• © 2006 by American Society of Clinical Oncology

Looking Back From Death: The Value of Retrospective Studies of End-of-Life Care

Cancer is projected to cause more than 570,000 deaths in the United States in 2005.¹ As patients with cancer approach death, the issues they face are often similar regardless of the original site of disease. Important practice variation has been documented near the end of life, with inadequate treatment of pain and other symptoms,² increasing use of overly aggressive anticancer treatments,³ and disparities in access to hospice services.⁴ As a result, the quality of end-of-life care delivered to patients who die as a result of cancer is a major public health concern.⁵

Two approaches have commonly been used to study the quality of end-of-life care. Many studies have retrospectively assessed the care received by patients in the time frame leading up to death,^3,4,6 whereas others have identified patients entering the terminal phase of disease and prospectively observed their care forward in time.^7,8 Each of these approaches has its theoretical and practical advantages and disadvantages, yet it has been suggested recently that retrospective designs are particularly biased and thus should be “laid to rest.”⁹ This sweeping recommendation, however, risks the loss of important opportunities to understand and improve end-of-life care in large representative cohorts of patients with cancer.

In a perfect world, the optimal way to study end-of-life care, assuming unlimited resources, cooperation, and time, would be to assemble a cohort of all patients with incident disease, regardless of initial stage or severity, and collect data on their care until all of them have died. This study design provides an unbiased look at care before death among those who died, and also allows study cohorts to be defined by sentinel events among those who went on to die as a result of their disease and those who did not. However, obtaining information from a representative sample is particularly challenging. If one is able to identify patients who are entering the terminal phase and approach them quickly enough, they are often by definition very ill and unable to provide information. Although such studies have been conducted successfully,^7,10,11 they are expensive and laborious, making alternative study designs attractive. There are two alternatives: less-than-optimal prospective approaches to define study cohorts and retrospective approaches to this task.

In retrospective studies of end-of-life care, the cohort consists of decedents. These participants can be either patients who died as a result of a particular disease (as determined by death certificate, for example) or patients with a particular disease who died as a result of any cause. Investigators then look back over the time leading up to death to observe preterminal care. Retrospective designs have several advantages over prospective designs. First, they allow for easy identification of cohorts of relevant patients. In contrast, prospectively identifying patients who are approaching the end of life is often not feasible—especially from administrative data sources.

Second, a retrospective approach allows all patients who come to the end of life to be studied, not just a nonrandom subset defined by a disease characteristic or event that identifies them as dying. For example, assembling a prospective cohort based on hospice enrollment will only capture those patients whose physicians recognized that they were approaching the end of life, whose physicians were able to raise this difficult issue with them, and who themselves philosophically accepted hospice care. In another example, it has been suggested that analysis of end-of-life care be restricted to patients presenting with metastatic breast cancer, rather than looking at all patients who die of breast cancer.⁹ Less than 20% of the patients who die as a result of breast cancer initially present with clearly terminal disease, however, and these patients have been shown to have different characteristics from those who initially present with earlier stage cancer and then later relapse.¹² Unfortunately, relapse of cancer is not accurately or completely captured by most tumor registries. A prospective cohort study of patients with metastatic breast cancer at diagnosis would yield no information about the majority of patients who die as a result of breast cancer. In addition, there is great variability in the natural history and effectiveness of therapies among patients with metastatic cancer, such that some patients with incurable metastatic breast cancer may be “terminally ill” for several years. In those situations, their initial care will not be focused on end-of-life issues. Without research using retrospective designs, the end-of-life experiences of patients not easily defined as dying will often be overlooked.

Lastly, by only assessing the end-of-life care of patients who actually die, efficient retrospective designs facilitate the development of performance measures that can be used with readily available data to monitor end-of-life care across providers, geographic areas, demographic groups, and time periods.¹³ Such performance measures may become key elements of new national systems to monitor and improve the quality of cancer care.^14-16

In this issue of the Journal of Clinical Oncology, Gagnon et al¹⁷ present a valuable advance in retrospective methods for studying end-of-life care. To select more refined and appropriate population-based cohorts in which to retrospectively assess the quality of terminal care delivered, they developed and validated an algorithm for use with administrative data to classify decedents who were dying as a result of breast cancer, as opposed to dying as a result of either the complications of treatment or unrelated causes. This study and its potential applications still suffer from the recognized limitations inherent in using administrative data to assess the quality of care, such as a lack of control over data quality and superficial clinical detail.¹⁵ However, as clinical and administrative data converge with the increasing adoption of electronic medical records, the quality of such data will continue to improve. The methods assessed by Gagnon et al will allow researchers to go beyond death certificates as the usual source of cause of death information to identify patients more accurately who were in need of terminal cancer care before their death. By advancing the methods for assembling cohorts using administrative data, the investigators have made an important contribution to population-based surveillance of the quality of terminal cancer care.

Because medicine is practiced prospectively, retrospective studies may seem less relevant to clinicians who are advising patients in real time. However, physicians usually know when their patients are nearing death. Still, the longer the time before death examined, the less the certainty that all the care observed is truly in the context of dying.¹⁸ As a result, retrospective designs are best used when characterizing what happens to patients in time periods very close to death, such as a few weeks or months. Even so, some patients may die unexpectedly and not be perceived as being near the end of life until the time of death. In those cases, care leading up to death would not have had an end-of-life intent. The approach outlined by Gagnon et al can help researchers to avoid including such patients.

End-of-life care is complex and difficult to study. Studies of this phase of care are critical, however, to identify deficiencies in quality and areas for improvement. Prospective and retrospective methods to study end-of-life care identify different groups of patients and are not intended to be interchangeable approaches. The prospective design is best used for research questions about the care given to patients who can readily be identified as having a terminal illness. The retrospective approach is optimal for asking, “what happens shortly before death to patients who die?” Retrospective designs are convenient, generalizable, and efficient for evaluating and implementing performance indicators related to end-of-life care. Rather than calling for the exclusive use of only one of these two broad approaches,⁹ we recommend that prospective and retrospective methods be considered complementary designs that can each provide useful insights, that assessment of the biases of each be explored through sensitivity analysis (eg, secondary analyses restricting or expanding the study populations), and that the relative limitations of each approach be acknowledged in study reports. Neither design is a gold standard. Rather, they each have an important place in end-of-life research and quality assessment.