Michael Grossman

BA, PhD

CUNY and NBER

Distinguished Professor

PLCO (Learn more about this study)

PLCO-108

Oct 20, 2014

An Econometric Evaluation of Prostate Cancer Treatments

Although thousands of men undergo treatment for prostate cancer every year, very little is known about the effectiveness of prostatectomies, irradiation, or hormonal therapy. Large scale randomized trials tend to be unfeasible. Even the largest trial to date, the Prostate Cancer Intervention Versus Observation Trial (PIVOT) has only had a few hundred participants (Wilt et al, 2012). Statistically significant effects are thus hard to detect. Observational studies, by contrast, have larger sample sizes, but do not take self-selection into account (e.g. Alibhai et al, 2005). It is conceivable that more health conscious men are more likely to seek treatment, so observational studies might over-estimate the effectiveness of cancer treatments. It is equally conceivable that men who live less healthy lives to begin with, might be more likely to receive treatment. In that instance, the effectiveness of an intervention would actually be underestimated. In other words, it has so far been impossible to obtain reliable causal estimates for the effect of prostate cancer treatments on mortality outcomes. Our study intends to change this by using two variations of the econometric technique of instrumental variables (IV).

The IV approach uses the variation in an instrumental variable that only affects the probability of receiving treatment but has otherwise no effect on the outcome to isolate a causal treatment effect. The assignment to the screening group in the PLCO Cancer Screning Trial constitutes just such an instrument. Assignment is random, so it will in itself have no effect on mortality but will affect the treatment probability. A proof that this method indeed yields consistent treatment effects with a causal interpretation is provided by Angrist et al (1996).

A variation of this technique is the so called regression discontinuity design (RDD). It exploits the fact that rules sometimes quasi-randomly divide a population into a treatment and control group. In the context of prostate cancer, PSA test results have this very effect. PSA levels above 4 ng/mL are considered to be indicative of prostate cancer. As a result, men with PSA levels of 4.1 ng/mL will be somewhat more likely to eventually receive a prostatectomy than men with PSA levels of 3.9 ng/mL. At the same time, the populations of men just above and of men just below the threshold will be virtually identical in terms of both observable and unobservable characteristics. Since this “randomization through threshold” is not deterministic (i.e. men below the threshold might still receive treatment) but merely probabilistic (i.e. men below the threshold are less likely to receive treatment), an RDD is a local version of the IV technique.

Both methods will allow us to measure CAUSAL effects of prostate cancer treatments on mortality in general and death from prostate cancer in particular. We believe that this is an important issue. Since prostatectomies, irradiation, and hormonal therapy all have substantial side effects, it is key to quantify the upside of these interventions in terms of prolonged lives.

Our main objective is to isolate the causal effects of prostate cancer treatments on mortality outcomes. We would like to assess the effectiveness of three different interventions: radical prostatectomies, radiation treatments and hormone treatments. Information on whether a trial participant has received either of these interventions is contained in the Prostate Treatments dataset. Mortality outcomes we would like to evaluate include:
• Probability of death from prostate cancer within the entire follow-up period
• Probability of death (in general not just prostate cancer specific) within the entire follow-up period
• Probability of death from prostate cancer within different time intervals (1 year, 2 years, 5 years,...)
• Probability of death (in general not just prostate cancer specific) within different time intervals (1 year, 2 years, 5 years,...)
• Time to death after intervention

Using the supplemental questionnaire data from 2006-2008 (i.e. post-treatment), effects of the above treatments on subsequent health measures and quality of life indicators can be evaluated. Our goal is to estimate the effect of prostate cancer treatments on the following outcomes:
• Body weight
• Use of pain killers and anti-inflammatory drugs
• Physical activity and fitness
• Smoking incidence
• Frequency of urination (very rough proxy of the side effects)

Mortality and all of the above indicators should be evaluated across socio-economic groups (e.g. by age group or ethnicity).

A useful 'by-product' is that we will be able to evaluate the degree to which physicians adhere to the PSA threshold of 4 ng/mL. For instance, a big jump at the 4 ng/mL in the probability of receiving treatment for prostate cancer within a certain time period, would be indicative of adherence to the threshold rule. A small jump - or now jump at all - would indicate that physicians do not rely as much on the threshold value.