WorkingMemoryAnalysis

#Load the packages and set the theme
library("tidyverse")
library("tidyr")
library("dplyr")
library("ggplot2")
library("cowplot")
library("afex")
library("ggbeeswarm")
library("emmeans")
theme_set(theme_bw(base_size = 15) + theme(legend.position="bottom"))

##################################################################
#                        TASK 2                                  #
##################################################################

##---------------------------------------------------------------------------------------------
## RQ1: "Does a longer versus a shorter gap lead to a retroactive or proactive memory benefit?" 
##---------------------------------------------------------------------------------------------


##### OPERATIONALISATION OF THE RESEARCH QUESTIONS ####
# RQ1: a) "Does the duration of unfilled gaps (long vs short) significantly impact serial recall accuracy?" 
# b) "Is the effect of gap length on serial recall accuracy moderated by whether the item appears before or after the gap (i.e., relative position)? 
# c) "Does a specific gap length provide a retroactive benefit (better accuracy for items before the gap), or a proactive benefit (better accuracy for items after the gap)?"


# RQ2: a) "Does engaging in a secondary task during the gap impact serial recall accuracy VS unfilled gaps of equal length (2.5s)?"
# b) "Is this effect moderated by the relative position of the item to the gap?"


#Operationalising the Research Question:
#Retroactive Benefit = higher accuracy for pre-gap items in conditions with long unfilled gaps compared to short gaps or filled gaps
#Proactive Benefit = higher accuracy for post-gap items in conditions with long unfilled gaps compared to short gaps or filled gaps
#Control condition: filled gap helps control for the passage of time - if performance differs between filled and long gaps, it suggests that free time (not just time passing) is the critical driving force of the effect

# Within-subjects 2 x 3 ANOVA 
#IV1 = amount of free time between trials (3 levels: long-gap, short-gap, filled gap)
#IV2 = relative position of serial position to the gap (2 levels: pre-gap or post-gap)
#DV = serial recall accuracy (0 = error, 1 = correct)

#Operationalisations of RQ:
#H1: type of gap (long, 2.5s vs short, 0.5s) has a significantly different effect on serial recall memory
  #Test via main effects: 
    #whether gap length has sig effect on serial recall accuracy, holding relative serial position (retroactive vs proactive memory) constant
    #whether or not a gap has sig effect on proactive or retroactive memory holding gap length constant
#H1: the effect of gap type depends on whether the gap is proactive vs retroactive?
  #Test via interaction:
    #whether the effect of gap length on serial recall accuracy (if there is one) significantly differs between retroactive vs proactive conditions
    #simple eff/ contrasts to see where the diff of the diff is 
#sign interaction would indicate that the effect of gap type on accuracy differs depending on whether the item is pre- or post-gap

##------------------------------------------------------------------------------
## RQ2: effect of filled vs unfilled gap type on proactive vs retroactive memory
##------------------------------------------------------------------------------

#### SORTING OUT THE DATA #####

#Loading in the data
data("wmgaps")
t2_data <- read_csv("data/wmgaps.csv")
glimpse(t2_data)


#Changing categorical variables into factors and re-labeling
t2_data <- t2_data %>% 
  mutate(
    id = factor(id),
    gap_type = factor(gap_type, levels = c("filled", "long", "short"),
                      labels = c("Filled", "Long", "Short")),
    rel_pos = factor(rel_pos, levels = c("pre-gap", "post-gap"),
                     labels = c("Before Gap", "After Gap")),
  )
glimpse(t2_data)
str(t2_data)
 
# Checking the factor levels:
  # unique(t2_data$gap_type)
  # unique(t2_data$rel_pos)
  # unique(t2_data$id)
  # unique(t2_data$accuracy) # I didn't change accuracy to a factor, even though it is binary categorical info, so proportion of correct responses can be calculated from it

#Count number of participants
t2_data %>% 
  summarise(n_ppnts = n_distinct(id))
#n_ppnts = 111

#Count total number of observations of accuracy
t2_data %>% 
  summarise(total_observations = n())
#there are 49950 observations, indicating each ppnt had multiple observations (within-subjects design)

# See how many non-missing value observations there are 
t2_data %>% 
  filter(!is.na(accuracy)) %>% 
  summarise(non_missing_observations = n())
#there are no missing observations (non_missing_observations = 49950)

# Check there are the same number of observations per participant
t2_data %>%
  group_by(id) %>%
  count() %>% 
  ungroup() %>% 
  summarise(all_n_450 = all(n == 450))
#All participants had 450 observations each

# Check if each combination of condition has the same number of observations
t2_data %>%
  group_by(gap_type, rel_pos) %>%
  summarise(n = n()) %>%
  summarise(equal_n_observation = all(n == first(n)))
#returns TRUE for all gap types

# Count number of data entries per subject for each serial position
t2_data %>%
  group_by(id, serial_pos) %>%
  summarise(n = n())
#All = 90; every participant provides 90 observations for each serial position 

##### DESCRIPTIVE ANALYSES #### 

# ANOVA models assume continuous data... so accuracy needs to be recalculated into something continuous no??
# Add columns for mean accuracy and standard deviation for each combination of gap type and relative position
t2_desc <- t2_data %>%
  group_by(gap_type, rel_pos) %>% 
  summarise(
    mean_accuracy = mean(accuracy),
    sd_accuracy = sd(accuracy),
    se_accuracy = sd(accuracy)/sqrt(length(accuracy)),
    median_accuracy = median(accuracy)) %>% 
    ungroup()
t2_desc
glimpse(t2_desc)
# 
# A tibble: 6 × 6
# gap_type   rel_pos    mean_accuracy sd_accuracy  se_accuracy  median_accuracy
# <fct>      <fct>              <dbl>       <dbl>       <dbl>            <dbl>
# 1 Filled   Before Gap         0.625       0.484     0.00531               1
# 2 Filled   After Gap          0.511       0.500     0.00548               1
# 3 Long     Before Gap         0.665       0.472     0.00517               1
# 4 Long     After Gap          0.558       0.497     0.00544               1
# 5 Short    Before Gap         0.672       0.469     0.00514               1
# 6 Short    After Gap          0.484       0.500     0.00548               0      

#Now lets visualise this data with a barchart, as the DV is discreet data
ggplot(t2_desc, aes(x = gap_type, 
                    y = mean_accuracy,
                    fill = rel_pos)) +
        geom_col(position = "dodge") + 
        scale_fill_manual(values = c("grey30", "grey70")) + #change fill of each bar to APA format (no colour)
        labs(title = "Mean Response Accuracy by Gap Type and Relative Position of Items",
             x = "Gap Type",
             y = "Mean Accuracy",
             fill = "Relative Position",
             caption = "*Error bars show SE of accuracy") + #Rename title, x axis, y axis and legend title
        geom_errorbar(aes(ymin = mean_accuracy - se_accuracy, #adding error bars (SE)
                          ymax = mean_accuracy + se_accuracy),
                      position = position_dodge(0.9),
                      width = 0.03,
                      alpha = 0.8,
                      size = 0.3) #making error bars not look silly


##### SET UP 2 X 3 REPEATED MEASURES ANOVA MODEL #### 
a1_t2 <- aov_ez("id", "accuracy", t2_data, 
                 within = c("gap_type", "rel_pos"))
a1_t2

  # Response: accuracy
  # Effect           df  MSE          F  ges p.value
  #            gap_type 1.64, 180.32 0.01  12.90 *** .012   <.001
  #             rel_pos       1, 110 0.01 348.36 *** .143   <.001
  #    gap_type:rel_pos 1.91, 210.52 0.00  25.56 *** .012   <.001
  #   Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘+’ 0.1 ‘ ’ 1
  # Sphericity correction method: GG 

#Very sig main effects and interaction. 
#As interaction is sig, main effects are not clearly interpretable 
#Lets look at the afex plot and follow-up tests

####### Plotting ANOVA results
plot1_t2 <- afex_plot(a1_t2, "gap_type", "rel_pos", 
                      error = "within",  
                      legend_title = "Relative Position",
            data_geom = ggbeeswarm::geom_beeswarm,
            data_arg = list(dodge.width = 0.5,
                            cex = 0.9)) +
            labs(title = "Effect of Relative Position and Gap Type on Mean Accuracy",
                 x = "Gap Type",
                 y = "Accuracy")
plot1_t2

##### FOLLOW-UP ANALYSES ######

##### ESTIMATED MARGINAL MEANS #####
em_t2 <- emmeans(a1_t2, c("rel_pos", "gap_type"))
em_t2

# rel_pos    gap_type emmean     SE  df lower.CL upper.CL
# Before.Gap Filled    0.625 0.0158 110    0.593    0.656
# After.Gap  Filled    0.511 0.0177 110    0.475    0.546
# Before.Gap Long      0.665 0.0145 110    0.636    0.693
# After.Gap  Long      0.558 0.0170 110    0.525    0.592
# Before.Gap Short     0.672 0.0135 110    0.645    0.699
# After.Gap  Short     0.484 0.0166 110    0.451    0.517
# Confidence level used: 0.95 

#Omnibus test of gap type on accuracy, accounting for rel pos
joint_tests(a1_t2, "gap_type")

# gap_type = Filled:
# model term df1 df2 F.ratio p.value
# rel_pos      1 110 145.630  <.0001
# 
# gap_type = Long:
# model term df1 df2 F.ratio p.value
# rel_pos      1 110 119.278  <.0001
# 
# gap_type = Short:
# model term df1 df2 F.ratio p.value
# rel_pos      1 110 261.362  <.0001


# Compare Before Gap vs After Gap for each gap type
pairs(em_t2, by = "gap_type", adjust = "holm")

# gap_type = Filled:
#   contrast               estimate      SE  df t.ratio p.value
# Before.Gap - After.Gap    0.114 0.00946 110  12.068  <.0001
# 
# gap_type = Long:
#   contrast               estimate      SE  df t.ratio p.value
# Before.Gap - After.Gap    0.106 0.00974 110  10.921  <.0001
# 
# gap_type = Short:
#   contrast               estimate      SE  df t.ratio p.value
# Before.Gap - After.Gap    0.188 0.01170 110  16.167  <.0001

#Sig effect of relative position of item to gap for long, short and filled gaps

##### RQ1 CONTRAST TIME ##### 

#Retroactive benefit = Long - short, before gap
#Proactive effect = Long - short, after gap

#   rel_pos    gap_type emmean     SE  df lower.CL upper.CL
# 1 Before.Gap Filled    0.625 0.0158 110    0.593    0.656
# 2 After.Gap  Filled    0.511 0.0177 110    0.475    0.546
# 3 Before.Gap Long      0.665 0.0145 110    0.636    0.693
# 4 After.Gap  Long      0.558 0.0170 110    0.525    0.592
# 5 Before.Gap Short     0.672 0.0135 110    0.645    0.699
# 6 After.Gap  Short     0.484 0.0166 110    0.451    0.517
# Confidence level used: 0.95 

con_list1 <- list(con1_retroactive = c(0,0,1,0,-1,0), #long before gap - short before gap 
                  con1_proactive = c(0,0,0,1,0,-1)) #long after gap - short after gap

con1 <- contrast(em_t2, con_list1, adjust = "holm")
con1

# contrast        estimate      SE  df t.ratio p.value
# con_retroactive -0.00769 0.00923 110  -0.833  0.407
# con_proactive    0.07423 0.00917 110   8.091  <.0001
# P value adjustment: holm method for 2 tests

#no significant difference between long vs short gaps for items appearing before the gap (p = 0.407) indicating no retroactive memory benefit 
#this suggests that extra free time isn't being used to reherse or consolidate previously seen items
#long gaps improved recall accuracy of items after the gap significantly compared to short gaps (p < .0001)
#suggests a proactive memory benefit: long gaps improved memory of serial position only for items appearing after the gap
#this proactive benefit suggests that the extra time is likely used to recharge cognitive resources, preparing participants better for subsequent information 

##### RQ2 CONTRAST TIME ##### 

#Retroactive Benefit: long - filled for items before gap
#Proactive Benefit: long - filled for items after gap

con_list2 <- list(con2_retroactive = c(-1,0,0,1,0,0), #long before - filled before
                  con2_proactive = c(0,-1,0,1,0,0)) #long after - filled after
  
con2 <- contrast(em_t2, con_list2, adjust = "holm")
con2

# contrast         estimate     SE  df t.ratio p.value
# con2_retroactive  -0.0664 0.0121 110  -5.503  <.001
# con2_proactive     0.0477 0.0124 110   3.856  <.001
# P value adjustment: holm method for 2 tests 

#significant difference between long - filled gaps for items appearing both before and after the gap (p < .0001)
#however, the direction of this difference flips across the two conditions
#there is a strong negative difference between long - filled gaps on recall accuracy (p < .001): suggests that for items presented before the gap, recall accuracy was better for filled gaps than unfilled gaps
#not only do having longer gaps not affect memory performance for items before the gap, filling those gaps with another task actually increases this retroactive memory performance compared to unfilled gaps
#this suggests that having free time after an item actually impairs retroactive memory of it
#Surprisingly, engaging in a secondary task (filling the gap) might help maintain better recall of earlier items, countering the idea that more free time allows for better memory via memory consolidation processes
#This is counterintuitive if assumed that free time is expected to help memory via rehearsal/ consolidation, indicating a different latent memory process is at play here which does not require attentional resources, but infact benefits when attention is engaged in something else. 
#In line with the previous results for RQ1, the difference in accuracy between long - filled was positive for items appearing after the gap, suggesting an unfilled gap improves performance only for items appearing after the gap
#this indicates that it is indeed the presence of free time, rather than purely the passing of time, contributing to the proactive memory benefit
#this indicates that there is a latent process at play during the gap, where having more time to restore memory resources for the upcoming items improves memory of them 
#suggests that the latent processes involved during the gap is involved in cognitive restoration (that is, recharging resources for the upcoming items), rather than rehersal or consolidation processes of items previously seen
#interestingly, filling the 2.5s of free time with a secondary task actually improves retroactive memory... this is counterintuitive and not in line with the memory consolidation hypothesis
#one possible reason for this is... ????
#overall, results suggest that memory consolidation latent processes are not involved in serial recall memory performance for short temporal time scales like this one, and might require longer than 2.5s to come into play... 
#if the gap was long enough e.g., overnight, for example, such that hippocampal replay can occur, then potentially a retroactive memory benefit will be seen. Not seen in short term working memory tasks like this one, more to do with long term memory.