Robust Evaluations of Intensive Care Unit Length of Stay Using Observational Data

Dear Editor:

In their recent special report, Khandelwal et al. highlight important methodological considerations in analyses of the impact of palliative care interventions on intensive care unit (ICU) length of stay (LOS) and healthcare costs. ¹ They make several sound recommendations for improving our ability to make a business case for palliative care, focusing on randomized trials. However, there are some additional considerations to note when these outcomes are analyzed with observational data, which is often the case in palliative care research.

First, in an observational study, the timing of an intervention is outside of the investigators' control and needs to be accounted for in analyses of an intervention's impact on ICU LOS. If an intervention is provided during an ICU admission, but the timing of this intervention is variable, distinguishing between total LOS and postintervention LOS is critical. Consider a study of the impact of inpatient palliative care consultations on ICU LOS. A long ICU LOS could itself be a trigger for referral to inpatient palliative care. ² If this relationship is not accounted for, the true impact of palliative care on LOS and associated costs will be obscured. To estimate the potential impact of a palliative care consultation on ICU LOS, analyses should account for the fact that initial LOS is correlated with both palliative care receipt and total LOS.

Alternatively, if the intervention was provided in a hospital before ICU admission, analyses should distinguish between the intervention's impact on risk of any ICU admission and its impact on ICU LOS. Consider an intervention that reduces the risk of ICU admissions, but does so mostly among patients with lower illness acuity. Patients with greater acuity will still have ICU admissions, and their LOS may be longer than what the patients with lower acuity would have experienced had they been admitted to the ICU. If a single regression model is run on only those with an admission, an analyst may erroneously conclude that the intervention led to a longer LOS. Instead, models that account for a large number of zeros or two-part models that adjust for the probability of ICU admission should be considered. ³

Second, Khandelwal et al. suggest accounting for censoring due to death by assigning an LOS value to death and testing the sensitivity of results to different LOS values. However, when used in conjunction with an observational study that incorporates adjustment for confounders, this method requires the assumption that the factors leading to long LOS and death are the same. Otherwise, to the extent that patient preferences or institutional factors are associated with ICU LOS but not mortality, and to the extent these confounders are associated with receipt of treatment, treatment effect estimates will be biased. Another alternative to consider is a competing risks model that treats censoring due to death as informative. ⁴

I wholeheartedly agree with Khandelwal et al.'s call for robust economic evaluations of palliative care interventions. Careful analyses—and detailed data—are needed to optimize both the value and quality of care provided to seriously ill patients. Additional caution is necessary when analyzing observational data.