Table 3.

Use of lasso logistic regression to select variables best at predicting the outcome of moderate-to-severe pruritus among 1668 Chronic Renal Insufficiency Cohort Study participants

Selected Predictors (Listed in Descending Order of Coefficient Magnitude)	Method of Lasso Lambda (λ) Selectiona
	Cross-Validation	Minimize BIC	Adaptive
Change in eGFR, absolute value	✓	✓	✓
10% decrease in weight from baseline	✓	✓	✓
Sex	✓	✓	✓
10% increase in weight from baseline	✓	✓	✓
Body mass index at baseline	✓	✓	✓
BDI score ≥11	✓	✓	✓
Baseline eGFR	✓		✓
Serum phosphate at baseline	✓		✓
New lung disease	✓		✓
New congestive heart failure	✓		✓
New peripheral vascular disease	✓		✓
New prescription opioid use	✓		✓
New hypertension	✓
Prescription opioid use at baseline	✓
Diabetes at baseline	✓
Peripheral vascular disease at baseline	✓
CRP at baseline	✓
PTH at baseline	✓
New diabetes	✓
Urine protein	✓
Race/ethnicity	✓
New coronary artery disease	✓

Variables are listed in order of the absolute values of their coefficients, with largest values first. All variables were standardized to facilitate comparisons of relative size of regression coefficients. Variables selected repeatedly by lasso regression models using different lambda tuning parameters demonstrate a more robust association with the outcome. In addition to comorbidities at baseline and baseline laboratory parameters, variables were created to represent a new diagnosis of diabetes, hypertension, coronary artery disease, peripheral vascular disease, congestive heart failure, or new use of opioid medications during follow-up. A 10% increase or decrease in weight refers to the percent change in weight as compared with the weight at the baseline visit. The absolute change in eGFR was calculated by subtracting the eGFR at the time of the outcome or censoring event from the baseline value. Because we wanted to create variables to capture change in eGFR during follow-up, this analysis was restricted to the 1668 participants with at least two measurements of eGFR. BDI, Beck's Depression Inventory; CRP, C-reactive protein; PTH, parathyroid hormone.

^aCross-validation chooses lambda on the basis of the model that minimizes the 10-fold cross-validation function. Minimize BIC refers to the lambda that results in the model with the smallest value for the Bayesian information criterion (BIC). Adaptive lasso also uses cross-validation but uses a different penalization criterion in which coefficients receive different data-driven weights.