John W. Tukey

1
John W. Tukeys Multiple Contributions to
Statistics at Merck

• Joseph F. Heyse
• Merck Research Laboratories
• Third International Conference on Multiple
  Comparisons
• Bethesda, Maryland
• August 5, 2002

2
Overview

• Professor John W. Tukey began consulting with
  Merck Sharp and Dohme Research Laboratories in
  1953 and continued until 2000.
• Prior to 1953 John was a consultant to Merck in
  the area of manufacturing.
• Through the years John made major contributions
  to the statistical aspects of all major research
  disciplines
• His consultations led to the establishment of
  Merck and industry standards for several
  statistical approaches

3
Areas of Involvement

• Safety assessment
• Clinical trials
• Laboratory quality control
• Clinical safety analyses
• Health economics
• Gene expression and microarray data
• Use of graphics

4
Agenda for June 1, 2000 Meeting

• 1. Multiple comparisons Applications of the
  False Discovery Rate to Vaccine Adverse
  Experience Data
• 2. Transformations for analyzing parasite count
  data with many zero counts
• 3. Use of TaqMan assay for gene expression
• 4. Error models for microarray data

5
Examples

• Trend testing in safety assessment
• Adjusting for multiplicity in rodent
  carcinogenicity studies
• Multiplicity applied to estimated variances

6
Trend Test for Dose Response(Tukey et al., 1985)

• Trend defined as progressiveness of response with
  increasing dose
• Three sets of carriers for the candidate set
• Arithmetic
• Ordinal
• Arithmetic-Logarithmic
• Statistical assessment for trend is taken as most
  extreme P-value computed from candidate set
• NOSTASOT Dose - No Statistical Significance of
  Trend Dose - Highest dose through which test for
  trend is N.S.

7
Properties of Trend Test

• Trend test inflates P-values slightly in
  conservative direction for safety assessment
• Adjusted trend test reported by Capizzi et al.
  (1992) favorable to other tests against ordered
  alternative hypothesis
• NOSTASOT is closed sequential procedure
• Tukey et al. also proposed an adjustment
  procedure for multiple safety assessment
  parameters with unknown correlation

8
Example
Summary statistics for toxicity study in dogs
s2 0.039 with 31 d.f.
9
Example
Trend Test Results
10
Example
NOSTASOT Analysis
NOSTASOT dose is D2 1.0 mg/kg/day
11
Example
Adjusted P-value for Dunnetts procedure
12
Multiple Significance Testing in Rodent
Carcinogenicity Experiments

• Mantel (1980) credits Tukey with proposal to
  adjust multiple P-values in carcinogenicity
  experiments
• where P1 is the smallest observed P-value, k1 is
  the number of tumor types that could have
  attained P1
• These methods have been improved by several
  authors and now are commonly applied

13
Grouping Based on Estimated Variances

• The naïve procedure of weighting the results of
  different experiments inversely to their
  estimated variance is unsatisfactory
• Cochran (1954) introduced the idea of partial
  weighting in which ½ to ²/3 of the studies that
  appear less variable are assigned equal weight
• Mosteller and Tukey (1984) treated the more
  realistic case with the possible presence of
  interaction
• Ciminera et al. (1993) applied those methods in
  the multicenter clinical trial setting.

14
Grouping of Centers Based on Estimated
Variances(Ciminera et al., 1994)
15
Insights on Statistics

• Randomization is the only thing you can safely
  assume when analyzing clinical trial data
• There is no such thing as a null effect
• There is always interaction
• Having only two points is the only time you
  should pretend that you have a linear
  relationship and in these cases, you should get
  more data

16
Insights on Character

• The best thing about being a statistician is
  that you get to play in other peoples
  backyards. (J.W.T.)
• Remember that you are a guest and need to bring
  your manners and respect.
• Its all about relationships.

17
What would John think of these remarks?

• Thank you for the kind words, but . . .

you could have said them using fewer slides.
18
References

• Capizzi T, Survill TT, Heyse JF, and Malani H
  An empirical and simulated comparison of some
  tests for detecting progressiveness of response
  with increasing doses of a compound. Biometrical
  Journal, 34275-289, 1992.
• Ciminera JL, Heyse JF, Nguyen HH, and Tukey JW
  Evaluation of multicentre clinical trial data
  using adaptations of the Mosteller-Tukey
  procedure. Statistics in Medicine, 121047-1061,
  1993.
• Ciminera JL, Heyse JF, Nguyen HH, and Tukey JW
  Tests for qualitative treatment-by-centre
  interaction using a pushback procedure.
  Statistics in Medicine, 121033-1045, 1993.

19
References (cont.)

• Cox JL, Heyse JF, and Tukey JW Efficacy
  estimates from parasite count data that include
  zero counts. Experimental Parasitology, 961-8,
  2000.
• Heyse JF and Rom D Adjusting for multiplicity
  of statistical tests in the analysis of
  carcinogenicity studies. Biometrical Journal,
  30883-896, 1988.
• Mantel N Assessing laboratory evidence for
  neoplastic activity. Biometrics, 36381-399,
  1980.
• Mantel N, Tukey JW, Ciminera JL, and Heyse JF
  Tumorigenicity assays, including use of the
  jackknife. Biometrical Journal, 24579-596, 1982.

20
References (cont.)

• Tukey JW, Ciminera JL, and Heyse JF Testing the
  statistical certainty of a response to increasing
  doses of a drug. Biometrics, 41295-301, 1985.