Page 141 - An Evidence Review of Active Surveillance in Men With Localized Prostate Cancer

be impossible to separate out the effects of individual tests that are conducted together (e.g., PSA
and DRE).
Consideration should be given to conceptualizing AS monitoring strategies as dynamic
treatment regimes (i.e., rules for sequential decision making based on the evolution of patient or
tumor characteristics over time). Such approaches formalize the process of choosing between
competing monitoring strategies based on expected responses to treatment and related
intermediate and long-term outcomes using appropriate statistical models. Compared to standard
research methods (e.g., directly comparing two monitoring strategies in a parallel group study),
dynamic treatment modeling may be better at identifying the optimal monitoring regime while
accounting for the temporal structure of the data (e.g., multiple monitoring visits) and the fact
that treatment decisions at each visit are determined by the measurements performed (e.g., PSA,
repeat biopsy). Indeed, statistical methods exist that can use observational or randomized study
data to determine the factors that should be considered as triggers for intervention, as well as the
optimal cut-off values of these factors. 243,244
At a minimum, future study reports should be very explicit and clear about what their
definitions of AS (or WW) were, what were the goals of the intervention, what were the exact
protocols, what were the exact definitions of progression, how and when protocols or standards
changed during their study (and why), and why and how often patients and clinicians chose to
not follow the protocols.

Key Question 3. Factors That Affect Offer of, Acceptance of,
and Adherence to AS

As described under the findings for Key Question 3, there are two major categories of studies
that address this Key Question: quantitative analyses of databases and registries, and more
qualitative analyses of surveys of men diagnosed with prostate cancer and their clinicians. To
date, both types of analyses have limitations that preclude strong conclusions. The databases tend
to have data only about what treatment patients received and when. Therefore, whether different
treatment options were offered to them, whether they accepted those options, and whether they
adhered to their initial choices could only be inferred. Even the best analysis of predictors of
initial treatment cannot adequately address the Key Question of interest to this conference’s
sponsors, since the three treatment stages of interest (offer, acceptance, and adherence) are not
described in the database. Thus, full statistical analyses of predictors will require the prospective
collection of data specifically about what interventions were offered to each patient, which
treatments the patients accepted, and when they chose to receive curative treatment despite lack
of evidence of progression. Ideally, data would also be collected on what a priori definition of
progression was used for each patient to allow the analysis of lack of adherence. These datasets
will need to be sufficiently large to allow for testing of multiple predictor variables. In addition,
future studies should only perform complete analyses of all treatment options (AS or WW,
surgery, radiation, ADT, and combinations) without arbitrarily grouping treatments (e.g., AS and
ADT) or selectively excluding treatments (e.g., by pairwise comparisons). This will minimize
bias and increase clarity about what is being tested.
We believe that, where possible, future database analyses and prospective observational
studies should focus on those predictors that are amenable to change or that can be acted upon.
For example, if it is shown that men who receive educational materials are more likely to accept
AS, this intervention can be implemented. Or if it is found that black men are less likely to be
offered AS, then training of physicians to minimize implicit bias may be warranted. However,

87