HOW MUCH DOES SCREENING REDUCE CANCER ... - medicine.mcgill.caReprints... · Screening & Prostate-Ca Mortality in Randomized European Study ’92-’08 (“ERSPC” nejm2009.04) As

HOW MUCH DOES SCREENING REDUCE CANCER MORTALITY?

James A. Hanley1, Zhihui (Amy) Liu1,2, Nandini Dendukuri1,3, Erin Strumpf1,4

1Dept. of Epidemiology, Biostatistics & Occupational Health2Cancer Care Ontario

3Dept. of Medicine, and Technology Assessment Unit4Dept. of Economics

McGill University

51ST ANNUAL ANDRÉ AISENSTADT MEMORIAL CLINICAL DAYCANCER SCREENING – UPDATE 2014

A Symposium in Honour of Dr. André LisbonaJewish General Hospital

October 22, 2014

Summary

• Harms have been (well) measured; benefits have been mis-measured

• By ignoring the delay until the reductions in mortality are expressed, theprevailing interpretations of the results of cancer screening trialsunder-estimate the mortality reductions that would be produced by asustained screening program

• P-value-driven RCT stopping/reporting rules exacerbate the problem

• Ways we might be able to avoid such misleading estimates

• Lung, Prostate, Colon: re-analysis of data from trials

• Breast : data from outdated trials population-screening

Outline

• Why do so many trials yield a 20% ‘mortality reduction’ ? [Theorem]

• The mortality reductions produced by a cancer screening program

• A way ahead? (impact of N-round program:∑i=N

i=1 impact of roundi )

• Illustrations: cancer of the prostate, breast, colon

• Comments: cancer of the breast

20% MORTALITY REDUCTION

A UNIVERSAL CONSTANT IN CANCER SCREENING TRIALS?

For many RCTs,single rate (hazard) ratio or risk difference is OK

• A single (overall) Rate Reduction (i.e., single Rate Ratio),based on all events that have occurred (regardless ofwhen) up to end of available follow-up time on each subject

• ‘Regardless of when’ implies proportional hazards, i.e.,reduction is immediate & sustained (if need be, bycontinuing to take medications)

• Numbers of events matter, but not their timing:Q: how to have sufficient events for desired precision?

more persons, less time? ↔ more time, fewer persons?• As amount of person time (number of events) increases,

updated single Rate Reduction traces out a random walk

Reductions in ‘event rates’ as follow-up time unfolds

P = 0.05

50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800(Cumulative) Number of Events

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

Percentage Reduction in Average Event Rate, if data are analyzed after indicated no. of events

Examples of ‘prevention’ / ‘early detection’ studies

HIV: if ‘intervention’ ineffective

P = 0.05


0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%


---> Time 0

0.5

1 [ref.]RateRatio

HIV: Adult circumcision

P = 0.05


0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%


---> Time 0

0.5

1 [ref.]RateRatio

Paralytic or non-paralytic poliomyelitis: Salk Vaccine

P = 0.05


0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%


2) Paralytic/non-paralytic poliomyelitis[Salk Vaccine]

HPV6,11,16,18 infection: Quadrivalent HPV Vaccine

P = 0.05


0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%



1) Human Papilloma Virus (HPV) infection[Quadrivalent HPV vaccine]

Death from ruptured abdominal aneurym: Ultrasound screening

P = 0.05


0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%




3) Death from ruptured abdominal aortic aneurysm[Ultrasound screening]

Cancer Screening Trial - theoretical

P = 0.05


0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%


[0%]

[2%]

[14%]

[35%][50%]

[50%]

[50%]

7)

---> Time 0

0.5

1 [ref.]RateRatio

3 actual cancer screening trials

P = 0.05


0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%




3) Death from ruptured abdominal aortic aneurysm[Ultrasound screening]

5) COLON

4) PROSTATE

6) LUNG

What payers would like to know about a PROGRAM

(a) Age-specific numbers of prostate cancer deaths in a steady state population with a given age-structure, if screening had not been available, and if screening had been available from ages 50 to 70

(b) The corresponding age-specific prostate cancer mortality rate ratios

Deaths in absence of screening

Deaths despite screening

Deaths averted by screening

Population

No.

pro

stat

e ca

ncer

dea

ths

per 1

-yea

r age

-ban

d

0

5

10

15

20

25

30

35

0K

5K

10K

15K

20K

25K

30K

35K

40K

45K

Population per 1-year age-band

50 55 60 65 70 75 80 85 Age

Screening

0 100%0.25 75%0.5 50%0.75 25%1 0%

No Screening

ScreeningMortalityRateRatio

MortalityReduction (%)

50 55 60 65 70 75 80 85 Age

WebFigure 2. Age-specific numbers of prostate cancer deaths and prostate cancer mortality rate ratios.Age-specific numbers from Quebec in the early 1990s are used to represent the (steady-state) annual numbers of prostate cancer deaths in the absence of screening.The numbers of annual deaths that there would have been in these same population had a screening program been available [from when men reach the age of 50 untilthey turn 70] are hypothetical. Note that these two sets of numbers are age-specific, not cumulative – they decrease if the age range is extended past 85 – and merelyreflect the exponential rise in prostate cancer death rates with age. The rate ratio graph in panel (b) is modeled after Figure 2-5(b) in Morrison and is designed to illustrate (from left to right) its three features: the time-lag until the deathsaverted by screening become apparent, the 20 years of full benefit that follow – after this lag -- the 20 years of screening, and the disappearance of the effect (i.e., areversion to late-age mortality rates in the unscreened scenario) at some point after the last age at which men are screened.

or (b) the Rate Ratio (or %Reduction) Function ...

(a) Age-specific numbers of prostate cancer deaths in a steady state population with a given age-structure, if screening had not been available, and if screening had been available from ages 50 to 70

(b) The corresponding age-specific prostate cancer mortality rate ratios

Deaths in absence of screening

Deaths despite screening

Deaths averted by screening

Population

No.

pro

stat

e ca

ncer

dea

ths

per 1

-yea

r age

-ban

d

0

5

10

15

20

25

30

35

0K

5K

10K

15K

20K

25K

30K

35K

40K

45K

Population per 1-year age-band

50 55 60 65 70 75 80 85 Age

Screening

0 100%0.25 75%0.5 50%0.75 25%1 0%

No Screening

ScreeningMortalityRateRatio

MortalityReduction (%)

50 55 60 65 70 75 80 85 Age

WebFigure 2. Age-specific numbers of prostate cancer deaths and prostate cancer mortality rate ratios.Age-specific numbers from Quebec in the early 1990s are used to represent the (steady-state) annual numbers of prostate cancer deaths in the absence of screening.The numbers of annual deaths that there would have been in these same population had a screening program been available [from when men reach the age of 50 untilthey turn 70] are hypothetical. Note that these two sets of numbers are age-specific, not cumulative – they decrease if the age range is extended past 85 – and merelyreflect the exponential rise in prostate cancer death rates with age. The rate ratio graph in panel (b) is modeled after Figure 2-5(b) in Morrison and is designed to illustrate (from left to right) its three features: the time-lag until the deathsaverted by screening become apparent, the 20 years of full benefit that follow – after this lag -- the 20 years of screening, and the disappearance of the effect (i.e., areversion to late-age mortality rates in the unscreened scenario) at some point after the last age at which men are screened.

‘% Reduction function’ (bathtub shape)

• The asymptote is the ultimate estimand

• It is determined by ...

– number and spacing of rounds, and

– the contribution of each round of screening

• From published trials, can one ..

– estimate the % Reduction function ?

– estimate contribution of each round ?(?? function shape if different schedule or if a program)

PROSTATE CANCER

Screening & Prostate-Ca Mortality in Randomized European Study ’92-’08 (“ERSPC” nejm2009.04)

As of December 31, 2006, with an average follow-up time of 8.8 years, there were 214 prostate-cancer deaths in thescreening group and 326 in the control group. (...) The adjusted rate ratio for death from prostate cancer in thescreening group was 0.80 (95% CI, 0.65 to 0.98; P=0.04).

“PSA-based screening reduced the rate of death from prostate cancer by 20%. ”

RE-ANALYSIS OF ERSPC DATAusing

year-specific prostate cancer mortality ratios

(A) Overall vs. (B) Year-specific mortality ratios

67%

(B)Prostate Cancer Mortality Rate Ratio (S / C)

1 2 3 4 5 6 7 8 9 10 11 12

0

0.25 75%

0.5 50%

0.75 25%

1 0%

1.25

67%

Percentage Reductionin Year-Specific Prostate Cancer

Mortality Rate

( [C - S] as % of C )

Follow-Up Year:

Yearly Numbers of Prostate Cancer Deathsin Control (C) and Screening (S) Arms . . .

Numbers of Men Being Followed at Mid-Yearin Control (C) and Screening (S) Arms . . .

C:

C:

2

89K

6

88K

21

87K

27

84K

26

82K

39

79K

29

76K

59

71K

40

55K

40

38K

21

22K

11

9K

S:

S:

5

73K

5

72K

10

71K

20

68K

21

66K

28

64K

27

61K

33

57K

25

44K

24

31K

8

18K

3

8K

(A)Cumulative Prostate Cancer Mortality

By the End of Follow-Up Year: 1 2 3 4 5 6 7 8 9 10 11 12

0

0.002

0.004

0.006

0.008

Control Arm (C)

Screening Arm (S)

Hanley, J Medical Screening, 2010.

Articles

www.thelancet.com Published online August 7, 2014 http://dx.doi.org/10.1016/S0140-6736(14)60525-0 1

Screening and prostate cancer mortality: results of the European Randomised Study of Screening for Prostate Cancer (ERSPC) at 13 years of follow-upFritz H Schröder, Jonas Hugosson, Monique J Roobol, Teuvo L J Tammela, Marco Zappa, Vera Nelen, Maciej Kwiatkowski, Marcos Lujan, Liisa Määttänen, Hans Lilja, Louis J Denis, Franz Recker, Alvaro Paez, Chris H Bangma, Sigrid Carlsson, Donella Puliti, Arnauld Villers, Xavier Rebillard, Matti Hakama, Ulf-Hakan Stenman, Paula Kujala, Kimmo Taari, Gunnar Aus, Andreas Huber, Theo H van der Kwast, Ron H N van Schaik, Harry J de Koning, Sue M Moss, Anssi Auvinen, for the ERSPC Investigators*

SummaryBackground The European Randomised study of Screening for Prostate Cancer (ERSPC) has shown signifi cant reductions in prostate cancer mortality after 9 years and 11 years of follow-up, but screening is controversial because of adverse events such as overdiagnosis. We provide updated results of mortality from prostate cancer with follow-up to 2010, with analyses truncated at 9, 11, and 13 years.

Methods ERSPC is a multicentre, randomised trial with a predefi ned centralised database, analysis plan, and core age group (55–69 years), which assesses prostate-specifi c antigen (PSA) testing in eight European countries. Eligible men aged 50–74 years were identifi ed from population registries and randomly assigned by computer generated random numbers to screening or no intervention (control). Investigators were masked to group allocation. The primary outcome was prostate cancer mortality in the core age group. Analysis was by intention to treat. We did a secondary analysis that corrected for selection bias due to non-participation. Only incidence and no mortality data at 9 years’ follow-up are reported for the French centres. This study is registered with Current Controlled Trials, number ISRCTN49127736.

Findings With data truncated at 13 years of follow-up, 7408 prostate cancer cases were diagnosed in the intervention group and 6107 cases in the control group. The rate ratio of prostate cancer incidence between the intervention and control groups was 1·91 (95% CI 1·83–1·99) after 9 years (1·64 [1·58–1·69] including France), 1·66 (1·60–1·73) after 11 years, and 1·57 (1·51–1·62) after 13 years. The rate ratio of prostate cancer mortality was 0·85 (0·70–1·03) after 9 years, 0·78 (0·66–0·91) after 11 years, and 0·79 (0·69–0·91) at 13 years. The absolute risk reduction of death from prostate cancer at 13 years was 0·11 per 1000 person-years or 1·28 per 1000 men randomised, which is equivalent to one prostate cancer death averted per 781 (95% CI 490–1929) men invited for screening or one per 27 (17–66) additional prostate cancer detected. After adjustment for non-participation, the rate ratio of prostate cancer mortality in men screened was 0·73 (95% CI 0·61–0·88).

Interpretation In this update the ERSPC confi rms a substantial reduction in prostate cancer mortality attributable to testing of PSA, with a substantially increased absolute eff ect at 13 years compared with fi ndings after 9 and 11 years. Despite our fi ndings, further quantifi cation of harms and their reduction are still considered a prerequisite for the introduction of populated-based screening.

Funding Each centre had its own funding responsibility.

IntroductionThe European Randomised study of Screening for Prostate Cancer (ERSPC) has shown signifi cant re-ductions in prostate cancer mortality after 9 years1 and 11 years of follow-up.2 Despite these results, screening for prostate cancer is controversial because of adverse eff ects such as overdiagnosis, which is estimated to include 40–50% of screen-detected cases and often results in overtreatment with subsequent side-eff ects.3–5 However, a modelling study, partly based on ERSPC data, showed that with a 4-year screening interval a gain of 52 life-years and a gain of 41 quality-of-life-adjusted life-years (QALYs) was achieved per 1000 men, despite some reduction in quality of life

due to overdiagnosis and long-term side-eff ects of treatment.5

We report updated results of mortality from prostate cancer with follow-up to 2010, with analyses truncated at 9 years, 11 years, and 13 years of follow-up. For the fi rst time, we include France in the analysis of incidence of prostate cancer at 9 years of follow-up, but not in the analysis of mortality because of incomplete follow-up to the end of 2010.

MethodsStudy design and participantsThe ERSPC is a multicentre, randomised, screening trial with the main aim to compare mortality from prostate

Published OnlineAugust 7, 2014http://dx.doi.org/10.1016/S0140-6736(14)60525-0

See Online/Commenthttp://dx.doi.org/10.1016/S0140-6736(14)61008-4

*For the full study group see appendix

Department of Urology, Erasmus University Medical Center, Rotterdam, Netherlands (Prof F H Schröder MD, M J Roobol PhD, Prof C H Bangma MD); Department of Urology, Sahlgrenska Academy at Goteborg University, Goteborg, Sweden (Prof J Hugosson PhD, S Carlsson MD); Department of Urology, Tampere University Hospital, Tampere, Finland (Prof T L J Tammela MD); School of Medicine, University of Tampere, Tampere, Finland (Prof T L J Tammela); Unit of Clinical and Descriptive Epidemiology, ISPO, Florence, Italy (M Zappa MD, D Puliti MSc); Provinciaal Instituut voor Hygiene, Antwerp, Belgium (V Nelen MD); Department of Urology, Kantonsspital Aarau, Aarau, Switzerland (M Kwiatkowski MD, Prof F Recker MD); Department of Urology, Academic Hospital Braunschweig, Braunschweig, Germany (M Kwiatkowski); Department of Urology, Hospital Infanta Cristina, Parla, Madrid, Spain (M Lujan MD); Department of Urology, Hospital Universitario de Getafe, Getafe, Madrid, Spain (M Lujan); Universidad Complutense de Madrid, Madrid, Spain (M Lujan); Finnish Cancer Registry, Helsinki, Finland (L Määttänen PhD,

Prof M Hakama PhD);

Articles

4 www.thelancet.com Published online August 7, 2014 http://dx.doi.org/10.1016/S0140-6736(14)60525-0

Department of Clinical Chemistry, Helsinki University

Central Hospital Laboratory Division (HUSLAB), Helsinki,

Finland (Prof U-H Stenman PhD); FIMLAB, Department of

Pathology, Tampere, Finland

analysis of incidence of prostate cancer with 1–9 years’ follow-up. Appendix pp 13–16, 20, 21 shows the analysis considering all available ages. Appendix p 17 shows a further secondary analysis of the results per centre for the core age group excluding France. No adjustment of signifi cance for α-spending in sequential analyses was

applied because the present analysis is protocol based and not driven by statistical signifi cance.17,18 Cumulative prostate cancer mortality by group was calculated with the Nelson-Aalen method.17 Number needed to invite (NNI) to avert one prostate cancer death was calculated as the inverse of the absolute risk reduction, and the number needed to detect (NND) as the NNI multiplied by the excess incidence of prostate cancer in the intervention group. Analyses were done with Stata version 12.1.

This trial is registered with Current Controlled Trials, number ISRCTN49127736.

Role of the funding sourceThe funders of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. Access to data was limited to the independent data centre led by SMM. None of the investigators had access to outcome data outside the planned offi cial reports of the data centre. FHS produced the primary version and was responsible for submitting the report.

ResultsIn the core group of men aged 55–69 years, excluding France, 162 388 were randomly assigned, of whom 145 died between randomisation and screening. With data truncated at 13 years of follow-up, 7408 prostate

Intervention group Control group Rate ratio*(95% CI)

p value Rate diff erence per 1000 person-years*(95% CI)

Rate diff er-ence per 1000 men*

Adjusted rate ratio in attenders*(95% CI)

p value

Prostate cancer deaths(n)

Person-years

Rate per 1000 person-years

Prostate cancer deaths(n)

Person-years

Rate per 1000 person- years

Years 1–9 193 614 590 0·31 278 751 777 0·37 0·85 (0·70 to 1·03) 0·10 −0·06 (−0·12 to 0·01) −0·46 ·· ··

Years 1–11 265 732 133 0·35 415 896 367 0·46 0·78 (0·66 to 0·91) 0·002 −0·10 (−0·17 to −0·04) −1·02 0·71 (0·58 to 0·88) 0·001

Years 1–13 355 825 018 0·43 545 1 011 192 0·54 0·79 (0·69 to 0·91) 0·001 −0·11 (−0·18 to −0·05) −1·28 0·73 (0·61 to 0·88) 0·0007

*Adjusted by centre and for the randomisation ratio 1:1·5 intervention group versus control group in Finland.

Table 3: Prostate cancer mortality in the intervention and control groups during three time periods truncated (all centres, core age group, France excluded except for years 1–9)

Figure 2: Nelson–Aalen estimates of cumulative prostate cancer mortality (all centres, excluding France)

Intervention groupControl group

1 3 5 7 9 11 130

0·002

0·004

0·006

0·008

0·010N

elso

n-Aa

len

cum

ulat

ive

haza

rd

Time since randomisation (years)

Intervention group Control group Rate ratio*(95% CI)

Rate diff erence per 1000 person-years*(95% CI)

Rate diff erence per 1000 men*

Prostate cancer(n)

Person-years

Rate per 1000 person- years

Prostate cancer(n)

Person-years

Rate per 1000 person-years

Years 1–9 including France 7902 835 353 9·46 5726 984 993 5·81 1·64 (1·58–1·69) 3·69 (3·42–3·95) 26·5

Years 1–9 6147 585 627 10·50 4127 736 688 5·60 1·91 (1·83–1·99) 5·00 (4·68–5·32) 39·0

Years 1–11 6797 692 186 9·82 5262 873 415 6·02 1·66 (1·60–1·73) 3·90 (3·61–4·20) 35·5

Years 1–13 7408 775 527 9·55 6107 980 474 6·23 1·57 (1·51–1·62) 3·44 (3·16–3·72) 34·8

*Control group for Finland weighted by 1:1·5.

Table 2: Prostate cancer incidence in the intervention and control groups during three time periods truncated (all centres, core age group, France excluded except for years 1–9)

Articles

www.thelancet.com Published online August 7, 2014 http://dx.doi.org/10.1016/S0140-6736(14)60525-0 5

(P Kujala MD); Department of Urology, Helsinki University Central Hospital and University of Helsinki, Helsinki, Finland (K Taari MD); Department of Urology, Carlanderska Sjukhuset Göteborg, Sweden (G Aus MD); Centre of Laboratory Medicine, Kantonsspital Aarau, Aarau, Switzerland (Prof A Huber MD); Department of Pathology, Erasmus University Medical Center, Rotterdam, Netherlands Prof T H van der Kwast PhD MD); Department of Clinical Chemistry, Erasmus University Medical Center, Rotterdam, Netherlands (Prof R H N van Schaik PhD); Department of Public Health, Erasmus University Medical Center, Rotterdam, Netherlands (Prof H J de Koning MD); and Centre for Cancer Prevention, Queen Mary University of London, London, UK (Prof S M Moss PhD)

Correspondence to:Prof Fritz H Schröder, Department of Urology, Erasmus University Medical Center, Rotterdam, PO Box 2040, [email protected]

cancer cases were diagnosed in the intervention group and 6107 cases in the control group (fi gure 1).

The median age at randomisation was 60·2 years (table 1). The overall compliance with biopsies was 85·6%, 20 188 of 23 574 screen-positive tests. On average, men in the intervention group were screened 2·3 times (ranging from 1·6 times in Belgium with a 7-year interval to 3·5 times in Sweden with a 2-year interval). Of the screen-positive men who underwent a biopsy, 4883 (24·2%) were diagnosed with prostate cancer within 12 months after testing (table 1).

With follow-up truncated at 13 years, prostate cancer incidence was 9·55 per 1000 person-years in the intervention group and 6·23 in the control group (table 2).

With follow-up truncated at 13 years, prostate cancer mortality was 0·43 per 1000 person-years in the inter-vention group and 0·54 per 1000 person-years in the control group (RR of 0·79, 95% CI 0·69–0·91, p=0·001; table 3, fi gure 2). We recorded a similar RR after 11 years (table 3). After adjustment for non-participation, we noted an RR of 0·71 (95% CI 0·58–0·88, p=0·001) after 11 years and 0·73 (0·61–0·88, p<0·0007) after 13 years (table 3).

The absolute risk reduction in prostate cancer mortality at 13 years of follow-up in the intervention group compared with the control group, after adjustment for the randomisation ratio of 1:1·5 in Finland was 0·11 prostate cancer deaths per 1000 person-years or 1·28 prostate cancer deaths per 1000 men, which yielded an NNI of 781 (95% CI 490–1929) and an NND of 27 (17–66). The NNI and NND were substantially decreased from follow-up to 9 years (NNI 1410, NND 48) and 11 years (NNI 979 [95% CI 594–2770], NND 35 [21–96]).1,2 All-cause mortality did not diff er between the two trial groups (table 4).

In addition to the core age group, we noted a signifi cant reduction in prostate cancer mortality for all 181 999 men aged 50–74 years at entry (excluding France; table 4). The

eff ect of screening did not signifi cantly diff er across 5-year bands in the core age group or across the entire age range, but, most likely by chance, we noted a signifi cant reduction in prostate cancer mortality in the 65–69 year age group. We recorded a non-signifi cant increase in prostate cancer mortality in the 70 year and older screening group (table 4); however, men in this age group were screened only once, which might explain the absence of an eff ect of starting to screen late in life.

Figure 3 shows prostate cancer mortality for the two trial groups in 4-year intervals from date of randomisation. At 0–4 years the RR was 0·88 (95% CI 0·58–1·34), which decreased to 0·82 (0·64–1·06) at 4–8 years, and further decreased to 0·72 (0·59–0·88) at 8–12 years.

An analysis of prostate cancer mortality in the inter-vention and control groups in the core age group of individual centres showed signifi cant RRs only for Sweden (0·62 [95% CI 0·41–0·92]) and the Netherlands

Figure 3: Nelson-Aalen estimates of cumulative prostate cancer in both groups by 4-year periods (all centres, excluding France)

0 4 8 120

20

40

60

80

100

Rate

per

100

000

pers

on-y

ears

Years in trial

Control groupIntervention group

RR 0·72

RR 0·82

RR 0·88

Intervention group Control group Rate ratio (95% CI) p value

Deaths (n) Person-years Rate per1000 person-years

Deaths (n) Person-years Rate per1000 person-years

All-cause mortality

Core age group 15 369 825 018 18·6 19 108 1 011 192 18·9 1·00 (0·98–1·02) 0·82

All ages 18 251 935 185 19·5 21 992 1 120 432 19·6 1·00 (0·98–1·02) 0·98

Prostate cancer mortality

Age groups (years)

≤54 6 64 265 0·09 7 62 312 0·11 0·84 (0·28–2·49) 0·75

55–59 114 411 834 0·28 174 524 314 0·33 0·81 (0·93–1·03) 0·09

60–64 121 240 895 0·50 159 280 404 0·57 0·90 (0·71–1·15) 0·41

65–69 120 172 289 0·70 212 206 474 1·03 0·69 (0·55–0·87) 0·002

70≥ 66 45 903 1·44 58 46 928 1·24 1·17 (0·82–1·66) 0·40

Core age group 355 825 018 0·43 545 1 011 192 0·54 0·79 (0·69–0·91) 0·001

All ages 427 935 185 0·46 610 1 120 432 0·54 0·83 (0·73–0·94) 0·004

Test for heterogeneity for prostate cancer mortality: all ages χ²₄=6·26 p=0·18; core age group: χ²₂=2·31 p=0·32.

Table 4: All cause and prostate cancer mortality by age at randomisation (France excluded)

See Online for appendix

BREAST CANCER

EVERY TRIAL & META-ANALYSIS:and (nejm2010) REPORT on NORWAY NATIONAL SCREENING PROGRAM:

REDUCTION UNDER-ESTIMATED

• Miettinen et al., Lancet 2002.• Hanley, Epidemiologic Reviews 2011.• Hanley JA, Z Liu Z, McGregor M. The [ratio of] benefits [to] harms of

breast cancer screening. Letter re the Report The Independent UKPanel on Breast Cancer Screening (Lancet Nov 17, 2012)

• Hanley JA, McGregor M, Liu Z, Strumpf EC, Dendukuri N.“Measuring the Mortality Impact of Breast Cancer Screening”.Can J Public Health. 2013 Sep 19;104(7):e437-42.(Response to 2011 Canadian Task Force on Preventive Health Care)

Observed breast cancer mortality deficits in 5 Mammography TrialsStudyAges at entryRatio of No. in Experimental arm : No. in Control armParticipation Rate in screens that were part of trial

1 2 3 4 5 6 7 8 9 10 11 12 13 14Year

H.I.P.

40-69

1 : 1

65%

Legend

E: Experimental armC: Control arm.M : MammographyPE : Physical ExamSS : Service Screening

54

64

612

819

1524

1625

1420

1510

610

Reduction

20%

40%

60%Yearly no. Breast

Cancer Deaths

C:E:

in arm

E:

C:Screeningcontent

M-PE M-PE M-PE M-PE

- - - -

CTFPH

C,2011

101130

Malmo

55-70

1 : 1

70% 00

45

01

62

65

45

77

410

28

21

612

Reduction

20%

40%

60%

C:E:

E:

C:

M M M M M M

- - - - - -

134162

Gothenburg

39-59

0.7 : 1

82% 15

24

38

38

1015

78

46

516

610

613

59

58

45

22

Reduction

20%

40%

60%

C:E:

E:

C:

M M M M

- - - M

SS

SS

SS

SS

SS

SS

SS

SS

54103

2Counties

40-74

1.4 : 1

89% 32

1017

1917

1717

1736

1325

811

Reduction

20%

40%

60%

C:E:

E:

C:

M M M

- - -

SS

SS

195206

Stockholm

50-64

2 : 1

80% 11

21

43

33

42

45

54

54

44

33

51

22

Reduction

20%

40%

60%

C:E:

E:

C:

M M -

- - M

4837

.

• Year-specific data: trials used by Task Force.

• 20 years of screening, 50–69, would be followedby 20 years (55–74) in which the breast cancermortality reduction in these years would be ≥40%, with smaller deficits in other years.

• Fewer than 200 women would need toparticipate in such a program in order to avert abreast cancer death in the age range 50-80.

Corresponding Task Force estimates:

Mortality reduction: 21% ;Number of women: 720.

COLON CANCER

FOBT screening for colon cancer – Minnesota Trial 1976-2008

FOBT screening for colon cancer – Minnesota Trial 1976-2008

Long-Term Mortality after Colorectal-Cancer Screening

n engl j med 369;12 nejm.org september 19, 2013 1109

adjusted relative-risk estimates for death from colorectal cancer for the annual-screening and biennial-screening groups were 0.65 (95% CI, 0.52 to 0.80) and 0.76 (95% CI, 0.61 to 0.95), respectively.

Annual or biennial screening with fecal occult-blood testing had no apparent effect on all-cause mortality. The relative risk of death from any cause was 1.00 (95% CI, 0.99 to 1.01) with an-nual screening, 0.99 (95% CI, 0.98 to 1.01) with biennial screening, and 1.00 (95% CI, 0.98 to 1.01) with annual and biennial screening com-bined (Fig. 2 and Table 1). No effect was seen on deaths from causes other than colorectal cancer; the relative risk of death from causes unrelated to colorectal cancer was 1.00 (95% CI, 0.99 to 1.02) with annual screening, 1.00 (95% CI, 0.98 to 1.01) with biennial screening, and 1.00 (95% CI, 0.99 to 1.01) with annual and biennial screening com-bined (Fig. S5 in the Supplementary Appendix). The causes of death are provided in Table S1 in the Supplementary Appendix.

SUBGROUP ANALYSESFigure 3 shows the numbers of participants who underwent randomization, the numbers of those who died from colorectal cancer, and the relative risks for the subgroups of age and sex, according to each study group and the combined screening groups. Graphs of cumulative colorectal-cancer mortality and corresponding relative risks for the subgroups are shown in Figures S6 and S7 in the Supplementary Appendix. The reduction in colorectal-cancer mortality was larger for men than for women in both screening groups and in the two groups combined; the relative risk of death from colorectal cancer was 0.61 (95% CI, 0.47 to 0.80) for men vs. 0.75 (95% CI, 0.57 to 0.97) for women in the annual-screening group, 0.63 (95% CI, 0.48 to 0.82) vs. 0.92 (95% CI, 0.72 to 1.18) in the biennial-screening group, and 0.62 (95% CI, 0.50 to 0.78) vs. 0.83 (95% CI, 0.67 to 1.04) in the combined screening groups. The in-teraction between sex and screening, as mea-sured by the ratio of the relative risk for men to that for women, was significant in the biennial-screening group (P = 0.04 for interaction) but not in the annual-screening group or the two groups combined (P = 0.30 and P = 0.06, respectively, for interaction).

The relative risks of death from colorectal cancer among participants who were less than

Cum

ulat

ive

Col

orec

tal-C

ance

r Mor

talit

y

0.03

0.02

0.01

0.000 5 10 15 20 25 30

Years since Randomization

No. at RiskControlBiennial screeningAnnual screening

14,49714,63514,658

13,10313,24313,294

11,32011,44511,437

915793239219

674168026802

445045834498

ControlBiennial screeningAnnual screening

0.03 (0.03–0.03)0.02 (0.02–0.03)

0.02 (0.02–0.02)

Cumulative Colorectal-Cancer Mortalityat 30 Yr (95% CI)

Control

Biennial

Annual

Figure 1. Cumulative Colorectal-Cancer Mortality.

Cumulative colorectal-cancer mortality was assessed on the basis of Kaplan–Meier estimates, evaluated at monthly time points. Point estimates and 95% confidence intervals at 30 years are also shown.

Cum

ulat

ive

All-C

ause

Mor

talit

y

0.8

0.4

0.6

0.2

0.00 5 10 15 20 25 30


No. at RiskControlBiennial screeningAnnual screening

14,49714,63514,658

13,10313,24313,294

11,32011,44511,437

915793239219

674168026802

445045834498

ControlBiennial screeningAnnual screening

0.71 (0.70–0.72)0.71 (0.70–0.71)0.71 (0.70–0.72)

Cumulative All-Cause Mortalityat 30 Yr (95% CI)

Figure 2. Cumulative All-Cause Mortality.

Cumulative all-cause mortality was assessed on the basis of Kaplan–Meier estimates, evaluated at monthly time points. Point estimates and 95% con-fidence intervals at 30 years are also shown.

The New England Journal of Medicine Downloaded from nejm.org by James Hanley on September 19, 2013. For personal use only. No other uses without permission.

Copyright © 2013 Massachusetts Medical Society. All rights reserved.

Radiologists as Statisticians2014-10-15, 1:21 PMThe 1970 Draft Lottery in R

Page 2 of 10http://msenux.redwoods.edu/math/R/lottery.php

Figure 1. Rep. Alexander Pirnie, R-NY, draws the first capsule in the lottery drawing held on Dec. 1, 1969. The capsule containedthe date, Sept. 14.

The last capsule drawn contained the date December 31. It was estimated by the Pentagon that men with draftnumbers in the last third, numbers 200 to 366, would escape the draft entirely. In fact, no man with a draftnumber higher than 195 was called to duty.

The fairness of the draft lottery was immediately debated. Critics contended that the process was not trulyrandom. A New York Times article quoted a White House source as saying "discussions that the lottery was notrandom are purely speculative." In that same New York Times article, Senator Edward Kennedy was quoted asasking the National Sciences the "apparent lack of randomness" in the selection.

The Data

The data is publicly available on the internet. One source is the Data and Story Library. The draft lottery data islocated at the following URL:

http://lib.stat.cmu.edu/DASL/Datafiles/DraftLottery.html

If you have not imported data into R from external sources, you might want to first work through the activityImporting Data in R.

One technique, as explained in Importing Data in R, suggests copying the data into a plain text file. Open asimple text editor (e.g., Notepad on Windows or Textedit on the Mac). Copy and paste the lottery data from theabove URL, including headers (but not the descriptive information above the headers), and save the file as

2014-10-15, 1:21 PMThe 1970 Draft Lottery in R


In the activity Boxplots in R we learned how to use R's boxplot command to produce a boxplot of a data set. Toexamine the "fairness" of the Selective Service's draft lottery, we will produce "side-by-side" boxplots for eachmonth of the year. That is, we will produce 12 boxplots, one for each month of the year, each containing ananalysis of the associated draft numbers for that month. The following command will produce these "side-by-side" boxplots shown in Figure 4.

> boxplot(Draft_No. ~ Month, data=lottery)

Figure 4. Side-by-side boxplots of draft numbers for each month.

Because the data in Month is categorical (you can see this by typing lottery$Month), the model formulaDraft_No. ~ Month causes the boxplot command to group the numerical data in Draft_No. according to thecategories in Month. Therefore, the command boxplot(Draft_No. ~ Month, data=lottery) creates 12 boxplots,one for each month. For example, the boxplot for April (see Apr in Figure 4) contains an analysis for only thosedraft numbers that were assigned to birth-dates in April. Similar comments are in order for the remainingmonths.

Unfortunately, the months are sorted in alphabetical order (the default behavior). It would be more appropriateif they were sorted in chronological order, January first, February second, etc. One solution would be to boxplotthe draft numbers versus the month number.

> boxplot(Draft_No. ~ Mo.Number, data=lottery)

This command produces the side-by-side boxplots shown in Figure 5.



Figure 2. A scatterplot of Draft_No. versus Day_of_year.

One could also "attach" the dataframe lottery (see Dataframes in R). When we "attach" a dataframe, we canaccess the columns without using the dollar notation. Thus, we can plot Draft_No. versus Day_of_year withthe following commands.

> attach(lottery)> plot(Day_of_year,Draft_No.)

It is good practice to "detach" the dataframe when finished.

> detach(lottery)

Readers should check that these commands produce a scatterplot identical to that shown in Figure 2.

Efficient Use of Dataframes

R's plot command, coupled with a "model formula," it the most efficient way to produce a scatterplot. Withoutfurther explanation, enter the following code. Note: Remember that ~ is a "tilde", not a minus sign, and islocated to the immediate left of the 1 key on the second row from the top of your keyboard.

> plot(Draft_No. ~ Day_of_year, data=lottery)

This command will produce the scatterplot shown in Figure 3. Note that it is identical to the scatterplot shownin Figure 2.



Figure 6. Side-by-side boxplots of draft numbers sorted by month.

Interpretation of Results

The image in Figure 6 is perfect. The months are now sorted in chronological order. But now, what does theimage of side-by-side boxplots tell us?

Remember, the heavy horizontal bar in each box is the median of the data set. The median draft number for themonth of December is very disconcerting. Remember, the lower the draft number, the more likely you would beinducted to serve in Vietnam. Why does the month of December have a median that is significantly lower thanmost of the other months. It seems that the men with birthdays in December are being unfairly selected. Indeed,with the exception of October, the last remaining months of the year all have medians that are significantlylower than the medians of the previous months. Something strange is going on!

One story offers a hint of an explanation. It seems that the capsules containing birthdays for January wereplaced in a shoe-box, thoroughly mixed, then poured into the glass container shown in Figure 1. Then the sameprocedure was followed for the capsules containing birthdays in February, stirring them thoroughly in a shoe-box, then pouring them into the glass container. This same procedure was followed for the remaining months.December was the last month processed, or so the story goes.

However, this is quite disturbing. If capsules were selected from the top of the glass container, they were morelikely to be a December birthday. According to the story, the person making the draws did not always reachdeep into the pile of capsules. This may be one explanation for why so many December birthdays were selectedearly in the process and assigned low draft numbers (which correlates to a higher chance of being drafted).

This story may be an oversimplification. Readers are encouraged to explore the reasons for why this process

Time-split versus time-lumped Rate Ratios

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

1:30

1-year binsRate Ratio

Year

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

1:30

rep(

1, 3

0)


Year

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

1:30

rep(

1, 3

0)


Year

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5


Year

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

rep(

1, 3

0)


Year

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

rep(

1, 3

0)


Year

annual

biennial

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

Time-split versus time-lumped Rate Ratios

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

1:30

1-year moving binsRate Ratio

Year

annual

biennial

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

1:30

rep(

1, 3

0)


Year

annual

biennial

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

1:30

rep(

1, 3

0)


Year

annual

biennial

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

1:30

rep(

1, 3

0)


Year

annual

biennial

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5


Year

annual

biennial

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

rep(

1, 3

0)


Year

annualbiennial

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

rep(

1, 3

0)


Year

annualbiennial

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

rep(

1, 3

0)


Year

annual

biennial

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

Dear Editor

• Shaukat et al. report reductions of 32% and 22% in colon cancer mortality in

those offered 11 annual and 6 biennial FOB screens, respectively. These

reductions were achieved despite a 4-year hiatus in screening, and averaging

over all 30-years of follow-up.

• What would the reductions have been without such an interruption? To answer

this, we extracted the yearly numbers of deaths from the published Figure 1, and

instead calculated yearly mortality reductions. Because of the unusual schedule,

the resulting reduction curve has a ‘W’ shape, showing the lagged responses to

the two phases of screening: after a delay of some years, mortality reductions

reached a nadir of around 40% before reverting to what they would be in the

absence of screening; this pattern is repeated when screening is resumed.

• Without the (funding related) hiatus, the reductions would have been around

40% for each year affected, which is substantially larger than those estimated.

− − − − − − − − − − − − − − − − − − − − − − − − − − − − − −

0 10 20 30

● ●

●

● ●

● ●

●

● ●●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●●

Annual

Biennial

S S S S S S

S S S S S S S S S S S

0%

20%

40%

60%

Red

uctio

n in

Col

on C

ance

r M

orta

lity


Yearly reductions in colon cancer mortality in two screening arms. Each dot is based

on number of deaths in a three year moving window; smooth curves were fitted though

them. Because the hiatus was in calendar-time rather than follow-up time, and entries

were staggered, the timing of the screens (each denoted by an ‘S’) is only approximate.

From Trial to Program

STATISTICAL MODEL

Convolution of reductions produced by individual roundsYearly

Number ofDeaths

(a)

0

20

40

60

80

100

YearlyMortalityReduction

RemarksUnscreened

Screened

S1 S2 S3

1YEAR: 2

1

3

1

2

4

1

2

3

5

35%

1

2

3

6

1

2

3

7

2

3

8

3

9 10 11 12 13

TRIAL

60%

40%

20%

If one round of screening reduces mortality in each of 5 future years, then

in a trial, 3 rounds of screening -- S1, S2 and S3 -- would produce 3 'waves'

of mortality reductions ('1', '2', '3'), each 5 years wide, over 7 years (Y3-Y9).

In such a trial, the maximal reduction (35%, year 6) would be smaller than

the sustained (46%) reductions produced by a 20-year screening program.

The average reduction, computed over 13 years of follow-up in such a trial would

be an even more serious underestimate of the impact of a 20-year program.

(b)

0

20

40

60

80

100

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S13 S14 S15 S16 S17 S18 S19 S20

50AGE:

1 1

2

1

2

3

1

2

3

4

55

46%

1

2

3

4

5

2

3

4

5

6

3

4

5

6

7

60

etc.

65 70

46%

16

17

18

19

20

17

18

19

20

18

19

20

19

20

20

75

PROGRAM YearlyMortalityReduction

60%

40%

20%

Annual mortality reductions produced by once-a-year screening that begins when women reach 50, and ends when they reach 69

Fitted Model (each round) & Resulting Fits for 6 and 11 Rounds (JH)

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

0 5 10 15 20 25 30

0.0

0.5

1.0

1.5

Rate Ratio

Year

annual

biennial

4

4

3

7

6

3

5

8

9

5

15

4

3

6

6

9

7

2

13

8

6

8

8

9

12

10

7

13

9

6

16

6

10

14

14

10

11

10

5

10

8

7

9

7

11

10

6

7

7

9

10

6

6

5

9

12

4

10

4

8

17

12

8

13

6

6

8

5

9

8

6

8

13

5

4

11

10

10

10

9

8

6

4

3

7

8

6

6

5

3

Effect of 1 round6.5

0 5 10 15 20 25 30 35

Red

uction

100 %

80 %

60 %

40 %

20 %

0 %

S S S S S S S S S S SS S S S S S Biennial (78% compliance)

Annual (75% compliance)

A D2

D1

D0

3

4

4

3

6

7

9

8

5

4

15

7

6

11

1

2

7

9

6

8

13

9

8

8

8

10

13

6

9

13

10

6

17

10

14

14

5

9

11

7

8

10

11

7

9

7

6

10

10

9

12

5

5

14

4

12

9

8

4

10

8

12

17

6

6

13

9

5

8

8

6

8

4

5

13

10

10

11

8

9

10

2

3

6

6

10

7

6

5

6

0 5 10 15 20 25 30 35

Follow−up year

Re

du

ction

100 %

80 %

60 %

40 %

20 %

0 %B

S S S S S S S S S S S S S S S SS S S S S S S S

Biennial (78% compliance)

95% confidence bands (biennial)

Annual (75% compliance)

Figure 5–4: Panel A: Empirical and fitted mortality reductions based on the yearlynumbers of colorectal cancer deaths in the two screening arms of the MinnesotaColorectal Cancer Screening Study, with the 4-year hiatus. The size of each dot isproportional to the information contribution of the empirical year-specific mortalityratio. Because the hiatus was in calendar-time rather than follow-up time, and entrieswere staggered, the timing of the screens, each denoted by an S, is only approximate.Panel B: Projection of yearly mortality reductions in colorectal cancer that wouldbe generated by 15 years of uninterrupted annual and biennial fecal occult bloodscreening. The grey area represents time-specific 95% confidence bands under thebiennial screening regimen.

98

LUNG CANCER

T h e n e w e ngl a nd j o u r na l o f m e dic i n e

10.1056/nejmoa1102873 nejm.org 1

original article

Reduced Lung-Cancer Mortality with Low-Dose Computed Tomographic Screening

The National Lung Screening Trial Research Team*

The members of the writing team (who are listed in the Appendix) assume re-sponsibility for the integrity of the article. Address reprint requests to Dr. Christine D. Berg at the Early Detection Research Group, Division of Cancer Prevention, National Cancer Institute, 6130 Execu-tive Blvd., Suite 3112, Bethesda, MD 20892-7346, or at [email protected].

* A complete list of members of the Na-tional Lung Screening Trial research team is provided in the Supplementary Appendix, available at NEJM.org.

This article (10.1056/NEJMoa1102873) was published on June 29, 2011, at NEJM.org.

N Engl J Med 2011.Copyright © 2011 Massachusetts Medical Society.

A BS TR AC T

BackgroundThe aggressive and heterogeneous nature of lung cancer has thwarted efforts to reduce mortality from this cancer through the use of screening. The advent of low-dose helical computed tomography (CT) altered the landscape of lung-cancer screen-ing, with studies indicating that low-dose CT detects many tumors at early stages. The National Lung Screening Trial (NLST) was conducted to determine whether screening with low-dose CT could reduce mortality from lung cancer.

MethodsFrom August 2002 through April 2004, we enrolled 53,454 persons at high risk for lung cancer at 33 U.S. medical centers. Participants were randomly assigned to un-dergo three annual screenings with either low-dose CT (26,722 participants) or sin-gle-view posteroanterior chest radiography (26,732). Data were collected on cases of lung cancer and deaths from lung cancer that occurred through December 31, 2009.

ResultsThe rate of adherence to screening was more than 90%. The rate of positive screen-ing tests was 24.2% with low-dose CT and 6.9% with radiography over all three rounds. A total of 96.4% of the positive screening results in the low-dose CT group and 94.5% in the radiography group were false positive results. The incidence of lung cancer was 645 cases per 100,000 person-years (1060 cancers) in the low-dose CT group, as compared with 572 cases per 100,000 person-years (941 cancers) in the radiography group (rate ratio, 1.13; 95% confidence interval [CI], 1.03 to 1.23). There were 247 deaths from lung cancer per 100,000 person-years in the low-dose CT group and 309 deaths per 100,000 person-years in the radiography group, representing a relative reduction in mortality from lung cancer with low-dose CT screening of 20.0% (95% CI, 6.8 to 26.7; P = 0.004). The rate of death from any cause was reduced in the low-dose CT group, as compared with the radiography group, by 6.7% (95% CI, 1.2 to 13.6; P = 0.02).

ConclusionsScreening with the use of low-dose CT reduces mortality from lung cancer. (Funded by the National Cancer Institute; National Lung Screening Trial ClinicalTrials.gov number, NCT00047385.)

The New England Journal of Medicine Downloaded from nejm.org on June 29, 2011. For personal use only. No other uses without permission.

Copyright © 2011 Massachusetts Medical Society. All rights reserved.

NLST

Age at entry : 55−74

CT : X−ray allocation = 1 : 1

Compliance = 94%

1 2 3 4 5 6 7 8 9 10

Year

31

37

57

68

67

82

84

95

72

84

42

73

Reduction

20%

40%

60%

No. Lung Cancer Deaths

CT CT CT

X−ray X−ray X−ray

Figure 6–1: NLST yearly numbers of lung cancer deaths, extracted from publishedNEJM report.

advantage of occasionally rediscussing statistical conclusions, by starting

from the same documents as their author. I have begun to think that no

one ought to publish biometric results, without lodging a well arranged

and well bound manuscript copy of all his data, in some place, where it

should be accessible under reasonable restrictions, to those who desire to

verify his work.

In screening trials, the cumulative measures often hide the reduction patterns

over time, and because of this, we have been on a ‘campaign’ calling on trialists to

report (at least) the yearly numbers of cancer-specific deaths (as opposed to just a

cumulative mortality reduction over some arbitrary follow-up window), if disclosing

the individual-level data is not possible [53]. From our experience, the aggregated

numbers are in fact ‘near-sufficient’ statistics. Moreover, even if yearly counts are not

105

NLST



Compliance = 94%

1 2 3 4 5 6 7 8 9 10

Year

31

37

57

68

67

82

84

95

72

84

42

73

50

80

55

90

70

95

80

100

Reduction

20%

40%

60%


CT CT CT


Figure 6–2: NLST yearly numbers of lung cancer deaths, with relatively large hypo-thetical reductions in years 7-10.

obtainable, we proposed to combine information from multiple trials by adding up

the trial-specific log-likelihoods to obtain an overall log-likelihood for more accurate

parameter estimates, which avoids sharing neither the individual-level data or the

aggregated data. This idea was presented by Hanley at the Statistical Society of

Canada Annual Meeting in 2012.

Having said this, the NCI generously reached out and made their individual-

level data available to qualified investigators in early 2013. We immediately took

advantage of the offer. Information on all of the 53,452 randomized persons (26,722

in the CT arm and 26,730 in the X-ray arm) was recorded in the patient file. Using

the following variates: number of days from randomization to the end of follow-up

(fup days), number of days from randomization to death (death days) and cause

106

NLST



Compliance = 94%

1 2 3 4 5 6 7 8 9 10

Year

31

37

57

68

67

82

84

95

72

84

42

73

65

80

75

90

80

95

90

100

Reduction

20%

40%

60%


CT CT CT


Figure 6–3: NLST yearly numbers of lung cancer deaths, with relatively small hy-pothetical reductions in years 7-10.

of death (finaldeathLC==1 for death from lung cancer), we can make a population-

time plot to illustrate the study base and some key features of the trial. Figure

6–4 shows that (i) the randomization ratio was 1:1, and the amount of population

time was similar between the two arms, (ii) the median length of follow-up was

about 6.5 years and most people were alive by the end of the follow-up; and (iii)

although these smokers (compared to the general population) may have an elevated

risk of dying from lung cancer, in absolute terms, lung cancer mortality was still

quite low in both arms – there were a total of 1,019 deaths from lung cancer over the

entire follow-up, 467 in the CT arms and 552 in the CXR arm. Thus, the empirical

6.5-year risk ratio of (467/26722)/(552/26730) = 0.846 and the mortality rate ratio

of (467/171,412)/(552/170,355) = 0.841 are very close. The cumulative mortality

reduction from lung cancer over 6.5 years is 1− 0.846 = 15.4%.

107

0 2 4 6 8

Follow−up Year

Po

pu

latio

n

CT arm

X−ray arm

467 lung ca. deaths

171,412 PYs

552 lung ca. deaths

170,355 PYs

0

5000

10000

15000

20000

25000

0

5000

10000

15000

20000

25000

Figure 6–4: NLST number at risk for the two arms, along with lung cancer deaths,using the individual-level data provided by the NCI.

Checking the yearly numbers that we extracted in Table 6–1(a) against those

calculated from individual-level data in Table 6–1(b) was one of our first tasks, by

including lung cancer deaths before the cutoff date only. They were almost identical,

only differing by a few deaths. Next we included lung cancer deaths also after the

108

Table 6–1: Yearly numbers of lung cancer deaths in the NLST. Part (a) was based onour extraction from the NEJM report, (b) and (c) are based on the individual-levelNLST data; in (b) only deaths that occurred before the cut-off (i.e. January 15th,2009) were included, and in (c) all deaths occurred before and after the cutoff datewere included.

(a) Year-specific data extracted from figure in NEJM reportFollow-up Year: 1 2 3 4 5 6 7 Total

Screens ↑ ↑ ↑X-ray Arm: 37 68 82 95 84 73 4 442

CT Arm: 31 57 67 84 72 42 3 354Reduction: 16% 16% 18% 12% 14% 42% 25% 20%

(b) Year-specific data including deaths before the cutoff onlyX-ray Arm: 38 70 83 91 88 74 4 448

CT Arm: 31 57 67 84 72 45 3 359Reduction: 18% 19% 19% 8% 18% 39% 25% 20%

(c) Year-specific data including deaths before and after the cutoffX-ray Arm: 38 70 83 91 89 116 65 552

CT Arm: 31 57 67 84 73 85 70 467Reduction: 18% 19% 19% 8% 18% 27% -8% 15%

cutoff in Table 6–1(c), which confirms our earlier suspicion that the counts were

incomplete starting in year 6. Now with the additional mortality data in year 7, the

reduction in year 6 turns out to be less dramatic – 27% instead of 42%. Although

not all deaths in year 7 had been adjudicated, if a similar fraction of deaths were

adjudicated between the two arms, then it would suggest that the signal had been

fading away by year 7.

The description of all the variables can be found in participant.dictionary.

d091412.rtf, while the ones we use in this chapter are included in Table 6–2.

109

NLST



Compliance = 94%

1 2 3 4 5 6 7 8 9 10

Year

31

38

57

70

67

83

84

91

73

88

85

117

70

65

Reduction

20%

40%

60%


CT CT CT


Figure 6–5: NLST yearly numbers of lung cancer deaths, corresponding to table6–1(c).

Table 6–2: These are the only variables needed for our model fitting, with descriptionsprovided by the NCI participant dictionary.pid: patient ID, one per patientrndgroup: a binary variable indicating to which arm the participant

was randomized, 1=CT and 2=X-ray.age: age at randomization, in years.death days: days from randomization to death.finaldeathLC: the authoritative variable indicating whether lung cancer was

the cause of death.deathcutoff: a binary variable indicating whether deaths occurred before

the cutoff for the official final analysis of lung cancer mortality.

110

be around 30%, which doubles the 15% reduction achieved with three rounds of

screening in the trial.

One could easily study whether a younger or older age group would benefit

more from early detection, by choosing data on those aged, say, 65 years or younger

at randomization. The fitted curve and the corresponding projection based on 10

rounds of annual screening are shown in Figure 6–6. Our choice of the age group is

rather arbitrary, but this serves as an illustration for other subgroup analyses, such

as splitting by gender, ethnicity group, medical history and so on.

Follow−up year

Reduction

100 %

80 %

60 %

40 %

20 %

0 %

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

S S S

S S S S S S S S S S

Fitted (3 rounds)

Projection (10 rounds)

Figure 6–6: Fitted reduction curve (dotted, black) based on the NLST data forpersons aged below 65 at onset of screening and projected curve based on 10 roundsof annual screenings.

113

Summary

• By ignoring the delay until the reductions in mortality are expressed, theprevailing interpretations of the results of cancer screening trialsunder-estimate the mortality reductions that would be produced by asustained screening program

• P-value-driven RCT stopping/reporting rules exacerbate the problem

• We might be able to avoid such misleading estimates if we . . .(i) distinguish a trial from a program(ii) run trials with sufficient rounds of screening and sufficient follow-up(iii) spend major portion of career waiting to measure real reductions(iv) analyze the data using time-specificity / non-proportional hazards(v) focus on parameters describing impact of 1 round of screening(vi) mammography: use data from population-screening, not old trials

FUNDING, CO-ORDINATES, DOWNLOADSNatural Sciences and Engineering Research Council of Canada

Le Fonds québécois de la recherche sur la nature et les technologies

Canadian Institutes of Health Research (2011-2014)........................................................................

[email protected]@Mail.McGill.CA

www.med.mcgill.ca/epidemiology/hanley

→ r e p r i n t s / talks

BIOSTATISTICS

http:/p: /wwwwwww.mw.mw.mmcgill.ca/ca/a epiepiepiepi-bibbiostosts at-at-aa occh/g/ggrad/bib ostatistit cs/

Why do statisticians commonly limit their inquiries to Averages?

F. Galton, Natural Inheritance, 1889.

“It is difficult to understand why statisticians commonly limittheir inquiries to Averages, and do not revel in morecomprehensive views.

Their souls seem as dull to the charm of variety as that of thenative of one of our flat English counties, whose retrospect ofSwitzerland was that, if its mountains could be thrown into itslakes, two nuisances would be got rid of at once.”

Timing of cholesterol reductions produced by statins

3 dogs at 20 mg/kg/day; 3 at 50 mg/kg/day

Fig. 6. Hypolipidemic effects of mevastatin in dogs. Three dogs received mevastatin for 13 days (from day 0 to day 12) at a dose of 20 mg/kg per day (A) or 50 mgikg per day (B) (Replotted from Fig. 1 of ref. 6). (Used with permission, Atherosclerosis. 1979. 32: 307-313.)

We felt that mevastatin should be evaluated more perti- nently in animal models comparable to FH in humans, since in patients with FH, regulation of HMG-CoA reductase is partially or completely lost, resulting in high reductase activity (42). At that time, however, such an animal model was not available.

The nonionic detergent Triton WR-1339 was shown to produce hypercholesterolemia in rats (66). Using this model, several groups suggested that the elevated levels of hepatic HMG-CoA reductase were responsible for the increase in plasma cholesterol (67-69). Mevastatin was found to be slightly effective in these animals, giving up to 21% reduction of plasma cholesterol at 100 mg/kg (70). These results aroused a glimmer of hope, but were still not sufficient.

Commercial eggs contain - 300 mg of cholesterol, and according to our preliminary analyses, two-thirds of this amount of cholesterol is derived from diet and the re- mainder is supplied by de novo synthesis. We expected that the level of cholesterol synthesis in hens that were ac- tively producing eggs would be higher than that in roosters. We fed hens a commercial diet supplemented with 0.1% mevastatin for 30 days. As expected, plasma cholesterol was reduced by as much as 50%, while body weight, diet consumption, and egg production were not significantly changed throughout the experiments (71).

The success in the experiments in hens opened up an opportunity to conduct experiments in dogs and monkeys. In dogs, mevastatin reduced plasma cholesterol by 30% at a dose of 20 mg/kg and as much as 44% at 50 mg/kg (Fig. 6) (6). &Lipoprotein (LDL) was markedly reduced by mevastatin while a-lipoprotein (HDL) was

not lowered but, rather, increased slightly. In early 1977, we gave mevastatin to monkeys for 11 days. The reduction of plasma cholesterol was 21% at a dose of 20 mg/kg and 36% at 50 mg/kg (Fig. 7) (7). Plasma triglyceride levels were not changed significantly in either dogs or monkeys. Fecal excretion of bile acids was slightly elevated in dogs but not significantly changed in monkeys (6, 7).

Monkey (50 mg/kg/day) 200

"1 , I ; 0

-16 -8 0 8 16 24

Days

Fig. 7. HypoJipidemic effects of mevastatin in cynomolgus monkeys. Three monkeys received mevastatin at a dose of 50 mg/kg per day for 11 days (from day 0 to day 10) (Reproduced from Fig. 1 of ref. 7). (Used with permission, Lipids. 1979. 14: 585-589.)

1574 Journal of Lipid Research Volume 33, 1992

by on April 3, 2010

ww

w.jlr.org

Dow

nloaded from

3 monkeys at 50

Fig. 6. Hypolipidemic effects of mevastatin in dogs. Three dogs received mevastatin for 13 days (from day 0 to day 12) at a dose of 20 mg/kg per day (A) or 50 mgikg per day (B) (Replotted from Fig. 1 of ref. 6). (Used with permission, Atherosclerosis. 1979. 32: 307-313.)

We felt that mevastatin should be evaluated more perti- nently in animal models comparable to FH in humans, since in patients with FH, regulation of HMG-CoA reductase is partially or completely lost, resulting in high reductase activity (42). At that time, however, such an animal model was not available.

The nonionic detergent Triton WR-1339 was shown to produce hypercholesterolemia in rats (66). Using this model, several groups suggested that the elevated levels of hepatic HMG-CoA reductase were responsible for the increase in plasma cholesterol (67-69). Mevastatin was found to be slightly effective in these animals, giving up to 21% reduction of plasma cholesterol at 100 mg/kg (70). These results aroused a glimmer of hope, but were still not sufficient.

Commercial eggs contain - 300 mg of cholesterol, and according to our preliminary analyses, two-thirds of this amount of cholesterol is derived from diet and the re- mainder is supplied by de novo synthesis. We expected that the level of cholesterol synthesis in hens that were ac- tively producing eggs would be higher than that in roosters. We fed hens a commercial diet supplemented with 0.1% mevastatin for 30 days. As expected, plasma cholesterol was reduced by as much as 50%, while body weight, diet consumption, and egg production were not significantly changed throughout the experiments (71).

The success in the experiments in hens opened up an opportunity to conduct experiments in dogs and monkeys. In dogs, mevastatin reduced plasma cholesterol by 30% at a dose of 20 mg/kg and as much as 44% at 50 mg/kg (Fig. 6) (6). &Lipoprotein (LDL) was markedly reduced by mevastatin while a-lipoprotein (HDL) was

not lowered but, rather, increased slightly. In early 1977, we gave mevastatin to monkeys for 11 days. The reduction of plasma cholesterol was 21% at a dose of 20 mg/kg and 36% at 50 mg/kg (Fig. 7) (7). Plasma triglyceride levels were not changed significantly in either dogs or monkeys. Fecal excretion of bile acids was slightly elevated in dogs but not significantly changed in monkeys (6, 7).

Monkey (50 mg/kg/day) 200

"1 , I ; 0

-16 -8 0 8 16 24

Days

Fig. 7. HypoJipidemic effects of mevastatin in cynomolgus monkeys. Three monkeys received mevastatin at a dose of 50 mg/kg per day for 11 days (from day 0 to day 10) (Reproduced from Fig. 1 of ref. 7). (Used with permission, Lipids. 1979. 14: 585-589.)

1574 Journal of Lipid Research Volume 33, 1992

by on April 3, 2010

ww

w.jlr.org

Dow

nloaded from

Timing of cholesterol reductions produced by statins

Humans

The loneliness of the long-distance trialist

1 2 3 4 5 6 7 8 9 10 11 120

0.002

0.004

0.006

0.008

0.01

Timing of Screening Effects(as seen in cumulative cause-specific mortality curves)

Abdominal Aortic Aneurysms

(One-off Screening, MASS)

Control

Arm

Screening

Arm

Prostate Cancer

(q 4y, ERSPC )

Control

Arm

Screening

Arm

Cumulative Cause-Specific Mortality

Follow-Up Year Supp Fig. A

Documents

HOW MUCH DOES SCREENING REDUCE CANCER ... - medicine.mcgill.caReprints... · Screening & Prostate-Ca Mortality in Randomized European Study ’92-’08 (“ERSPC” nejm2009.04) As