My Methods write up for this is below, let me know if anyone has some feedback. Thanks, and I know this is kind of a boring study, but a good initial platform for me to learn from.

**METHODS**

The primary study focus of patient experience was measured using Consumer Assessment of Healthcare Providers and Systems (CAHPS®) Clinician and Group surveys. Data came from surveys returned between January 2014 through April 2017 for patients seen at a pediatric hematology-oncology outpatient clinic located in a Midwestern city. The clinic was a mid-size semi-private practice with five full-time physicians, no fellows or mid-level providers, who saw approximately sixty to seventy new oncology patients per year, as well as a significant volume of patients with a variety of hematologic conditions. Other clinic staff included four nurses, two patient access associates, three research assistants, one patient care technician, two social workers, one child life specialist, one psychologist, one nurse coordinator, and one scheduler.

The CAHPS® instrument included 25 items scored on a 5-point Likert scale in addition to a “Rate This Provider” 0 to 10 point scale. The primary study outcome was to determine predictors of a Top-Box score for “Rate This Provider” (defined as score of 9 or 10) and predictors of “Likelihood of Your Recommending This Practice to Others” (defined as score of 5, “Very Good”). Additional survey questions were also dichotomized based on a response of “Very Good” or not. Patient demographic information was not provided due to the anonymity of the surveying process, though the physician provider seen at the visit was available.

*Analysis*

Data was preliminarily reviewed to examine item response missingnesss and need to control for patients nested in providers. Approximately 0% to 2% of individual response question data were missing with no discernible patterns (e.g., missing response combinations or monotonicity); with these data classified as missing completely at random. Approximately 6% to 8% of data were missing for the two questions on visit delays and patient medications. The greater missingness rates in these questions were assumed to be related to respondents interpreting questions as “not applicable” if there were no visit delays or patient medications. An empty multilevel model was used to examine the need to control for provider level random effects. Model results revealed that the provider level did not explaining a significant amount of response variability in regards to study outcomes. As a result analyses were based on complete case data without the use of random effects.

Study data were partitioned into two random subsets based on a 60/40 split. The larger subset was used for predictor selection related to the two study outcomes using least absolute shrinkage and selection operator (LASSO) based on multiple logistic regression (glmnet package in R, Vienna Austria, see supplemental Figure Sa and Sb), due to possible multicollinearity in item responses and sparse outcome concerns. Predictors selected via LASSO were then used within the smaller dataset for estimate calculation using Bayesian logistic regression (PROC GENMOD, DIST=BIN, BAYES option in SAS 9.4, Cary NC). Models incorporated normal prior distributions (mean = 1; variance = 0.5) using multiple chain Monte Carlo (i.e., 3 chains, thinning = “3”) with 100,000 iterations and discarding first 50,000 as burn-ins. Model convergence was monitored and checked via traces plots, autocorrelation, posterior samples, and Gelman-Rubin statistics and Raftery-Louis diagnostics. Results were presented using odds ratios (OR) with 95% credible intervals (CI). A sensitivity analysis was conducted by rerunning the Bayesian analyses with non-informative prior distributions. Institutional Review Board approval was obtained for this study.