Predicting the prevalence of chronic kidney disease in the English population: a cross-sectional study

The Quality Improvement in CKD (QICKD - ISRCTN56023731) trial [10, 11] population includes a representative large sample of patients with CKD stages 3 to 5 in England [12]. The primary aim of the QICKD trial is to compare quality improvement interventions aimed at lowering systolic blood pressure in patients with CKD in primary care; ethical approval has been given for secondary analyses of the data. Ethics approval was received from the Oxford Research Ethics Committee (Committee C) (ref: 07/H0606/141). Here, the QICKD dataset was used to model the association between the prevalence of CKD and its potentially explanatory factors. This dataset contained patient-level data, extracted from the computer systems of 129 English general practices (GP), based in London, Surrey, Leicester, Birmingham, Cambridge and Sussex. The full dataset contained information on 930,997 people, of whom 743,935 are aged 18 or over. We only included people aged 18 or over to be consistent with the P4P scheme. The data are cross-sectional, extracted in 2009.

Cases of CKD stages 3 to 5 were strictly defined in accordance with the 2002 K-DOQI classification [13] on the basis of an estimated glomerular filtration rate (eGFR) of less than 60 ml/min/1.73 m² for at least 90 days. Laboratories in England report eGFR using the four-variable modified diet in renal disease formula [14], with correction factors applied under the guidance of the National External Quality Assessment Service [15] to account for differences in local creatinine assays. People without a serum creatinine measurement in their electronic record were assumed for the purposes of the analysis not to have CKD.

We considered 11 potentially explanatory variables. These may be loosely classified as socio-demographic variables and variables about the presence or absence of cardiovascular disease. The socio-demographic variables were: age of subject (in years), gender, ethnicity, smoking status and deprivation score. Ethnicity was based on the 2001 England and Wales Census ‘5 + 1’ categories: ‘Asian’, ‘Black’, ‘Mixed’, ‘White’ ‘Other’, and ‘Not Stated’ [16]. There were two additional categories: ‘Not recorded’ occurred when there was an explicit code stating that ethnicity was not recorded, and ‘Missing’ was for missing ethnicity data. Smoking status was recorded as ‘never smoked’, ‘ex-smoker’, ‘smoker’ or it may be missing Ethnicity and smoking status were both modelled using dummy variables. Deprivation score was a continuous variable from the 2007 index of multiple deprivation [17], and based on the patient’s postcode [18].

The cardiovascular diseases considered were: diabetes, ischaemic heart disease (IHD), heart failure, hypertension, peripheral vascular disease (PVD) and stroke. These were modelled using dichotomous indicators. Data on systolic and diastolic blood pressure were available, but were not used due to high levels of missing data (23% of values were missing).

Multivariable logistic regression was used to model the dependency of having CKD on the potentially explanatory factors. Model building was mainly based on the recommendations of Hosmer and Lemeshow [19]. Briefly, variables that showed a significant univariate association with CKD were included in a multivariable model. Manual backwards elimination was then applied; dropping non-significant variables one at a time. When no more variables could be deleted a check was made to see if any variables could be included. This gave a preliminary main-effects model. We checked if transformations were required for any continuous variables in this model, and considered possible interactions. We only considered interactions with age as this was known to be a strong predictor of CKD prevalence [5‐7] and important interactions with age are often identified [20]. We used sample splitting to validate this approach; the sample was randomly split into two sub-sets (of approximately equal size), and the model-building process applied to both sub-sets.

Due to the large sample size nearly all variables were statistically significant with p-values less than 0.001. Instead we used clinical significance; for our study a variable was defined as clinically significant if its odds ratio (OR) was either above 1.49 or below 0.67. These values are derived from published CKD guidelines [1], in which it is stated that a rise in serum creatinine of over 20% should be considered significant. We checked for interactions by plotting the prevalence of CKD against age and stratifying by the levels of each factor (deprivation was categorised into quintiles). We observed a non-linear interaction between age and the cardiovascular diseases. This was modelled by introducing a dummy variable ‘below age 50’ which takes the value 0 if patients are below the age of 50 years and 1 otherwise, and including its interaction with the cardiovascular diseases.

We constructed a ‘clinical’ model (considering all the potentially explanatory variables) and a ‘parsimonious’ model which considered just age and gender as it was noted that sometimes these are the only variables for which data are available. We also present the results from the full main-effects model, as the use of this is sometimes recommended in the literature [20, 21]. We used STATA version 10.1 [22] for all analyses.

Missing values for smoking status and ethnicity were treated as separate categories. Individuals with missing blood pressure readings (23%) were assumed not to have hypertension. Missing data for deprivation (19%) were due to a computer error during data collection, and so are assumed to be missing completely at random and were imputed by using a single implementation of the ICE procedure [23]. There were no missing data for any of the other variables.

Models were compared based on both their Akaike’s and the Bayesian information criteria (AIC and BIC respectively) [24], models with lower values were preferred. We performed a residual analysis using the deviance residuals to check goodness of fit. The Hosmer-Lemeshow test [19] may also be used to formally check goodness of fit. However, this measure is known to be of limited use for large sample sizes [25], so a graphical alternative was used: predicted and observed CKD prevalence were compared using deciles of predicted values [26, 27]. The ability of the models to correctly classify patients was summarised by their ‘area under the receiver-operating-characteristic curve’ (AUROC) [28], along with their sensitivity, specificity, positive predictive value and negative predictive value. These values are known to be biased (give values that are too optimistic about model performance) when calculated based on the same data to which the model was built. To avoid this, we used a model built on one sub-set of the data, and calculated the statistics on the remaining sub-set.

The model-building process was repeated separately for CKD that had and had not been identified under the P4P scheme. Descriptive statistics were also produced for these two CKD classifications.

Using cross-sectional data is a limitation, as it is known that rates of progression vary by patient characteristics [1]. The results of this analysis may be used to identify areas with a high prevalence of CKD, where early identification will be beneficial in reducing both progression to renal failure and morbidity from cardiovascular disease. However, when targeting resources for CKD, consideration should also be given to variations in rates of progression across populations. We also made no distinction between varying levels of kidney disease. The available literature suggests that the risk profile for CKD may vary as kidney disease progresses; for example it has been shown that the proportion of males with CKD increases with worsening stage [37], and a recent study found that non-white ethnicity was a significant predictor of renal replacement therapy [38]. As renal failure can be devastating for the patient and very expensive [1], more research is required into rates of progression.

We have assumed that people without a serum creatinine measurement did not have CKD. Whilst this is consistent with previous approaches [39], there was no measurement recorded for 56% of the sample. Hence the prevalence of CKD reported here is likely to be an under-estimate.

The choice to use clinical significance instead of statistical significance posed some problems. In particular the choice of whether or not to include the multi-categorical variables ethnicity and smoking status was slightly arbitrary. The importance of all the omitted variables warrants further research. For the continuous variables the value of the odds ratio (and hence their clinical significance) depends on the units reported. We used the odds ratio per 10-year increase in age, which is commonly employed in the literature [5, 8, 29, 30, 34]. For deprivation we used a 10-point increase (deprivation values range between 0.75 and 77.37). Using the results from the full main-effects model, we would need to use a 45-point increase in deprivation for it to become clinically significant, and a 5-year increase in age for it to become not clinically significant.