|
|||||
|
|
||||||
© 2002 American Society for Clinical Oncology
Screening for Lung CancerTampere School of Public health, University of Tampere, Tampere, Finland To the Editor:The recent article by Strauss1 concludes that, in the Mayo Lung Project, survival was superior in the screened population and provided unbiassed surrogate for cure. The conclusion is based on methodologic errors, and the paper should not be taken as evidence for screening for lung cancer. The multivariate analysis under the heading of "Overdiagnosis Bias" does not address the issue of overdiagnosis at all. The overdiagnosis bias as an explanation of the inconsistency between no mortality reduction but better survival among those intended to be screened was eliminated in the article by multivariate Cox proportional hazards regression with mode of screening and mode of treatment as variables. Screen-detected cases were more often resected than the others. Statistically significant effect was found only of resectability, which led the author to conclude that detection was only a surrogate for resection. However, detection was not a surrogate, but resection was a chain in causal pathway between detection and death. Hence, multivariate methods cannot be used in this case at all. The section "Mortality, Randomization, and Confounding" accounts the no-mortality reduction by failure of randomization (the experimental arm having higher risk of lung cancer than the control arm). The reasoning based on biology and large number of confounders is indirect, and the evidence presented is weak. We found the biologic characteristics of lung cancer to be consistent with the overdiagnosis hypothesis.2 A large number of confounders is evidence against rather than for the confounder explanation. It is most unlikely that imbalance between the arms could have occurred with, and only with, the most critical confounderincidence of lung cancerand not only in the Minnesota trial but in other trials as well. The other explanations, including overdiagnosis, have much higher credibility. Randomized experiments were introduced to eliminate and not to create such biasses as discussed by Strauss. The evidence from a randomized trial is stronger than from any other combination of design or analysis. The analysis and conclusions in the article1 are erroneous and misleading. Unless better evidence surfaces, screening for lung cancer should not be recommended. REFERENCES
1. Strauss GM: The Mayo Lung Cohort: A regression analysis focusing on lung cancer incidence and mortality. J Clin Oncol 20: 1973-1983, 2002 2. Hakama M, Holli K, Visakorpi T, et al: Low biological aggressiveness of screen-detected lung cancers may indicate over-diagnosis. Int J Cancer 66: 6-10, 1996[CrossRef][Medline]
ResponseRoger Williams Medical Center, Providence, RI In Reply:Dr Hakama takes issue with the conclusion that survival was unbiased in the Mayo Lung Project (MLP) and thus provided an accurate surrogate for cure.1 He believes that "this conclusion is based on methodologic errors" and "should not be taken as evidence for screening for lung cancer". Specifically, he objects to the conclusion that MLP data are inconsistent with the contention that screening led to the overdiagnosis of lung cancer. The question of overdiagnosis is critical because the paper also demonstrated that randomization provides an opportunity to rigorously eliminate confounding by both length bias and lead-time bias in the setting of a randomized population trial (RPT).1 Length bias is eliminated by using an intent-to-treat analysis, whereas lead-time bias is eliminated by measuring survival from randomization, not from diagnosis.1 Accordingly, overdiagnosis is the only conventional screening bias that potentially confounds survival in a mature RPT. Dr Hakama asserts that the analysis of overdiagnosis bias "does not address the issue of overdiagnosis at all." He objects to the use of multivariate Cox proportional hazards regression as a statistical tool to assess the extent to which the data are consistent with the hypothesis that chest x-ray screening leads to the detection of pseudodisease in lung cancer.2 The article demonstrated that resection was the only multivariate predictor of lung cancer mortality among cases, whereas screen detection was not important when adjusted for differences in resection rates.1 Because screen detection was only important insofar as it predicted resection, I concluded that screen detection was not a surrogate for pseudodisease. Dr Hakama argues that "resection was a chain in the causal pathway between detection and death. Hence, multivariate methods can not be used in this case at all." Dr Hakama is absolutely correct that if a potential confounder is on the causal pathway between exposure and outcome, then adjustment for confounding is inappropriate. For example, if one were interested in studying the relationship between cigarette smoking and lung cancer mortality, it would be inappropriate to adjust for the dramatic increase in lung cancer incidence related to smoking. Because cigarette smoking causes lung cancer, and because lung cancer causes lung cancer mortality, adjusting for higher lung cancer incidence among smokers would lead to the spurious conclusion that cigarette smoking does not increase lung cancer mortality. The decision about whether a confounder is on the causal pathway must be based on biologic rather than statistical considerations. The question to be addressed here is whether resection is part of the causal pathway between screen detection and lung cancer mortality. It is difficult to understand how screen detection can be inferred to cause resection. Resection was not required in the face of screen detection in MLP, and other therapies, as well as no treatment, were available options. Nonetheless, screen detection facilitated resection, because 68% (123 of 181) of screen-detected cancers were resected compared with 11% (22 of 185) of symptom-detected cases. Accordingly, it is appropriate to use Cox regression to determine whether screen detection or resection was the important independent predictor of survival. The causal pathway argument is particularly curious, because in his own article, Hakama et al3 argues for the existence of overdiagnosis in lung cancer. If screen-detection were a surrogate for pseudodisease, then adjusting for resection would not change the relationship between screen detection and lung cancer mortality, because resection would not be a predictor of survival. If screen-detected cases were comprised of both true cancers and pseudodisease, then screen detection would remain significant, even after adjusting for resectability. The fact that screen detection lost its predictive value when adjusted for resection is inconsistent with the overdiagnosis hypothesis. Nonetheless, rather than debating whether a particular statistical technique is useful to address the overdiagnosis question, it is perhaps more informative to rely on a more descriptive analysis of the data. Table 1 defines four subgroups, based on resectability and method of detection. It demonstrates very strong differences in risk of lung cancer death when comparing resected with unresected groups, but no differences between screen-detected and symptom-detected cases within resection categories. Table 1 also demonstrates no significant differences in death from other causes when comparing subgroups and that lung cancer caused 89% of all deaths among those with lung cancer. Accordingly, competing causes of death affected only a small minority of lung cancer patients.
Figure 1 compares survival among the four subgroups and supports similar, though not identical, conclusions. There is a highly significant overall difference when comparing all four groups (P < .0001). Among the two subgroups that underwent resection, a statistically significant difference was not seen (P = .20), although, with only 22 symptom-detected cases, there was little power to detect a difference.
On the other hand, there was a significant survival difference when comparing subgroups not undergoing resection (P = .0075). Among 163 symptom-detected cases, every patient succumbed within 44 months of diagnosis; whereas among 58 screen-detected cases, the last patient did not succumb until 70 months. However, this survival advantage represents a classic example of length bias because not a single individual among 221 unresected patients achieved long-term survival. (In dramatic contrast, 50% of 145 resected patients were cured.) Dr Hakama believes that his article provides evidence in support of overdiagnosis.3 However, to borrow a phrase from his letter, it is my position that his article "does not address the issue of overdiagnosis at all." He reports on lung cancer cases diagnosed at Tampere University Hospital between 1983 and 1987, during which time chest x-ray (CXR) screening was widely available in Finland for the detection of tuberculosis.3 Among 506 cases, 26% were screen-detected and 74% symptom-detected. The major finding was that survival was significantly superior among screen-detected cases that were diploid or of low S phase, compared either with aneuploid or high S-phase screen-detected tumors, or to symptom-detected cases. He interprets this observation as supporting the conclusion that favorable sur-vival among these screen-detected subgroups reflected their "low malignant potential." Moreover, he suggests that the favor-able subgroups "possibly would not have surfaced clinically in the absence of screening, or would not have caused death even if left untreated."3 Hakama et als observation of better survival for tumors with favorable biologic characteristics provides another example of length bias. However, it provides no evidence for overdiagnosis. Hakama et al assert that "overdiagnosis is at the extremity of the biologic spectrum of length bias, making it difficult to distinguish between these two biases."3 However, on this point, we strongly disagree. Length bias is integrally related to screening for cancer and must always be addressed when interpreting screening effectiveness. On the other hand, if pseudodisease does not exist, then overdiagnosis cannot occur. Although there is considerable evidence that aneuploid and high S-phase tumors are more rapidly fatal than diploid or low S-phase lesions, there is also abundant evidence that these more favor-able lesions are fully capable of causing lung cancer death, even among those with stage I disease.4 There is absolutely no biologic evidence that diploid or low S-phase lung cancers represent pseudodisease. On the other hand, Dr Hakama has the data to address this question. Among 131 screen-detected cases, only 34% were resected, while 41% received other treatment, and 24% were untreated. Among 375 symptom-detected cases, the respective figures were 13%, 49%, and 37%. Overall, 34% (170 of 506) of screen- or symptom-detected cases received no treatment whatsoever. If Hakama et al are correct that many "low biologic aggressiveness" tumors "would not have caused death even if left untreated",3 they should report how many untreated lung cancer patients were long-term survivors. Unfortunately, the article included no analysis of survival as a function of treatment. Dr Hakama asserts that "the evidence from a randomized trial is stronger than from any other combination of design or analysis." On this point, we completely agree. However, MLP was a randomized trial, and the question is not whether randomization is superior to observational designs, but rather whether survival or mortality provides the best measure of screening effectiveness in the RPT setting. Dr Hakama concludes that "unless better evidence will appear, screening for lung cancer should not be recommended." My article did not raise the policy question of whether screening for lung cancer should be recommended. This was primarily because of its limited scope. Although survival was unbiased in MLP, the survival/mortality question needs to be addressed from a more global perspective. Moreover, my Journal of Clinical Oncology article contained very little consideration of problems that exist with regard to randomization and mortality in the RPT setting. All these issues will be fully explored elsewhere. However, if it can be demonstrated that survival provides the most accurate measure of screening effectiveness in the RPT setting, my analysis would provide a rationale to consider a change in public policy. Randomization to screening in MLP was associated with a greater than two-fold improvement in long-term survival and cure. Given the fact that more than a million people die from lung cancer every year, early detection could provide an opportunity to dramatically reduce lung cancer mortality on a global basis. REFERENCES
1. Strauss GM: The Mayo Lung Cohort: A regression analysis focusing on lung cancer incidence and mortality. J Clin Oncol 20: 1973-1983, 2002
2. Black WC: Overdiagnosis: An unrecognized cause of confusion and harm in cancer screening. J Natl Cancer Inst 92: 1280-1282, 2000 3. Hakama M, Holli K, Visakorpi T, et al: Low biological aggressiveness of screen-detected lung cancers may indicate over-diagnosis. Int J Cancer 65: 1-5, 1996[CrossRef][Medline] 4. Strauss GM, Kwiatkowski D, Harpole DH, et al: Molecular and pathologic markers in stage I non-small cell carcinoma of the lung. J Clin Oncol 13: 1265-1279, 1995[Abstract]
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|||||||||||
|
Copyright © 2002 by the American Society of Clinical Oncology, Online ISSN: 1527-7755. Print ISSN: 0732-183X
|