d hypothesis, errors, bias, confouding rss6 2014
TRANSCRIPT
Research Errors
Hashem Alhashemi
Objectives
• Hypothesis
• Hypothesis testing
• Types of error
• Confounding
• Bias
Hypothesis
An expected answer to a question
Do you have to start with a hypothesis all the time?
No
If you do not have enough background Info
Null hypothesis
A defendant is innocent until proven guilty
No relationship identified between variables tested
Alternative hypothesis
The opposite of Null hypothesis (your expectation).
Without torturing the Defendant (The Data).
𝛂 error/Type 1 error (P value)
• Set by agreement at 5%.
• To minimize the Probability of chance being the explanation for the findings.
• When Null hypothesis is true, we would tolerate the mistake of rejecting it 5 times, if the trial was repeated 100 times.
• The American error: America is Number 1.
• Finding a difference when there is no difference.
Confidence interval or margin of error (1-𝛂):
• The result of that is a 95% confidence interval.
• Means that if the trial was repeated 100 times the estimate would fall in the confidence range 95 times.
• The margin is a result of measurement error (what/How).
Hypothesis Testing
No Difference ?
Difference ! Reject null when True Find a Diff when ɸ Diff Type I error (American) (α) 0.05 due to Chance
Clinical Trial
No Difference Accept null when True Find ɸ Diff when ɸ Diff Confidence (1-α) 95% Truth CI 95%
Evidence of + effect True Effect =
Type 1 error
5%
?
𝛃 error/Type 2 error:
• Set by agreement at 20%.
• The British error: Blinding people about the difference when it actually exists (Palestinians').
Power of the study (1-𝛃):
• 𝛃 : When the alternative hypothesis is true, we would tolerate the mistake of rejecting it 20% of the time, if the trial was repeated 100 times.
• (1-𝛃): Means that if the trial was repeated 100 times the difference/effect would be found 80 times.
Difference
Hypothesis Testing
? Clinical Trial
No Difference ! Accept Null when False Find φ Diff when a Diff Type II Error (British) B = 20% (? Design)
Reject Null when False Find a Diff when a Diff Truth 1-B = 80% (Power)
Difference
lack of evidence lack of effect =
For thousands of years we did not have an evidence
about bacteria
(𝛃 error) Type 2 error
Lack of evidence did not prevent bacteria from killing
millions
• As you can see we are less tolerant to 𝛂 error because it may cause people to change their practice to a new way that is not superior (harmful?).
• On the other hand, we tolerate 𝛃 error more because its real side-effect is more tolerable than 𝛂 error.
• 𝛃 side effect is that the investigator may need to repeat the experiment under better conditions like larger sample size, or strict exclusion criteria.
𝛂 or 𝛃
Hypothesis Testing
No Difference ?
Difference ! Reject null when True Find a Diff when ɸ Diff Type I error (American) (α) 0.05 due to Chance
Clinical Trial
No Difference Accept null when True Find ɸ Diff when ɸ Diff Confidence (1-α) 95% Truth CI 95%
? Clinical Trial
No Difference ! Accept Null when False Find φ Diff when a Diff Type II Error (British) B = 20% (? Design)
Reject Null when False Find a Diff when a Diff Truth 1-B = 80% (Power)
Difference
Risk Matrix
Difference
Confounding
Is over estimation or under estimation of the relationship between two variables as a result of their relation to a third variable (confounder).
Confounding
C (Confounder i.e. smoking)
(Exposure i.e. coffee) E -- -- -- -- -- -- -- -- -- -- D (Disease i.e. lung cancer)
Confounding causes bias; however, bias is not limited to confounding.
Confounding in Reality
• We can adjust for it in the design and/or analysis.
• Design : Randomization, matching, & or restriction.
• Analysis : stratified analysis (MH), or multivariate analysis (Regression).
Unknown Confounders
known Confounders
Bias
Wrong/Invalid Conclusion
Bias
Any factor that can invalidate your conclusion by giving an alternative explanation for the results.
It can be due chance (external) or due systematic/design error (internal).
Bias
• an oblique or diagonal line of direction, especially across a woven fabric.
• a tendency or inclination, that prevents
someone from having a right judgment
(prejudice/wrong judgment).
Bias 101
Straight vs bias The bias runs at an angle to the straight and crossgrains. The true bias running at an exact 45-degree angle
Bias Bias
Truth
Bias Bias
Truth
Selection/sampling, Misclassification , Recall, Pt Response, Observer/assessor, loss of F/U, Publication, Language
Type of Bias
Misclassification Proper Case definition
Selection/sampling Randomization/ Allocation Concealment
Confounding Randomization& Regression
Patient Response Single blinding
Observer/assessor Double blinding
Crossover/contamination Intention to treat analysis
Recall Previous records & Documents
loss of F/U Visits, phone calls, e-mail
Questionnaire/interviewer Construct, pilot, & training
Non-response Restrict your population &
minimize non-response rate
Method of Prevention
Bias Prevention
Central Agency Next allocation
unknown Equal chance
each allocation
Regardless of intervention received Analysis by intention
not intervention
Bias
• General and design specific causes.
• Can not be accepted nor adjusted for.
• Better Avoided/prevented at any cost.
• Can happen in any stage of the study:
Design &, or conduct of the study (Pt selection,
Data Collection/entry).
Conclusion
Research Errors
• Type 1 & 2 we accept them as game roles from step one (sample size calculation).
• Confounding we can adjust/control for it in the design and/or analysis.
• Any thing else (we can not accept nor adjust for) is Bias. It should be avoided/prevented at any cost.
END
Conclusion
Sample size estimation
Statistical consideration
• Prevalence/treatment effect (from literature review)
• Maximum/margin of error
• Type I error 𝛂
• Type II error 𝛃
• Design effect
Practical consideration
• Personnel
• Resources
Being Wrong
Basics in Epidemiology & Biostatistics
Hashem Alhashemi MD, MPH, FRCPC Assistant Professor, KSAU-HS