This is probably the easiest way to keep everyone updated about what we’ve found so that we can then discuss the story we want to tell. These are the results from the mixed effects model analysis, the check for whether we find the hypothesized pattern in the likelihood of a risky choice given the points earned and the modelling results.
Let’s first visualize the finding and then check with the regression analysis. In order to be able to plot things I’ve had to create some aggregate measure of the data because on trial level the predictions are binary. So I created ten bins along the range of points a subject had earned and used the mean of the points in that bin as value to plot on the x-axis and the mean of the binary data of risky options chosen to get a likelihood to plot on the y-axis. The finer lines are individual data, the bolder mean lines are fitted with the loess method to depict the data, so it’s not an actual model that was fitted and depicted but an aggregatoin over subjects.
In the goal condition we would expect an increase in the likelihood of a risky choice over points (e.g. house money effect, i.e. the more I’ve won so far the riskier I can play with it; but note that this is a post hoc explanation), with a drop as soon as the goal is reached. In the no goal there should be no such drop.
Let’s first look at the plots over all variance conditions to see the general trend.
Likelihood of a risky choice by number of points earned, over all conditions.
Now this looks like overall we find the hypothesized pattern, i.e. we find that in the Goal condition there is a dip in the likelihood of a risky choice when subjects reach (or are very close to reach and probably have enough trials left) the goal. This is optimal, because it minimizes the chance of falling below the goal again. The U-shape of the functions can be explained by good or bad luck, i.e. if they were unlucky they took the risky option and suffered high losses, resulting in negative total points. On the other hand if they were lucky, they could reach a very high total points number by picking the risky option and earning high positive amounts. Because towards the endpoints the data becomes less reliable, the plot window only shows the intervall [-25, 150].
Now let’s look at the same plot but separated for the different environments:
Likelihood of a risky choice by number of points earned, separate environments.
The nice thing is that we see the dip around the goal in the goal conditions in all environments, and also the slopes go in the right direction, i.e. in the Equal environment the slope in the no goal and in the beginning of the goal condition is pretty close to 0, i.e. they may have learned that the options have similar returns (in fact equal returns) but were risk averse. In the High environment they may have learned that the risky option had the higher average return and thus became risky over time. The opposite is true for the Low environment.
This is what we see if we look at the same plot but now with trial instead of points on the x-axis.
Likelihood of a risky choice by trial number, separated for environments.
The dependent variable was whether the risky option was chosen (1) or not (0) at a given trial. Predictors were the goal condition (no goal = 0, goal = 1), whether the goal of 100 points was reached (below goal = 0, above goal = 1), the points earned by a given trial and in the second model, the environment. Random intercepts were included for every person and every game.
We used a generalized mixed effects model. This is how it looked like:
# mixed effects model with random intercepts for subjects and games
model <- lme4::glmer(high.var.chosen ~ goal.condition.bin * overGoal.f + points.cum +
(1|id.f/game), data = df_trial, family = "binomial")
summary(model)
r2(model)
And this is the result of the regression:
Estimate | Std. Error | z value | Pr(>z) | |
---|---|---|---|---|
(Intercept) | -0.3923004 | 0.0665101 | -5.898 | 3.67e-09 *** |
goal.condition.bin1 | 0.0865610 | 0.0960116 | 0.902 | 0.367 |
overGoal.f1 | 0.2216200 | 0.0488954 | 4.533 | 5.83e-06 *** |
points.cum | -0.0026771 | 0.0003284 | -8.151 | 3.62e-16 *** |
goal.condition.bin1:overGoal.f1 | -0.9684303 | 0.0653238 | -14.825 | < 2e-16 *** |
The coding is NoGoal = 0, Goal = 1; and belowGoal = 0, overGoal = 1.
It seems that in general, in the NoGoal condition, the probability of a risky choice becomes larger, once a subject is above the goal, and decreases slowly with the number of points earned. In the Goal condition once the goal is reached, the likelihood of a risky choice decreases. This confirms the patterns shown in the plots. However, we should say that the model was nearly unidentifiable (thus e.g. entering yet another predictor, e.g. the environment may be a problem, but we did it anyway so here’s the table).
Estimate | Std. Error | z value | Pr(>z) | |
---|---|---|---|---|
(Intercept) | -0.511628 | 0.076620 | -6.677 | 2.43e-11 *** |
goal.condition.bin1 | 0.112715 | 0.077717 | 1.450 | 0.1470 |
overGoal.f1 | 0.231346 | 0.043935 | 5.266 | 1.40e-07 *** |
points.cum | -0.002274 | 0.000292 | -7.788 | 6.81e-15 *** |
variance.conditionHigh | 0.488517 | 0.094269 | 5.182 | 2.19e-07 *** |
variance.conditionLow | -0.307820 | 0.095010 | -3.240 | 0.0012 ** |
goal.condition.bin1:overGoal.f1 | -0.929184 | 0.057912 | -16.045 | < 2e-16 *** |
We would expect that subjects from the goal condition are best classified by the model that accounts for the goal. The subjects from the no goal model should be best classified by either the mean or the reinforcement learning model.
The next section provides a short overview over the models.
Model | Description | Impression parameter | Choice Parameter | Total Free Parameters |
---|---|---|---|---|
Random | Baseline model to compare the others with. Random choice in every trial | - | - | 0 |
NaturalMean | For each option takes the mean of all seen outcomes. Then put’s these means into a softmax | - | \(\phi\) | 1 |
RL | Standard reinforcement learning model with one learning parameter for impression updating. Impressions are then put into a softmax. | \(\alpha\) (learning rate) | \(\phi\) | 2 |
SampEx_Int_Goal | For each option calculates the mean and standard deviation of the last N (free discrete parameter that was estimated) outcomes, then creates a cummulative (for the number of trials left) normal density and from this derives the probability of sampling a value at least as big as the difference from the current point total to the goal. Then put’s these probabilities into a softmax | \(N\) (memory capacity) | \(\phi\) | 2 |
Here are tables showing the proportion of people from the Goal and NoGoal condition were classified by which model:
Overall
Goal | NoGoal | |
---|---|---|
Random | 0.180 | 0.154 |
NaturalMean | 0.148 | 0.244 |
RL | 0.360 | 0.421 |
SampEx_Int_Goal | 0.312 | 0.181 |
Low Environment
Goal | NoGoal | |
---|---|---|
Random | 0.182 | 0.072 |
NaturalMean | 0.197 | 0.275 |
RL | 0.455 | 0.420 |
SampEx_Int_Goal | 0.167 | 0.232 |
High Environment
Goal | NoGoal | |
---|---|---|
Random | 0.197 | 0.234 |
NaturalMean | 0.115 | 0.247 |
RL | 0.328 | 0.364 |
SampEx_Int_Goal | 0.361 | 0.156 |
Equal Environment
Goal | NoGoal | |
---|---|---|
Random | 0.161 | 0.147 |
NaturalMean | 0.129 | 0.213 |
RL | 0.290 | 0.480 |
SampEx_Int_Goal | 0.419 | 0.160 |
The results of two environments (High and Equal) suggest, that the model that accounts for the goal is better in classifying people from the Goal than from the NoGoal condition. Thus also from a modeling perspective we have some evidence that, at least for some subjects, accounting for a person’s goal is an important aspect.
Here are some summary tables to show model fits and parameter value statistics.
Over all environments:
model_best | bic_mean | bic_sd | par_Imp_mean | par_Imp_sd | par_Choice_mean | par_Choice_sd | N |
---|---|---|---|---|---|---|---|
NaturalMean | 284.6419 | 57.11349 | NA | NA | 0.1021640 | 0.0743124 | 82 |
Random | 346.5740 | 0.00000 | NA | NA | NA | NA | 68 |
RL | 271.3271 | 60.62504 | 0.7193733 | 0.2604189 | 0.0953607 | 0.1166225 | 161 |
SampEx_Int_Goal | 307.5428 | 30.28930 | 2.8787879 | 3.6902318 | 1.7736558 | 0.7426561 | 99 |
Separated for the three environments
model_best | variance.condition | bic_mean | bic_sd | par_Imp_mean | par_Imp_sd | par_Choice_mean | par_Choice_sd | N |
---|---|---|---|---|---|---|---|---|
NaturalMean | Equal | 288.9867 | 54.12785 | NA | NA | 0.0951604 | 0.0701292 | 24 |
NaturalMean | High | 289.3647 | 51.91944 | NA | NA | 0.1104961 | 0.0897531 | 26 |
NaturalMean | Low | 277.5460 | 63.95490 | NA | NA | 0.1006469 | 0.0646944 | 32 |
Random | Equal | 346.5740 | 0.00000 | NA | NA | NA | NA | 21 |
Random | High | 346.5740 | 0.00000 | NA | NA | NA | NA | 30 |
Random | Low | 346.5740 | 0.00000 | NA | NA | NA | NA | 17 |
RL | Equal | 276.3863 | 54.87985 | 0.6979056 | 0.2566165 | 0.0922045 | 0.1094558 | 54 |
RL | High | 280.6715 | 48.69127 | 0.7632377 | 0.2534000 | 0.0737897 | 0.0605315 | 48 |
RL | Low | 259.0944 | 72.23123 | 0.7033355 | 0.2694299 | 0.1157988 | 0.1509413 | 59 |
SampEx_Int_Goal | Equal | 311.3188 | 29.13052 | 3.1842105 | 3.8963272 | 1.6562858 | 0.8051688 | 38 |
SampEx_Int_Goal | High | 305.8229 | 29.23471 | 2.6176471 | 3.4816811 | 1.9621618 | 0.7205251 | 34 |
SampEx_Int_Goal | Low | 304.3941 | 33.62546 | 2.7777778 | 3.7553381 | 1.7014653 | 0.6521926 | 27 |
Here are the Plots from the previous analyses that are also interesting and could be used here.
Proportion of risky options chosen, separated for goal condition and the state. 100 Points corresponded to the goal in the goal condition. The effect was roughly the same over environments. Bold horizontal lines indicate the mean. Boxes indicate 95% bayesian highest density intervals.
Probability of choosing the risky option, given the RST evidence strengths for that option. Only data from goal condition was used. Point sizes and numbers above the points indicate the proportion of occurrences within these bins of evidence strengths.
Differences in prediction accuracy rates (\(pA_{RST} - pA_{EV}\)) in trials in which the model prediction differed, separated for both goals and environments. A value above 0 (above the dashed line), indicates higher prediction accuracy rate of RST, a value below 0 a higher prediction accuracy rate of EV. Bold horizontal lines indicate the mean. Boxes indicate 95% Bayesian highest density intervals.
First the plots for the likelihood of a risky choice given the number of points:
Simulation Data: Likelihood of a risky choice by number of points earned, separated for environments.
And now the same but with trial on the x-axis
Simulation Data: Likelihood of a risky choice by trial number, separated for environments.
Plot of simulated data. Proportion of risky options chosen, separated for goal condition and the state. 100 Points corresponded to the goal in the goal condition. The effect was roughly the same over environments. Bold horizontal lines indicate the mean. Boxes indicate 95% bayesian highest density intervals.