Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013 A causal relationship between cigarette smoking and lung cancer was first suspected in the 1920s on the basis of clinical observations. To test this apparent association, numerous epidemiologic studies were undertaken between 1930 and 1960. Two studies were conducted by Richard Doll and Austin Bradford Hill in Great Britain. The first was a case-control study begun in 1947 comparing the smoking habits of lung cancer patients with the smoking habits of other patients. The second was a cohort study begun in 1951 recording causes of death among British physicians in relation to smoking habits. This case study deals first with the case-control study, then with the cohort study. Data for the case-control study were obtained from hospitalized patients in London and vicinity over a 4-year period (April 1948 - February 1952). Initially, 20 hospitals, and later more, were asked to notify the investigators of all patients admitted with a new diagnosis of lung cancer. These patients were then interviewed concerning smoking habits, as were controls selected from patients with other disorders (primarily non-malignant) who were hospitalized in the same hospitals at the same time. Data for the cohort study were obtained from the population of all physicians listed in the British Medical Register who resided in England and Wales as of October 1951. Information about present and past smoking habits was obtained by questionnaire. Information about lung cancer came from death certificates and other mortality data recorded during ensuing years. Question 1: What makes the first study a case-control study? The first study is a case-control study.It meets the definition of a case-control study as defined by Friis and Sellers: an analytic study of persons diagnosed with the disease of interest and a suitable control group of patients without the disease (Friis and Sellers, 2014). In this case, this was study of persons diagnosed with lung cancer (with the disease of interest) and patients of other disorders who were hospitalized at the same time (a suitable control group of patients without the disease). References: Friis R.H. and Sellers T. A. (2014) Epidemiology for Public Health Practice, Fifth Edition. Jones and Bartlett Learning, LLC. Page 303. Question 2: What makes the second study a cohort study? The second study is a cohort study. It meets the definition of a cohort study as defined by Friis and Selers: “a population group or subset therof (distinguished by a common characteristic), that is followed over a period of time” (Friis and Sellers, page 325, 2014). In this case , this was a study of the past smoking habits of persons who resided in England and Whales (population) as of October 1951 and died during the ensuing years (period of time). References:
Page 1 of 12
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013 Friis R.H. and Sellers T. A. (2014) Epidemiology for Public Health Practice, Fifth Edition. Jones and Bartlett Learning, LLC. Page 325. The remainder of Part I deals with the case-control study. Question 3: Why might hospitals have been chosen as the setting for this study? According to Friis and Sellers, a preferred approach is to conduct case-control studies in which cases and controls are population based. They explain, when the selection of population-based study groups are not feasible, cases may need to be derived from one or more major hospitals. They also explain that although hospital-based studies are “inherently subject to greater potential for errors than population based studies, their use is certainly justified when little information has be reported about a particular exposure-disease association or when a population based case registry is not available” (Friis and Sellers, page 308, 2014). In this case, there was little information and no population based case registry available, so the use of the hospital was justified. References: Friis R.H. and Sellers T. A. (2014) Epidemiology for Public Health Practice, Fifth Edition. Jones and Bartlett Learning, LLC. Page 308. Question 4: What other sources of cases and controls might have been used? According to Friis and Sellers, the ideal situation is to identify and enroll all incident cases in a defined population in a specified time period is using a disease registry or a complete listing of all available cases from a source such as the Vital Statistics Bureau. They also explain that the best way to ensure that the distribution of exposure among the controls represents the exposure levels in the population is to select population-based controls. A suggested method to identify such controls is to obtain a list that contains names and addresses of most residents in the same geographic area as the cases (Friis and Sellers, 2014) References: Friis R.H. and Sellers T. A. (2014) Epidemiology for Public Health Practice, Fifth Edition. Jones and Bartlett Learning, LLC. Page 305-307. Question 5: What are the advantages of selecting controls from the same hospitals as cases? The advantages to selecting controls from the same hospitals and cases include that the cases and controls are from the same geographic area. Additionally, Friis and Sellers explains that other advantages to selecting controls from the same hospital are that “the study personnel who are already in the hospital to interview cases may achieve time efficiency by also interviewing controls. The time saving, plus the fact that hospital controls may be more likely than population controls to participate, ultimately equates to cost savings” (Friis and Sellers, pages 308-309, 2014). Page 2 of 12
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013 References: Friis R.H. and Sellers T. A. (2014) Epidemiology for Public Health Practice, Fifth Edition. Jones and Bartlett Learning, LLC. Page 308-309. Question 6: How representative of all persons with lung cancer are hospitalized patients with lung cancer? At the time of the study using hospitalized patients with lung cancer was probably fairly representative of all persons with lung cancer, as outpatient treatments and diagnostic methods were not available. Currently, there is more advanced knowledge of cancer including knowledge about the differing stages of the disease that may not require hospitalization. Advancements in outpatient treatment coupled with early identification may exclude those diagnosed with the disease from being represented by the cases making the sample not very representative of all persons with lung cancer. Question 7: How representative of the general population without lung cancer are hospitalized patients without lung cancer? Not very, as persons who are hospitalized generally have poorer health conditions than those of the general population. They may also engage in more risk-taking behavior or high-risk activities such as smoking, drinking, drug use, or working in situations where exposure to other carcinogens may increase their propensity to develop lung cancer. Generally, people who engage in healthier activities stay out of hospitals. Question 8: How may these representativeness issues affect interpretation of the study's results? Friis and Sellers explain that one of the major limitations of case-control studies is indeterminate representativeness of the cases and controls. If the cases and controls are not representative of the population, the results of the study will be biased. This would be a case of selection bias. Selection bias occurs when the relation between exposure and disease is different from those participating and those who would be theoretically eligible to participate but do not (Friis and Sellers, 2014). References Friis R.H. and Sellers T. A. (2014) Epidemiology for Public Health Practice, Fifth Edition. Jones and Bartlett Learning, LLC. Pages 316, 442. Over 1,700 patients with lung cancer, all under age 75, were eligible for the case-control study. About 15% of these persons were not interviewed because of death, discharge, severity of illness, or inability to speak English. An additional group of patients were interviewed but later excluded when initial lung cancer diagnosis proved mistaken. The final study group included 1,465 cases (1,357 males and 108 females). The following table shows the relationship between cigarette smoking and lung cancer among male cases and controls.
Page 3 of 12
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013 Table 1. Smoking status before onset of the present illness, lung cancer cases and matched controls with other diseases, Great Britain, 1948-1952.
Cigarette smoker Non-smoker Total
Cases
Controls
1350
1296
7 1357
61 1357
Question 9: From this table, calculate the proportion of cases and controls who smoked.
Cigarette smoker Non-smoker Total 1350/1357 = Proportion Smoked =
Cases 1350 7 1357 0.9948 99.5%
Controls 1296 61 1357 0.9550 95.5%
Proportion smoked, cases: 99.5% Proportion smoked, controls: 95.5% Question 10: What do you infer from these proportions? Smoking is a behavior that occurs in more than 95% of both the cases and controls. Question 11a: Calculate the odds of smoking among the cases.
Cigarette smoker Non-smoker Total 1350/1357 = Proportion Smoked = smokers/non-smokers Odds
Cases 1350 7 1357 0.9948 99.5% 192.9 192.9/1
Controls 1296 61 1357 0.9550 95.5% 21.2 21.2/1
The odds of smoking among the cases is 192.9/1 Page 4 of 12
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013 Question 11b: Calculate the odds of smoking among the controls.
Cigarette smoker Non-smoker Total 1350/1357 = Proportion Smoked = smokers/non-smokers Odds
Cases 1350 7 1357 0.9948 99.5% 192.9 192.9/1
Controls 1296 61 1357 0.9550 95.5% 21.2 21.2/1
The odds of smoking among the controls is 21.2/1 Question 12: Calculate the ratio of these odds. How does this compare with the cross-product ratio?
Cigarette smoker Non-smoker Total 1350/1357 = Proportion Smoked = smokers/non-smokers Odds [(smokers/nonsmokers)cases]/(smokers/nonsmokers)(controls) Odds Ratio (OR) The Odds Ratio is 9.1
Cases 1350 7 1357 0.9948 99.5% 192.9 192.9/1
Controls 1296 61 1357 0.9550 95.5% 21.2 21.2/1
9.1 9.1
Question 13: What do you infer from the odds ratio about the relationship between smoking and lung cancer? The Odds Ratio (OR) is 9.1 and according to Friis and Sellers “the OR literally measures the odds for exposure to a given disease” (Friis and Sellers, page 310, 2014). In this case the OR is saying that persons who smoke are 9X more likely to develop lung cancer than those who do not smoke References
Page 5 of 12
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013 Friis R.H. and Sellers T. A. (2014) Epidemiology for Public Health Practice, Fifth Edition. Jones and Bartlett Learning, LLC. Page 310. Table 2 shows the frequency distribution of male cases and controls by average number of cigarettes smoked per day. Table 2. Most recent amount of cigarettes smoked daily before onset of the present illness, lung cancer cases and matched controls with other diseases, Great Britain, 1948-1952.
Daily Number of Cigarettes
Number of Cases
Number of Controls
0 1-14 15-24 25+ All smokers
7 565 445 340 1350
61 706 408 182 1296
Odds Ratio referent
Question 14: Compute the odds ratio by category of daily cigarette consumption, comparing each smoking category to nonsmokers.
Daily Number of Cigarettes
Number of Cases
Number of Controls
Odds Ratio
0 1-14 15-24 25+ All smokers
7 565 445 340 1350
61 706 408 182 1296
referent 7.0 9.5 16.3 9.1
smokers (cases) * non-smokers (controls) A 34465 27145 20740 82350
smokers (controls) * nonsmokers (cases) B 4942 2856 1274 9072
The Odds Ratio for 1-14 cigarettes is 7.0; 15-24 cigarettes is 9.5; 25+ cigarettes is 16.3; and all smokers is 9.1. Question 15: Interpret these results. These results could be interpreted to mean that the odds of developing lung cancer increases with the number of cigarettes a person uses daily. Although the study demonstrates a clear association between smoking and lung cancer, cause-andeffect is not the only explanation. Page 6 of 12
A/B
7.0 9.5 16.3 9.1
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013 Question 16: What are the other possible explanations for the apparent association? There could be simply a correlation but not causation; chance; bias including selection, information, or confounding, or investigator error. The next section of this case study deals with the cohort study. Data for the cohort study were obtained from the population of all physicians listed in the British Medical Register who resided in England and Wales as of October 1951. Questionnaires were mailed in October 1951, to 59,600 physicians. The questionnaire asked the physicians to classify themselves into one of three categories: 1) current smoker, 2) ex-smoker, or 3) nonsmoker. Smokers and ex-smokers were asked the amount they smoked, their method of smoking, the age they started to smoke, and, if they had stopped smoking, how long it had been since they last smoked. Nonsmokers were defined as persons who had never consistently smoked as much as one cigarette a day for as long as one year. Usable responses to the questionnaire were received from 40,637 (68%) physicians, of whom 34,445 were males and 6,192 were females. Question 17: How might the response rate of 68% affect the study's results? This response rate could suggest selection bias and skew the results of the study. The next section of this case study is limited to the analysis of male physician respondents, 35 years of age or older. The occurrence of lung cancer in physicians responding to the questionnaire was documented over a 10-year period (November 1951 through October 1961) from death certificates filed with the Registrar General of the United Kingdom and from lists of physician deaths provided by the British Medical Association. All certificates indicating that the decedent was a physician were abstracted. For each death attributed to lung cancer, medical records were reviewed to confirm the diagnosis. Diagnoses of lung cancer were based on the best evidence available; about 70% were from biopsy, autopsy, or sputum cytology (combined with bronchoscopy or X-ray evidence); 29% were from cytology, bronchoscopy, or X-ray alone; and only 1% were from just case history, physical examination, or death certificate. Of 4,597 deaths in the cohort over the 10-year period, 157 were reported to have been caused by lung cancer; in 4 of the 157 cases this diagnosis could not be documented, leaving 153 confirmed deaths from lung cancer. The following table shows numbers of lung cancer deaths by daily number of cigarettes smoked at the time of the 1951 questionnaire (for male physicians who were nonsmokers and current smokers only). Person-years of observation ("person-years at risk") are given for each smoking category. The number of cigarettes smoked was available for 136 of the persons who died from lung cancer.
Table 3. Number and rate (per 1,000 person-years) of lung cancer deaths by number of cigarettes smoked per day, Doll and Hill physician cohort study, Great Britain, 1951-1961.
Page 7 of 12
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013
Daily Number of Cigarettes 0 1-14 15-24 25+ All smokers Total
Deaths 3 22 54 57 133 136
Person-years at Risk 42800 38600 38900 25100 102600 145400
Mortality Rate per 1,000 personyears 0.07
Rate Ratio referent
Rate difference per 1000 personyears referent
Question 18: Compute lung cancer mortality rates, rate ratios, and rate differences for each smoking category. What do each of these measures mean?
Daily Number of Cigarettes 0 1-14 15-24 25+ All smokers Total
Deaths 3 22 54 57 133 136
Person-years at Risk 42800 38600 38900 25100 102600 145400
Mortality Rate per 1,000 personyears 0.07 0.57 1.39 2.25 1.30 0.94
Rate Ratio referent 8.10 19.80 32.40 18.60
Rate difference per 1000 personyears referent 0.50 1.32 2.20 1.23
These results could be interpreted to mean that mortality rates increase with the number of cigarettes a person uses daily. Question 19: What proportion of lung cancer deaths among all smokers can be attributed to smoking? What is this proportion called? The proportion of lung cancer deaths among all smokers that can be attributed to smoking is 94.6%. This proportion is called Absolute Risk.
Page 8 of 12
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013 Absolute Risk = [(1.30/0.07)/1.30]*100 = 94.6
Rate difference Daily Number of Person-years at Mortality Rate per 1,000 per 1000 personCigarettes Deaths Risk person-years Rate Ratio years 0 3 42800 0.07 referent referent 1-14 22 38600 0.57 8.10 0.50 15-24 54 38900 1.39 19.80 1.32 25+ 57 25100 2.25 32.40 2.20 All smokers 133 102600 1.30 18.60 1.23 Total 136 145400 0.94 Question 20: If no one had smoked, how many deaths from lung cancer would have been averted? Based on the proportion in the previous question, 95 percent of deaths from lung cancer can be attributed to smoking. Using this information, approximately 126 deaths of the 133 could have been averted if no one had smoked. The cohort study also provided mortality rates for cardiovascular disease among smokers and nonsmokers. The following table presents lung cancer mortality data and comparable cardiovascular disease mortality data.
Table 4. Mortality rates (per 1,000 person-years), rate ratios, and excess deaths from lung cancer and cardiovascular disease by smoking status, Doll and Hill physician cohort study, Great Britain, 1951-1961.
Smokers Lung Cancer Cardiovascular Disease
Mortality rate per 1,000 person-years Excess deaths Non-smokers All Rate Ratio per 1000 personyears
Attributable risk percent among smokers
1.3
0.07
0.94
18.5
1.23
95%
9.5
7.32
8.87
1.3
2.19
23%
Question 21: Which cause of death has a stronger association with smoking? Why? Deaths resulting from lung cancer have a stronger association with smoking. The rate ratio demonstrates that there is a higher association between smoking and lung cancer at 18.5 than there is between smoking and cardiovascular disease at 1.3. In calculating the attributable risk percent, the excess lung cancer deaths attributable to smoking is expressed as a percentage of all lung cancer mortality among all smokers. The attributable risk Page 9 of 12
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013 percent of 95% for smoking may be interpreted as the proportion of lung cancer deaths among smokers that could have been prevented if they had not smoked. A similar measure, the population attributable risk ercent expresses the excess lung cancer deaths attributable to smoking as a percentage of all lung cancer mortality among the entire population. From a prevention perspective, the population attributable risk percent for a given exposure can be interpreted as the proportion of cases in the entire population that would be prevented if the exposure had not occurred. The population attributable risk percent is often used in assessing the cost-effectiveness and costbenefit of community-based intervention programs. One formula for the population attributable risk percent is: PAR% = (Incidence in entire population * Incidence in unexposed) / Incidence in entire population. Question 22: Calculate the population attributable risk percent for lung cancer mortality and for cardiovascular disease mortality. How do they compare? How do they differ from the attributable risk percent?
Smokers
Mortality rate per 1,000 person-years Excess deaths Non-smokers All Rate Ratio per 1000 personyears
Lung Cancer 1.3 0.07 Cardiovascular 9.5 7.32 Disease PAR = (All *Non-smokers)/All*100
Attributable risk percent among smokers
0.94
18.5
1.23
95%
8.87
1.3
2.19
23%
PAR for Lung Cancer is 92.5% PAR for Cardiovascular disease is 17.4% The compare in that death as a result of lung cancer and cardiovascular disease can be attributed to smoking. They differ in that a much higher percentage of deaths from lung cancer (92.5 percent) can be attributed to smoking than they can be for cardiovascular disease (at 17.4 percent). Question 23: How many lung cancer deaths per 1,000 persons per year are attributable to smoking among the entire population? How many cardiovascular disease deaths? Lung Cancer: 0.87 deaths per 1,000 person-years (PAR*All) Cardiovascular Disease: 1.54 deaths per 1.000 person-years (PAR*All) The following table shows the relationship between smoking and lung cancer mortality in terms of the effects of stopping smoking.
Table 5. Number and rate (per 1,000 person-years) of lung cancer deaths for current smokers and exsmokers by years since quitting, Doll and Hill physician cohort study, Great Britain, 1951-1961.
Page 10 of 12
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013
Cigarette Smoking Status
Lung Cancer Deaths
Rate per 1000 personyears
Rate Ratio
Current smokers
133
1.3
18.5
For ex-smokers, years since quitting < 5 years 5-9 10-19 20+
5 7 3 2
0.67 0.49 0.18 0.19
9.6 7 2.6 2.7
non-smokers
3
0.07
1
Question 24: What do these data imply for the practice of public health and preventive medicine? The data implies that smokers are the highest risk group for lung cancer and while quitting reduces the risk, those who never smoke are in the lowest risk group. Public health initiatives and preventive medicine efforts to prevent people from smoking would have the most benefit. Additionally, educating those who smoke to quit smoking will help to reduce deaths from lung cancer. As noted at the beginning of this case study, Doll and Hill began their case-control study in 1947. They began their cohort study in 1951. The odds ratios and rate ratios from the two studies by numbers of cigarettes smoked are given in the table below. Table 6. Comparison of measures of association from Doll and Hill’s 1948-1952 case-control study and Doll and Hill’s 1951-1961 physician cohort study, by number of cigarettes smoked daily, Great Britain. Daily number of Cigarettes Smoked
Rate Ratio from Cohort Study
0 1-14 15-24 25+ All Smokers
1.0 (ref) 8.1 19.8 32.4 18.5
Odds Ratio from case-control study 1.0 (ref) 7 9.5 16.3 9.1
Question 25: Compare the results of the two studies. Comment on the similarities and differences in the computed measures of association. Page 11 of 12
Jill Marshall MPH 510 – Applied Epidemiology Case Study – Cigarette Smoking and Lung Cancer June 8, 2013 Both studies demonstrate that lung cancer deaths are associated with smoking. Both studies also demonstrate that the risk of a lung cancer death increases with the daily number of cigarettes smoked. The difference is in comparing the rate ratio to the odds ratio. The odds ratios from the case-control are not as strong (high of a number) as the rate ratios from the cohort study. Question 26: What are the advantages and disadvantages of case-control vs. cohort studies? Answer 26 Sample size Costs Study time
Case-control smaller/ disadvantage inexpensive/advantage quick/advantage
Cohort larger/advantage expensive/disadvantage lengthy/disadvantage
Rare disease Rare exposure Multiple exposures Multiple outcomes
good for this/advantage not good for this/disadvantage good for this/advantage not good for this/disadvantage
not good for this/disadvantage good for this/advantage not good for this/disadvantage good for this/advantage
Progression, spectrum of illness Disease rates
not good for this/disadvantage not good for this/disadvantage
good for this/advantage good for this/advantage
Recall bias
potential for bias/disadvantage
Loss to follow-up Selection bias
less potential for loss/advantage potential for bias/disadvantage
less potential for bias/advantage potential for loss/disadvantage less potential for bias/advantage
Question 27: Which type of study (cohort or case-control) would you have done first? Why? Why do a second study? Why do the other type of study? I would have conducted a case-control study first especially if there were an accessible disease registry available. A case-control study is cost effective and less time consuming than a cohort study. I might do a cohort study if there is evidence to support the findings from the case-control study. Question 28: Which of the following criteria for causality are met by the evidence presented from these two studies? Answer 28 Strong association Consistency among studies Exposure precedes disease Dose-response effect Biologic plausibility
Page 12 of 12
Yes X X X X
No
X
