1. Henrik Singmann David Kellen Karl Christoph Klauer
Investigating the Other-Race Effect of Germans towards Turks and Arabs using Multinomial Processing Tree Models
Henrik Singmann
David Kellen
Karl Christoph Klauer

13. Other-Race Effect (ORE)
Expectations (Meissner & Brigham, 2001):
Pick the wrong turkish/arabic individual (1.6 × more likely).
Pick the correct white individual (1.4 × more likely)
75% of 239 individuals wrongfully convicted because of eyewitness testimonies (Innocence Project, 2009):
53% cross-race errors
Alleged perpetrators served an average of 12 years prison

14. Research Questions Is there an ORE towards Turks/Arabic individuals?
largest ethnic minority in Germany (Statistisches Bundesamt, 2011)
increasing prejudices towards people with Middle Eastern descent since the terrorist attacks on 9-11 (e.g., Morgan, Wisneski, & Skitka, 2011)

16. Research Questions Is there an ORE towards Turks/Arabic individuals?
largest ethnic minority in Germany (Statistisches Bundesamt, 2011)
increasing prejudices towards people with Middle Eastern descent since the terrorist attacks on 9-11 (e.g., Morgan, Wisneski, & Skitka, 2011)
ORE: Memory or Response effect?
Mathematical modeling using the 2HTM model

17. Recognition Memory Paradigm (e.g., Malpass & Kravitz, 1969)
Study phase:
Two classes of items:
central-european faces
turkish/arabic faces
Test phase:
old? new?
old? new?
old? new?
old? new?
old? new?
old? new?
Two item types:
old faces
new faces

18. Recognition Memory Paradigm (e.g., Malpass & Kravitz, 1969)
Study phase:
Old face
old? new?
New face
old? new?
Test phase:
old? new?
old? new?
old? new?
old? new?
old? new?
old? new?
Correct Rejection
False Alarm
Hit
Miss

19. Recognition Memory Paradigm: Data
Responses per item class - 2 × 2 table:
Relative frequency per item type = 100%:
80% hits → 20% misses
35% false alarms → 65% correct rejections
response
"old"
"new"
item type
old
hit
miss
new
false alarm
correct rejection
two DVs

20. Example Data Two cognitive processes (Snodgrass & Corwin, 1988):
Hits
False Alarms
P 1
80%
50%
P 2
P 3
Two cognitive processes (Snodgrass & Corwin, 1988):
Memory processes: How good one can differentiate between studied and non-studied items.
Response processes: Tendency to use a specific response option (e.g., tendency to respond with "old").
Mathematical modeling to disentangle the two processes.

21. Materials 100 pictures each (turkish/arabic and white):
from websites of turkish/arabic and central european football teams (no first divisions)
Pretest: ethnicity, valence, distinctiveness
Singmann, Kellen, & Klauer (2013), CogSci Proceedings

22. Methods 42 (white) students from University of Freiburg
2 study phases:
100 pictures - 50 randomly selected pictures per item class (arabic and white)
2 seconds presentation time
Test phase:
200 pictures - 50 old and 50 new pictures per item class
response options:
old face
skip
new face

23. "The person viewing a lineup should be told that the perpetrator may not be in the lineup" (Innocence Project Website)
"… the instruction given to witnesses should be tempered quite carefully with a warning indicating the perpetrator's possible absence." (Tredoux, Meissner, Malpass, & Zimmermann, 2004)
"These data suggest that utilisation of an explicit don't know option may be of practical value" (Weber & Perfect, 2012)
Methods
42 (white) students from University of Freiburg
2 study phases:
100 pictures - 50 randomly selected pictures per item class (arabic and white)
2 seconds presentation time
Test phase:
200 pictures - 50 old and 50 new pictures per item class
response options:
old face
skip
new face

24. Results: Response Proportions
*** † *
*** *
d alte Items ~ 0.3, d neue Items ~ .8
Error bars show +/- 1 SD

25. Is the ORE a memory-based or response-based effect?
Modeling the data: Multinomial Processing Tree (MPT) models

26. MPT Models
mathematical measurement model for categorical data: latent cognitive all-or-nothing processes (Riefer & Batchelder, 1988)
Parameters of the model: Probability with which a certain process is reached (0 ≤ parameter ≤ 1)
extended 2-high-threshold model (2HTM; Snodgrass & Corwin, 1988):
Do & Dn = memory processes
g1 & g2 = response processes

27. 2-high-threshold model (2HTM)Do
"old face"
old faces g1
"old face"
1 - Do g2
"new face"
1 – g1
1 – g2
"skip" Dn
"new face"
new faces g1
"old face"
1 - Dn g2
"new face"
1 – g1
1 – g2
"skip"

28. 2-high-threshold model (2HTM)
detection state Do
"old face"
old faces g1
"old face"
1 - Do g2
"new face"
1 – g1
1 – g2
"skip" Dn
"new face"
new faces g1
"old face"
1 - Dn g2
"new face"
1 – g1
1 – g2
"skip"

29. 2-high-threshold model (2HTM)Do
"old face"
old faces g1
"old face"
1 - Do g2
"new face"
detection state
1 – g1
1 – g2
"skip"
uncertainty state Dn
"new face"
new faces g1
"old face"
1 - Dn g2
"new face"
1 – g1
1 – g2
"skip"

30. 2-high-threshold model (2HTM)Do
"old face"
old faces g1
"old face"
1 - Do g2
"new face"
1 – g1
1 – g2
"skip"
P(hits) = P("old face"|old faces) = ? Dn
"new face"
new faces g1
"old face"
1 - Dn g2
"new face"
1 – g1
1 – g2
"skip"

31. 2-high-threshold model (2HTM)Do
"old face"
old faces g1
"old face"
1 - Do g2
"new face"
1 – g1
1 – g2
"skip"
P(hits) = P("old face"|old faces) = ? Dn
"new face"
new faces g1
"old face"
1 - Dn g2
"new face"
1 – g1
1 – g2
"skip"

32. 2-high-threshold model (2HTM)Do
"old face"
old faces g1
"old face"
1 - Do g2
"new face"
1 – g1
1 – g2
"skip"
P(hits) = P("old face"|old faces) = ? Dn
"new face"
new faces g1
"old face"
1 - Dn g2
"new face"
1 – g1
1 – g2
"skip"

33. 2-high-threshold model (2HTM)Do
"old face"
old faces g1
"old face"
1 - Do g2
"new face"
1 – g1
1 – g2
"skip"
P(hits) = P("old face"|old faces) = Do + (1 – Do) × g1 Dn
"new face"
new faces
Estimated individual parameters using MPTinR (Singmann & Kellen, 2013, Behavior Research Methods) g1
"old face"
1 - Dn g2
"new face"
1 – g1
1 – g2
"skip"

34. 2HTM Ergebnisse † *** detection of old faces: Do
detection of new faces: Dn
tendency for "old": g1 g2 †
***

35. Model Selection MPT parameters capture entities of primary interest:
Probability with which a certain cognitive state is reached (e.g., Dn = distractor rejection)
Psychological hypothesis correspond to relationships among model parameters
8 versions of model correspond to different hypothesis about ORE:
no restrictions Do Dn
Do & Dn g
Do, g
Dn, g
Do, Dn, g
ORE: mainly response process

36. Model Selection MPT parameters capture entities of primary interest:
Probability with which a certain cognitive state is reached (e.g., DN = distractor rejection)
Psychological hypothesis correspond to relationships among model parameters
8 versions of model correspond to different hypothesis about ORE:
no restrictions Do Dn
Do & Dn g
Do, g
Dn, g
Do, Dn, g
ORE: mainly response process
ORE: memory processes

37. Model Selection MPT parameters capture entities of primary interest:
Probability with which a certain cognitive state is reached (e.g., DN = distractor rejection)
Psychological hypothesis correspond to relationships among model parameters
8 versions of model correspond to different hypothesis about ORE:
no restrictions Do Dn
Do & Dn g
Do, g
Dn, g
Do, Dn, g
ORE: memory & response processes
ORE: mainly response process
ORE: memory processes

38. Model Selection MPT parameters capture entities of primary interest:
Probability with which a certain cognitive state is reached (e.g., DN = distractor rejection)
Psychological hypothesis correspond to relationships among model parameters
8 versions of model correspond to different hypothesis about ORE:
no restrictions Do Dn
Do & Dn g
Do, g
Dn, g
Do, Dn, g
ORE: memory & response processes
ORE: mainly response process
ORE: memory processes
no ORE

39. Model Selection
8 versions of model correspond to different hypothesis about ORE:
no restrictions Do Dn
Do & Dn g
Do, g
Dn, g
Do, Dn, g
memory parameters (Do & Dn) order restricted (e.g., Do Arabic < Do White) representing prior knowledge
ORE: memory & response processes
ORE: mainly response process
ORE: memory processes
no ORE

40. Model Selection: FIA Restricted models provide worse fit a priori.
Model selection requires weighting of:
Model fit: Ability to acount for the obtained data
Model flexiblity: Ability to account for data in general
Goal: Select model with best trade-off between fit and flexibility
Fisher Information Approximation (FIA):
Based on Minimum Description Length Principle (Pitt, Myung, & Zhang, 2002): Model and data as code to be compressed
Accounts for the functional form (in contrast to AIC and BIC)
Accounts for order restrictions (in contrast to AIC and BIC)
Implemented in MPTinR (Singmann & Kellen, 2013)

41. Model Selection: Flexibility
Roberts & Pashler, 2000, PsychRev

42. FIA: Results # restricted df G² p FIA best 1 None 19.19 516.41 2 Do 42
19.19
516.41 2 Do 42
48.29
.23
486.55 3 Dn
71.07
.003
497.69 4
Do, Dn 84
96.91
.16
503.66 5 g
182.71
<.001
477.10 6
Do, g
126
210.97
454.46 16 7
Dn, g
356.87
527.66 8
Do, Dn, g
168
385.13
504.25 14

43. FIA: Results # restricted df G² p FIA best 1 None 19.19 516.41 2 Do 42
19.19
516.41 2 Do 42
48.29
.23
486.55 3 Dn
71.07
.003
497.69 4
Do, Dn 84
96.91
.16
503.66 5 g
182.71
<.001
477.10 6
Do, g
126
210.97
454.46 16 7
Dn, g
356.87
527.66 8
Do, Dn, g
168
385.13
504.25 14

44. Summary & Conclusion
1. Analysis – ORE towards arabic/turkish individuals mainly based on memory processes (parameter Dn)
2. Analysis – model selection using FIA:
ORE based on memory processes (n = 16)
no ORE (n = 14)
First data showing ORE (of Germans) towards individuals of middle eastern decent (cf. Sporer & Horry, 2011)
Important for justice system
Memory effect

45. Thank you for your attention!
Thanks to: Dr. David Kellen and Prof. Karl Christoph Klauer
Alina Arnhold Johannes Falck Felicitas Flade Hannah Kammüller Eva Maaßen Jasmyn Touchstone