Biostatistics in Practice

1
Biostatistics in Practice
Session 1 Design and Fundamentals of Inference
Peter D. Christenson Biostatistician http//gcrc
.humc.edu/Biostat
2
Local Biostatistical Resources

• Biostatistician Peter Christenson
• Assist with study design, protocol development
• Minor, limited analysis of data
• Major analysis as investigator with FTE on
  funded studies
• GCRC and non-GCRC studies
• Biostatistics short courses 6 weeks 2x/yr
• GCRC computer laboratory in RB-3
• For GCRC studies
• Statistical, graphics, database software
• Webpage www.gcrc.humc.edu/Biostat

3
Readings for Session 1from StatisticalPractice.co
m

• Is statistics hard?
• An all-important foundation
• Cause and effect
• Study design
• Random samples / randomization

4
Is Statistics Hard?

• Statistics is used for
• Description.
• Inference.
• Inference requires a peculiar way of thinking
• It is backwards.
• The statistical methods are convoluted.
• Lack of evidence is not evidence of the opposite.

5
Backwards and Convoluted
Example We measure cholesterol in 100 subjects
before and after drug therapy. They all
decrease. Common sense conclusion Therapy
probably works. Statistical approach is
backwards Step 1 First, assume a conclusion of
no therapy effect. Step 2 Does the data
contradict this assumption? If so, then conclude
that therapy works.
6
Statistics Can Work AGAINST Good Science
Good science We again assume no therapy effect,
where this effect is pre-specified and specific,
i.e. change in LDL cholesterol. Data has a 5
chance of occurring under this assumption, so we
conclude that therapy works. OK (5 chance we are
wrong) Bad science We again assume no therapy
effect, but this effect is defined by looking
around all of the measured characteristics after
performing the study. Assumption is contradicted
for one characteristic, and we conclude that
therapy works. NOT OK (This is almost guaranteed
to happen if many characteristics are examined.)
7
Lack of Evidence ? Proof of the Opposite
Example, continued Cholesterol changes range
from -20 to 15 in 100 subjects before and after
drug therapy. Conclusion The data does not
contradict the assumption of no therapy effect.
Lack of proof of therapy effectiveness does not
imply the drug is not effective. Cholesterol
changes range from -2 to 1 in 100 subjects
before and after drug therapy. Here, lack of
proof of effectiveness does imply that the drug
is practically ineffective.
8
Cause and EffectSee p. 1 of readings.

• Association is not causation.
• Examples. See p. 1 of readings
• Crimes ? as church membership ?.
• Welfare causes paupers?
• Another example
• Pre-marital cohabitation causes divorce? See next
  slide. The only hard data is that the divorce
  rate is greater for those who cohabitated than
  for those who did not. The rest is speculation.

9
Cause and Effect No, Association only
10
Cause and Effect Limitations of Data ItselfSee
pp 2-3 of readings.

• Data analysis can often only lead, not prove.
• Need mechanisms, randomization, independent
  sources of evidence (See p. 3 of readings,
  smoking and lung cancer).
• Scientific fields with decreasing ability to
  assign causality
• Physical sciences.
• Clinical trials.
• Epidemiology and sociology.
• History.

11
Study Design General RequirementsSee pp 2-7 of
readings.

1. Clearly stated research question and primary
   outcome measure.
2. Project must be feasible.
3. Carefully considered all facets of protocol and
   data items.
4. Keep it simple.
5. Research has consequences.

12
Study Design TypesSee pp. 7-13 of readings.

• Observational studies
• Surveys
• Cross sectional / longitudinal
• Cohort / case-control
• Interventional / controlled clinical trials

13
Study Design IssuesSee pp 13-18 of readings.

1. Comparison groups (none, self, internal,
   external)
2. Study size considerations
3. Many subjects vs. many measurements
4. Paired vs. unpaired data
5. Parallel groups vs. cross-over studies
6. Repeated measures
7. Intention-to-treat and meta-analysis

14
Random Samples / RandomizationSee pp 1-2 of
readings.

• Random samples of subjects
• Randomization of treatment
• Validity
• Generalizability

15
Review Questions and Reference
Questions on the next five slides are from an
excellent biostatistics textbook Martin Bland,
Introduction to Medical Statistics, 3rd ed.,
Oxford University Press, 2000. The author states
that some questions are quite hard. If you
score 1 for a correct answer, -1 for an
incorrect answer, and 0 for a part you omitted, I
would regard 40 at the pass level, 50 as good,
60 as very good, and 70 as excellent. Some
questions may be ambiguous, so you will not score
100.
16
Question AAnswer True or False for each part (5
answers)

• When testing a new medical treatment, suitable
  control groups include patients who
• Are treated by a different MD at the same time.
• Are treated by a different hospital.
• Are not willing to receive the new treatment.
• Were treated by the same doctor in the past.
• Are not suitable for the new treatment.

17
Question BAnswer True or False for each part (5
answers)

• In an experiment to compare 2 treatments,
  subjects are allocated using random numbers so
  that
• The sample may be referred to a known population.
• When deciding whether to include the subject, we
  do not know which treatment the subject will
  receive.
• The subjects will get the treatment best suited
  to them.
• The two groups will be similar, apart for
  treatment.
• Treatments are assigned to meet subject
  characteristics.

18
Question CAnswer True or False for each part (5
answers)

• In a double-blind clinical trial
• The patients do not know which treatment they
  receive.
• Each patient receives a placebo.
• The patients are blind to the fact that they are
  in a trial.
• Each patient receives both treatments.
• The clinician making assessment does not know
  which treatment the patient receives.

19
Question DAnswer True or False for each part (5
answers)

• Cross-over designs for clinical trials
• May be used to compare several treatments.
• Involve no randomization.
• Require fewer patients than do designs comparing
  independent groups.
• Are useful for comparing treatments intended to
  alleviate chronic symptoms.
• Use the patient as his or her own control.

20
Question EAnswer True or False for each part (5
answers)

• Placebos are useful in clinical trials
• When 2 apparently similar active treatments are
  to be compared.
• To guarantee comparability in non-randomized
  trials.
• Because the fact of being treated may itself
  produce a response.
• Because they may help to conceal the subjects
  treatment from clinicians who assess them.
• When an active treatment is compared to no
  treatment.

21
Answers to Questions

1. F,F,F,F,F.
2. F,T,F,T,F.
3. T,F,F,F,T.
4. T,F,T,T,T.
5. F,F,T,T,T.