Extended Lung Cancer Incidence Follow-up in the Mayo Lung Project and Overdiagnosis

Abstract

Background: A troubling aspect of cancer screening is the potential for overdiagnosis, i.e., detection of disease that, in the absence of screening, would never have been diagnosed. Overdiagnosis is of particular concern in lung cancer screening because newer screening modalities can identify small nodules of unknown clinical significance. Previously published analyses of data from the Mayo Lung Project, a large randomized controlled trial conducted among 9211 male cigarette smokers in the 1970s and early 1980s indicated that overdiagnosis might exist in lung cancer screening. At the end of follow-up (July 1, 1983), no difference in lung cancer mortality was observed, but an excess of 46 cases in the intervention arm suggested overdiagnosis. Because that excess could instead have resulted from short follow-up time, we investigated this possibility by conducting long-term lung cancer incidence follow-up. Methods: We investigated the lung cancer status through 1999 of the 7118 participants in the Mayo Lung Project who were alive and without diagnosed lung cancer in 1983 by use of medical records, surveys mailed to participants or next-of-kin, and state death certificates. Results: Information was available for 6101 participants, including 811 with inconclusive lung cancer status. From November 1971 through December 31, 1999, 585 participants in the intervention arm and 500 in the usual-care arm were diagnosed with lung cancer. Conclusions: The persistence of excess cases in the intervention arm after 16 additional years of follow-up provides continued support for overdiagnosis in lung cancer screening.

A troubling aspect of cancer screening is the potential for overdiagnosis. Overdiagnosis refers to the identification, through screening, of disease that would never have been diagnosed otherwise during a person's lifetime ( 1 ) . In the absence of screening, overdiagnosed cancers would not present symptomatically because the disease is indolent or, for potentially lethal disease, death from another cause preceded the theoretical date of symptomatic detection. Overdiagnosis wastes health care resources, and the evaluation and treatment of overdiagnosed lesions can lead to morbidity and even to premature mortality ( 2 ) .

A randomized controlled trial that uses a stop-screen design ( 1 ) (i.e., one in which screening is terminated after a prespecified number of years but follow-up continues for ascertainment of cases of disease and deaths) provides the best setting in which to assess whether overdiagnosis accompanies screening. Because randomization produces study arms with the same expected number of cases, the experience of the control arm can be viewed as the counterfactual experience of the intervention arm (i.e., the experience that the intervention arm participants would have had if they had not received the intervention). It is this principle that allows assessment of overdiagnosis in a stop-screen randomized controlled trial. During the screening phase of a stop-screen randomized controlled trial, an excess of cases will exist in the intervention arm that is caused by the introduction of lead time, the phenomenon in which screening advances the date of a cancer diagnosis so that diagnosis occurs earlier than if the cancer were detected symptomatically. If overdiagnosis has not occurred, the cumulative number of cases in each arm will equalize with time after screening stops (i.e., catch-up) as the counterparts of the earlier screen-detected cancers are detected symptomatically in the control arm. If overdiagnosis has occurred, the number of cases in both arms will never equalize because the excess cases in the intervention arm will have no counterparts in the control arm.

Overdiagnosis is of particular concern in lung cancer screening, because new imaging technologies can detect very small lung nodules. Although these nodules are considered to be abnormal, their clinical significance remains uncertain. Furthermore, persons most likely to be screened for lung cancer (i.e., heavy or long-term cigarette smokers) often have clinically significant cardiac or pulmonary comorbid conditions, which increases the risk of serious adverse events, including death, during diagnostic evaluation and treatment. Identification of overdiagnosed cancers through screening would unnecessarily put these persons at potentially serious risk.

Previously published analyses of data from the Mayo Lung Project, a large randomized controlled trial conducted among 9211 male cigarette smokers in the 1970s and early 1980s ( 3 , 4 ) , indicated that overdiagnosis might exist in lung cancer screening. At the end of the follow-up phase in 1983, 206 lung cancers had been diagnosed in the intervention arm, and 160 had been diagnosed in the usual-care arm ( 4 ) . Although some researchers concluded that overdiagnosis was responsible for the difference ( 5 ) , the short follow-up time after screening (on average, approximately 3 years) led others to wonder whether catch-up would occur with additional follow-up ( 6 ) . To determine whether lung cancer frequency would equalize in the two groups over time and to obtain a better understanding of overdiagnosis of lung cancer in the Mayo Lung Project and of lung cancer screening in general, we analyzed data on lung cancer incidence from an extended follow-up through 1999 of the 7118 Mayo Lung Project participants who, as of 1983, were alive, were not known to have been diagnosed with the disease, and had not requested that their data be excluded from future research.

S UBJECTS AND M ETHODS

The Mayo Lung Project

The goal of the Mayo Lung Project was to assess whether an intense regimen of chest x-rays and sputum cytology could reduce lung cancer mortality. Details of study design and trial operations ( 4 , 7 – 13 ) and commentary, additional analyses, and reinterpretations of study findings ( 6 , 14 – 16 ) have been published previously. In brief, the study began in November 1971 and enrolled 9211 older male smokers who tested negative for lung cancer on chest x-ray and sputum cytology screening (i.e., the eligibility screen) and were judged to have a life expectancy of at least 5 years and sufficient respiratory reserve to undergo lobectomy if necessary. Participants were randomly assigned to one of two study arms: the intervention arm ( n = 4618), in which participants received a chest x-ray and sputum cytology examination every 4 months, or the usual-care or control arm ( n = 4593), in which participants received a recommendation at trial entry only for an annual screening chest x-ray and sputum cytology examination (reflecting the standard of care at the Mayo Clinic in 1970). No screening examinations (other than the eligibility screen) were provided to participants in the usual-care arm. All participants were encouraged to quit smoking.

The 4618 men randomly assigned to the intervention arm were screened for 6 years and then followed for at least 1.5 years. Compliance with screening examinations averaged 75%, and only 12 participants were lost to follow-up. In the usual-care arm, only 14 participants were lost to follow-up. To ascertain lung cancer diagnoses and deaths, the Mayo Lung Project used a standard questionnaire that was mailed to all participants. Participants in the usual-care arm received the questionnaire annually, and those in the intervention arm received the questionnaire every 4 months during their 6-year screening period and annually thereafter. An additional questionnaire was mailed to participants in the usual-care arm in the last year or two of the study. That questionnaire requested the year in which the participant had his last chest x-ray.

The Mayo Lung Project ended on July 1, 1983. As of that date, lung cancer mortality in the two arms was similar: 3.2 per 1000 person-years in the intervention arm and 3.0 per 1000 person-years in the usual-care arm. However, 206 lung cancers had been diagnosed in the intervention arm, and 160 had been diagnosed in the usual-care arm.

Assessment of Lung Cancer Status

Our goal was to identify any lung cancer diagnoses in the Mayo Lung Project cohort that occurred between the end of the earlier follow-up period (July 1, 1983) and December 31, 1999. The cohort totaled 9189 participants after the removal of 22 participants (eight in the intervention arm and 14 in the usual-care arm) who requested, after the close of the Mayo Lung Project, that their data not be used for other research. All study activities were approved by the U.S. National Institutes of Health Office of Management and Budget and by the Mayo Clinic Institutional Review Board. Data collection began in November 2001 and was completed in September 2003. Medical record ascertainment was completed by April 14, 2003, the date that the Privacy Rule of the U.S. Health Insurance Portability and Accountability Act of 1996 became legally mandated. Given stipulations concerning release of personal health information in the Privacy Rule, data collection after April 14, 2003, would have proved more costly and difficult.

Given the likelihood that many Mayo Lung Project participants would be deceased, we began with an assessment of the ability of next-of-kin (including friends or others familiar with the medical history of the participant, in addition to blood relations) to accurately report lung cancer diagnoses for their deceased male relatives. A standardized questionnaire that queried next-of-kin about participants' medical history, including history of lung cancer, was mailed to the next-of-kin of 133 of the 140 deceased participants known to have had lung cancer diagnosed at the Mayo Clinic since the end of the follow-up phase of the Mayo Lung Project (seven participants were excluded for whom no next-of-kin could be identified). Next-of-kin were asked to choose from among the following five possibilities for each medical condition: “I know for sure he did not have it,” “I don't think he had it,” “I don't know if he did or didn't have it,” “I think he had it,” and “I know for sure he had it.” Because the accuracy was high, as measured by the high percentage of correctly reported lung cancers (sensitivity) ( Table 1 ), we continued with study activities.

We used the Mayo Lung Project database to identify those participants who required no additional investigation. This group included 366 participants with a lung cancer diagnosis before July 1, 1983, and 22 participants with a lung cancer diagnosis after July 1, 1983, as noted in the database but not previously reported. An additional 1683 participants received no additional investigation because the database indicated death before July 1, 1983, and no lung cancer diagnosis. Medical records, surveys mailed to participants or next-of-kin, and state death certificates were used to ascertain lung cancer status for the remaining 7118 participants.

We began our search for lung cancers with the Mayo Clinic computerized record system. Participants were considered to have had lung cancer if a lung cancer diagnosis before December 31, 1999, was identified in that source. Participants were considered not to have had lung cancer if they had been examined at the Mayo Clinic after December 31, 1999, and had no record of lung cancer in the computerized record system. Participants whose lung cancer status could not be determined by these two sources were slated for follow-up with mailed self-administered questionnaires. There were two short, standardized questionnaires: one for participants and one for next-of-kin. Both questionnaires contained questions on lung cancer diagnosis and month and year of diagnosis. The Participant Questionnaire also included questions on general health, recent smoking, and chest scans, and the Next-of-Kin Questionnaire also included questions on the participant's smoking status in the last 12 months of life and the relationship between the participant and the person completing the questionnaire. Categorical answer choices were provided for all questions except amount smoked, year quit, and dates of diagnoses and survey completion.

The Mayo Clinic Survey Research Center was responsible for locating participants or their next-of-kin. A fee-based Internet research and location service (Accurint, http://www.accurint.com ), approved for use by the Mayo Clinic, was used to find mailing addresses and vital status of participants and to identify next-of-kin and their addresses. If information was incorrect or unavailable from that source, other sources were used, including telephone directories, obituaries, and the U.S. Postal Service.

Multiple attempts were made to contact participants and next-of-kin with mailings or, if necessary, by telephone calls. Questionnaires were frequently administered over the phone by Survey Research Center personnel at the preference of the participant or his next-of-kin. Because the Participant Questionnaires were mailed to participants with unknown vital status, a number of these questionnaires were completed by a next-of-kin who received the mailing intended for his or her deceased relative. Approximately 12% of next-of-kin provided information on Participant Questionnaires.

For participants who had no next-of-kin and were known to be dead, death certificates were obtained to ascertain lung cancer diagnoses from information on the cause of death. Other sources, such as medical records, obituaries, and correspondence from the U.S. Postal Service, were used occasionally to determine vital status as well as history of lung cancer where possible. Participants for whom death certificates or other sources that explicitly mentioned lung cancer were available were classified as having had lung cancer. Participants for whom available medical records did not mention lung cancer were classified as not having had lung cancer. Participants for whom death certificates and sources other than medical records did not explicitly mention lung cancer were classified as “inconclusive.” Relevant information from death certificates and other sources was recorded on Next-of-Kin Questionnaires by Survey Research Center staff.

During data collection, study personnel who could potentially interact with participants or their next-of-kin or who handled death certificates, medical records, or other sources were not told of participants' randomization assignment. Data presented to the National Institutes of Health Project Officer, Mayo Principal Investigator, and other supervisory staff during the course of collection activities provided no information on study arm. Collected data were entered in a computerized database and ultimately merged with previous data from the Mayo Lung Project.

Reliability of Lung Cancer Report

Our preliminary data collection activity, i.e., assessment of next-of-kin lung cancer recall, indicated that these reports of lung cancer were highly reliable and that reliability did not differ by study arm. Although we were confident that a similar situation would hold for lung cancer reports obtained during the course of the study, we explored the reliability of these reports further by acquiring a small sample of medical records and death certificates. Although budgetary constraints and known difficulties in obtaining these materials precluded collection of sufficient records and certificates to make the analysis statistically rigorous, we felt that it was important to collect these data.

To explore reliability among living participants, we attempted to collect medical records for 10 living study participants who reported recent lung cancer, although we mailed 19 medical record release forms because we expected limited willingness to release records and difficulty in obtaining them. We were also pessimistic about the ability to obtain medical records for deceased participants, and we instead attempted to obtain death certificates for 20 deceased participants whose next-of-kin reported lung cancer, acknowledging the limitation of this source (that not all persons diagnosed with lung cancer are reported to have died of the disease). We requested 27 death certificates in total to allow for difficulties in obtaining such materials. In both instances, we used the most recently diagnosed lung cancers to improve the likelihood of obtaining records and death certificates. Data collected through reliability activities were not used to change data collected through lung cancer assessment activities.

Statistical Analysis

In our preliminary next-of-kin lung cancer recall study, any next-of-kin who chose “I know for sure he had it” or “I think he had it” was considered to have correctly recalled the lung cancer diagnosis. We examined sensitivity (i.e., percentage of lung cancers correctly reported) in each arm using the Mayo Clinic record as the “gold standard” and examined whether agreement differed by study arm, in a statistically significant manner, by use of a Pearson chi-square test ( 17 ) . Ninety-five percent confidence intervals for sensitivity were calculated using a normal approximation for the binomial distribution ( 18 ) .

The primary goal of this analysis was to determine whether more lung cancers had been diagnosed in participants in the intervention arm than in participants in the usual-care arm. Because the exact year of diagnosis was not available for approximately 16% of reported cases (because of the use of death certificates and information from distant next-of-kin), we used a Pearson chi-square test, as opposed to life-table methods, to examine whether an association existed between study arm and lung cancer diagnosis ( 17 ) . Similar conclusions were obtained with life-table methods, after excluding cases with unknown year of diagnosis.

Because lung cancer status proved to be unavailable for 20% of the Mayo Lung Project cohort, we conducted sensitivity analyses that addressed the expected number of lung cancers among participants with missing lung cancer status, as well as the effect on our conclusions if that expected number was varied. In this analysis, we assumed, for each arm, that the lung cancer proportion among participants with missing lung cancer status was the same as that among participants whose lung cancer status was known. Therefore, the arm-specific lung cancer proportion was calculated by use of a numerator that equaled the number of lung cancers and a denominator that equaled the numerator plus the number of participants who were known not to have lung cancer. After calculating the expected number of additional lung cancers separately for each study arm (by applying the lung cancer proportion to the number of participants with missing lung cancer status and rounding to the nearest digit), we varied the usual-care arm proportion to determine the change necessary to eliminate statistical significance, as well as the change necessary to equalize the number of lung cancers across arms.

Secondary analyses focused on whether other characteristics could explain any observed differences in the number of lung cancers detected. Source and availability of information were examined by comparing distributions, and statistical significance was tested with a Pearson chi-square test ( 17 ) . To examine whether general health, chest imaging use, and cigarette smoking habits differed across arms, we compared the answer distributions. Because possible responses to non–lung cancer questions had an inherent quantitative nature, we first used rank sum (Wilcoxon) tests ( 18 ) to assess whether the median rank varied by study arm. Pearson chi-square tests ( 17 ) also were used to avoid any assumption of a linear relationship across the response categories. Analyses of these variables were restricted to the Participant Questionnaires completed by living participants (as opposed to those completed by next-of-kin) and Next-of-Kin Questionnaires completed by next-of-kin (as opposed to information obtained from death certificates or other sources). Analyses concerning lung cancer reliability did not involve statistical tests because of the small sample size. We used two-sided tests and a statistical significance level of .05 for all statistical tests. SAS statistical software (version 8.2) was used for data analyses ( 19 ) .

R ESULTS

Ability of Next-of-Kin to Report Lung Cancer

One hundred seventeen (88%) of the 133 questionnaires mailed to the next-of-kin of deceased participants known to have had lung cancer diagnosed at the Mayo Clinic between July 1, 1983, and December 31, 1999, were completed and returned ( Table 1 ). The answer choice “I know for sure he had lung cancer” was selected on 98 (84%) of the 117 returned questionnaires and the choice “I think he had lung cancer” was selected on another four (3%), for a sensitivity of 87% (95% confidence interval [CI] = 81.1% to 93.2%). Sensitivity did not differ by study arm (for intervention group, 59/66 = 89%, 95% CI = 82.0% to 96.8%; for control group, 43/51 = 84%, 95% CI = 74.3% to 94.3%; P = .42).

Source of Information

The source of information for the study data is presented in Table 2 . Lung cancer status was available from the Mayo Lung Project database for 2071 participants. Information from study questionnaires was available for 6101 (86%) of the 7118 participants whose lung cancer status could not be determined by use of the Mayo Lung Project database, although information for 811 of the 6101 participants was insufficient to assign lung cancer status. Next-of-kin provided information for 3260 (46%) of the 7118 participants. Information was provided by 1086 participants (15%) and was obtained from Mayo Clinic records for 864 (12%) of the 7118 participants. We obtained information for 858 (12%) of the 7118 participants from death certificates and other sources. Source of information was not meaningfully different across study arms, although the associated P value nearly attained statistical significance ( P = .059, Pearson chi-squared test), perhaps because of the large sample size. Follow-up through December 31, 1999, was available for 8172 participants (89%) of the entire Mayo Lung Project cohort of 9189 participants, and availability did not vary by study arm ( P = .201). Information on lung cancer status was available for 7361 (80%) of the cohort of 9189 participants, and again availability did not vary by study arm ( P = .094).

Lung Cancer Diagnoses

Extended incidence follow-up results are found in Table 3 . We identified another 719 lung cancers among the 9189 participants of the Mayo Lung Project during the extended follow-up (379 in the intervention arm and 340 in the usual-care arm). Thus, there were a total number of 1085 lung cancers in the cohort (585 in the intervention arm and 500 in the usual-care arm) ( P = .009). Sixty-eight percent in each study arm (3140 participants in the intervention arm and 3136 participants in the usual-care arm) had sufficient information to be classified as having no lung cancer, and lung cancer status for the remaining 20%—19% (885 participants) in the intervention arm and 21% (943 participants) in the usual-care arm—could not be determined, primarily because of refusal, loss to follow-up, and limitations of death certificate information. Among the 3725 intervention arm and 3636 usual-care arm participants with known lung cancer status, roughly 16% and 14%, respectively, had a diagnosis of lung cancer ( P = .018). The source of lung cancer report varied somewhat across study arms, with the most pronounced absolute difference occurring for reports by next-of-kin (next-of-kin reported 43% of lung cancers in the intervention arm versus 54% in the usual-care arm).

Sensitivity Analyses

If the arm-specific proportions of lung cancer did not vary by missing lung cancer status, approximately 139 lung cancers ([585/3725] × 885) were missed in the intervention arm, for a total of 724 (15.7%) people developing lung cancer among 4610 participants. By use of the same approach, approximately 130 ([500/3636] × 943) were missed in the usual-care arm, for a total of 630 (13.8%) people developing lung cancer among 4579 participants ( P = .009). To eliminate the statistically significant difference between the two arms, an additional 23 missed cases (over and above the 130 expected missed cases) would be necessary in the usual-care arm (i.e., 153 missed cases in total). In other words, the proportion of participants developing lung cancer among usual-care participants with unknown or inconclusive status would have to be approximately 18% higher than that found among usual-care arm participants with known lung cancer status (153/943 versus 500/3636). To equalize the absolute number of people developing lung cancer, a total of 224 (724 expected in the intervention arm − 500 observed in the usual-care arm) missing lung cancers in the usual-care arm would be necessary. The total of 224 is 94 over and above the 130 expected missed cases in the usual-care arm and would require a lung cancer proportion about 75% greater than that found among usual-care arm participants with known lung cancer status (224/943 versus 500/3636).

Questions Other than Lung Cancer Status

Table 4 presents responses to non–lung cancer questions asked on the surveys. Reports of general health, chest imaging, and smoking behavior among living participants did not vary by study arm; previous smoking behavior of deceased participants, as reported by next-of-kin, did not vary by study arm as well. All P values, regardless of statistical test used (rank sum or Pearson chi-square test) were greater than .10, with more than half being greater than .50.

Reliability of Lung Cancer Report

Of the 19 living participants selected for the lung cancer report reliability substudy, nine gave permission for us to review their medical records. We requested medical records for all nine participants but received records for only six. These records confirmed lung cancers in all six. Of the 27 deceased participants selected for this reliability substudy, we received death certificates for all but one. Of the 26 deceased participants, 18 (69%) death certificates noted lung cancer. That percentage varied somewhat by study arm (57% among 14 intervention arm participants versus 83% among 12 usual-care arm participants), although the difference was not statistically significant ( P = .15).

D ISCUSSION

Extended follow-up from July 1, 1983, through December 31, 1999, for lung cancer incidence among Mayo Lung Project participants has resulted in the identification of an additional 719 lung cancers, for a total of 1085 lung cancers diagnosed in the cohort from November 1971 through December, 31, 1999. These cancers were not equally distributed by study arm: 585 were diagnosed in the intervention arm and 500 in the usual-care arm. The statistically significant excess of 85 cancers in the intervention arm compared with the usual-care arm provides continued support for the existence of overdiagnosis in lung cancer screening.

To conclude from these data that overdiagnosis exists implies that other explanations for the continued case excess can be dismissed. For example, an excess could have resulted if randomization failed to produce two study arms with similar profiles of lung cancer risk factors ( 20 ) or profiles, that were similar at randomization, changed over time. The first possibility has been examined ( 15 ) , and the results, in concert with theoretical aspects of randomization ( 21 ) , did not support such a possibility ( 2 , 6 , 15 , 22 , 23 ) . Regarding the second possibility, data in Table 4 do not indicate that smoking behavior differed markedly across study arms as time passed. Screening utilization after the close of the Mayo Lung Project also could have produced an excess if intervention arm participants received more screening examinations than participants in the usual-care arm throughout the follow-up period and beyond. However, data concerning time since last chest x-ray and chest imaging in the period from 1999 through 2001 suggest that recent patterns were quite similar in the two arms ( Table 4 ).

Similar modest rates of reporting error in each arm should not have affected results appreciably or differentially, although an excess of cases in the intervention arm could have occurred if the report of a lung cancer, true or erroneous, were much more likely to occur in that arm. We found little evidence for that possibility in the next-of-kin and medical record substudies, although the latter was small and diagnoses in the usual-care arm were under-represented. Agreement with death certificates varied somewhat by study arm, but death certificates are not a reliable or consistent source for past lung cancer diagnoses. An excess of cases in the intervention arm also could occur in certain situations if lung cancer rates among participants with missing lung cancers status differed from rates among participants with known status and did so differentially. One calculation in our sensitivity analysis suggested that the rate would need to be much higher (approximately 75% higher) among participants in the usual-care arm with missing lung cancer status compared with known status to equalize the number of cases across arms. However, the other calculation in our sensitivity analysis revealed that a less pronounced increase (18%) and an accompanying small number of missing cases over and above what would be expected if the lung cancer rate did not vary by missing status (23 and 130, respectively) would have been necessary to eliminate statistical significance. We are hesitant to draw conclusions from these calculations, however, given the assumptions and minimal information on which they were based.

The existence of overdiagnosis in the Mayo Lung Project is supported by the findings of other analyses of data for the cohort. For example, results of extended lung cancer mortality follow-up through 1996 ( 6 ) indicated no reduction in lung cancer mortality but statistically significantly longer survival for patients in the intervention arm with lung cancers diagnosed through July 1, 1983. Because a true improvement in case survival must be accompanied by a reduction in mortality, it was concluded ( 6 ) that screening biases were responsible for the discordance and that overdiagnosis was more likely than lead-time bias or length bias. Expert pathology review of 105 Mayo Lung Project lung cancers ( 24 ) —77 from the intervention arm and 28 from the usual-care arm—also supported the occurrence of overdiagnosis, as all lesions were confirmed histologically to be carcinoma. That finding helped to refute the possibility that the excess cases had been misdiagnosed as lung cancer, even though seven (7%) of the 105 pathology samples were classified unanimously by members of the review panel as carcinoma in situ. Although the authors note that carcinoma in situ might therefore explain the excess intervention arm cases, we hesitate to extrapolate that percentage to all Mayo Lung Project lung cancers because the 105 specimens were not chosen at random; furthermore, early resected disease was overrepresented. If extrapolation were appropriate, though, only approximately 41 cases—half of the excess—actually might have been preinvasive disease. Finally, a project that estimated tumor doubling times for 44 Mayo Lung Project lung cancers ( 25 ) supports the existence of indolent lung tumors as one source of overdiagnosed cases, as the authors calculated that two tumors had doubling times of more than 300 days, more than twice that of the median doubling time (i.e., 144 days).

Strong evidence of overdiagnosis in lung cancer screening also is provided by a Czechoslovakian randomized clinical trial ( 26 ) . In that trial participants were randomly assigned to an intervention arm that received semiannual screening for 3 years or a control arm that received screening during the third year only. Intervention and control arm participants then received annual chest x-ray for 3 years after the initial 3-year study period. Although no difference in mortality was observed between the arms at 6 years, there were 108 cases in the intervention arm and 82 cases in the control arm. This trial, although not a stop-screen trial, should have experienced an equalization of cases across arms if overdiagnosis did not exist because the screening regimens ultimately became the same. That equalization never occurred, however ( 26 ) . Furthermore, the majority of the excess occurred during the first period of the study—the period in which the control arm did not receive screening as part of the study.

It is perplexing that the excess number of lung cancer cases in the Mayo Lung Project cohort increased over time—from 46 cases at the close of the project in 1983 to 85 at the end of 1999 ( Table 5 ). Radiation exposure occurring as part of the intense screening regimen could have increased lung cancer risk years after the exposure because the latency period of radiation-induced cancer is long, but such an increase could explain only a small part of difference, if any. Recent estimates of lung cancer risk elevation accompanying long-term low-radiation-dose spiral computed tomography screening, which confers an effective radiation dose greater than chest x-ray ( 27 – 29 ) , indicate that risk elevation is unlikely to be greater than approximately 5% ( 30 ) .

Another possibility—i.e., that after the close of the Project, participants in the intervention arm were statistically significantly more likely to receive chest x-ray than participants in the usual-care arm—was not supported by the data available on chest imaging ( Table 4 ). Responses concerning use of chest x-ray in the years 1999, 2000, and 2001 suggest that recent imaging was similar across arms, but they do not speak to imaging in the late 1970s and early 1980s. The excess could have grown if intervention arm participants continued their screening regimens after their study examinations ceased. That scenario may be supported by the data in Table 5 , because growth of the excess was most pronounced in the first half of the 1980s. But our data suggest that imaging ultimately became similar in the two arms, which should have resulted in catch-up if overdiagnosis had not occurred in the cohort. Another possible explanation is that, for lung cancers identified through sources other than the participant or next-of-kin, these sources may have been more likely to provide information for participants in the intervention arm with lung cancer than for those in the usual-care arm. Mayo Clinic records identified 30 more lung cancers and death certificates identified 18 more cancers among participants in the intervention arm than among participants in the usual-care arm. It is possible that more intervention arm participants continued their health care at Mayo and that death certificates reports of lung cancer reflect sticky diagnosis bias, the phenomenon whereby accuracy of cause of death on the death certificate is affected by previous cancer diagnoses ( 2 ) .

We have alluded to limitations of our study and their likely impact, but it is also worthwhile to enumerate those limitations. First, the majority of our data concerning lung cancer status came from self-report. Self-report can be less reliable than medical records, and reliability is probably further lacking because much information was provided by next-of-kin and some information was more than 20 years old. It was necessary to use more than one source to evaluate lung cancer status, and, given that yield of certain sources varied by study arm, the possibility of differential ascertainment of lung cancers is real. Some sources, such as death certificates, could not be used to rule out lung cancer, leaving approximately 25% of the cohort with an inconclusive or unknown lung cancer status. Despite these limitations, we feel the data are informative; furthermore, we believe that the data were collected in the best manner possible given the difficulty of the question at hand.

It has been suggested that overdiagnosis in lung cancer—in particular, screen-detected nonlethal lung cancers—is biologically implausible because of the typically short survival of lung cancer patients ( 20 ) . Such statistics are based overwhelmingly on cancers that present symptomatically rather than through screen detection, however. Incidental findings of lung cancer on autopsy ( 31 – 33 ) and examination of lung tissue removed to improve pulmonary function ( 34 ) also support the existence of indolent lesions. Indolent lesions almost certainly exist in prostate ( 35 ) and breast ( 36 ) cancer, with much of the evidence of their existence coming to light as a result of screening programs. The concept of indolent lesions is consistent with current theories of carcinogenesis—a process with periods of quiescence, regression, and progression, but not necessarily with all three ( 37 ) .

Although the magnitude of overdiagnosis in chest x-ray screening appears to be modest, the very real and deleterious role that overdiagnosis plays in mass screening cannot be discounted. The newest imaging technologies can detect very small lung abnormalities, but these abnormalities may be clinically unimportant. The question thus remains as to whether early detection of lung cancer through mass screening results in a net benefit to the public's health.

References