Subjects and procedures
Data was obtained from KTX patients enrolled at the kidney transplant facilities of the Renal Unit at the Cardiff and Wales NHS Trust in Cardiff, UK (n=209) and Saint Louis University Hospital, St. Louis, MO (n=233);
Description of US sample and data collection procedures
Patients were identified from the renal departmental database of the Saint Louis University Hospital from January 2008 to June 2008. All adult patients (18–74 years old) with a documented kidney transplantation were asked to answer a self-administered questionnaire during a regular visit at the transplant clinic (n=282). Patients providing informed consent were 233 (82% of the original sample). We excluded patients with multi-organ or multiple transplant, and those with no serum creatinine measurements in the 3 months prior to the interview. Since severe acute health events may affect quality of life and determine transient variation in Glomerular Filtration Rate (GFR) we also excluded patients who underwent major surgery in the months prior to enrollment and those with markers of acute cellular damage (Creatine-Kinase > 200 U/L). The final sample resulted in 137 patients. Information including biochemistry assessments results obtained in the 3 months prior to interview and lifetime medical history was abstracted from clinical charts, transplant coordinators records and electronic medical records. We evaluated the accuracy of data reporting across the 3 different sources by evaluating the agreement of laboratory test results performed on the same date for each subject. Since the agreement was almost perfect (ρ=0.99 for all laboratory test results considered) we merged the information into one common clinical database in order to maximize data completeness. The Saint Louis University Institutional Review Board approved the study protocol.
Description of UK sample and data collection procedures
All patients registered in the renal departmental database of the Cardiff and Vale NHS Trust in September 2002 (n=1251) were asked to answer a self-administered questionnaire. Of them 157 were on Continuous Ambulatory Peritonale Dialysis, 268 were on hemodialysis, 115 were on CKD pre-dialysis stage, and 711 received a transplantation. All patients treated at the renal clinic received a postal-survey at their home while those on chronic dialysis at the time of survey were asked to complete the questionnaire during a regular dialysis session . . Patients providing the informed consent were 33.3% of the original sample (29,4% among transplant recipients). For the present analysis we included all adult patients with a documented kidney transplant who completed the survey questionnaire (n=209 of whom 14 were on dialysis after graft failure at the time of survey). and had at least one serum creatinine measurement in the 3 months prior to the interview (n=144). No patient with markers of acute cellular damage in the 3 months prior to survey return date were identified (Creatine-Kinase > 200 U/L). Clinical data including serum creatinine concentrations and medical history were abstracted from the clinical database of the Cardiff and Vale NHS Trust.
Predictor and covariates
We obtained clinical information from patients’ charts. We collected all laboratory test results recorded in the 3 month period prior to the date of questionnaire administration (serum creatinine, albumin, hemoglobin, Alanine Transaminase (ALT), Aspartate Transaminase (AST), creatine kinase, glucose, phosphorus and calcium). We estimated Glomerular Filtration Rate (GFR) with the MDRD equation (4-variable) . We classified patients according to renal function with the National Kidney Foundation Chronic Kidney Disease (CKD) staging system  (CKD 1–2, eGFR ≥ 60 ml/min/1.73 m2; CKD 3, 60 ml/min/1.73 m2 > eGFR ≥ 30 ml/min/1.73 m2; CKD 4, 30 ml/min/1.73 m2 > eGFR ≥ 15 ml/min/1.73 m2; CKD 5, eGFR < 15 ml/min/1.73m2 or patient on dialysis). Both questionnaire administered in US and UK included a section on socio-demographic characteristics (age, gender, education, ethnicity, employment status). For the US sample, depression was defined by self-reported medical diagnosis in the 12 months prior to the interview or prescription of antidrepssant/anti-anxiety medication as reported in clinical charts. For the UK sample the same comorbidity were defined using ICD9 codes listed in Quan H. et al. , and Li, B., et al. . Lifetime diagnoses of diabetes and cardiovascular diseases were abstracted from clinical charts (electronic records and hardcopies) in both samples. Diabetes was defined by ICD-9 (or ICD-10) codes (complicated and uncomplicated as defined in Quan H. et al. ) anytime, prescription of insulin or any anti-diabetic drug in the 90 days screening period. Cardio-vascular disease was defined by ICD-9 (or ICD-10) codes anytime in patient’s history (myocardial infarction, congestive heart failure, peripheral vascular diseases and cerebrovascular diseases (Quan H. et al. ).
Descriptive statistics were calculated and reported as mean ± standard deviation (or median with interquartile range) for continuous variables and frequency for categorical variables. Differences in socio-demographic and clinical characteristics across CKD stages were tested by χ2 for categorical variables and Analysis of Variance (or Kruskall-Wallis test when appropriate) for continuous variables. We evaluated the unadjusted associations between CKD severity, glucose, AST, hemoglobin, phosphorus, calcium, and Albumin serum concentrations with Spearman’s correlation coefficient. Spearman's correlation was also used to test unadjusted associations between eGFR (or CKD stages) and quality of life.
Ordinary Least Square (OLS) models are traditionally adopted to analyze quality of life data. Under ideal conditions they have attractive properties: the conditional mean is an easy-to-interpret, parsimonious representation of the relationship between a continuous outcome and a predictor variable. Additionally, economic models used in cost-effectiveness analysis adopt adjusted means from OLS models in Quality-Adjusted Life Years (QALY) calculation. For this reason we initially used general linear models to obtain adjusted association estimates. Since EQ-5Dindex
scores were strongly skewed and OLS regression assumptions were not satisfied, we identified the power transformation maximizing model R2
with the SAS Proc Transform routine:
where QL is the transformed dependent variable of the general linear model and QOL is the raw health utility score.
We specified 2 consecutive steps for the GLM analyses. In the first step we included renal function alone to assess the unadjusted coefficients of association. Variables included in the second step were age, gender, ethnicity, months since transplant, diagnosis of diabetes, hypertension, cardiovascular diseases, anxiety/depression, ALT, AST, glucose levels and center of enrollment. From the second step, we obtained adjusted means for each CKD category (as defined above). Results were back-transformed into the original scale, correcting for back-transformation bias . Significance of trend across CKD categories was assessed by partial Spearman’s correlation including all variables entered at each consecutive step.
Since the distribution of the EQ-5Dindex score was strongly skewed and a relevant ceiling effect was observed, we used Quantile Regression to evaluate the consistency of the association between renal impairment and HRQOL at the 15th, 30th, 50th, 70th and 85th quantile of the outcome distribution  Quantile regression minimizes mean absolute distance at a given quantile, rather than modeling the conditional mean as in standard regression. Quantile regression is robust to departures from ordinary least square assumptions. In quantile regression, a quantile, such as the median, depends on the ranks of the Y values, and not on specific values in the tails of the distribution. Quantile regression was introduced by Koenker and Bassett (1978) , for the analysis of linear and non-linear response models. Useful features of quantile regression include (a) the models can be used to characterize the entire conditional distribution of a dependent variable; (b) the resulting estimated coefficients from quantile regression are robust to outlier observations on the dependent variable and violation of normality and homoscedasticity of the error term; (c) the resulting estimators are more efficient than those from OLS in the case that the error term is non-normal; (d) potentially different solutions at different quantiles may be interpreted as differences in the response of the dependent variable to change in the regressors at various points in the conditional distribution of the dependent variable; e) parameter estimates can be interpreted as change in the dependent variable per unit change in the independent variable, allowing direct comparison with OLS parameter estimates. We modeled the relationship of CKD stages (categorical variable) with HRQOL and calculated adjusted medians using parameters estimates obtained with Quantile Regression. We modeled the relationship between eGFR with HRQOL using the raw outcome data. Models have been adjusted for age, gender, education, ethnicity, time since transplant, diagnosis of diabetes, hypertension, cardiovascular diseases, glucose levels and AST. In all models (OLS and Quantile Regression) renal function was included alternatively as a continuous (eGFR) or categorical variable. Center effect was accommodated by including in the regression models an indicator variable denoting the center of enrollment and its interaction term with the main predictor of the analysis. Parameter estimates in all analysis on the continuous predictor refer to a 10 ml/min./1.73 m2 change in eGFR.
Blood concentrations of Hemoglobin, Albumin, Calcium and Phosphorus were not considered as confounders in the statistical models since they may be part of the causal pathway linking CKD severity and HRQOL. In order to evaluate whether the association between eGFR and HRQOL was partially independent from the hypothesized mediators we included those variables to the statistical models in a secondary analysis.
We considered P values < 0.05 as statistically significant and p<0.10 as marginally significant. We used SAS 9.2® to conduct all the analyses.