Washington, DC (UroToday.com) As has been highlighted during the symposium at this year’s Society of Urologic Oncology Annual Meeting, research provides the opportunity for urologists and urologic oncologists to expand their sphere of influence beyond clinical catchment areas and beyond their years of practice. However, to do so requires methodologically robust research. To this end, the Society of Urologic Oncology arranged a plenary session focused on the “Statistical Implications on Urologic Cancer”. While an understanding of statistical methodology is important for those undertaking research, it is also important for the critical interpretation of the published literature, a task that all urologists will undertake.
Dr. Sebastien Haneuse wrapped up the session on statistics in urologic oncology with a discussion of the role of the p-value in clinical trial reporting, and research more generally.
First, he began by outlining the null hypothesis testing paradigm. Statistical analyses, as he summarized, can be divided into the estimation of some rate or association, hypothesis testing, and the prediction of future outcomes. For hypothesis testing, we must begin with a null hypothesis, typically of “no association.” Then data is acquired and analyzed to assess the validity of the null hypothesis. The p-value is used to quantify the strength of the evidence, specifically with respect to the initial presumption of the null hypothesis, given the assumptions of the statistical models used.
However, beyond a numerical value between 0 and 1, p-values are often operationalized as a dichotomous measure using a threshold value, traditionally 0.05.
Dr. Haneuse then outlined two key questions relating to the role of the p-value:
- Is the p-value a reasonable value to compute?
- Is the way in which p-values are actually used reasonable?
He highlighted that this has been an ongoing debate for more than 100 years, both among statisticians and others, with a large volume of resulting literature and yet no consensus has been achieved.
Addressing the first question, he suggested that we could simply report p-values in their numerical form. However, clinicians, patients, and humans in general, typically like binary decisions both for their simplicity and as an impetus for future actions. The value of 0.05 was arbitrarily selected and much of the discussion regarding the role of p-values in clinical research has focused on whether this is a suitable threshold for modern clinical and public health research. Historically, and in its original intent, a “significant” p-value was intended to prompt “further scrutiny.” However, currently, this is not how it is used. Instead, a statistically significant p-value is often used as an indication of the importance of results or the basis of action.
Dr. Haneuse then highlighted a recent position statement from the American Statistical Association, dating from 2016 and 2019. Taken together, these address the aforementioned questions as follows:
- Is the p-value a reasonable value to compute? YES
- Is the way in which p-values are actually used reasonable? NO
The 2016 editorials laid out six principles for the use and interpretation of p-values as follows:
The 2019 editorial went even further to suggest that the expression “statistically significant” should not be used.
He concluded with a few personal perspectives:
- Our interpretation of p-values, rather than p-values themselves, is the problem.
- There is little value in debating/changing the threshold utilized for dichotomizing p-values.
- Research should move away from a prescriptive approach to how conclusions are drawn and reported.
Presented by: Sebastien Haneuse, PhD, Associate Professor of Biostatistics, Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, Massachusetts, USA.
Written by: Christopher J.D. Wallis, MD, PhD, FRCSC Contact: @WallisCJD on Twitter at the 20th Annual Meeting of the Society of Urologic Oncology (SUO), December 4 - 6, 2019, Washington, DC