Supplemental Materials
Memory States Influence Value-Based Decisions
by K. D. Duncan & D. Shohamy, 2016, JEP: General
http://dx.doi.org/10.1037/xge0000231
Trial Sequence Design:
Because decision-making is a complex and dynamic process, participants had partial
control over the sequences of trials that they were exposed to, for example which
specific cards were repeated. We nevertheless were able to account for several
extraneous and potentially confounding factors through the experimental design and
data analysis.
Novel and familiar scenes: Exposure to novel and familiar scenes was the primary
manipulation in these experiments. Our hypothesis was that the novel/familiar
quality of the scene itself, regardless of the actual content of the scene, would
influence how people use memory to make choices. Because each scene was
presented immediately before specific cards (and concurrently with cards in
Experiments 1 & 3), the scenes could additionally become associated with cards in
memory and, thus, their content could also prime memories for card values. To avoid
this possibility, the trial sequences were designed such that the same scene was never
present during both the learning of a card's value and the subsequent use of that value.
In Experiment 2, the novelty/familiarity of a scene was determined by whether it was
the scenes first or second presentation. Because the first presentation necessarily
occurs before the second, this resulted in a greater density of novel scenes early on in
the experiment. To adjust for this potential confound, we removed the first 25 trials
from each session. There was a similar distribution of novel and familiar scenes over
the remaining trials (Figure S1)
Figure S1: Distribution of familiar and novel scenes across Experiment 2
Additionally, when possible, novel/familiar scene status was reversed for pairs of
subjects such that similar choices for one subject would be made in the opposite
experimental condition as another subject. This insured that reported effects were
not unintentionally influenced by idiosyncratic properties of choices/choice
sequences. The exception to this procedure was the early trials in Experiment 2, when
all subjects saw novel scenes. As described above, these trials were removed from all
analyses.
Delay between value-learning and value-decision: The delay between the initial
learning of a cards value and the repetition of that card was constrained to a limited
range (5-24 intervening trials). The inclusion of delay as a covariate in each of the
mixed models did not influence the any of the effects of interest.
Model Details:
Mixed-effects models were preformed using the lme4 package (Bates & Maechler,
2009) in the R programming language and were estimated by optimizing restricted
maximum likelihood. All models that predicted choice (Old Card=1; New Card=0)
used a logistic linking function. Old Card Value was coded in dollars and the predictor
was centered to take values that ranged between -$0.5 and $0.5 (mean $0). The
Preceding Scene was coded as 0/1 (0=Familiar Scene; 1=Novel Scene). All models also
included random effects terms for all coefficients (including interactions) across
subjects. Confidence intervals were estimated using a nonparametric boot-strapping
procedure implemented in the lme4 package (confint.merMod function, 1000
iterations). Planned contrasts between coefficients were performed using the esticon
function (DoBy package for R, Hojsgaard & Halekoh, 2009).
Table S1: Experiment 1 coefficients
Predictor
Intercept
Old Card Value
Decision Scene
Learning Scene
Old Card Value x Decision Scene
Old Card Value x Learning Scene
Estimate
-0.04
2.64
-0.01
-0.11
-0.52
0.44
CI 95%
-0.24 , 0.15
2.02 , 3.25
-0.13 , 0.14
-0.25 , 0.05
-0.94 ,-0.07
-0.11 , 1.07
p
0.7
<2e-16
0.88
0.18
0.02
0.14
Table S2: Experiment 2 coefficients
Predictor
Intercept
Old Card Value
Decision Scene
Learning Scene
Old Card Value x Decision Scene
Old Card Value x Learning Scene
Estimate
0.04
3.01
0.08
-0.06
-0.57
0.14
CI 95%
-0.22 , 0.29
2.42 , 3.64
-0.11 , 0.28
-0.27 , 0.15
-1.08 ,-0.01
-0.36 , 0.70
p
0.79
<2e-16
0.37
0.56
0.02
0.58
Table S3: Experiment 2 coefficients for primed scenes
Predictor
Intercept
Old Card Value
Decision Scene
Learning Scene
Old Card Value x Decision Scene
Old Card Value x Learning Scene
Estimate
0.12
3.09
-0.08
-0.09
-1.03
1.16
CI 95%
-0.19 , 0.48
2.30 , 4.01
-0.44 , 0.21
-0.41 , 0.22
-1.94 ,-0.06
0.17 , 2.29
p
0.46
<3e-14
0.62
0.57
0.03
0.02
Table S3: Experiment 3 coefficients
Predictor
Intercept
Old Card Value
Decision Scene
Learning Scene
Old Card Value x Decision Scene
Old Card Value x Learning Scene
Estimate
-0.11
2.42
-0.01
-0.12
-0.47
0.36
CI 95%
-0.24 , 0.15
2.02 , 3.25
-0.13 , 0.14
-0.25 , 0.05
-0.79, -0.15
0.04 , 0.68
p
0.22
<2e-16
0.98
0.03
0.003
0.03
Reaction Time Analyses:
To investigate whether contextual novelty and familiarity also influenced the speed
with which participants made decisions we predicted reaction time (RT) on critical
trials based on the value of the old card, whether the decision-scene was familiar or
novel and whether the learning-scene was familiar or novel using a Mixed Generalized
Linear Model. The analysis only included trials on which subjects made an objectively
optimal decision (i.e. choosing the 'old' card when it was worth >$.5 and not choosing it
when it was worth <$.5) because it is on these trials that participants are most likely to have
used their memories to make a choice. In Experiment 1, we found that subjects were indeed
faster at making these memory-dependent decisions when the options were presented in
the context of a familiar scene (t=2.54, p=.01, =17.6; Table S3 & Figure S2). This
indicates that subjects were not only more likely to use their memory to make choices in
familiar contexts, but that they were also faster to recall the value and make a well informed
choice. Conversely, the scene that served as a context for the original learning of the old
card's value did not influence participants' RTs (p=.87). Additionally, the preceding scenes
in Experiment 2 did not influence RTs (p>.69; Table s4 & Figure S3).
Table S4: Experiment 1 reaction time coefficients
Predictor
Intercept
Old Card Value
Decision Scene
Learning Scene
Estimate
855
-70
18
1
95% CI
814, 893
-88, -47
0, 33
-12, 16
p
<5e-13
0.01
0.87
Figure S2: Experiment 1 reaction time
Experiment 1 Reaction Time: Average reaction time (RT) taken to make 'correct'
choices (choosing old cards worth >50¢ and not choosing old cards worth <50¢). A.
Trials are binned according to the scene presented in the background while the
decision was made (familiar in green and novel in blue) and the value of the old card.
B. Trials are binned according to the scene presented on the trial on which the value
of the old card was originally learned (familiar in green and novel in blue) and the
value of the old card. Error bars are 95% confidence intervals estimated across
subjects.
Table S5: Experiment 2 reaction time coefficients
Predictor
Intercept
Old Card Value
Decision Scene
Learning Scene
Estimate
894
-59
3
3
95% CI
854, 930
-94, -19
-17, 23
-12, 19
p
<3e-08
0.73
0.70
Figure S3: Experiment 2 reaction time
Experiment 2 Reaction Time: Average reaction time (RT) taken to make 'correct'
choices (choosing old cards worth >50¢ and not choosing old cards worth <50¢). A.
Trials are binned according to the scene that preceded the decision (familiar in green
and novel in blue) and the value of the old card. B. Trials are binned according to the
scene that preceded the trial on which the value of the old card was originally learned
(familiar in green and novel in blue) and the value of the old card. Error bars are 95%
confidence intervals estimated across subjects.
Table S6: Experiment 3 reaction time coefficients
Predictor
Intercept
Old Card Value
Decision Scene
Learning Scene
Estimate
832
-79
-2
3
95% CI
795, 869
-100, -58
-12, 9
-8, 13
p
<9e-09
0.76
0.62
Figure S4: Experiment 3 reaction time
Experiment 3 Reaction Time: Average reaction time (RT) taken to make 'correct'
choices (choosing old cards worth >50¢ and not choosing old cards worth <50¢). A.
Trials are binned according to the scene that preceded the decision (familiar in green
and novel in blue) and the value of the old card. B. Trials are binned according to the
scene that preceded the trial on which the value of the old card was originally learned
(familiar in green and novel in blue) and the value of the old card. Error bars are 95%
confidence intervals estimated across subjects.