Design and Analysis of Clinical Study 2. Bias and Confounders

1
Design and Analysis of Clinical Study 2. Bias
and Confounders

• Dr. Tuan V. Nguyen
• Garvan Institute of Medical Research
• Sydney, Australia

2
Biases Confounding

• Bias means difference from the truth
• There are 3 types of bias
• Selection bias
• Information bias
• Confounding

3
Selection Bias

• Non-representativeness
• Patients referred for specialist care are
  different from those in the community
• Used hospitalized smokers as the exposed and
  healthy volunteer non-smokers as the unexposed
• Migration bias.
• People with chronic lung disease tend to move out
  of urban areas those with psychiatric problems
  seek the anonymity of cities
• High dropout rates.
• Those who drop out of a study tend to be
  different from those continuing

4
Selection bias - Berkson
(a) General population odds ratio 1.06
Respiratory disease Diseases of bones and joints Diseases of bones and joints Total
Respiratory disease Yes No Total
Yes 17 207 234
No 184 2376 2560
(b) Hospitalized population Odds ratio 4.06
Respiratory disease Diseases of bones and joints Diseases of bones and joints Total
Respiratory disease Yes No Total
Yes 5 15 20
No 18 219 237
Ref Roberts RS, et al. J Chron Dis 31119-28
5
Bias by Indication

• Whenever we compare a group of patients who use a
  drug to those who dont in a non experimental
  observational study (cohort, not randomized).
• The 2 groups differ in many respects Bias by
  indication.
• Comparison of hypertensive patients who are on
  minoxidil or hydralazine and those on other
  agents
• That patients on those agents have higher BP
• Is it because they dont work as well ?
• No, the opposite. They are reserved for those
  with severe resistant hypertension.
• That is the indication for those agents.

6
Survivor Treatment Bias

• Patients who received statin during admission for
  MI had much lower in-hospital mortality.
• Statin?
• The ones who died are different.
• Some died very soon after admission (no statin).
• Some were so sick that they were treated with
  multiple drugs, modalities, ICU etc.
• No statin

7
Information Bias

• Response Bias occurs when subjects give
  inaccurate responses.
• Measurement Bias occurs when instruments are
  faulty
• Observer error
• A process tends to show improvement when being
  observed. (Hawthorne Effect)

8
Confounders

• Confounders act by being associated with both a
  risk factor and outcome in a way that makes the
  two seem related.

Poor Maternal Nutrition
Low Birth Weight
Low Socioeconomic Class
9
Example of Confounder - Sex
Treatment Outcome Outcome RR 0.2
Treatment Poor Good RR 0.2
New Rx 30 1970 RR 0.2
Standard Rs 150 1850 RR 0.2
Treatment Outcome Outcome RR 0.4
Treatment Poor Good RR 0.4
New Rx 20 980 RR 0.4
Standard Rs 50 950 RR 0.4
Males
Treatment Outcome Outcome RR 0.1
Treatment Poor Good RR 0.1
New Rx 10 990 RR 0.1
Standard Rs 100 900 RR 0.1
Females
10
Strategies for Reducing Biases

• Have clear and precise definitions (e.g. for
  cases controlsexposurecriteria for
  inclusion/exclusion)
• Blinding where appropriate
• Reduce measurement error by quality control
• Careful check of study design choice of
  subjects ascertainment of disease and
  exposureplanning of questionnaires methods of
  data collection.

11
How to Deal with Confouders 1

• Think about possible confounders at the design
  stage, and gather data on all possible
  confounders.
• A quick test about a possible confounder is to
  check whether it is unevenly distributed between
  study and comparison groups.
• Suspect confounding if the odds ratio gets
  altered after adjusting for another factor.

12
How to Deal with Confouders 2

• Design stage
• Strict inclusion criteria
• Matching
• Randomization
• Analysis stage
• Do analysis by adjusting for several strata of
  the confounding variable
• Multiple regression analysis

13
How to Check for Confouders

• First calculate Odds Ratio for the exposure
  variable.
• Next calculate odds ratio for different strata of
  the confounding variable
• If the odds ratios are not materially different
  then there is no confounding.

14
Validity

• Are the conclusions true?
• Common threats to validity
• Selection bias
• Measurement bias
• Differential loss of subjects
• Confounders
• Unexpected events
• Hawthorne effect

15
How to Ensure Validity

• Have a control group. Helps against confounding,
  unexpected events, Hawthorne effect.
• Random assignment of subjects to different
  groups.
• Before / After measurements.
• Carefully prepared research designs.
• Quality control of equipment
• Knowledge of environmental events especially if
  the study is of long duration.
• Unobtrusive methods of observation.

16
Cause-and-Effect Relationship

• Strength of Research Design is most important

1. Well - conducted randomized controlled trials
(adequate sample size blinding
standardized methods of measurement and
analysis) 2. Cohort studies - next best
(minimize selection measurement bias check for
confounders)
17
Evidence for cause-and-effect

• Reversible association (removal of cause
  decreases risk)
• Consistency (several studies come up with same
  findings)
• Biological plausibility
• Specificity
• Analogy

• Temporal sequence (cause must precede effect)
• Strength of association (Relative risk or odds
  ratio)
• Dose-Response relationship

18
Flow chart for cause-and-effect inference
Association (O.R. R.R. Pearsons r)
Yes
No
Bias
Not likely
Likely
Chance
Excluded
Possible
No
Possible
Error
CAUSE