9  Simpson’s paradox

Example 9.1 (Simpson’s paradox). There are two doctors, Dr. Lee and Dr. Wong, performing two types of surgeries — heart surgery (hard) and band-aid removal (easy). Dr. Lee has higher overall surgery success rate. Is Dr. Lee necessarily a better doctor than Dr. Wong?

No. Consider the following example:

Dr. Lee Dr. Wong
Heart Band-Aid Total Heart Band-Aid Total
Success 2 81 83 70 10 80
Failure 8 9 17 20 0 20
Success rate 20% 90% 83% 78% 100% 80%

The truth is Dr. Lee has overall higher success rate because he only does easy surgeries (band-aid removal). Dr. Wong does mostly hard surgeries and thus has a lower overall success rate. Yet, he is better at each single type of surgery. To formalize the argument, let \(S\): successful surgery; \(D\): treated by Dr. Lee, \(D^{c}\): treated by Dr. Wong; \(E\): heart surgery, \(E^{c}\): band-aid removal. Dr. Wong is better at each type of surgery, \[\begin{aligned}P(S|D,E) & <P(S|D^{c},E)\\ P(S|D,E^{c}) & <P(S|D^{c},E^{c}); \end{aligned}\] But, Dr. Lee has a higher overall successful rate, \[P(S|D)>P(S|D^{c}).\] This is because there is a “confounder” \(E\): \[P(S|D)=\underbrace{P(S|D,E)}_{<P(S|D^{c},E)}\underbrace{P(E|D)}_{\textrm{weight}}+\underbrace{P(S|D,E^{c})}_{<P(S|D^{c},E^{c})}\underbrace{P(E^{c}|D)}_{\textrm{weight}}.\]

A confounder is a variable that influences with both explanatory variable and the outcome variable, which therefore “confounds” the correlation between the two. In our example, the type of surgery (\(E\)) is associated with both the doctor and the outcome. Without the confounder being controlled, it is impossible to draw valid conclusions from the statistics.

In general terms, Simpson’s paradox refers to the paradox in which a trend that appears across different groups of aggregate data is the reverse of the trend that appears when the aggregate data is broken up into its components. It is one of the most common sources of statistical misuse. Here is another example.

Example 9.2 (UC Berkeley gender bias). One of the best-known examples of Simpson’s paradox comes from a study of gender bias among graduate school admissions to University of California, Berkeley. The admission figures for the fall of 1973 showed that men applying were more likely than women to be admitted, and the difference was so large that it was unlikely to be due to chance.

Male Female
Applicants Admitted Applicants Admitted
Total 8,442 44% 4,321 35%

However, when taking into account the information about departments being applied to, the conclusion turns to the opposite: in most departments, the admission rate for women is higher than men. The lower overall admission rate is caused by the fact that women tended to apply to more competitive departments with lower rates of admission, whereas men tended to apply to less competitive departments with higher rates of admission.

Department Male Female
Applicants Admitted Applicants Admitted
A 825 62% 108 82%
B 560 63% 25 68%
C 325 37% 593 34%
D 417 33% 375 35%
E 191 28% 393 24%
F 373 6% 341 7%
Total 2691 45% 1835 30%

See https://setosa.io/simpsons for a really good illustration of the Simpson’s paradox.

R demostration

# R has a built-in dataset `UCBAdmissions`
# we convert it to data frame for analysis
data <- as.data.frame(UCBAdmissions)

# browse the first a few rows
head(data)
     Admit Gender Dept Freq
1 Admitted   Male    A  512
2 Rejected   Male    A  313
3 Admitted Female    A   89
4 Rejected Female    A   19
5 Admitted   Male    B  353
6 Rejected   Male    B  207
# subset of the data with only admissions
data <- subset(data, Admit == 'Admitted')

# number of admissions by Gender
aggregate(Freq ~ Gender, data = data, FUN = sum)
  Gender Freq
1   Male 1198
2 Female  557
# number of admissions by Gender and Department
aggregate(Freq ~ Gender + Dept, data = data, FUN = sum)
   Gender Dept Freq
1    Male    A  512
2  Female    A   89
3    Male    B  353
4  Female    B   17
5    Male    C  120
6  Female    C  202
7    Male    D  138
8  Female    D  131
9    Male    E   53
10 Female    E   94
11   Male    F   22
12 Female    F   24
The importance of conditional thinking

Whenever we talk about probability or statistics, always remind ourselves what we are the conditioning on. Any statistical reasoning without specifying the conditions can be misleading. We are prone to such fallacies everyday everywhere.

  • “10 millions new jobs were added during the term of President X.” But it doesn’t tell you this was achieved conditioned on that the economy had just had the worst recession.
  • “Private schools’ graduates earned 50% more than those graduated from public schools.” But it doesn’t tell you the background of those students who enrolled in private schools.

Be vigilant to these claims when you see them next time.