuminuria
3–30 mg/mmol	−0.180 (−1.207 to 0.847), 0.731	0.056 (−0.449 to 0.562), 0.828
> 30 mg/mmol	0.138 (−1.033 to 1.308), 0.818	0.651 (0.073 to 1.228), 0.027
BMI (kg/m2)	−0.085 (−0.173 to 0.003), 0.058	−0.0001 (−0.040 to 0.040), 0.995
Beta-blocker (yes)	−0.434 (−1.460 to 0.592), 0.407	−0.222 (−0.741 to 0.297), 0.402
A2RB (yes)	−0.695 (−1.665 to 0.275), 0.160	−0.669 (−1.160 to −0.177), 0.008
Slope constant (progression rate of change) (ml/minute/1.73 m2/year)	−1.082 (−1.749 to −0.414), 0.001	−1.197 (−1.498 to −0.895), < 0.001
Diabetes * slope	1.058 (0.153 to 1.964), 0.022	0.037 (−0.403 to 0.477), 0.870
Ethnicity group * slope
African-Caribbean	0.161 (−0.992 to 1.315), 0.784	−0.046 (−0.935 to 0.842), 0.919
South Asian	−1.011 (−2.249 to 0.227), 0.109	−0.581 (−1.419 to 0.256), 0.173
Albuminuria * slope
3–30 mg/mmol	−0.251 (−0.942 to 0.440), 0.477	0.048 (−0.276 to 0.371), 0.772
> 30 mg/mmol	−0.861 (−1.685 to −0.037), 0.040	−0.479 (−0.858 to −0.099), 0.013
Beta-blocker * slope	−0.598 (−1.348 to 0.152), 0.118	0.043 (−0.300 to 0.385), 0.808
A2RB * slope	−0.609 (−1.335 to 0.117), 0.100	−0.235 (−0.560 to 0.090) 0.156

In the substudy, intercept was associated with baseline CKD-EPI_creatinine, age, BMI and A2RB medication (p < 0.2). Intercept terms were also included in the model for covariates which were associated with slope. In the final model, the rate of change in CKD-EPI_creatinine over time is associated with ethnicity group, albuminuria and diabetes status, and beta-blocker and A2RB medication (p < 0.2).

Albuminuria decreased the progression slope (steeper decline) by −0.861 ml/minute/1.73 m²/year (95% CI −1.685 to −0.037) for those with ACR > 30 mg/mmol. South Asian ethnicity decreased the progression slope (steeper decline) by −1.011 ml/minute/1.73 m²/year (95% CI −2.249 to 0.227). Diabetes increased the slope (slower decline) by 1.058 ml/minute/1.73 m²/year (95% CI 0.153 to 1.964): this seems counterintuitive and may be partially due to the lower intercept for those with diabetes. The coefficient is the combined effect of diabetes status with all other covariates in the model and diabetes status is correlated with BMI which is also included in the model and could affect the coefficient for this covariate. The rate of decline was steeper for those prescribed beta-blocker medication, the slope decreasing by −0.598 ml/minute/1.73 m²/year (95% CI −1.348 to 0.152), and for those prescribed A2RBs, who had a decrease in slope of −0.609 ml/minute/1.73 m²/year (95% CI −1.335 to 0.117). The intercept also decreased by −0.695 ml/minute/1.73 m² (95% CI −1.665 to 0.275) for those taking A2RBs.

Age, baseline CKD-EPI_creatinine and BMI were associated with intercept only. There was a decrease of −0.038 ml/minute/1.73 m² (95% CI −0.080 to −0.003) for each year of age, an increase of 0.945 ml/minute/1.73 m² (95% CI 0.905 to 0.985) for each unit of baseline CKD-EPI_creatinine and a decrease of −0.085 ml/minute/1.73 m² (−0.173 to 0.003) for each BMI unit (kg/m²).

Using the full data set, diabetes status was associated with the intercept estimates of CKD-EPI_creatinine (p = 0.083), but not associated with progression, and the association of BMI with the intercept was no longer significant, but BMI and diabetes status were highly correlated so it is likely these effects are presenting in a different way in the model. Using the full data set, there was no longer an association between the use of beta-blocker medication and progression, but there was a statistically significant association between A2RB use and the intercept levels of CKD-EPI_creatinine. Other covariate estimates for the full data set were similar to those seen for the substudy.

Figures 8–10 illustrate CKD-EPI_creatinine progression over time in the substudy for diabetes, albuminuria and ethnicity groups, respectively. In Figure 8 the decline in the medians of those who have and those who do not have diabetes appears to be similar, the differences seen in the model could be due to the differences in the intercepts of those with and without diabetes, and the correlation between BMI and diabetes status. In Figure 9 the steeper decline over time is more obvious for those with albuminuria, compared to those who do not have albuminuria. The equivalent figure for the combined main and substudy data may be found in Appendix 1, Figure 21. Figure 10 suggests a steeper decline for people of South Asian ethnicity. The equivalent figure for the combined main and substudy data may be found in Appendix 1 (Figure 22).

FIGURE 8.

Median and IQR CKD-EPI_creatinine over time by diabetes status (substudy data).

FIGURE 9.

Median and IQR CKD-EPI_creatinine over time by albuminuria status (substudy data).

FIGURE 10.

Median and IQR CKD-EPI_creatinine over time by ethnicity group (substudy data).

Cystatin C estimated glomerular filtration rate random coefficients final covariate model

The individual covariate regression models for CKD-EPI_cystatin had coefficients with a p-value of < 0.2 for the intercept term for all covariates with the exception of albuminuria and diastolic BP. The individual regressions had coefficients with p < 0.2 for the rate of change (progression) for covariates gender, baseline CKD-EPI_cystatin, albuminuria, systolic and diastolic BP, BMI, waist circumference and smoking status (data not shown).

Table 23 shows the estimates from the random coefficients regression final covariate model for CKD-EPI_cystatin. For the substudy, intercept was associated with baseline CKD-EPI_cystatin, gender, ethnicity group, BMI, waist circumference and smoking status (p < 0.2). Intercept terms were also included in the model for covariates which were associated with slope. In the final model, the rate of change in CKD-EPI_cystatin over time was associated with baseline CKD-EPI_cystatin, diastolic BP, BMI and beta-blocker and A2RB medication (p < 0.2).

TABLE 23 - Estimated GFR CKD-EPIcystatin covariate model

Note

For the substudy, terms were retained in the model where p < 0.2 and intercept terms retained where the interaction with slope p < 0.2. Associations observed to be significant in the substudy were then assessed in the combined data set. Units for progression are ml/minute/1.73 m²/year. Values show regression coefficient (95% CI), p-value.

Model term/covariate	CKD-EPIcystatin (ml/minute/1.73 m2)
	Substudy data only	Main study and substudy data combined
Intercept constant (Caucasian, females, non-smoker, no beta-blockers, no A2RB)	1.862 (−3.446 to 7.190), 0.493	0.914 (−1.784 to 3.612), 0.507
Baseline CKD-EPIcystatin (ml/minute/1.73 m2)	0.991 (0.957 to 1.025), < 0.001	0.963 (0.946 to 0.980), < 0.001
Males	−0.768 (−1.895 to 0.358), 0.181	0.574 (0.016 to 1.132), 0.044
Ethnicity group
African-Caribbean	1.082 (−0.300 to 2.463), 0.125	0.975 (−0.151 to 2.101), 0.090
South Asian	−0.110 (−1.603 to 1.382), 0.885	−0.372 (−1.449 to 0.703), 0.497
Diastolic BP (mmHg)	0.008 (−0.034 to 0.050), 0.708	0.009 (−0.013 to 0.031), 0.429
BMI (kg/m2)	−0.220 (−0.366 to −0.073), 0.003	−0.004 (−0.078 to 0.071), 0.927
Waist circumference (cm)	0.046 (−0.011 to 0.103), 0.115	−0.006 (−0.036 to 0.023), 0.677
Smoking status
Current smoker	−1.547 (−3.203 to 0.108), 0.067	−0.078 (−1.024 to 0.868), 0.872
Ex-smoker	−0.012 (−1.077 to 1.053), 0.983	−0.535 (−1.060 to −0.009), 0.046
Beta-blockers (yes)	0.065 (−0.982 to 1.112), 0.903	−0.274 (−0.825 to 0.272), 0.323
A2RB (yes)	0.486 (−0.511 to 1.483), 0.339	−0.095 (−0.609 to 0.418), 0.717
Slope constant (progression rate of change) (ml/minute/1.73 m2/year)	−2.047 (−5.488 to 1.393), 0.244	−0.142 (−1.739 to 1.455), 0.861
Baseline CKD-EPIcystatin * slope	0.031 (0.003 to 0.060), 0.031	0.011 (−0.003 to 0.025), 0.118
Diastolic BP * slope	−0.024 (−0.052 to 0.004), 0.094	−0.016 (−0.029 to −0.002), 0.022
BMI * slope	0.077 (0.005 to 0.148), 0.035	0.006 (−0.027 to 0.038), 0.727
Beta-blockers * slope	−0.997 (−1.826 to −0.168), 0.018	0.018 (−0.354 to 0.390), 0.924
A2RB * slope	−0.855 (−1.633 to −0.077), 0.031	−0.389 (−0.739 to −0.038), 0.030

Baseline CKD-EPI_cystatin increased the intercept by 0.991 ml/minute/1.73 m² (95% CI 0.957 to 1.025) for each unit of baseline CKD-EPI_cystatin, and increased the progression slope (slower decline) by 0.031 ml/minute/1.73 m²/year (95% CI 0.003 to 0.060) for each unit. The rate of decline was also steeper for those prescribed beta-blocker medication, the slope decreasing by −0.997 ml/minute/1.73 m²/year (95% CI −1.826 to −0.168), and for those prescribed A2RBs, who had a decrease in the slope of −0.855 ml/minute/1.73 m²/year (95% CI −1.633 to −0.077). Diastolic BP decreased the progression slope (steeper decline) by −0.024 ml/minute/1.73 m²/year (95% CI −0.052 to 0.004) for each mmHg unit of BP. BMI decreased the intercept by −0.220 ml/minute/1.73 m² (95% CI −0.366 to −0.073) for each unit (kg/m²), and increased the progression slope (slower decline) by 0.077 ml/minute/1.73 m²/year (95% CI 0.005 to 0.148).

Gender, ethnicity group, waist circumference and smoking status were associated with the intercept only. There was an increase of 0.046 ml/minute/1.73 m² (95% CI −0.011 to 0.103) for each unit change in waist circumference (cm), a decrease of −0.768 ml/minute/1.73 m² (95% CI −1.895 to 0.358) for males, an increase of 1.082 ml/minute/1.73 m² (95% CI −0.300 to 2.463) for African-Caribbean ethnicity and a decrease in intercept of −1.547 ml/minute/1.73 m² (95% CI −3.203 to 0.108) for those who currently smoke.

Using the full data set, BMI and waist circumference were no longer associated with the intercept estimates of CKD-EPI_cystatin, and BMI was no longer associated with progression. The effect of smoking status changed: using the substudy data only, there was an association between current smokers and the intercept, whereas using the full data set there was an association between ex-smokers and the intercept. This change may be due to the increased number of ex-smokers in the full data set (n = 367), compared to the number of current smokers (n = 70). Using the full data set, there was no longer an association between the use of beta-blocker medication and progression. Other estimates for the full data set were similar to those seen for the substudy data.

Figure 11 illustrates the changes in CKD-EPI_cystatin progression over time by ethnicity group. Those of African-Caribbean ethnicity have a different profile over time compared to the other ethnicity groups, although there was little evidence to suggest a difference in progression over time, but the intercept was higher for African-Caribbeans in the final model. The equivalent figure for the combined main and substudy data may be found in Appendix 1, Figure 23.

FIGURE 11.

Median and IQR CKD-EPI_cystatin over time by ethnicity group (substudy data).

Urinary albumin-to-creatinine ratio random coefficients final covariate model

The individual covariate regression models for log-transformed urinary ACR had intercept coefficients with a p-value of < 0.2 for all covariates with the exception of diabetes, BMI, waist circumference and vascular disease. The individual regressions had coefficients with p < 0.2 for the rate of change (progression) for covariates diabetes status, ethnicity group, albuminuria, baseline log ACR, systolic and diastolic BP and waist circumference (data not shown).

Table 24 shows the estimates from the random coefficients regression final covariate model for log ACR. For the substudy, intercept was associated with baseline log ACR, albuminuria (ACR category), age, systolic BP, diabetes, smoking status and loop diuretic medication (p < 0.2). Intercept terms were also included in the model for covariates associated with slope. In the final model, the rate of change in log ACR over time was associated with baseline log ACR, diabetes status, ethnicity group, albuminuria (ACR category), diastolic BP, smoking status and loop diuretic and CCB medication (p < 0.2).

TABLE 24 - Urinary ACR (log-transformed) covariate model

For baseline ACR, coefficients show per cent change in ACR based on 1% change in baseline ACR.

Note

For the substudy, terms were retained in the model where p < 0.2 and intercept terms retained where the interaction with slope p < 0.2. Associations observed to be significant in the substudy were then assessed in the combined data set. Units for progression are percentage change in mg/mmol/year. Values show regression coefficient (95% CI) expressed as percentage change in ACR per one unit change in the covariate⁺, p-value.

Model term/covariate	Log-transformed ACR (mg/mmol)
	Substudy data only	Main study and substudy data combined
Intercept (Caucasian, no diabetes, no albuminuria, non-smoker, no loop diuretics, no CCB)	−4.88% (−31.2% to 60.8%), 0.816	1.21% (−19.4% to 27.3%), 0.916
Age (years)	−0.499% (−0.896% to 0.100%), 0.011	−0.200% (−0.399% to 0.030%), 0.101
Baseline ACR (log mg/mmol)+	0.605% (0.557% to 0.653%), < 0.001	0.533% (0.508% to 0.558%), < 0.001
Diabetes (yes)	10.1% (0.300% to 21.2%), 0.049	5.44% (0.300% to 10.7%), 0.038
Ethnicity group
African-Caribbean	2.02% (−13.8% to 11.4%), 0.756	3.15% (−7.69% to 15.1%), 0.587
South Asian	−7.87% (−19.7% to 5.76%), 0.243	−1.09% (−10.8% to 9.75%), 0.839
Albuminuria
3–30 mg/mmol	129% (101% to 160%), < 0.001	178% (160% to 198%), < 0.001
> 30 mg/mmol	472% (360% to 611%), < 0.001	713% (625% to 812%), < 0.001
Systolic BP (mmHg)	0.300% (0.100% to 0.602%), 0.002	0.200% (0.100% to 0.300%), 0.001
Diastolic BP (mmHg)	−0.200% (−0.598% to 0.300%), 0.470	−0.200% (−0.399% to 0.100%), 0.133
Smoking status
Current smoker	4.39% (−10.6% to 21.9%), 0.587	−3.82% (−12.1% to 5.34%), 0.404
Ex-smoker	−4.11% (−12.9% to 5.44%), 0.387	−4.21% (−8.88% to 0.702%), 0.092
Loop diuretic (yes)	12.5% (1.01% to 25.4%), 0.032	3.46% (−3.34% to 10.7%), 0.330
CCB (yes)	0.401% (−7.87% to 9.31%), 0.936	−1.49% (−6.29% to 3.56%), 0.557
Slope (progression rate of change) (% mg/mmol/year)	−12.5% (−27.9% to 6.18%), 0.175	−5.35% (−14.2% to 4.39%), 0.274
Baseline ACR * slopea	−0.118% (−0.146% to −0.092%), < 0.001	−0.090% (−0.103% to −0.077%), < 0.001
Ethnicity group * slope
African-Caribbean	12.0% (2.74% to 22.0%), 0.010	10.8% (3.87% to 18.3%), 0.002
South Asian	10.6% (1.01% to 21.0%), 0.030	8.98% (2.63% to 15.7%), 0.005
Albuminuria * slope
3–30 mg/mmol	25.6% (16.6% to 35.4%), < 0.001	15.4% (11.2% to 19.7%), < 0.001
> 30 mg/mmol	57.6% (39.2% to 78.6%), < 0.001	41.8% (33.2% to 50.7%), < 0.001
Diastolic BP * slope	0.170% (−0.060% to 0.411%), 0.147	0.100% (−0.020% to 0.220%), 0.103
Smoking status * slope
Current smoker	12.4% (1.31% to 24.7%), 0.028	6.82% (1.41% to 12.6%), 0.014
Ex-smoker	7.25% (0.702% to 14.1%), 0.029	4.71% (1.82% to 7.79%), 0.001
Loop diuretic * slope	−10.1% (−16.3% to −3.54%), 0.003	−3.15% (−6.67% to 0.501%), 0.094
CCB * slope	4.29% (−1.19% to 10.1%), 0.129	3.56% (0.803% to 6.50%), 0.012

There was a 0.605% increase in ACR intercept (95% CI 0.557% to 0.653%) for each 1% increase in baseline ACR, and a −0.118% decrease in slope (slower incline) (95% CI −0.146% to −0.092%) for each 1% increase in baseline ACR. Albuminuria (ACR category) increases both the intercept and the slope (steeper incline), larger increases were seen for those with ACR higher than 30 mg/mmol, ACR intercept increased by 472% (95% CI 360% to 611%) and slope increased by 57.6% (95% CI 39.2% to 78.6%) for those with ACR > 30 mg/mmol. Diastolic BP increased the slope (steeper incline) by 0.170% (95% CI −0.060% to 0.411%), for each increasing unit of BP (mmHg). For current smokers, there was an increase in progression (steeper incline) of 12.4% (95% CI 1.31% to 24.7%). For ex-smokers, the slope increased (steeper incline) by 7.25% (95% CI 0.702% to 14.1%). For African-Caribbean ethnicity, there was an increase in slope of 12.0% (95% CI 2.74% to 22.0%), compared to an increase in slope of 10.6% (95% CI 1.01% to 21.0%) for those of South Asian ethnicity. The rate of incline was slower for those prescribed loop diuretic medication, the slope decreasing by −10.1% (95% CI −16.3 to −3.54%); in contrast, the rate of incline was faster for those prescribed CCB who had an increase in slope of 4.29% (95% CI −1.19% to 10.1%).

Age and systolic BP were associated with the intercept only. Age decreased the ACR intercept by −0.499% (95% CI −0.896% to 0.100%) for each increasing year of age, and there was an increase of 0.300% (95% CI 0.100% to 0.602%) for each increasing unit of systolic BP (mmHg).

The parameter estimates of covariates associated with progression from the model using the full data set were similar to those from the substudy for all covariates in the final model. The parameter estimates for covariates associated with the intercept were different for some covariates using the full data set. Diastolic BP and smoking status were associated with intercept estimates of log ACR (p < 0.2) using the full data set, and there was no significant association between loop diuretic use and intercept levels of log ACR. Smoking status showed an association between ex-smokers and the intercept, which is likely a result of having more ex-smokers in the full data set (n = 367). Other estimates for the full data set were similar to those seen for the substudy data.

Figures 12 and 13 illustrate the changes in log-transformed ACR progression over time for albuminuria status and ethnicity groups, respectively. In Figure 12 the profiles of those with albuminuria incline more steeply compared to those who do not have albuminuria, although the data show high variability. The equivalent figure for the combined main and substudy data may be found in Appendix 1, Figure 24. Figure 13 supports a steeper incline overall for those with African-Caribbean and South Asian ethnicity. The equivalent figure for the combined main and substudy data may be found in Appendix 1, Figure 25.

FIGURE 12.

Median and IQR ACR over time by albuminuria status (substudy data).

FIGURE 13.

Median and IQR ACR over time by ethnicity group (substudy data).

The biological variability of measured and estimated glomerular filtration rate

Estimates of components of biological variation are given in Table 25. 97 The geometric exact CV_I value (95% CI) for mGFR was 6.7% (5.6 to 8.2). CV_I values for the eGFR equations were broadly equivalent: MDRD 5.0% (4.3 to 6.1), CKD-EPI_creatinine 5.3% (4.5 to 6.4), CKD-EPI_cystatin 5.3% (4.5 to 6.5), and CKD-EPI_{creatinine-cystatin} 5.0% (4.3 to 6.2) to each other. Modelling to investigate differences showed the CV_I for MDRD and CKD-EPI_{creatinine-cystatin} eGFRs to be significantly (at 5% level) lower than for mGFR (difference −1.8%, p = 0.027 and difference −1.8%, p = 0.022 respectively). Using the MDRD equation, positive and negative RCVs were 15.1% and 13.1%, respectively. For example, if the baseline MDRD GFR (ml/minute/1.73 m²) in an individual is 59, significant increases or decreases would be to values > 68 or < 51, respectively.

Sensitivity analyses were carried out without outlier detection and deletion. Data were similar to those obtained following outlier removal, with analyses after outlier removal estimating slightly reduced CVs.

Modelling to identify any trends over time resulted in non-significant slopes [coef = −0.005; 95% CI (−0.020 to 0.009); p = 0.488], thus providing no evidence of a change in disease state (kidney function) over the duration of the study.

Ability of glomerular filtration rate-estimating equations, together with albumin-to-creatinine ratio, or albumin-to-creatinine ratio alone, to predict those people that have progressive loss of kidney function (chronic kidney disease progression) and to predict mortality

We investigated the association between baseline factors and CKD progression. Models were developed for each of the main study equations using CKD progression defined as (1) a 25% decrease in mGFR and/or an increase in ACR category within the study period (Table 26) and (2) decrease in mGFR exceeding the RCV, and/or an increase in ACR category within the study period (Table 27). All models demonstrated an association between lower baseline eGFR and increased risk of renal progression within the study follow-up period.

Models including all of the GFR equations demonstrated an association between lower baseline eGFR and renal progression, defined as a decrease in GFR exceeding the RCV, and/or an increase in ACR category within the study follow-up period (see Table 27). For every 1 ml/minute/1.73 m² higher eGFR at baseline, there was a 4% reduction in the odds of progressing.

To evaluate the outcome of death in this cohort, two regression models were fitted: a logistic model (Table 28) with death within the study as an outcome and a Cox regression model (Table 29) to evaluate time to death.

Accuracy of more recent GFR-estimating equations including BIS1 and BIS2 equations, CAPA equation, LMR equation, FAS creatinine equation, FAS creatinine–cystatin equation, EKFC equation and 2021 CKD-EPI equations: baseline analysis

Performance characteristics of the more recent GFR-estimating equations are described in Table 5. With the exception of the CKD-EPI(2021)_creatinine and BIS1_creatinine equations, all of the newer GFR-estimating equations were negatively biased compared to mGFR (negative median bias with non-overlapping CIs). Biases of the cystatin C containing equations, before recalibration (see Impact of cystatin C calibration on the performance of more recent cystatin C containing glomerular filtration rate-estimating equations), tended to be greater than those of the creatinine containing equations, with equations incorporating both cystatin C and creatinine having intermediate bias. Several of the equations [BIS2_{creatinine-cystatin}, FAS_{creatinine-cystatin}, CKD-EPI(2021)_{creatinine-cystatin}] achieved P30 values in excess of 90%. The overall performance (P30) of the CKD-EPI(2021)_{creatinine-cystatin} and BIS2_{creatinine-cystatin} equations was superior to that of their respective creatinine-only equations [CKD-EPI(2021)_creatinine and BIS1_creatinine].

Ability of the newer glomerular filtration rate-estimating equations to reflect and detect changes in glomerular filtration rate over 3 years

The ability of the newer GFR-estimating equations to reflect and detect change in GFR over time is shown in Tables 9 and 12–15. As observed for the four primary equations, the newer equations, compared to mGFR, all estimated change in GFR within either 3 ml/minute/1.73 m²/years or difference in % observed change within 5%/years (absolute difference) for > 70% of patients (see Table 9).

The sensitivity and specificity of the newer equations to detect change in mGFR > 10 ml/minute/1.73 m² were similar to that of the primary study equations (see Table 12). Similar observations were seen for the other change metrics investigated: GFR change in excess of the RCV, change in excess of 25% and change in excess of 25% in addition to a change in the GFR category (see Tables 13–15).

Influence of body surface area adjustment method on accuracy (P30) of glomerular filtration rate-estimating equations: Haycock equation compared to the Du Bois equation

Accuracy of the GFR equations was unaffected by the method of BSA adjustment (Table 30). Bias and precision data for these analyses may be found in Appendix 1 (see Appendix 1, Table 41).

Impact of cystatin C calibration on the performance of more recent cystatin C containing glomerular filtration rate-estimating equations

As noted with the original CKD-EPI cystatin C containing equations (see Table 5), in all cases, substitution of recalibrated cystatin C values into the newer GFR-estimating equations was associated with a decrease in bias, with non-overlapping CIs, and with an increase in P30, in some cases with non-overlapping CIs compared to the equivalent non-cystatin C containing equation (e.g. BIS1 compared to BIS2). Relative improvement in performance tended to be more marked in the cystatin-only equations compared to the creatinine–cystatin equations (see Table 5).

Impact of creatinine method on accuracy of glomerular filtration rate-estimating equations: enzymatic compared to isotope dilution mass spectrometry method

There was no evidence of an effect of creatinine method on the performance of creatinine-based GFR-estimating equations; for example, CKD-EPI_creatinine using enzymatic creatinine P30 was 90.2 (88.4 to 91.9) compared to 89.0 (87.1 to 90.8) for ID-MS-creatinine (Table 31).

Creatinine results obtained using the enzymatic method were higher than those using the ID-MS assay, the relationship between the two methods being described by the linear regression equation; enzymatic = 3.23 + 1.01(ID-MS), R² 0.969 (see Appendix 1, Figure 26). Deming regression produced a similar equation [enzymatic = 1.20 + 1.03(ID-MS)].

Bias plot analysis demonstrated a constant mean positive bias of 4.7 μmol/l (CI 4.3 to 5.0) for the enzymatic compared to the ID-MS assays (see Appendix 1, Figure 27).

Methods