Rebecca Y Coley

PhD

Johns Hopkins Bloomberg School of Public Health

Postdoctoral Fellow

PLCO (Learn more about this study)

PLCO-118

Nov 12, 2014

Causal Inference for Prostate Cancer Screening and Treatment using a Bayesian Hierarchical Latent Class Approach

Screening and treatment for prostate cancer are highly controversial. Our research will use causal inference techniques to answer questions of primary interest to health care policy makers as well as clinicians and patients considering screening and treatment options including: What are the population- and subgroup-level causal effects of screening and various treatment regimes on survival and quality of life? Given an individual’s characteristics, what are the expected causal effects of screening on his survival and quality of life? Upon positive screen and biopsy, what is the probability an individual has lethal cancer given his particular biopsy results and which treatment options will optimize lifespan and quality of life taking into account his health priorities?

The benefits of prostate specific antigen (PSA)-based prostate cancer screening are unclear. While sustained decreases in prostate cancer mortality have been observed since the advent of PSA-based screening in the 1990’s, some reduction in mortality could also be explained by improvements in prostate cancer treatment (Etzioni et al, 2008). The largest-to-date randomized screening trial examining the impact of PSA-based screening on prostate cancer mortality in the United States, the Prostate, Lung, and Colorectal, and Ovarian (PLCO) cancer screening trial, did not find a significant decrease in prostate cancer mortality risk among men assigned to screening, but contamination was considerable and prevented causal inference about the effects of screening (Andriole et al, 2009).

Concurrently, widespread PSA-based screening has lead to increased diagnosis of prostate cancer and accompanying concerns of overtreatment (Miller, 2012). Upon diagnosis, early curative treatment with surgery, radiation, or androgen deprivation therapy is common even when tumors are suspected to pose lower risk (Cooperberg et al, 2010). Curative intervention can be physically, emotionally, and financially taxing for patients. In particular, at least 20-30% of men experience urinary incontinence and erectile dysfunction after surgery or radiotherapy (Chou et al, 2011). Meanwhile, the benefits of treatment vary for patients depending on the severity of their cancer. Specifically, men whose cancer would never become symptomatic have no potential to benefit from treatment.

Despite the risks associated with overtreatment, patients and doctors may often choose early treatment despite biopsy results indicating low risk due to uncertainty in initial diagnosis and, more specifically, an inability of existing biopsy techniques to distinguish with certainty between cancers that will remain indolent and those that are, or will become, aggressive. Prostate biopsies are only informative about the biopsied tissue, but features of non-biopsied tissues remain unobserved. As a result, doctors and patients must make treatment decisions in the face of this uncertainty.

For this project, we will apply a latent class model that reflects both uncertainty in diagnosis and differential effects of screening and treatment on survival and quality of life depending on an individual’s unobserved true cancer state. Via a Bayesian hierarchical framework, our model will incorporate information about prostate cancer prevalence and sensitivity of biopsy procedures in order to estimate average and subgroup effects.

The objective of this study is to use existing causal inference methods (see, for e.g., Robins, Hernan, and Brumback 2000; Angrist and Pischke 2009, Heran and Robins 2014) as well as a novel Bayesian hierarchical latent class model to estimate the following causal effects.

First, we will estimate the average causal effects of screening and treatment (prostatectomy, radiation, and hormone therapy) on prostate cancer and all-cause mortality. The mortality interval we consider (e.g., 10-year mortality) in our analyses will depend on the completeness of the data. In addition to prostate cancer-specific mortality, we will also consider all-cause mortality because surgery, radiation, and hormone therapy can have serious side effects which affect longevity.

Second, we will estimate the causal effects of screening and treatment within subgroups defined by prostate cancer status (i.e., none, indolent, or lethal). While true prostate cancer status is not observed for the majority of study participants, we will supplement available PSA, biopsy and prostatectomy results with published scientific findings about prostate cancer prevalence and sensitivity of screening and biopsy procedures.

Third, we will estimate overall and subgroup-specific effects of screening and treatment on quality of life among those who would survive regardless of treatment. Since quality of life as an outcome is inherently nested under survival- a patient who has died by definition has no quality of life- we will use principal stratification to restrict this analysis to participants who would survive with and without PSA testing (for the effect of screening analysis) or who would survive under all treatment options (for the effect of treatment analysis) (Frangakis and Rubin, 2002).

Finally, we will use this data to build a predictive model that enables more informed clinician and patient decision-making about screening and treatment under the umbrella of a larger project, the Johns Hopkins Individualized Health Initiative (Hopkins inHealth). Hopkins inHealth is a University-wide collaboration to more intelligently use electronic medical records data and, ultimately, develop a learning health care system. Within the context of prostate cancer, the types of predictions we will be able to make include: a man’s risk of lethal prostate cancer prior to screening, the likely severity of his prostate cancer given positive biopsy results, and the likely effect of various treatment options on his survival and quality of life.