We used multivariable linear regression with robust SEs26 to explore the linear association between the family-assessed outcomes (QODD-1, FS-ICU) and our three predictors of interest: average daily ICU costs, total ICU costs, and total hospital costs. We adopt a nonparametric interpretation of linear regression that assumes neither linearity nor heteroscedasticity.27 To understand the magnitude of the associations, under the assumption of linearity, the regression coefficients represent the difference in quality of death outcomes per doubling of the cost variable. Generalized linear estimating equations were used to examine the relationship between nurse-assessed QODD-1 and costs, which allowed us to account for correlation between observations attributable to nurses completing surveys for multiple patients.28 If both the nurse at time of death and the nurse for the prior shift responded, we selected the questionnaire that was more complete; if both were equally complete, we randomly selected one per patient.16,17 Baseline, minimally adjusted regression models included dummy-indicator variable for hospital and pre/post intervention adjustment. In addition, we adjusted models for any confounders that changed the minimally adjusted cost parameter estimate by an absolute value ≥ 10%, selecting from the following a priori-identified variables: patient age, race, sex, education, median household income by zip code, insurance type, and underlying cause of death. For models with family-assessed outcomes, we also examined family member’s age, sex, level of education, spouse vs other relationship, and presence at the time of death. For nurse-assessed outcomes, we examined nurses’ age and sex. In addition, a priori we identified patient race, age and underlying cause of death attributable to trauma as potential effect modifiers. For the family-reported outcomes (QODD-1 and FS-ICU), we identified patient insurance status, median household income by zip code, and family member level of education as potential effect modifiers. In sensitivity analyses, we compared available patient information between survey responders and nonresponders to assess the potential extent of selection bias using t tests with unequal variances or χ2 tests, as appropriate. A two-sided α level of ≤ 0.05 was considered statistically significant. Data were analyzed using STATA, version 12.0 (StataCorp LP) statistical software.