Measuring the effectiveness of androgen-deprivation therapy for prostate cancer in the medicare population: adequate data are neither the same as nor the enemy of perfect data.

Abstract

In the late 1970s, investigators learned that androgendeprivation therapy (ADT) decreased serum testosterone levels and,more importantly, reducedbonepainamongmenwith prostate cancer. Whereas ADT demonstrated benefit in patients with metastatic disease and as an adjunct to radiation therapy in patients with locally advanced disease, its use at the organ-confined stage has never been supported by evidence or expert guidelines. Most importantly, randomized data have shown that immediate ADT for nonmetastatic prostate cancernotonly lacks a survival benefit but maycauseharm, suchasanexcess riskofbone fractures.1Nevertheless, the use of ADT for localized prostate cancer increased greatly between the 1990s and early 2000swith compelling evidence that favorable reimbursement contributed to this trend.2 In this issue of JAMA InternalMedicine, Lu-Yao et al3 have confirmed and extended their previous observational cohort studies of prostate cancer treatment outcomes in the Medicare population residing in a region encompassed by the National Cancer Institute’s Surveillance, Epidemiology, andEnd Results (SEER) registries. InaMedicare cohort spanningnearly 2 decades of 66 717 men aged 65 years and older with localized prostate cancerwho receivednodefinitive local therapy, they found that primary ADT confers neither overall nor disease-specific advantage. Moreover, Potosky et al4 have recently reported a similar observational cohort study of more than 15 000 men receiving non–curative intent treatment within integrated health care systems. This study leveraged longitudinalmedical recorddatawith richer clinicaldetail than is available in SEER-Medicare and, like the study of Lu-Yao et al,3 demonstrated neither a survival advantage nor disadvantage from ADT. The analysis of Lu-Yao et al3 highlights both the strengths andpitfalls of usingobservational cohorts, suchas those identified from SEER-Medicare linked data, tomeasure effectiveness. Inferences from observational studies aremost reliable when the groups being compared are similar in all respects other than the intervention under scrutiny. Of course in practice, this is often not the case. Indeed, in the study of Lu-Yao et al,3 baseline disease characteristics differ substantially between men who did and did not receive ADT with regard to clinical stage and grade, among other attributes. As it became clear that there is little or no role for primaryADT in the treatment of nonmetastatic prostate cancer, contemporary clinicians have becomemore selective in recommending primary ADT, for example, reserving it for patients with poor prognosis or the inability to tolerate radiation therapy. Conversely, “conservative management” of prostate cancer as a concept has shifted fromwatchfulwaiting, apassiveapproach forolder menwith limited life expectancy focused on palliation rather than treatment, to active surveillance, in which healthy patients are intensively monitored with repeated prostatespecific antigen tests, digital rectal examinations, and prostate biopsies at systematic intervals and in which definitive treatment is advocated at any signof disease progression.5 Finally, prostate-specific antigen testing at diagnosis, an important prognostic and predictive marker, is imperfectly captured by SEER and is only available for themost recent years. Both the missing data and the imbalance from this registrybased study impede the ability to make definitive comparisonsbetween these treatment strategies.The limitationsofobservational cohort studies using SEER-Medicare datamust be balanced against the recognition that there will not be additional randomized trials and that the assessment of effectiveness fromSEER-Medicare data canbedone expeditiously and at nominal cost. One strategy to circumvent the classic selection bias that plagues the ability to make valid inferences from treatment comparisons using observational data is to use an instrumental variable to account for confounding by unmeasured differences between treatment groups. To be considered valid, an instrumental variable must satisfy 2 quintessential conditions: thevariablemustbecorrelatedwith the treatmentwhile not being associated with the outcome of interest except through the effect of the treatment itself. This statistical method therefore reduces treatment selectionbias and“pseudorandomizes” patients. Lu-Yao et al3 used the proportion of patientswho receivedADTwithinahealth service areaas their instrument. Knowinghowwell an instrument performs is the key to interpreting how successfully it accomplished the goal of pseudorandomization. TheF-statistic is commonlyused to assess the strength of the instrument and thereby the ability tomitigate potential selection bias. A robust instrument typically has an F-statistic greater than 10. The attempt to identify an instrumental variable is a strength of the study of LuYao et al3; nonetheless, inclusion of an assessment of the instrument’s performance would have helped readers gain a better understanding of howwell it addressed potential confounding. Indeed, physicians needmore guidance on how to interpret the results of instrumental variable analyses as they Related article page 1460 Research Original Investigation Androgen-Deprivation Therapy for Prostate Cancer