As Predicted:Advice-taking and social learning in paranoia (#44848)
Created: 07/20/2020 06:11 AM (PT)
This is an anonymized version of the pre-registration. It was created by the author(s) to use during peer-review.
A non-anonymized version (containing author names) should be made available by the authors when the work it supports is made public.
1) Have any data been collected for this study already?
No, no data have been collected for this study yet.2) What's the main question being asked or hypothesis being tested in this study?
We are interested in how people use social information to learn the payoffs associated with two options and how this varies with pre-existing paranoia (measured using the persecution subscale of the R-GPTS). Participants learn the payoffs associated with two options (framed as fishing lakes) over 20 trials. The experiment consists of three experimental conditions in which participants choose between two different pairs of options. After five learning trials participants either receive advice from another player, observe the choice of another player or receive no information, depending on the experimental condition. We expect that social information will generally aid learning, but that the tendency to take advice from another player will be reduced in paranoia. We expect an inverse relationship between paranoia and tendency to learn through eavesdropping but this relationship will not be as strong as for the paranoia-advice inverse association.
H1. Paranoia will be inversely associated with the tendency to follow advice from another player and the tendency to copy the observed decision of another player.
H1(a) The magnitude of the inverse association between paranoia and social learning will be larger for advice than for eavesdropping.
H2. People perform better on the task when they have social information, compared to control.
H2(a) Performance in advice condition inversely associated with paranoia.
H3. People learn the best option more quickly when they have social information, compared to control
H3(a) Learning speed inversely associated with paranoia in the advice condition.
H4. Paranoia positively associated with choice stochasticity.
H5. Paranoia inversely associated with tendency to post-hoc identify the best lake.
H5(a) The lake which was not promoted by an adviser will gain higher ratings in paranoid individuals compared with low-paranoia individuals.3) Describe the key dependent variable(s) specifying how they will be measured.
We will use the following DVs:
1. ‘Choose best lake’ - tendency to choose the best lake (1/0) in trial 6 after social information (compared to same trial number in the control condition). This response term is a proxy for social learning in the social information conditions.
2. ‘Task performance’ - the number of correct lakes chosen across all trials. An ordinal categorical response term.
3. ‘Learning speed’ - the number of trials before the best option is discovered. Best option is ‘discovered’ if the participant chooses it 3x in a row. An ordinal categorical response term. Participants’ learning rate will be estimated by fitting a q-learning model to their decisions.
4. ‘Choice stochasticity’ - will be estimated in two different ways. The first is using U-value, and the second is by examining the ratio between win-stay vs win-switch responses.
5. ‘Post-hoc identification of best lake’ - the number of times the lake was reported as best lake in a series of 12 post experiment pairwise comparisons of all lakes, an ordinal categorical response term.
6. ‘Post-hoc identification of worst lake’ - the number of times the lake was reported as the worst lake in the 12 post experiment pairwise comparisons of all lakes, an ordinal categorical response term.4) How many and which conditions will participants be assigned to?
All participants assigned to 3 conditions (control / advice / eavesdrop) in counter-balanced order. Social information (advice / eavesdropping) always indicates the best lake.
5) Specify exactly which analyses you will conduct to examine the main question/hypothesis.
We will use cumulative link or generalised linear mixed models (GLMMs) in R (see section 9). We will follow an information-theoretic approach with model averaging described in Grueber et al. 2011. Full model estimates and confidence intervals will be reported.
Paranoia = score from the persecutory subscale of the R-GPTS
Cognitive score = number of correct solutions to Hagen Matrices Tests (HMT-S) out of a total possible 6.
Condition is a 3-level categorical variable (control / advice / eavesdrop). Control will be the reference category & we will perform post-hoc tests that compare differences between eavesdrop and advice conditions.
H1 & H1a
Model <- Choose best lake (1/0) ~ paranoia + cognitive score + condition + paranoia: condition + cognitive score: condition + (1|participant id)
Tendency to choose best lake will be higher following social information compared to control (H1) but this relationship will be moderated by paranoia (H1a). We control for cognitive score as we expect tendency to follow social information to be inversely associated with cognitive score.
H2 & H2a
Model <- Task performance ~ paranoia + cognitive score + condition + paranoia: condition (1|participant id)
Task performance will be lower in control relative to social conditions (H2). Task performance will be inversely associated with paranoia (H2a) and positively associated with cognitive score. If paranoia reduces tendency to follow advice, we expect an interaction between paranoia and condition (H2a).
H3 & H3a
Model <- Learning speed ~ paranoia + cognitive score + condition + paranoia: condition (1|participant id)
Learning speed will be slower in control relative to social conditions (H3). Learning speed will be inversely related with paranoia (H3a) and cognitive score. If paranoia reduces tendency to follow advice, we expect an interaction between paranoia and condition (H3a).
Model <- choice stochasticity~ paranoia + condition + (1|participant ID)
Choice stochasticity will be higher in paranoia and lower in social information rounds relative to control.
H5 & H5a
Model <- Post-hoc identification of best/worst lake ~ paranoia + condition + cognitive score + paranoia : condition + (1|participant ID)
Post-hoc identification of best lake inversely associated with paranoia and with cognitive score. Identification of best lake higher in social information rounds compared to control, but moderated by paranoia as above.6) Describe exactly how outliers will be defined and handled, and your precise rule(s) for excluding observations.
Participants failing to complete the experiment will be excluded from analysis. Participants with consistently fast response time (median < 350 ms) will be removed from analysis. If participants do not discover the best lake, we will run the analyses above excluding these participants from the data.7) How many observations will be collected or what will determine sample size?
No need to justify decision, but be precise about exactly how the number will be determined.
We will recruit a sample of 1500 participants to ensure that we have sufficient representation across the full paranoia spectrum.
8) Anything else you would like to pre-register?
(e.g., secondary analyses, variables collected for exploratory purposes, unusual analyses planned?)
1. This exploratory analysis pertains to H2 above. Participants may not discover the best lake in at least one of the trials. Where > 10 % of participants do not discover the best lake in at least one trial, we will run another model with a binary response term (discovered best lake / did not discover best lake). The model will be fit as a logistic regression as below:
Discovered best lake ~ condition + paranoia + cognitive score + condition: paranoia + (1|participant id).
2. We have pre-registered a number of models above as cumulative link models, which allow the dependent variable to be specified as an ordinal categorical response term. We will check whether these models can be fit using more traditional generalised linear regression models (GLMMs) where the response term can be specified as a continuous variable rather than an ordinal categorical variable. If (i) the assumptions of GLMM are not violated and (ii) the results are qualitatively similar under both approaches, then we will report the results from the GLMM as these are more routinely used and therefore more widely understood.
3. We will conduct a model fitting procedure to characterise individual learning rates, decision noise (beta, inverse temperature), and the effect of social information. These models will use standard reinforcement learning mechanisms (q-learning), and will implement social information as a special reward, which will be characterised as a free parameter. Models will include a series of nested models with increased level of complexity, i.e. number of parameters, and a model comparison process will be conducted where BIC/DIC scores will be compared to identify the best fitting model. Models using hierarchical bayesian levels (estimating world volatility) will be used as an exploratory analysis as the number of choices and blocks per individual are low, and our design does not include volatility manipulation.
We expect that paranoia and cognitive score will be correlated. Since all of our models involve specifying paranoia and cognitive score as explanatory terms, we will also perform commonality analyses to determine how much of the variance in the dependent variable is explained by shared and unique variance (respectively) in these correlated terms.
4. We will check all models above modelling paranoia as the score from the social reference subscale of the R-GPTS (rather than the persecution subscale).