Supplementary Table S5: Learning Rate Estimates For High

Supplementary Table S5: Learning Rate Estimates For High- versus LowPerformers
Learning
Rate
(Successes)
Learning
Rate
(Failures)
Model Error
(degrees of
divergence)
AIC
(Asymmetric)
AIC
(Traditional)
High Performers
27
98
9
98
29
97
8
95
22
91
2
89
34
88
4
81
1
77
0.009
0.280
0.002
0.060
0.780
0.002
0.020
0.620
0.001
0.020
0.030
0.003
0.050
0.140
0.004
0.380
0.050
0
19
17
21
21
12
32
22
29
47
11.20
10.27
19.11
26.58
39.90
47.53
50.94
55.08
73.54
11.30
12.61
18.27
27.81
39.63
44.85
54.27
60.30
72.63
Low Performers
19
14
33
10
6
35
20
26
5
24
11
23
31
21
16
30
18
17
3
28
25
15
32
13
7
12
0.780
0.001
0.430
0.001
1.000
1.000
0.850
1.000
0.990
0.110
0.040
0.090
0.560
0.130
0.230
0.190
NaN
0.300
0.190
1.000
1.000
1.000
0.980
0.070
0.870
1.000
0.020
0
0.820
0
0.780
0.970
0.030
0.480
0.450
0
0
0
0
0
0
0
NaN
0.860
0
0.350
0.004
0.070
0.080
0
0.070
0.100
25
41
44
47
38
25
31
57
34
16
18
35
25
36
21
60
NaN
52
50
66
42
55
40
58
55
56
53.34
70.26
84.17
80.20
55.93
50.70
70.30
90.02
87.71
43.56
75.19
53.31
37.32
68.48
51.81
93.90
NaN
88.29
79.21
86.67
61.32
86.88
72.53
84.00
65.92
80.86
75.34
82.59
86.76
84.34
64.07
56.70
83.11
90.27
87.29
88.69
89.40
92.56
90.87
92.72
92.72
92.72
NaN
89.24
92.55
90.37
91.78
91.06
88.98
91.87
91.57
79.23
Subject
#
% Optimal
Choices
70
69
67
67
67
64
63
61
61
55
55
53
52
52
52
50
50
50
50
48
48
48
45
44
41
38
Estimated learning rates for high-performing and low-performing subjects, using the
modified Rescorla-Wagner (RW) model as described in Methods. Subjects are ordered
by percentage of optimal selections during the Testing Phase. All high performers
showed positive learning rates from treatment failures as well as successes. 50% of low
performers showed zero learning rates from failures. Model errors reflect the difference
between the treatment algorithm predicted by the RW model with the learning rates as
shown, and the actual algorithm as measured by the logistic regression model of
treatment choices during the Testing Phase. Errors are expressed as the angle between
normalized 7-dimensional vectors corresponding to the two algorithms, in degrees.
AIC’s are reported for both the adapted Rescorla-Wagner Model with asymmetric
learning and a traditional Rescorla-Wagner Model. Subject 18 selected the same
treatment for all patients, so learning rates could not be estimated.