Seminars Flashcards
What is the difference between selective accessibility and scale distortion?
Selective accessibility - a change in anchor, changes the entire representation of the thing being represented
Scale distortion - a change in anchor will result in a different scale used to describe the stimulus - does not change how you represent that stimulus; important as it means that objective scales can become subjective scales
What the three predictions from Frederick and Mochon’s scale distortion theory of anchoring?
Prediction 1 = participants are provided with a scale with which they are expected to choose a stimulus that fits that value, this then anchors responses when asked to estimate further values
Prediction 2 = participants provide a self-selected scale value following provision of a stimulus, this then anchors responses when asked to estimate further values
Prediction 3 = these scaling effects will only occur with values estimated in the same numerical scale, even when the scales pertain to the same category of information ie weight
What did experiment 1 test, which prediction did it support and what were its key findings?
(Frederick and Mochon)
Prediction 1
Select 1000lb animal/15 choices ordered by weight; half asked to estimate weight of wolf before
Wolf group made larger animal choices
Demonstrates scaling effect
What did experiment 2a test, which prediction did it support and what were its key findings?
Prediction 1
Select food closest to 400cal/13 foods ordered by calories;
half asked to estimate apple cal before
Apple group picked more calorific items
Replication of experiment 1 = reliability + not dependent on scale type
What did experiment 2b test, which prediction did it support and what were its key findings?
Prediction 1
Draw a line on a glass where 200 cal of hersheys syrup would be; half estimated cal in a hersheys kiss before
Kiss condition (smaller calorie anchor) reported 200 cal larger compared to control
Replication of 2a - reliability
What did experiment 3a test, which prediction did it support and what were its key findings?
Prediction 2+3
Estimate weight of giraffe alone (lbs); or estimate weight of raccoon then giraffe (lbs)
Raccoon anchor lowered estimations of giraffes in subsequent ratings
Didn’t show a scale first = self anchored, a more classical approach
What did experiment 3b test, which prediction did it support and what were its key findings?
Prediction 2+3
Weight of giraffe (lbs); OR first estimate blue whale (lbs) or blue whale (tonnes)
Blue whale lbs impacted giraffe estimations; blue whale tonnes did not
Distortion is scale specific
What did experiment 4 test, which prediction did it support and what were its key findings?
Prediction 2+3
Estimate 3 features of a giraffe: weight (lbs), height (ft), weight relative to a grand piano (1= piano much heavier, 7 = giraffe much heavier) ; half participants estimated raccoon (lbs) before
Giraffe = judged less when previously asked to estimate weight of raccoon; no effects on the other characteristics
Distortion is scale specific; demonstrated that through other metrics (giraffe height, comparative heaviness) that the mental representation of the item remains the same
What did experiment 5 test, which prediction did it support and what were its key findings?
Prediction 2+3
Estimate strawberry (cal) or domino’s pizza; estimate: a) cal b) g fat c) lbs lost if not eaten d) number per serving e) number of days a rat could survive off (all concerning McDonalds fries)
Only the calorie anchor had an effect
Replication of studies 3a/b/4
What do these results suggest about the selective availability theory?
One set of experiments does not nullify; still a degree of difference between subjective and objective scales
Where subjective scales ie small and large are used, meaning can be interpreted contextually – either clearly ie large mouse running up a small elephants trunk; or ambiguously ie lowering customer expectations about a product may increase satisfaction ratings but this is unsure whether it is because of a change in experience of the product or a mere scaling effect
Objective scales ie inches, pounds etc are classically thought to be immune to contextual effects and the changes that occur are due to changes in how the target stimulus is represented – this research challenges this – anchoring effects = response language effects
Still a compelling argument for when respondents have a large pool or relevant knowledge to draw from
Judging non scalable information may still be subject to anchoring ie viewing images or houses etc
What are the practical implications for Frederick and Mochon’s findings?
Question the validity of all ‘objective’ numerical scales
Response scale effects on objective scales may have different behavioural consequences compared to subjective scales - objective have a meaningful zero point; context effects = judgement relativity of labels
Can explain a function of the mind in terms of language ie its not architecture
Useful to be aware of when making financial decisions
What are multiple comparisons for fMRI and why are they important?
130,000 voxels = probability of at least one false positive almost certain; needs accounting for
What methods of correcting for multiple comparisons are there?
Correction holds false positives to a standard rate
Family-wise error rate (FEWR) ie Bonferroni correction = 5% chance of 1+ FP in whole data set (conservative measure - misses some data) OR Gaussian Random Field Theory
False discovery rate (FDR) = 5% of the results are expected to be false positives = less conservative than FEWR, balances statistical power better
Standard statistical thresholds ie P=<0.001 and low minimum clusters ie K=>8 are inefficient control for multiple comparisons - sometimes too conservative sometimes too liberal
What did the Salmon scanners find?
Used an uncorrected P value of <0.001 and found activation in several voxels in a dead salmons brain
When corrected for multiple comparisons (using FDR and FEWR separately) - no activation
What is the difference between multiple comparisons and a non-independence error?
Non-independence error = inflation of cluster-wise statistical estimates when constituent voxels were selected using the same statistical measure ie the correlation value of a voxel cluster will be inflated if the voxels were originally selected due to the criteria that they have a high correlation
Multiple comparisons = related to the prevalence of false positives present across the set of selected voxels at the first stage; not just with fMRI but all data gathered over multiple tests
What is the problem with low statistical power?
Reduced likelihood of finding true positive
True positives found will be of inflated significance
Increased likelihood of false positives
Absolute minimum recommendation for fMRI is 20 participants
What solutions are there for low statistical power?
Caution in extrapolating from effect sizes in small samples - likely inflated
Sample size justified by a priori power analysis = chooses a sample size needed to be informative (usually larger than can afford)
Learn from genetics - data sharing = key; meta-analysis = great
In necessarily smaller samples (ie patients) - collect more data and present at individual level rather than group; more liberal stat threshold (larger FDR?); restrict search space to minimise noise; Bayesian methods - stabilise low level estimates
What are some problems with flexibility and exploration in data analysis?
Research now is pushed as hypothesis driven - little allowance for exploration = nonsensical as fail to develop decent hypotheses in the first place
Leads to HARKing = Hypothesising After Results are Known = hiding data driven choices and overstate actual evidence
Same things apply to stats/processing packages - lots to choose from = no consistency
How do you solve some problems with flexibility and exploration in data analysis?
Pre-registration of methods and analysis plans ie sample size, analysis tools, predicted outcomes, ROI/localiser strategies for analysis etc
Encourage peer validation of published exploratory results in order to confirm effects found downstream in the research process
What are some problems with multiple comparisons?
People dont use it… Use mass univariate testing instead ie separate hypothesis applied to each voxel = inflate the false positive rate if not corrected for
People ‘shop around’ for statistical programming that inflates their findings = P hacking/selective reporting/inflation bias
How do you improve the problems with multiple comparisons?
Make sure people actually do it…
Report whole brain results (if available)
Justify any non-standard methods for correction
Share unthresholded statistics so can be reanalysed/interpreted in meta analyses
Non-parametric methods ie permutation tests = more accurate
Abandon univariate in= favour of multivaraite = whole brain as measurement
What are some problems with software?
Increasingly complex = increased likelihood of bugs
Some is widely used/open source = a degree of standardisation; others are written for specialised functions by non-professionals = no quality checking
How do we solve software errors?
Code reviews - check for inflation and minimise error rates
Dont fall for the ‘not invented here’ fallacy ie if we didnt make it its shit even if purposes are the same - errors more likely found in large user base
Custom code submitted with publishing of articles
What are some problems with study reporting?
Some skimp on methods including correction for multiple comparison correction techniques
Neuroimaging claims often reported without statistical groundings
How do we improve reporting?
Authors to follow accepted standards and journals should require for publishing
durrr…
What are some problems with replication?
Focus on novelty of findings rather than replicability - those that are replicated dont often find same results
How can we improve replication?
Use data from consortiums
Community needs to change its priorities - replication awards/grants given out
Findings with medical/political implications should be replicated before findings published
What are some general arguments to support neuroscience?
All of the processes have to be grounded somewhere - most likely the brain; even if we dont have all the pieces we still have something of significance
Nothing can be replicated exactly - superficial differences, even in apparent noise, might provide insight
Data we have is independent of the validity of the discipline