Priming and Anchoring Effects in Visualization · 2017. 10. 4. · Prime NSep Target 1 Prime NSep...

Priming and Anchoring Effects in Visualization

André Calero Valdez, Martina Ziefle, Michael Sedlmair

Priming and Anchoring Effects in Visualization. IEEE InfoVis: Phoenix, AZ, USA. Oct 4th, 2017.Michael Sedlmair 2

ThanksAndré!

What are priming and anchoring effects?

Priming and Anchoring Effects in Visualization. IEEE InfoVis: Phoenix, AZ, USA. Oct 4th, 2017.Michael Sedlmair 4https://crew.co/blog/the-priming-effect-why-youre-less-in-control-of-your-actions-than-you-think/

Priming and Anchoring Effects in Visualization. IEEE InfoVis: Phoenix, AZ, USA. Oct 4th, 2017.Michael Sedlmair 5https://crew.co/blog/the-priming-effect-why-youre-less-in-control-of-your-actions-than-you-think/

Priming and Anchoring Effects in Visualization. IEEE InfoVis: Phoenix, AZ, USA. Oct 4th, 2017.Michael Sedlmair

Priming effects ……describe phenomena in which human responses are influenced by a preceding perceptual stimulus.

6


Anchoring effects …… describe phenomena in which a previous stimulus provides a frame of reference.

7


Anchoring

8

Q: At which age did Gandhi die?

Was Gandhi’s age higher or lower than 9 years?

7168 75

7271

65

73

72

69 8878 85

92

7195

93

102 89

Was Gandhi’s age higher or lower

than 141 years?

Anchor Anchor*

* he died at the age of 78.https://commons.wikimedia.org/wiki/File:Portrait_Gandhi.jpg

Are priming and/or anchoring effects at

play in Visualization?


Example

10

Q: How separable are the classes?

Hypothesis:x ≠ y ⟹ some sort of anchoring/priming

y?

??

??x

? ?

???

?


Why is it important?• Current models assume perception to be

more or less constant• Some work on individual differences

[Kay and Heer 2016, Toker et al. 2013]

• We: temporal effects

11

−4 −2 0 2 4

−4−2

02

4

n500−d10−c3−spr0.1−out0 − GlimmerMDS: 2d ScatterplotDistance−99−100−99−100

Distribution−100−100−100−100−22−2.315−40

X1

X2

●

●●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●●

●

●

●

●

●

●

●

●

●●●

●

●

●●

●

●●

●●

●●●

●●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●●

●

●

● ●●

●●

●

●

●

●

●●

●

●●

●

●●

●

●

●●

●

●●

●

●

●

●

●

●●

●

●

●

●

●●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

● ●

●●

●

●

●

●

●●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●●

●

●

●●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●●

●●

●

●

●

●

●

●

●

●

●

●●

●

●●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●●

●●

●

●●

●

● ●

●

●

●

●●

●

●

●

●

●

●

●

●●●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●●●

class012


• A series of 5 studies on priming/anchoring in visualization …

• … using class separability in scatterplots as an example.

Today: Brief overview & main results

Contribution

12


(Pilot) Study 1• Q: Any priming/anchoring

effects visible in real data? • Data: from previous work

[Sedlmair et al., EuroVis 12, InfoVis 13, EuroVis 15]

• Task: rate separability on a scale 1-5

• MTurk Study- 180 participants

13* 200 overall, 180 after removing speeders, etc.


Study 1: Setup (in a nutshell)

14

unclear target

Separable primes

Non-Separable primes

Hypothesis: Priming on unclear targets.

Cat-eye plot w/ 95% CI

trainin

g,

etc.

trainin

g,

etc.

12

34

5

sep

nsep


Study 1: Example result

15

unclear target

Separable primes

Catseye plots of 95% confidence intervalls for between−subjects means (n=180)

Trial

Sepa

rabi

lity

Target 1 Target 2 Target 3 Target 4

12

34

5

Non-Separable primes

trainin

g,

etc.

trainin

g,

etc.


Study 1: Summary results • Maybe priming effect was at play • BUT: Too many uncontrolled variables to say definitely

16


(Pilot) Study 2• Goal: Identify well-controlled stimuli • How: Created 200 sample scatterplots

- only distance between classes is varied • MTurk Study

- 43 participants* - task: separability on scale 1-5

17* 47 overall

smallest dist highest dist


Study 2: Result —> Three suitable stimuli

18

2 primes / anchorsparticipants agreed

(a) clearly non-separable or

(b) separable

Study 3 (n=196)

2 + 4 = 6 judgments

2 Training Tasksforced choice

Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x

Legend:Sep separable stimulusNSep non-separable stimulusTarget "unclear" stimulus

Study 4 (n=243)4 judgments

A

B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

1 unclear targetlarge variance across participants

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B


Study 3

19

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

MTurk Study: 196 participants** 251 overall

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B


Study 3: Results

20

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B


clearly separable

clearly non-separable

21

12

34

5

Study 3: Results

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Catseye plots of 95% confidence intervall for within subject means (n=196)

Stimulus

Sepa

rabi

lity

Stimulus 1 Target 1 Stimulus 2 Target 2

0.0

0.2

0.4

0.6

0.8

1.0


Stimulus

Sepa

rabi

lity


0.0

0.2

0.4

0.6

0.8

1.0


Stimulus

Sepa

rabi

lity


0.0

0.2

0.4

0.6

0.8

1.0


Stimulus

Sepa

rabi

lity


0.0

0.2

0.4

0.6

0.8

1.0


clearly separable


22

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

12

34

5

• Visible effect in 2nd trial

• BUT: No effect in 1st trial

• Reason - training anchors - after masking small

priming effect

Study 3: Results


Stimulus

Sepa

rabi

lity


0.0

0.2

0.4

0.6

0.8

1.0 2nd trial1st trial


Study 4 = Study 3 without Training

23

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

MTurk Study: 243 participants** 351 overall


Study 4: Results

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B


Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B1

23

45

Study 4: Results

No bias from training tasks

clearly separable



Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B1

23

45

Clear anchoring effect for first target

Study 4: Results

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

clearly separable



Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B1

23

45

Study 4: Results

Also anchored by first target

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

clearly separable



Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B1

23

45

Study 4: Results

2/3 of a tick mark difference for the very same target

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

clearly separable



Study 4: Summary results

29

• Anchoring effect! • First anchor in small studies is very influential. • What not: within-subject priming effects*

* within-subject: does even the very same person judge scatterplots differently based on what they have seen before?


Study 5• Goal: understand

subtle, within-subject priming effects• Long term usage study

- data: 200 randomly generated scatterplots w/ centroid distance = [0;4] Stdev

- task: separability judgments: 1-5 scale up to 1000 judgments per participant

- Online study with 64 participants / 28,544 judgments*

30* 31,105 overall

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Stud

y 3

(n=1

96)

2 +

4 =

6 ju

dgm

ents

2 Tr

aini

ng T

asks

forc

ed c

hoice

Prim

e Se

p

Prim

e NS

ep

Targ

et 1

Prim

e NS

ep

Targ

et 2

Prim

e Se

p

Mas

king

Task

5x

Lege

nd:

Sep

se

para

ble

stim

ulus

NSep

non-

sepa

rabl

e st

imul

usTa

rget

"u

ncle

ar" s

timul

us

Stud

y 4

(n=2

43)

4 ju

dgm

ents

A B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

etc.

x 1000


Analysis• Regression model

predict current stimulus based on: (1) the current centroid distance (2) the separability of the previous stimulus.

• Hypotheses - H1: Mainly depends on centroid distance - H2: Small priming effect, i.e., it depends

also on previous stimulus

31

p-dist

Study 3 (n=196)

2 + 4 = 6 judgments


Prime Sep

Prime NSep

Target 1

Prime NSep

Target 2

Prime Sep

Masking Task 5x



A

B

c-dist


Results — Simplified Linear Regression

• rating = [1;5]

• c-dist, p-dist = [0;4]

32

rating = 0.6 + 1.0 × c-dist + 0.1 × p-dist

smallest dist highest dist


Results — Simplified Linear Regression

• Intercept • Influence of the

current distance [0;4] is high • Small influence of

previous distance [0;4]

33

rating = 0.6 + 1.0 × c-dist + 0.1 × p-dist

c-dist=1.5

p-dist=4

p-dist=0


Study 5: Summary results

34

• We do see within-subject priming effects in long-term usage

• They account for ~7% of the next judgment

Conclusions


Summary• priming & anchoring effects in VIS • 5 studies

- from application-driven - to well-controlled

• first evidence for anchoring & priming in VIS

36


Future work …• … just a beginning • How to use in vis design?

- counteract?

37

tasks

idioms

setups

Priming and Anchoring Effects in VisualizationAndré Calero Valdez, Martina Ziefle, Michael Sedlmair

Thanks!

email: [email protected]

A

B

Target

A

B

Target

Perceived separability

low

high

I

II

I II

slides: https://homepage.univie.ac.at/michael.sedlmair/talks/InfoVis17_anchoring.pdf

mailto:[email protected]?subject=

https://homepage.univie.ac.at/michael.sedlmair/talks/InfoVis17_anchoring.pdf

Priming and Anchoring Effects in Visualization · 2017. 10. 4. · Prime NSep Target 1 Prime NSep...

Documents

Transcript of Priming and Anchoring Effects in Visualization · 2017. 10. 4. · Prime NSep Target 1 Prime NSep...