APRIL 8 Flashcards
Ps happiness levels were best explained by model based on WHAT as opposed to WHAT?
model based on RPEs
as opposed to OVERALL EARNINGS
(happiness depends on whether things are going better than we expected)
paper: “depression is associated with reduced __________ ______ in a dual valence, magnitude ______ ______”
depression is associated with reduced OUTCOME SENSITIVITY in a dual valence, magnitude LEARNING TASK”
MDD
heterogeneous psychiatric condition associated with symptoms including:
- low mood
- feelings of worthlessness and guilt
- lack of interest in things previously deemed pleasurable (ie. anhedonia)
cognitive impairments associated with MDD
- judgment bias
- memory
- decision-making
MDD and decision-making can be probed using…
RL frameworks
useful in identifying SPECIFIC CAUSES of CHANGES in decision making
how can RL frameworks probe depression and decision making
can identify SPECIFIC CAUSES of changes in decision making
ie. look into OUTCOME SENSITIVITY and CHOICE STOCHASTICITY
choice stochasiticity
the randomness in decision making
how is estimated value of a choice updated?
based on the difference between the PREDICTED VALUE of the outcome and the ACTUAL OBSERVED OUTCOME
^ the RPE
what’s the RPE scaled by?
learning rate (alpha)
beta parameter
(inverse temperature parameter)
captures the degree to which value estimates influence choice probabilities
is a measure of RANDOMNESS/CHOICE STOCHASTICITY
results of many studies suggest the major impact of MDD in decision making is mediated by what?
by changes in OUTCOME SENSITIVITY (or choice stochasticity)
and simultaneously they think that LEARNING RATE IS UNCHANGED
but other studies report conflicting results - so literature is inconsistent with regards to the specific processes impacted in MDD related to changes in decision making
connection to anhedonia
change in outcome sensitivity/choice stochasticity is thought to be related to ANHEDONIA
(low beta = more randomness/stochasticity - this is associated with anhedonia)
what kind of outcomes do most studies use in probing MDD and decision-making?
- BINARY OUTCOMES
ie. 1 or 0
- only probe learning about POSITIVE and NEGATIVE outcomes on SEPARATE TRIALS (or even separate tasks)
it remains unknown how MDD impacts decision-making processes in tasks with…
- continuous outcomes (not 1 or 0)
- outcomes that are both negative and positive
central hypothesis and primary prediction
CENTRAL HYPOTHESIS:
^ reinforcement learning is DISRUPTED in depression
PRIMARY PREDICTION:
^ current MDD will show IMPAIRED REWARD SENSITIVITY but NO DIFF in learning rate
STATE or TRAIT:
^ no prediction
study setup
assessed RL in:
- current MDD
- remitted MDD
- healthy controls
probe RL parameters in task involving BOTH REWARD and LOSS outcomes that vary in magnitude (non-binary)
state or trait?
TRAIT: tendency to experience bouts of depression
STATE: currently experiencing an episode of depression
in study design:
^ current MDD have high state and trait depression
^ remitted MDD have only trait depression
^ healthy controls have neither
demographics
- CURRENT MDD ps had:
a) higher levels of self-reported depression symptoms
B) higher state & trait anxiety
c) higher anhedonia
- REMITTED MDD ps had:
a) higher levels of self-reported depression symptoms
b) higher state & trait anxiety
- but didn’t have higher anhedonia ratings than controls
drifting magnitude reinforcement learning task
after completing questionnaires, Ps complete RL task
on computer screen
task = designed to make Ps learn about VARYING AMOUNTS of LOSSES and REWARDS at the SAME TIME
initial results: influence of win and loss on choice
first they report the data without using computational modeeling
calculate probability of choosing each shape after experiencing a win or loss for each P group
non-significant trend for a REDUCTION in INFLUENCE OF MAGNITUDE in currently depressed participants
BUT THIS ISN’T A STRONG RESULT
computational modelling - step 1
compare 3 different RL models to determine which model fits the data best
OPTION 1: model with two learning rates (reward and loss) and single outcome sensitivity parameter
OPTION 2: model with single learning rate and independent reward and loss sensitivity parameters
OPTION 3: model with two learning rates (reward and loss) and independent reward and loss sensitivity parameters
FOUND THAT OPTION 2 BEST FIT THE EXPERIMENTAL DATA
OPTION 3:
computational modelling - step 2
fit the winning model to EVERY P’s data to estimate the INDIVIDUAL PARAMETERS that best explain their choices
compare parameter estimates BETWEEN GROUPS using ANOVA and pairwise comparisons
computational modelling results
- NO EFFECT ON LEARNING RATE
- SENSITIVITY to REWARD and LOSS is significantly LOWER in Ps with CURRENT MDD
- no diff between remitted MDD and healthy controls
^ suggests this is a STATE effect (not a trait effect)
was there a correlation between outcome sensitivity parameters and continuous symptom levels inferred from questionnaires?
no
