Selection Bias in Observational Studies of Palliative Care: Lessons Learned

Area of Science:

• Health Services Research
• Gerontology
• Epidemiology

Background:

• Palliative care (PC) programs are often assessed using observational data.
• This methodology raises concerns regarding potential selection bias due to unobserved patient characteristics.
• Accurate evaluation of PC effectiveness is crucial for improving serious illness care.

Purpose of the Study:

• To quantify selection bias in a palliative care demonstration program.
• To assess the impact of both observed and unobserved factors on confounding.
• To inform best practices for evaluating PC programs using observational designs.

Main Methods:

• Utilized administrative and Medicare claims data (2013-2017) for 2983 beneficiaries aged 65+ in a PC program.
• Employed three matched comparison cohorts: regional, two-state, and eight-state.
• Measured confounding by comparing baseline characteristics (observed) and follow-up duration/mortality rates (unobserved).

Main Results:

• Observed confounding was evident in baseline differences (race, morbidity, utilization).
• Unobserved confounding was significant, with longer follow-up in comparison cohorts (187-207 days vs. 162 days).
• The PC cohort exhibited higher 6-month and 1-year mortality rates compared to all comparison groups.

Conclusions:

• Selection of comparison groups substantially influences confounding and effect estimates in PC evaluations.
• The significant impact of confounding highlights challenges in evaluating novel care models without randomization.
• Lessons learned can improve future observational studies of palliative care programs.