Reply: End-of-Life Treatment Preferences in the Health and Retirement Study.

We read with interest the letter from Dr. Block and Dr. Mathay (1). They hypothesize that mortality differences by presepsis trajectory class may be explained, at least in part, by differences in goals of care between these different types of patients. We agree that treatment preferences may explain some of the differences in observed mortality. In the year before sepsis, patients in the high-use class spent more than twice as many days in a healthcare facility than those in the rising use class (2), suggesting both a willingness to accept intensive medical treatment and an ability to tolerate living with chronic illness.

Because most patients die after a decision to limit treatment, it is necessary to consider a patient’s prespecified treatment preferences rather than treatment limitations initiated shortly before death. As Dr. Block and Dr. Mathay note, the Health and Retirement Study collects information on written end-of-life (EOL) instructions through an “exit survey” of each decedent’s next of kin (3). These survey data are publically available at https://hrs.isr.umich.edu/data-products, and we encourage researchers to explore this rich resource.

Exit surveys have been completed for the majority (87%) of the derivation cohort. However, of the 1,320 patients with an exit survey completed, only 304 (23%) had written EOL instructions completed before their sepsis hospitalization, with an additional 55 (4%) patients having EOL instructions written at an unknown time. The proportion of patients with presepsis EOL instructions differed by presepsis trajectory class: low-use (24%), rising use (21%), and high-use (10%; P = 0.04). Among the 304 patients with presepsis EOL instructions, however, neither “desire to limit care in certain situations” (93% vs. 94% vs. 83%; P = 0.9) nor “desire to have certain treatments withheld” (78% vs. 74% vs. 83%; P = 0.6) differed by presepsis trajectory class assignment.

Incorporating data on EOL instructions into our multivariable models would require a careful approach to dealing with the high rate of missing information, and thus is beyond the scope of this letter. However, we encourage others to use the publically available Health and Retirement Study datasets to further explore this question and will certainly consider doing so ourselves in our future work.