Replication of ‘Negative Valence Widens Generalization of Learning’ Schechtman et al. 2010 J.Neuro

Justification for replication

Valence is among the principle constructs used to explain animal behavior; we tend towards things we like, away from what we don’t. Remarkably, not much is known about how valence effects processes upstream of decision-making–e.g. memory and perception. Schechtman et al. 2010 developed a paradigm in which auditory cues are paired with negative and positively valenced outcomes during an acquisition phase. During a generalization phase, they showed that subjects are more likely to confuse novel tones with the previous negative cue than the positive one. The authors discribe this result as differential perceptual generalization as a function of valence.

My hope is to build upon this core finding in order to study how these valenced processes interact with memory. By replicating this study in an online sample, I would be well positioned to iterate through paradigms, eventually converging on a study that will enable me to explore the relationship between valence and memory.

Description of procedures

There are two stages in this experiment which repeat; an acquisition stage and a generalization stage.

Differences from Original Study

Because this study is based on performance, and chance-level performance would result in subjects receiving $0.00. This design incentives subjects to remain engaged in a way that is well suited for an online setting. Additionally, the resulting pattern of behavioral evidence will allow us to identify subjects who were not engaged (e.g. performance around chance). Any subjects who are not performing above 80% accuracy within the first block will be excluded from further analysis.

In principle, the currect javascript implimentation should not deviate in a meaningful way from the original study. Most importantly, the distribution of trial types and overall experimental length have been largely perserved, even when it is not critical for the main hypothesis; for example, the proportion of control trials is within several percentage points of the original study, not only the distribution of positive and negative instrumental and pavlovian trials.

The biggest deviation is the exclusion of several of the original control stimuli in the generalization phase. Two different auditory frequencies (e.g. 100 and 900Hz) must be played at different amplitudes in order to sound like they are being played at the same volume; typically, lower tones sound quieter, so the amplitude has to be greater. There was no calibration precedure the original authors used to ensure that tones of different frequencies were equally audible, and this seems to be acceptable for their laboratory setting. However, generating stimuli following their procedure, and then playing those tones on laptops, it was no possible to hear the 80Hz tones. This group of controll tones around 100Hz (80, 100, 120) was excluded.

Data preparation

The data preparation pipeline is going to be simplified because of how I’m saving the data. For each trial, I’m saving the data to the server I’m running the experiment off of, and these data are already formatted for how the data will eventually be analyzed. In particular, in the generalization stage, the distance from the origina tone, valance, and key press are all recorded. And example trial’s data looks like this:

{"trial_data":{"rt":482,"stimulus":"sound/305","key_press":80,"stage":"generalization","correct_response":"space","valence":"positive","distance":5,"trial_type":"audio-keyboard-response","trial_index":136,"time_elapsed":35688961,"internal_node_id":"0.0-4.0-1.11","correct":false,"i_generalization_trial":11,"i_block":0},"data_type":"single_trial","dbname":"sleep_affect_memory","colname":"replication","iteration_name":"pilot0_251","context":"piloting","worker_id":"NONE","assignment_id":"NONE","hit_id":"NOPE","browser":"Chrome"}

Loading the data from the server in python:

import json, pymongo, pandas
import numpy as np
import matplotlib.pyplot as plt 
import warnings; warnings.simplefilter('ignore')
  
auth_path = "/Users/biota/memory/sleep_affect_memory/experiment/.credentials/auth.json"
  
def connect_to_database(): 
    # load credentials to access the database, connect, identify collection
    data = json.load(open(auth_path))
    mongo_tunnel = 'mongodb://' + data['user'] + ':' + data['password'] + '@127.0.0.1'
    connection = pymongo.MongoClient(mongo_tunnel)
    data_base = connection['sleep_affect_memory']
    collection = data_base['replication']
    
    return collection
    
def identify_workers(collection):
    # the second is my worker id
    exclude = ['NONE', 'A33F2FVAMGJDGG']
    all_workers = [i for i in collection.distinct('worker_id') if i not in exclude]
    # extract workers who've completed entire experiment -- not returned HIT early
    complete = [] 
    for i_worker in all_workers: 
        tmp_data = collection.find({'worker_id':i_worker})
        if 'worker_feedback' in tmp_data[tmp_data.count()-1]['trial_data']: 
            complete.append(i_worker)
            
    return complete
    
def extract_data(): 
    # connect with mongo
    collection = connect_to_database() 
    # identify workers who completed experiment
    worker_ids = identify_workers(collection)
    # pret to remove worker identifiers 
    subject_ids = {worker_ids[i]:i for i in range(len(worker_ids))}
    # initialize data frame
    subject_trial_data = pandas.DataFrame()
    # iterate over workers 
    for i_worker in worker_ids: 
        # extract worker's data from mongo database
        i_data = collection.find({'worker_id':i_worker})
        # extract trial data
        for one_trial in i_data: 
            # only extract data we want 
            if 'worker_feedback' not in one_trial['trial_data'].keys(): 
                q = {i:one_trial['trial_data'][i] for i in list(one_trial['trial_data'].keys())}
                # use anonymized worker identifier
                q['subject'] = subject_ids[i_worker]
                subject_trial_data = subject_trial_data.append(q, ignore_index=True)
                
    return subject_trial_data
  
# extract and format data from database
data = extract_data()
  
# identify the hypothesis space within data we're (not) interested in 
hypothesis_space = (data.stage=='generalization') & (data.valence != 'control').values
  
# annoyong: convert NaNs+NoneTypes from javascript into an R readable format after loading from python
distance = [abs(int(i)) if (type(i)==float or type(i)==int) * (i==i) else -1 for i in data['distance']]
key_press = [int(i) if (type(i)==float or type(i)==int) * (i==i) else -1 for i in data['key_press']]

Import data from python into R and clean up the data frame

# main import from python
data = data.frame(py$data)

# overwrite poorly formatted vectors
data['distance'] = py$distance
data['key_press'] = as.factor(py$key_press)

# convert key presses to their corresponding valence associations
data$association = revalue(data$key_press, c("-1"=NaN, '32'='novel','80'='positive', '81'='negative'))

The next step is critical for formatting the data: to determine whether the decision matches the nearest valence. For the tones that were in the acquisition stage, this is going to give an estimate of accuracy. For other tones, this is going to tell us whether subjects were confusing novel tones with their valenced neighbors–basically, making errors.

# determine whether the decision valence matches the nearest valenced tone
data$match = data$association == data$valence

Confirmatory analysis

To perform our primary confirmatory test, we will restrict our analysis to include the following information:

1. generalization trials
2. ‘positive’ or ‘negative’ valenced tones (not controls)
3. distance of this tone to tones in the acquisition stage (0, 5, 20, 60, or 100Hz)

hypothesis_data = select(data, valence, distance, match, subject)[py$hypothesis_space,]

Visualize the relationship between the proportion of errors distance from original tones and valence

hypothesis_data %>% 
  group_by( valence, distance) %>% 
  summarise(p_identification = mean(match), 
            sem = sd(match)/sqrt(length(match)),
            y_lower = p_identification-sem, 
            y_upper = p_identification+sem) %>% 
  ggplot(aes(x=distance, y=p_identification, color=valence)) + 
    geom_line() + 
    geom_errorbar(aes(ymin=y_lower, ymax=y_upper), width=2) + 
    ggtitle("'Generalization curve' averaged across 2 subjects in pilot B")

The primary question this replication hopes to address is “Are the slopes between the positive and negative key responses different?” Using an ANOVA this can be modeled as p(response) ~ distance * valence_type, where we expect the interaction term distance * valence_type to be significant. We also, expect the slope for the negative valence term to be less than the postive valence, signifying “wider generalization”.

summary(lm('match ~ valence * distance', hypothesis_data))

## 
## Call:
## lm(formula = "match ~ valence * distance", data = hypothesis_data)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -0.8887 -0.2228  0.1446  0.3169  0.7772 
## 
## Coefficients:
##                            Estimate Std. Error t value Pr(>|t|)    
## (Intercept)               0.6777867  0.0716013   9.466 1.05e-15 ***
## valencepositive           0.2108803  0.1012596   2.083  0.03974 *  
## distance                  0.0002663  0.0014876   0.179  0.85825    
## valencepositive:distance -0.0069250  0.0021038  -3.292  0.00136 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.4297 on 104 degrees of freedom
## Multiple R-squared:  0.1621, Adjusted R-squared:  0.1379 
## F-statistic: 6.706 on 3 and 104 DF,  p-value: 0.0003505

In this small sample (n=2), we replicate the interaction between the valence and distance.

Visualization summary statistics and engagement

Visualize summary statistics for each of the 2 subjects in pilot B, as well as some quick checks on their engagement:

acquisition_plot = data %>% 
  filter(stage=='acquisition') %>% 
  group_by(subject, i_block) %>% 
  summarise(mean_accuracy=mean(correct)) %>% 
  ggplot(aes(x=i_block, y=mean_accuracy, color=as.factor(subject))) + 
    geom_line() + 
    ggtitle("Block accuracy | Acquisition phase") + theme(legend.position="none")

generalization_plot = data %>% 
  filter(stage=='generalization') %>% 
  group_by(subject, i_block) %>% 
  summarise(mean_accuracy=mean(correct)) %>% 
  ggplot(aes(x=i_block, y=mean_accuracy, color=as.factor(subject))) + 
    geom_line() + 
    ggtitle("Block accuracy | Generalization phase") + theme(legend.position="none")

acquisition_keypress_plot = data %>% 
  filter(stage=='acquisition', key_press!=-1) %>% 
  ggplot(aes(x=i_acquisition_trial, y=key_press, color=as.factor(subject))) + 
    geom_jitter(height=.1, alpha=.5) + 
    scale_y_discrete(breaks=c("80", "81"), labels=c("positive", "negative")) + 
    ggtitle("Distribution of key presses") + theme(legend.position="none")

generalization_keypress_plot = data %>% 
  filter(stage=='generalization', key_press!=-1) %>% 
  ggplot(aes(x=i_generalization_trial, y=key_press, color=as.factor(subject))) + 
    geom_jitter(height=.1, alpha=.5) + 
    scale_y_discrete(breaks=c("80", "81", '32'), labels=c("positive", "negative", 'novel')) + 
    ggtitle("Distribution of key presses") + theme(legend.position="none")

grid.arrange(acquisition_plot, generalization_plot, acquisition_keypress_plot, generalization_keypress_plot, nrow = 2)

These data suggest that while subjects are not just distractedly responding with single key presses, their accuracy is far below ceiling.

Considering the overall performance rates for subjects in this pilot, I would like to increase the bonus for each correct trial from $0.02 to $0.04 in the full replication. The bonus earned by participants in pilot B was ~$1.00 for ~20 minutes (they were subsequently paid an extra $1.00, so their total compensation for ~20 minutes in the study was $2.50). I hope that increasing the bonus further incentivizes performance on the task, which will lead to more effective learning, better test of the hypothesis, and a more equitable wage for workers.