The Numbers are the Easy Part: Dr. Beverly Rockhill UNC

1
The Numbers are the Easy PartDr. Beverly
RockhillUNC
2
Questions of interpretation and usefulness of
population attributable fractions, with
illustrations from breast cancer Beverly
Rockhill, Ph.D.University of North Carolina,
Chapel Hill
3
Background

• History
• Definition(s) of PAF
• Summary PAF
• Distributive property of PAF

4
Three population attributable fraction estimates
for breast cancer

• PAF 1 25 (CPS-I)
• PAF 2 55 (BCDDP)
• PAF 3 41 (NHANES EFS)

5
Question 1 What does it mean when variables
other than modifiable causes are included as
variables in PAF calculations?
6

• PAF 1 family history, BBD, Jewish ethnicity,
  late menopause (50 yrs), early menarche (lt12
  yrs), ever married, late age at first birth (gt30
  yrs), college education, daily alcohol,
  overweight
• PAF 2 early menarche (lt14 yrs), late age at
  first birth (gt20 yrs), BBD, family history
• PAF 3 late age at first birth (gt20 yrs),
  family history, HH income in upper 2/3 of US
  population

7
Question 2 What does it mean when nearly
everyone in population is exposed?
8

• PAF 1 fam hist,BBD, Jewish, menopause 50 yrs,
  menarche lt12 yrs, ever married, AFB gt30 yrs,
  college, alcohol, overweight75-80 exposed
  (PAF 25)
• PAF 2 menarche lt14 yrs, AFB gt20 yrs, BBD, fam
  hist 98 exposed (PAF 55)
• PAF 3 AFB gt20 yrs, fam hist, income upper 2/3
  90 exposed (PAF 41)

9
Question 3 What does it mean when PAFs are
interpreted as amount of disease explained?
10

• PAF 1 (25) (au) Unable to identify the
  causes of more than about ¼ of cases. Shows
  that 75 of women who develop breast cancer have
  no known risk factors.
• PAF 2 (55) Another report estimates that 55
  of breast cancers have one or more risk factors.
  
• PAF 3 (41) (au) Suggests a substantial
  proportion of breast cancer cases in US are
  explained by well-established risk factors.

11
RR 2
12
General observations about PAF

• Magnitude of PAF estimate dependent on choice of
  exposure cutpoint(s) estimates can be made high
  (though very imprecise) by considering high
  proportion exposed
• Choice of cause(s) often arbitrary, sometimes
  meaningless

13
Implications of PAFs for prevention

• Population-level interpretation of PAF
  necessary number needed to change in terms of
  exposure very high relative to number of cases
  that will be averted

14
Implications of PAFs for prevention

• PAF says nothing about causes of specific
  individual cases, nor about which cases are (will
  be) attributable to exposure

15
Conclusions

• Plugging in RRs (IRRs) and exposure prevalences
  is the easy part of PAF estimation
• Critical consideration of implications of choices
  of exposures and exposure cutpoints, and of PAF
  estimate, for possible prevention strategies is
  the difficult part

16
Conclusions

• Key question to be addressed in PAF analyses
  often unasked, and thus unanswered Is there
  any implication for realistic and effective
  prevention strategy that will shift exposure in
  high proportion of population?

17
Estimated five-year risk of breast cancer,
according to breast cancer status at end of
follow-up
18
RRq1-5 20
19
RRq1-5 200