STATISTICAL
ERRORS CAN
CAUSE DEATHS
Dr A. Indrayan
PhD(OhioState), FAMS, FRSS, FASc
Medical Errors
• Surgeon is trained for years but is held
responsible for just one death on table
• Physicians are punished for wrong
diagnosis or wrong treatment in just one
case
Statistical Errors – 1
• Wrong statistical methods can lead to
wrong conclusions
• Use of wrong results on a large section of
upcoming patients can kill many or
jeopardise their health but none is
punished
• Ineffective treatment is widely used
(killing many) and effective treatment is
missed (killed, but could have been
saved)
Statistical Errors – 2
• Results wrong by just 1%, when used on
millions, can threaten life and health of
many
• We need to be extra careful in our
statistical help to medical research
Funny Examples
• Almost all of us have more legs than the
world average
• Head in an oven and leg in a freezer, the
person is comfortable on AVERAGE
• In a trial on two persons, one cured, the
other didn’t. The efficacy is 50%.
• Duration of hospital stay for 5 patients is
(days) 3, 5, 41, 6, 5. The mean is 12 days
and the SD is 16.25 days.
Real Examples – 1
• Mean survival period of patients with
prostate cancer is 4 years and of lung
cancer is 12 years (age, treatment)
• Area under the concentration curves –
same area but different curves. Same for
ROC curves
Real Examples – 2
Response Rate in Mild and Severe Hyperthyroid Cases
Group
Number
of Subjects
Number
Responded
Response
Rate (%)
I. Treatment
group
40 80.0
Mild 86.7
Severe 60.0
II. Control
group
40 62.5
Mild 87.5
Severe 56.2
Response rate nearly same in mild and severe cases
but different overall??
Real Examples – 2 contd.
• Severe cases more in control group
Response Rate in Mild and Severe Hyperthyroid Cases
Group
Number
of Subjects
Number
Responded
Response Rate
(%)
I. Treatment
group
40 32 80.0
Mild 30 26 86.7
Severe 10 6 60.0
II. Control
group
40 25 62.5
Mild 8 7 87.5
Severe 32 18 56.2
Real Examples – 3
• Mean vs. Proportion: Iron supplementation
Rise in Hb level(g/dL)
0.4, 0.7, -0.9, 0.3, 0.1, 0.5, 0.9, 0.2
Mean rise = 0.275 g/dL (P > 0.05, NS)
7 out of 8 show rise: P(x ≥ 7) = 0.035 Sig.
• One outlier can yield
a significant relationship
Problems with Sample
• Biased: Survivors, Volunteers, Clinic
subjects
• Size: Too small – medically significant
effect missed, Too big – overpowered,
trivial effect significant
• Selection: Random – rarely adopted,
Nonrandom – statistical methods not
applicable
• Informed consent, inclusion/exclusion
criteria
Problems with Data
• Errors in elicitation (interview),
recording, suspicious lab. results
• Missing values, incomplete info, dropouts
• Outliers
• Correct assessment (pain, stress/anxiety,
family support) – nonavailabilty of
appropriate scales
• Data manipulation
Problems with Analysis – 1
• Mean or proportion, mean or median
• Sensitivity or predictivity
• Looking for linearity
• Ignoring assumptions (Gaussian,
independence, homoscedasticity)
• Confounding, multicollinearity
Problems with Analysis – 2
• Categories of continuous variables
• Cherry picking variables and statistical
methods
• Forgetting baseline values
• Analysis done by semi-killed
professionals using software (everyone is
a statistician)
• Statisticians’ complicity
Problems with P-values
• Multiple P-values, multiple comparisons
• Accumulation of Type-I error
• Data dredging (torture till confess)
• Accepting a null
• Null not medically significant effect
• Probability statement –results remain inexact
• P-values are only for sampling fluctuations but
generally incorporate chance of errors due to
faulty design and faulty data by default
Problems with Interpretation
• Over-dependence on P-values
• Data skills not as much widely available
as the data
• Ignoring biological plausibility,
corroborative evidence
• Medically significant effect
• Association/correlation as cause-effect
• Probabilities work in the long run and
may miserably fail in individual cases
Conclusion
• Medical biostatisticians too are in the
profession of saving lives and reduce
suffering
• Need to extremely cautious and
professional in analysis, interpretation
and advice

Statistical errors can cause deaths

  • 1.
    STATISTICAL ERRORS CAN CAUSE DEATHS DrA. Indrayan PhD(OhioState), FAMS, FRSS, FASc
  • 2.
    Medical Errors • Surgeonis trained for years but is held responsible for just one death on table • Physicians are punished for wrong diagnosis or wrong treatment in just one case
  • 3.
    Statistical Errors –1 • Wrong statistical methods can lead to wrong conclusions • Use of wrong results on a large section of upcoming patients can kill many or jeopardise their health but none is punished • Ineffective treatment is widely used (killing many) and effective treatment is missed (killed, but could have been saved)
  • 4.
    Statistical Errors –2 • Results wrong by just 1%, when used on millions, can threaten life and health of many • We need to be extra careful in our statistical help to medical research
  • 5.
    Funny Examples • Almostall of us have more legs than the world average • Head in an oven and leg in a freezer, the person is comfortable on AVERAGE • In a trial on two persons, one cured, the other didn’t. The efficacy is 50%. • Duration of hospital stay for 5 patients is (days) 3, 5, 41, 6, 5. The mean is 12 days and the SD is 16.25 days.
  • 6.
    Real Examples –1 • Mean survival period of patients with prostate cancer is 4 years and of lung cancer is 12 years (age, treatment) • Area under the concentration curves – same area but different curves. Same for ROC curves
  • 7.
    Real Examples –2 Response Rate in Mild and Severe Hyperthyroid Cases Group Number of Subjects Number Responded Response Rate (%) I. Treatment group 40 80.0 Mild 86.7 Severe 60.0 II. Control group 40 62.5 Mild 87.5 Severe 56.2 Response rate nearly same in mild and severe cases but different overall??
  • 8.
    Real Examples –2 contd. • Severe cases more in control group Response Rate in Mild and Severe Hyperthyroid Cases Group Number of Subjects Number Responded Response Rate (%) I. Treatment group 40 32 80.0 Mild 30 26 86.7 Severe 10 6 60.0 II. Control group 40 25 62.5 Mild 8 7 87.5 Severe 32 18 56.2
  • 9.
    Real Examples –3 • Mean vs. Proportion: Iron supplementation Rise in Hb level(g/dL) 0.4, 0.7, -0.9, 0.3, 0.1, 0.5, 0.9, 0.2 Mean rise = 0.275 g/dL (P > 0.05, NS) 7 out of 8 show rise: P(x ≥ 7) = 0.035 Sig. • One outlier can yield a significant relationship
  • 10.
    Problems with Sample •Biased: Survivors, Volunteers, Clinic subjects • Size: Too small – medically significant effect missed, Too big – overpowered, trivial effect significant • Selection: Random – rarely adopted, Nonrandom – statistical methods not applicable • Informed consent, inclusion/exclusion criteria
  • 11.
    Problems with Data •Errors in elicitation (interview), recording, suspicious lab. results • Missing values, incomplete info, dropouts • Outliers • Correct assessment (pain, stress/anxiety, family support) – nonavailabilty of appropriate scales • Data manipulation
  • 12.
    Problems with Analysis– 1 • Mean or proportion, mean or median • Sensitivity or predictivity • Looking for linearity • Ignoring assumptions (Gaussian, independence, homoscedasticity) • Confounding, multicollinearity
  • 13.
    Problems with Analysis– 2 • Categories of continuous variables • Cherry picking variables and statistical methods • Forgetting baseline values • Analysis done by semi-killed professionals using software (everyone is a statistician) • Statisticians’ complicity
  • 14.
    Problems with P-values •Multiple P-values, multiple comparisons • Accumulation of Type-I error • Data dredging (torture till confess) • Accepting a null • Null not medically significant effect • Probability statement –results remain inexact • P-values are only for sampling fluctuations but generally incorporate chance of errors due to faulty design and faulty data by default
  • 15.
    Problems with Interpretation •Over-dependence on P-values • Data skills not as much widely available as the data • Ignoring biological plausibility, corroborative evidence • Medically significant effect • Association/correlation as cause-effect • Probabilities work in the long run and may miserably fail in individual cases
  • 16.
    Conclusion • Medical biostatisticianstoo are in the profession of saving lives and reduce suffering • Need to extremely cautious and professional in analysis, interpretation and advice