SPED102 Lecture Notes - Lecture 7: Comorbidity, Cold Fusion, False Positives And False Negatives
SPED102 Lecture
Week 7
VII: Use and Misuse of Statistics
Overview
• Multiple comparisons fallacy and the sharpshooter fallacy: research applications
• Other strategies to bias your research
• Some tricks to misrepresent your presentation of data
Biasing results without (outright) fraud
• For those who are amoral and unethical the question arises
• How can we do research in such a way to maximise a desired and predetermined
outcome without overt fraud (i.e. Fabricating data)
Revision: the (not so) magic p value
• The p value reported in research gives an approximation the probability that you have false
positive result (compared to a true negative)
More damn statistics!
• This is based on the assumption that you only do ONE comparison (test)
• If you do multiple comparisons, the chance of getting a false positive result is above the
allowable 5%
The multiple comparisons fallacy
• Many sneaky (or incompetent) ways to do lots of comparisons (and hide them)
Texas Sharpshooter Fallacy
• Choosing your outcomes after the experiment
• I have predicted the all lotto numbers with 100% accuracy for 6 consecutive weeks… what
are the chances?
• Don't believe me?
Multiple Comparisons and Sharpshooting Applied to Research
• Multiple outcomes
• Switching outcomes after the research
• Subgroup analysis
• Open-ended criteria
• Arbitrary termination of studies
• Selective publication
Multiple Outcomes
• Have lots of outcomes (dependent variables)
• Some of them are likely to be significant by chance
Humphries et al (1992)
• Compared Sensory Integration Therapy to
• Comparison intervention (perceptual-motor program)
• No treatment
How to Handle Multiple Outcomes
• Other outcomes are secondary
find more resources at oneclass.com
find more resources at oneclass.com
Document Summary
Overview: multiple comparisons fallacy and the sharpshooter fallacy: research applications, other strategies to bias your research, some tricks to misrepresent your presentation of data. Biasing results without (outright) fraud: for those who are amoral and unethical the question arises, how can we do research in such a way to maximise a desired and predetermined outcome without overt fraud (i. e. fabricating data) Revision: the (not so) magic p value: the p value reported in research gives an approximation the probability that you have false positive result (compared to a true negative) More damn statistics: this is based on the assumption that you only do one comparison (test) If you do multiple comparisons, the chance of getting a false positive result is above the allowable 5% The multiple comparisons fallacy: many sneaky (or incompetent) ways to do lots of comparisons (and hide them) Texas sharpshooter fallacy: choosing your outcomes after the experiment.