Beatriz R. Quineche1*,
M. Wyatt Toure1, Simon M. Reader1
1McGill University,
Department of Biology, 1205 Docteur Penfield, Montreal, Quebec H3A 1B1, Canada
*Project lead
Date of last update: Apr 20 2021
Brief Overview
This is the analysis for an experiment conducted by Beatriz Quineche as part of
an Honour’s thesis in the Reader Laboratory at McGill University. The data and R
code to produce this analysis are in the
beatriz-master-data_one.csv.csv
file and the
learning-chamber-analysis.Rmd
file respectively. A list of variables and their descriptions is given in the
metadata section of the README file. These files can all be accessed at the
GitHub repository for this project.
The goal of this project was to conduct of a proof of concept experiment to see
the feasibility of training guppies in a conditioning chamber which consisted of
a tank supplemented with raspberry pi controlled lights and feeders as well as a
raspberry pi camera. Guppies were given a simple classical conditioning task.
Four times a day for three consecutive days a light would briefly illuminate for
five seconds after which the light turned off and food was delivered from a
servo-powered feeder that was hung above the tank
(see Apparatus Design).
This resulted in a total of 12 reinforced trials being conducted in the span of
three days. On the fifth day guppies were food deprived to increase food
motivation and thus increase the likelihood of eliciting a demonstrable
behavioural effect for the probe trial which occurred on the sixth day.
The association expected to be formed was that the side of the tank that
produces a light predicts which feeder will have food delivered to it. Guppies
were expected to demonstrate evidence of successfully learning the association
by behaving in the following manner after the light turned on: 1) increasing
the amount of time spent in the side of the tank where the feeder was located,
2) increasing the number of visits (coming within 2 body lengths) to the
rewarded feeder, and/or 3) approaching the rewarded feeder quicker than the
unrewarded feeder.
Data check
Checking categorical variables
First we want to make sure the data is properly formatted. We will use the
describe_all_cat function from the package tidyext to do so. This will
summarise all of our categorical variables. Note ID was filtered out for this
table.
Table 1: Summary of the frequency of observations for all categorical variables.
The levels of each variable are given in the group column and the frequency of
observations within each level are given in the frequency column. The relative
contribution of a particular level of a variable to the total amount of
observations within that variable is givin in the % column.
Variable
Group
Frequency
%
batch
2
12
35
batch
3
12
35
batch
1
10
29
light.status
1before
17
50
light.status
2after
17
50
light.side
right
22
65
light.side
left
12
35
size
large
18
53
size
small
16
47
Of note is that we have slightly, but not significantly (Exact binomial test p
= 0.121), more individuals for which the light shone on the right side (65%)
versus the left side (35%). All other grouping levels are essentially equal.
Checking individual observations
Next we will check the data for individuals. All individuals should have two
measures, one for the ‘before light’ phase and one for the ‘after light’ phase.
Table 2: ID data check table caption
Variable
Group
Frequency
%
id
large_pi1_w1
2
6
id
large_pi1_w2
2
6
id
large_pi1_w3
2
6
id
large_pi2_w1
2
6
id
large_pi2_w2
2
6
id
large_pi2_w3
2
6
id
large_pi3_w1
2
6
id
large_pi3_w2
2
6
id
large_pi3_w3
2
6
id
small_pi1_w1
2
6
id
small_pi1_w2
2
6
id
small_pi1_w3
2
6
id
small_pi2_w1
2
6
id
small_pi2_w2
2
6
id
small_pi2_w3
2
6
id
small_pi3_w2
2
6
id
small_pi3_w3
2
6
Checking total time values
Since we used automated tracking we want to make sure that there are minimal
tracking errors. We measured the time a guppy was on either the left side or
right side of the tank so we can use this to check for any missing data. Since
trials lasted 3 minutes these measures should add up to around 180.
Table 3: Checking tracking data errors
id
light.status
total.time
large_pi2_w1
2after
175.6
small_pi2_w1
1before
175.6
small_pi3_w2
2after
177.8
small_pi2_w2
1before
178.0
small_pi1_w1
2after
178.2
small_pi1_w3
1before
178.2
large_pi3_w1
2after
178.6
large_pi3_w2
2after
178.8
large_pi2_w2
1before
179.2
small_pi1_w2
1before
179.2
small_pi2_w2
2after
179.2
large_pi3_w3
1before
179.4
small_pi3_w3
1before
179.6
large_pi1_w2
2after
179.6
small_pi1_w1
1before
179.6
large_pi1_w2
1before
179.8
large_pi3_w3
2after
179.8
small_pi1_w3
2after
179.8
large_pi1_w3
1before
180.0
large_pi3_w2
1before
180.0
large_pi1_w1
1before
180.2
large_pi1_w1
2after
180.2
large_pi1_w3
2after
180.2
large_pi2_w1
1before
180.2
large_pi2_w2
2after
180.2
large_pi2_w3
1before
180.2
large_pi3_w1
1before
180.2
small_pi1_w2
2after
180.2
small_pi2_w1
2after
180.2
small_pi2_w3
2after
180.2
small_pi3_w2
1before
180.2
small_pi3_w3
2after
180.2
large_pi2_w3
2after
180.2
small_pi2_w3
1before
180.2
Models
We analysed the data using linear mixed effect and generalized linear mixed
effect models with the lmer() and glmer() functions from lme4 package.
P-values and effective degrees of freedom were obtained using the lmerTest
package. Model residuals were checked they met distributional assumptions with
the DHARMa package, you can click the ‘See Model Residuals’ button below the
model formulas to see the residual diagnostic plots produced by DHARMa for
that particular model.
Model 1 - Time on the rewarded side
To determine whether individuals increase their time spent on the side of the
tank the light shone from, we fit a linear mixed effects model with fixed effect
of light status (before or after) as well as a random effect of individual id.
Our response variable, ‘rewarding side preference’, is the amount of time a
guppy spent on the side which the light shone from subtracted by the time spent
on the other side of the tank. This model asks whether the preference for the
rewarding side of the tank changed between baseline and test and whether this
differs with rewarded object colour.
Table 4: Summary of a linear mixed effect model (Model 1) estimating the time
spent on the side of the tank where light had been activated (model estimates
± S.E.) where the fixed effect is the factor light status (‘before the light’
or ‘after the light’) and the random effect of individual id. The effect of
light status is non-significant, but the estimate indicates guppies approached
the rewarding feeder quicker than the unrewarding feeder after the light came on.
Factor
Estimate
Std. Error
T statistic
df
P value
Intercept
-5.918
21.084
-0.281
29.473
0.781
Light status
27.671
25.075
1.104
16.000
0.286
There is a non-significant effect of light status (p = 0.286). Guppies
non-significantly increase their preference for the side with the light after
the light has shone by 27.7 seconds.
Figure 1: Data are means ± SE. Lines connect individuals across light periods. The dashed line represents the value of the pre-light baseline behavioural measure.
Model 2 - Latency to the rewarded side
To determine whether individuals increase the speed at which they come within
two body lengths of the rewarded feeder relative to the unrewarded feeder after
the light turns on we fit a linear mixed effects model. Our response variable
latency.difference is the latency to approach the unrewarded feeder subtracted
by the latency to approach the rewarded feeder. Positive values indicate that
the rewarded feeder was approached quicker than the unrewarded feeder. A value
of 0 would indicate both feeders were reached at the same time (in this case
this means both feeders were never approached and thus both values are given a
max score of 180 as it is impossible for guppies to be in two places at once).
The fixed effect is the light status which is either ‘before the light turns on’
or ‘after the light turns on’. We additionally fit a random effect of individual
id to account for repeated measures.
Table 5: Summary of a linear mixed effect model (Model 2) estimating how fast
guppies approached the unrewarded feeder for which the light had been activated
over the unrewarded feeder with no lights activated (model estimates ± S.E.)
where the fixed effect is the factor light status (‘before the light’ or ‘after
the light’). The effect of light status is non-significant, but the estimate
indicates guppies approached the rewarding feeder quicker than the unrewarding
feeder after the light came on.
Factor
Estimate
Std. Error
T statistic
df
P value
Intercept
-9.941
21.928
-0.453
32
0.653
Light status
39.612
31.011
1.277
32
0.211
After the light came on guppies non-significantly approached the rewarded feeder
on average 39.6 seconds faster than
the non-rewarded feeder.
Figure 2: Data are means ± SE Bold line connects means across trials.
Model 3 - Visits to the rewarded feeder
To determine whether individual increase the frequency with which they come
within two body lengths of the rewarded feder, we fit a binomial generalized
linear mixed effects model. Our response variable, denoted by
cbind(rewarding.feeder.visits,unrewarding.feeder.visits) is the proportion of
visits to the rewarded feeder. This model asks whether the proportion of visits
to the rewarded feeder differs between light phases. We fit a random effect of
individual id to account for repeated measures.
visits_model <-
glmer(cbind(rewarding.feeder.visits, unrewarding.feeder.visits) ~ light.status + (1 | id),
data = full_data,
family = "binomial"
)
Results
Table 6: Summary of a binomial generalized linear mixed effects model
(Model 3) estimating the proportion of visits guppies made to the activated
light side’s feeder (model estimates ± S.E.) where the fixed effect is the
factor light status (‘before the light’ or ‘after the light’). The effect of
light status is non-significant, but the estimate indicates guppies increase
their visits to the unrewarding feeder after its corresponding light had been
activated.
Factor
Estimate
Std. Error
T statistic
P value
Intercept
-0.108
0.238
-0.455
0.649
Light status
0.399
0.254
1.568
0.117
There is a non-significant increase in the proportion of visits to the rewarded
feeder after the light turns on. Guppies go from making 47% of their visits to
the rewarded feeder before the light shines to making 57% of their visits being
to the rewarded feeder after the light shines, an increase of 10%.
Figure 3: Data are means ± SE Bold line connects means across trials.
Main findings
Guppies non-significantly increase the amount of time they spend on the side
of the tank where the light shone (Figure 4A)
Guppies non-significantly increase the speed at which they visit the
rewarded feeder over the unrewarded feeder after the light shines
(Figure 4B)
Guppies non-significantly increase their proportion of visits to the
rewarded feeder after the light shines (Figure 4C)
Figure 4: Data are means ± SE. Dashed lines represent mean values for the pre-light baseline of the behaviour being displayed. Squares represent pre-light baselines and circles represent post-light measures. (A) Plot for rewarding side preference in seconds (B) Plot for rewarding feeder latency bias (C) Plot for proportion of visits to the rewarding feeder
We see that while the effect of the light turning is not significant for any one
preference measure, they all have an effect size that is in the direction
consistent with the hypothesis that guppies can learn a food-light association
in the automated chamber.
Our sample of guppies has a high amount of variation leading to estimates with
wide confidence intervals. Moreover, our sample is relatively small so the
decrease in estimate precision due to sampling error is amplified. There could
be several reasons for this variation. Given that we did not explicitly quantify
performance during training it may be that guppies that did not perform any of
the behaviours suggesting learning of the association did not feed as much as
guppies which did, leading to a difference in performance on the test trial. If
we assume guppies did receive similar reinforcement, there may still be
variation due to individual differences in the expression of learned behaviour.
Individual guppies may express learning in different manners. Some may choose to
spend more time on the side with the feeder while others may choose to make more
visits to the feeder. In this case, the strongest and most common responses are
more likely to produce statistically significant estimates as they will have
less variance. Latency we might expect to be a particularly noisy measure
because it depends on where an individual was in the tank when the light came
on. By chance, individuals may have been further away or closer to the feeder
when the light came on which could produce additional variation on the latency
metric. The low precision on the estimate of the effect of light status on time
spent on the rewarding side of the tank may be due to side biases. While
non-significant, guppies that had the light shine on the right side of the tank
increased their preference for the right side of the tank more than the guppies
that had the light shine on the left side of the tank increased their preference
for the left side of the tank. Problematically this could be an artefact of
sample size—there were nearly double the guppies for which the light shone on
the right side of the tank versus the left (11 on the right vs 6 on the left).
Follow up data exploration
Following the results for our a priori hypotheses we conducted some post-hoc
explorations into the data which are detailed in the sections below.
Side bias
From Simon: “I’d suggest also calculating time on left - right, and plotting
against side illuminated - this will help to illustrate any side bias”.
There is not a statistically clear side bias either way. For the ‘light on the
left’ guppies there appears to be a slight bias but this may be an artefact of
sample size (the confidence interval around the mean is very wide), there are 6
guppies for which the light shone on the left while there are almost double the
guppies for which the light shone on the right (11 guppies). Investigating this
in a model reveals no significant effects.
Time bins
From Simon: “One possibility is that an initial preference is wiped out once
fish discover no food. So you could look at 1st minute or whatever time period
seems sensible”
Behaviour might be structured temporally so we wanted to see whether there were
differences in behaviour that were apparent by binning the data into 1 minute
bins.
Tools used and References
A complete list of the tools used is produced below:
Package
Version
Reference
broom
0.5.5
David Robinson and Alex Hayes (2020). broom: Convert Statistical Analysis Objects into Tidy Tibbles. R package version 0.5.5. https://CRAN.R-project.org/package=broom
John Fox, Sanford Weisberg and Brad Price (2019). carData: Companion to Applied Regression Data Sets. R package version 3.0-3. https://CRAN.R-project.org/package=carData
Florian Hartig (2020). DHARMa: Residual Diagnostics for Hierarchical (Multi-Level / Mixed) Regression Models. R package version 0.3.3.0. http://florianhartig.github.io/DHARMa/
dplyr
1.0.3
Hadley Wickham, Romain François, Lionel Henry and Kirill Müller (2021). dplyr: A Grammar of Data Manipulation. R package version 1.0.3. https://CRAN.R-project.org/package=dplyr
effects
4.1.4
John Fox and Sanford Weisberg (2019). An R Companion to Applied Regression, 3rd Edition. Thousand Oaks, CA http://tinyurl.com/carbook
Yihui Xie (2020). knitr: A General-Purpose Package for Dynamic Report Generation in R. R package version 1.30.
lme4
1.1.21
Douglas Bates, Martin Maechler, Ben Bolker, Steve Walker (2015). Fitting Linear Mixed-Effects Models Using lme4. Journal of Statistical Software, 67(1), 1-48. doi:10.18637/jss.v067.i01.
lmerTest
3.1.1
Kuznetsova A, Brockhoff PB, Christensen RHB (2017). “lmerTest Package:Tests in Linear Mixed Effects Models.” Journal of StatisticalSoftware, 82(13), 1-26. doi: 10.18637/jss.v082.i13 (URL:https://doi.org/10.18637/jss.v082.i13).
R Core Team (2019). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/.
report
0.2.0
Makowski, D., Ben-Shachar, M.S., Patil, I. & Lüdecke, D. (2020). Automated reporting as a practical tool to improve reproducibility and methodological best practices adoption. CRAN. Available from https://github.com/easystats/report. doi: .
