Evaluation of Reliability and Correlations of Quality Measures in Cancer Care

This cross-sectional study assesses measures of care process, utilization, end-of-life care, and survival at oncology practices that treat older Medicare beneficiaries with newly diagnosed lung, breast, or colorectal cancer.

Note that models did not include registry because they included random effects for practice. eAppendix 3. Quality Measures Specification eTable 5 includes details on the specification of the various measures. Note that the process measures are based on care that has been shown in prior studies (typically randomized clinical trials) to be associated with improved outcomes (including survival, disease-free survival, and/or progression-free survival). We used multi-level hierarchical linear models with practice-level random effects to calculate adjusted practice-level rates for each quality measure or summary measure. Models took the form:

eTable 5. Specification of Measures
where Y ij is receipt of the quality measure of interest for the patient i in practice j β 0j is the intercept for the practice j; we assume β 0j ~ N(0,  2 between ) X ij a vector of covariates for patient i in practice j ꞓ ij is the random error associated with the patient i in practice j; we assume ꞓ ij ~ N(0,  2 within ) Models adjusted for age (65-74, 75-84, ≥85), sex, race/ethnicity (White, Black, other/unknown), marital status (unmarried, married, unknown), census-tract median household income (quartiles), census-level proportion of residents without a high school education (quartiles), Charlson Comorbidity Index 12 (0, 1, 2, ≥3), and year of diagnosis. We adjusted for American Joint Commission on Cancer, 7 th edition stage (1, 2, 3, 4, unknown) when patients of more than one stage were included in a measure.
Estimated practice-level random effects, β 0j , represent the difference in performance for practice j compared to an average practice, holding patient characteristics constant. We use these estimated random effects to compute adjusted practice-level rates as: where Y ave is the mean rate for the measure averaged across all patients in the cohort.
Finally, we compute reliability for a practice with median number of attributed patients as the ratio of the between variance to the total variance: R med =  2 between /( 2 between + 2 within /n med ) Here we provide SAS code for the models for an example measure:

. Estimation of Sample Sizes for All Newly Diagnosed Cancer Patients in Practices
As described in the methods, we recalculated reliability after estimating the total number of newly diagnosed patients with lung, colon/rectal, or breast cancer that a practice would be expected to treat, assuming data on quality could be extracted from a comprehensive electronic medical record.
We first estimated the number of patients per practice expected if we also had data for Medicare Advantage patients (based on SEER-Medicare data) and individuals aged <65 years (based on the age distribution of each cancer type). Our first estimates assumed 5 years of data, as for the main analyses. We first identified the total number of patients aged 65 and older in the SEER Medicare data including both fee-for-service Medicare patients and Medicare Advantage patients. This provided an estimate of the total number of patients aged 65 and older. We then used data on the median age at diagnosis for each cancer type to estimate the total number of patients aged 65+ and thus estimate the total number of patients of all ages. For example, the median age at breast cancer diagnosis is 62, thus we estimated that 45% of all breast cancer patients are aged 65+ at diagnosis. The median age at diagnosed of lung cancer is 70, so we estimated that 67% of all lung cancer patients are diagnosed at age 65+. The median age of diagnosis for colon cancer is ~70 (68 for men, 72 for women) and rectal cancer is 63 years, thus we estimated that approximately 60% of rectal cancer patients are diagnosed at age 65 and older.
We then estimated the proportion of all patients represented in our cohort (column 6 in eTable 3. For each practice, we then divided the number of patients by this factor. Chemotherapy in last 2 weeks of life 6587 0 (0.0%) 0 (0.0%) 0 (0.0%)

All Cancers (450 total practices)
Proportion of patients who died who did not enroll in hospice more than 3 days before death