Puzzles, Paradoxes and Pitfalls

Download Report

Transcript Puzzles, Paradoxes and Pitfalls

Statistical Pitfalls
Stephen Senn
(c) Stephen Senn
Statistical Pitfalls
1
Four to watch out for
• Regression to the mean
• Simpson’s paradox
• Invalid inversion
– The error of the transposed conditional
• Selective sampling
(c) Stephen Senn
Statistical Pitfalls
2
There are Three Kinds of
Statistician
• Those who can count
• Those who can’t
(c) Stephen Senn
Statistical Pitfalls
3
Regression to the mean
• Powerful phenomenon causing apparent
change over time
• If individuals are selected for treatment
because extreme when measured again
on average they will be closer to the mean
• Discovered by Francis Galton (1822-1911)
(c) Stephen Senn
Statistical Pitfalls
4
(c) Stephen Senn
Statistical Pitfalls
5
(c) Stephen Senn
Statistical Pitfalls
6
(c) Stephen Senn
Statistical Pitfalls
7
Consequences
• Spontaneous improvement over time easy
to produce
• Is a consequence of the way data are
studied not the phenomenon being studied
• Always compare to the control group
• Quite possibly the explanation of the
placebo effect
(c) Stephen Senn
Statistical Pitfalls
8
Examples
• Remedial treatment for accident blackspots
– Sacrificing a chicken would work
• The placebo effect
– How do you know that nothing at all would not work
• Horace Secrist’s discovery of the decline in
profitability of the most profitable US firms
– Harold Hotelling put him right
• And he originally studied journalism!
(c) Stephen Senn
Statistical Pitfalls
9
How to Prove Spells Against
Rain Work
• Wait for a very rainy day
• Say these words
– “Rain rain go away, come again another day”
• By some mystical pluvioincantory process the
weather will be drier at some stage in the future
than it was on the day the spell was uttered
• Conclusion
– The spell works against rain
(c) Stephen Senn
Statistical Pitfalls
10
Simpson’s Paradox
Simpson, E. H. (1951). "The interpretation of interaction in contingency
tables." Journal of the Royal Statistical Society, Series B 13: 238-241.
Graduate Admissions to Berkeley 1973
A Bias against Women?
Per cent admission by sex
Male
Female
(c) Stephen Senn
Admitted Denied
Total
44
56
100
35
65
100
Statistical Pitfalls
11
Graduate Admissions to Berkeley 1973
The Bias Disappears?
Faculty
A
B
C
D
E
F
(c) Stephen Senn
Male
Female
Admitted Denied
Admitted Denied
62
38
82
18
60
40
68
32
37
63
34
66
33
67
35
65
28
72
24
76
6
94
7
93
Statistical Pitfalls
12
Simpson’s Paradox
The Berkeley Data
• Women were more likely to target arts
faculty departments
• These department had lower admission
rates
• Hence, admission rates for women were
lower overall
• Despite fact that department by
department they were not
(c) Stephen Senn
Statistical Pitfalls
13
The Origin of Babies
• Rival theories
– Mulberry bushes
– Doctors’ bags
– Storks
• We shall begin a statistical investigation of
the last of these
(c) Stephen Senn
Statistical Pitfalls
14
Storks and Births in Europe
1500
1000
Births
Source
Matthews, 2000
Teaching
Statistics, 22, 3628
500
0
0
10000
20000
30000
Storks
(c) Stephen Senn
Statistical Pitfalls
15
Storks and Babies
• Larger countries tend to have more storks
• Larger countries tend to have more babies
• Hence the size of the country may be a
third factor responsible for the correlation
(c) Stephen Senn
Statistical Pitfalls
16
Storks per Area and Birth Rates in Europe
Births
30
20
10
0.00
0.05
0.10
Storks
(c) Stephen Senn
Statistical Pitfalls
17
Correlations: S per A, B Rate
Pearson correlation of S per A and B Rate = 0.161
P-Value = 0.536
(c) Stephen Senn
Statistical Pitfalls
18
Morals
• Watch out for confounding variables
• Where possible (it is not always possible)
design the study so that these are
accounted for
– For example in experiment have controls and
randomise
• Take care in jumping to conclusions
(c) Stephen Senn
Statistical Pitfalls
19
Invalid Inversion
• Most women do not get breast cancer
• However most breast cancer victims are
women
• You cannot reverse probability statements
• It is not generally true, for example, that
the probability of the evidence given
innocence is the same as the probability of
innocence given evidence
(c) Stephen Senn
Statistical Pitfalls
20
OJ Simpson’s Paradox
“Let me begin with a refrain constantly repeated by attorney Alan
Dershowitz during the trial. He declared that since fewer than 1 in a 1000
women who are abused by their mates go on to be killed by them, the
spousal abuse in the Simpsons' marriage was irrelevant to the case.”
John Allen Paulos
“the issue is whether a history of spousal abuse is necessarily a prelude to
murder”.
Alan Dershowitz
(c) Stephen Senn
Statistical Pitfalls
21
That Calculation
About 2000 women are murdered annually by a current or former mate in the
USA
About 2 million spousal assaults occur annually.
The ratio of one to the other is one in thousand.
Therefore a woman in an abusive relationship has only a 1 in 1000 chance of
being murdered by their mate each year
(c) Stephen Senn
Statistical Pitfalls
22
Mariage and Murder
‘Dershowitz had stated in the L.A. Times article that “the issue is whether a
history of spousal abuse is necessarily a prelude to murder”.
He’s Wrong - The issue is not whether abuse leads to murder but whether a
history of abuse helps identify the murderer.’
Kevin Hayes, University of Limerick
http://www.ul.ie/elements/Issue5/Oj.htm
http://www.maths.ul.ie/KH.htm
The following data taken from Haye’s website show data on women murdered in
the USA in 1992
(c) Stephen Senn
Statistical Pitfalls
23
‘Marriage’ and Murder
Current/Former
Husband or Mate
Other
Total
History
Abuse
715
175
890
No Abuse
715
3330
4045
1430
3506
4936
Totals
(c) Stephen Senn
Statistical Pitfalls
24
Abuse and Murder
Current/Former
Husband or Mate
Other
Total
History
Abuse
715
175
890
No Abuse
715
3330
4045
1430
3506
4936
Totals
(c) Stephen Senn
Statistical Pitfalls
25
OJ Simpson Revisited
“Given certain reasonable factual assumptions, it can be easily shown using
probability theory that if a man abuses his wife and she is later murdered, the
batterer is the murderer more than 80% of the time. (A nice demonstration of
this by Jon Merz and Jonathan Caulkins appeared in a recent issue of
Chance magazine.) Thus, without any further evidence, there was
mathematical warrant for immediate police suspicion of Mr. Simpson.”
John Allen Paulos
(c) Stephen Senn
Statistical Pitfalls
26
Selective Sampling
• We often make an assumption that the
data arrive without ‘side’
• This is not necessarily true
• One may have to think carefully about the
data-generation process
(c) Stephen Senn
Statistical Pitfalls
27
Abraham Wald (1902-1950)
• Rumanian/Hungarian/American,
mathematical statistician
• Inventor of decision theory
– brilliant and seminal paper of 1939
• Also innovator for sequential analysis
• Died in a plane crash in India
• Ironically, was employed by US military to
advise on plane safety in World War II
(c) Stephen Senn
Statistical Pitfalls
28
Wald’s Problem
• Returning planes were examined to see
where they had been hit
• Engines were rarely hit
• Fuel tanks very often
• Extra armour could be placed but not
everywhere
• Where should it be placed?
(c) Stephen Senn
Statistical Pitfalls
29
Wald
The Military and The Aircraft
• The US Military decided to reinforce the
fuel tanks
• That was where the most shots were
• They argued that therefore the fuel tanks
needed protection
(c) Stephen Senn
Statistical Pitfalls
30
Wald and the Aircraft
• Wald argued that the pattern of shots received
ought to be random
• The fact that it was not, indicated that this sample
was not random
• If the shots hit the fuel tank, the plane returned
safely
• If it hit the engine, it did not
• Solution: reinforce the engines not the fuel tanks!
(c) Stephen Senn
Statistical Pitfalls
31
Moral
• Think carefully about the process that has
led to the data in hand
• There may be subtle effects at work
• Don’t jump to conclusions
(c) Stephen Senn
Statistical Pitfalls
32
Question
• Studies have shown that if popes from the
13th to 19th century are compared to artists
of the same era, they died older
• It has been claimed that this shows the
effect of status in society on longevity
• Is there a snag?
Carrieri, M. P. and D. Serraino (2005). "Longevity of popes and
artists between the 13th and the 19th century.“
Int J Epidemiol 34(6): 1435-1436.
(c) Stephen Senn
Statistical Pitfalls
33
Key Questions
You should always ask
• What happened to the controls?
• Are there any hidden confounders?
• Has the probability statement been framed
the right way round?
• Is there a bias in the way the data are
collected?