Zero acquaintance accuracy = guessing traits right just by seeing someone briefly
Type I error = false positive false hopes, unnecessary side effects
Type II error = miss delay medication
Conservative is avoiding type I error
Alpha value determines how conservative (lower alpha) you are
Low alpha (conservative), less type I error but more type II
Alpha is the type I error rate
Power = likelihood not making type II error = finding effect when there is effect
Lower alpha less power
Higher sample size more power
Greater effect size more power
When effect size is small, bigger sample is needed
Power = 1 – kans op type 2 fout)
Sources of unsystematic variability:
- Measurement error: variables measured with less precise instrument or less careful
coding
- Irrelevant individual differences: can obscure difference between 2 groups and thus
weaken power, solution: repeated measures design to reduce impact of individuals
differences increased power
- Situation noise: weakens power
One-tailed test = whether schizophrenia drug reduces symptoms
Two-tailed test = whether schizophrenia drug reduces or increases symptoms
More power if:
- Larger alpha level
- Effect size is large
- Sample is large
- Lower levels of unsystematic variability
- Most appropriate statistical test
t-test = whether differences between two group means in independent-groups design is
statistically significant:
- Larger differences means larger t
- Larger SD lower t
- Larger sample size n larger t
In t-test not sample size but degrees of freedom = (n1-1) + (n2-1)
Confidence interval P495
, Hoorcollege 1
2 manieren om hypothesen te evalueren:
- NHST
- Bayesiaanse hypothese evaluatie
Cohen’s d = gestandaardiseerde effectsize = difference in means / sd
Effect size cohen’s d:
- Small = .20 medium = .50 large = .80
Effecten kunnen vaak in replicatieonderzoek niet opnieuw gevonden worden, oorzaken:
- Sloppy science
- Publication bias
Bayesiaanse hypothese evalutatie als alternatief voor NHST als reactie op replicatiecrisis,
want veel onderzoekers bedreven sloppy science om maar significance van 0.05 te halen,
p-waarde vervangen door Bayes factor (BF) = geeft relatieve steun in data voor H0 versus Ha
BF0a = 5 betekent steun in data 5x groter voor H0 dan voor Ha
Dus BF0a kleiner dan 1 wel effect
Als BF0a groter dan 1 geen effect
BF0a = 2 is zelfde als BFa0 = 0.5
f0
BF0a =
c0
f0 = fit (wordt letterlijk en figuurlijk kleiner als verschil in gemiddelden toeneemt)
c0 = specificiteit ( ‘=’ heel specifiek, ‘>’ beetje specifiek, ‘is niet teken’ niet specifiek)
Bayes factor heeft geen grenswaarde, dus remedie tegen:
- Questionable research practices, want incentive om analyses te manipuleren is weg
- Publication bias
Type I error = false positive false hopes, unnecessary side effects
Type II error = miss delay medication
Conservative is avoiding type I error
Alpha value determines how conservative (lower alpha) you are
Low alpha (conservative), less type I error but more type II
Alpha is the type I error rate
Power = likelihood not making type II error = finding effect when there is effect
Lower alpha less power
Higher sample size more power
Greater effect size more power
When effect size is small, bigger sample is needed
Power = 1 – kans op type 2 fout)
Sources of unsystematic variability:
- Measurement error: variables measured with less precise instrument or less careful
coding
- Irrelevant individual differences: can obscure difference between 2 groups and thus
weaken power, solution: repeated measures design to reduce impact of individuals
differences increased power
- Situation noise: weakens power
One-tailed test = whether schizophrenia drug reduces symptoms
Two-tailed test = whether schizophrenia drug reduces or increases symptoms
More power if:
- Larger alpha level
- Effect size is large
- Sample is large
- Lower levels of unsystematic variability
- Most appropriate statistical test
t-test = whether differences between two group means in independent-groups design is
statistically significant:
- Larger differences means larger t
- Larger SD lower t
- Larger sample size n larger t
In t-test not sample size but degrees of freedom = (n1-1) + (n2-1)
Confidence interval P495
, Hoorcollege 1
2 manieren om hypothesen te evalueren:
- NHST
- Bayesiaanse hypothese evaluatie
Cohen’s d = gestandaardiseerde effectsize = difference in means / sd
Effect size cohen’s d:
- Small = .20 medium = .50 large = .80
Effecten kunnen vaak in replicatieonderzoek niet opnieuw gevonden worden, oorzaken:
- Sloppy science
- Publication bias
Bayesiaanse hypothese evalutatie als alternatief voor NHST als reactie op replicatiecrisis,
want veel onderzoekers bedreven sloppy science om maar significance van 0.05 te halen,
p-waarde vervangen door Bayes factor (BF) = geeft relatieve steun in data voor H0 versus Ha
BF0a = 5 betekent steun in data 5x groter voor H0 dan voor Ha
Dus BF0a kleiner dan 1 wel effect
Als BF0a groter dan 1 geen effect
BF0a = 2 is zelfde als BFa0 = 0.5
f0
BF0a =
c0
f0 = fit (wordt letterlijk en figuurlijk kleiner als verschil in gemiddelden toeneemt)
c0 = specificiteit ( ‘=’ heel specifiek, ‘>’ beetje specifiek, ‘is niet teken’ niet specifiek)
Bayes factor heeft geen grenswaarde, dus remedie tegen:
- Questionable research practices, want incentive om analyses te manipuleren is weg
- Publication bias
