2.1 Participants

sample_info <- df_sub_info %>%
  distinct(subject,gender,age) #only one row per subject

n_total <- nrow(sample_info)
n_female <- sum(sample_info$gender == "female")
age_mean <- mean(sample_info$age, na.rm = TRUE)
age_sd <- sqrt(sum((sample_info$age - age_mean)^2, 
                   na.rm=TRUE)/sum(!is.na(sample_info$age))) # population SD
age_min <- min(sample_info$age, na.rm = TRUE)
age_max <- max(sample_info$age, na.rm = TRUE)
mean_T1 <- mean(df_AB$T1)

Fourty-eight participants took part in total, 8 of whom were excluded after the first session. One participant was excluded excluded as a precaution because they developed an atypical headache after the first session, and we could not rule out this was related to the tDCS. Another stopped responding to our requests to schedule the second session. The remaining six participants were excluded because their mean T1 accuracy in the first session was too low, which would leave too few trials to analyze, because our T2 accuracy measure included only trials in which T1 was seen. We used a cut-off of 63% T1 accuracy as an exclusion criterium, which was two standard deviations below the mean of a separate pilot study (n = 10).

This left a final sample of 40 participants (29 female, mean age = 20.94, SD = 2.45, range = 18–28). This sample size was determined a priori to slightly exceed London and Slagter (2015) (n = 34). Mean T1 accuracy in the remaining 40 participants was 82%, which is comparable to previous studies using this task (86% in London and Slagter 2015; in 82% in Slagter and Georgopoulou 2013).

The experiment and recruitment took place at the University of Amsterdam. All procedures for this study were approved by the ethics review board of the Faculty for Social and Behavioral Sciences, and complied with relevant laws and institutional guidelines. All participants provided their written informed consent and were compensated with course credit or €10 per hour (typically €65 for completing two full sessions).

2.2 Procedure

# Session intervals in days
ds <- df_sub_info %>%
  select(subject,session,time) %>%
  spread(session,time, sep = "_") %>% # times for session_1 and session_2
  # count the number of integer days in the interval between sessions
  mutate(d_d = (session_1 %--% session_2) %/% days(1)) %>%
  count(d_d, sort = TRUE)

# determine session order
so <- df_sub_info %>%
  count(session,polarity) %>%
  filter(session == 1)

The study procedures were identical to London and Slagter (2015): participants received anodal and cathodal tDCS in separate sessions (Figure 2.1), which typically took place exactly one week apart (for 29 participants; sessions were separated by 6 days for 6 participants; 8 days for 3 participants; 4 days for 1 participant; 10 days for 1 participant). The time in between served to keep the sessions as similar as possible, and to minimize the risk of tDCS carry-over effects. 18 participants received anodal tDCS in the first session and cathodal tDCS in the second, and vice versa for the remaining 22 participants.

First, participants experienced the sensations induced by tDCS in a brief trial stimulation (see the tDCS section). Next, sEBR was measured for 6 minutes (see the sEBR section), after which participants completed 20 practice trials of the task (see the Task section). For the main portion of the experiment, participants performed three blocks of the task (Figure 2.1): before tDCS (baseline), during anodal/cathodal tDCS (tDCS), and after tDCS (post).

Within each block of the task, participants took a self-timed break every 50 trials (~5 minutes); between the blocks, the experimenter walked in. Participants performed the task for exactly 20 minutes during the baseline and post blocks. During the tDCS block, the task started after the 1-minute ramp-up of the current was complete, and continued for 21 minutes (constant current, plus 1-minute of ramp-down).

knitr::include_graphics("figures/figure_2_procedure.png", auto_pdf = TRUE)

Figure 2.1: Experimental design. Spontaneous eye blink rate was measured for 6 minutes prior to the start of the task. Then (following a short practice block), participants performed three 20-minute blocks of the attentional blink task: a baseline block without stimulation, a tDCS block during 20 minutes of anodal (red) or cathodal (blue) stimulation, and finally a post-test block (also without stimulation). The second session (typically 7 days later at the same time of day) was identical, except that the tDCS polarity was reversed.

2.3 Task

The attentional blink task (Figure 2.2) was almost identical to the one used in London and Slagter (2015) and Slagter and Georgopoulou (2013), which in turn was based on a task designed by Dux and Marois (2008). A rapid serial visual presentation stream of 15 letters (cf. 17 letters in London and Slagter (2015)) was shown on each trial, using Presentation software (Neurobehavioral Systems, Inc.). Each letter was displayed for 91.7 ms (11 frames at 120 Hz) on a dark gray background. The letters were presented in font size 40 (font: Courier New) at a viewing distance of 90 cm. On each trial, the letters were randomly sampled without replacement from the alphabet, excluding the letters I, L, O, Q, U and V, as they were too similar to each other. All distractor letters were mid-gray, whereas T1 and T2 were colored. T1 was red and always appeared at position 5 in the stream. T2 was green and followed T1 after either 2 distractors (lag 3) or 7 distractors (lag 8) (cf. lags 2, 4 and 10 in London and Slagter (2015)).

The letter stream was preceded by a fixation cross (same color as the letters) presented for 1750 ms and followed by another fixation cross (presented for 1000 ms). Finally, the participant was prompted to type in (using a standard keyboard) the letter they thought was presented as T1 (“Which letter was red?”), followed by T2 (“Which letter was green?”).

trial_counts <- df_AB %>%
  group_by(lag) %>%
  summarise_at(vars(trials), funs(mean, min, max, sd))

Trial duration varied slightly because both the T1 and T2 response questions were self-paced, so some participants completed more trials than others depending on their response times. On average, participants completed 130 short lag trials (SD = 17; range = 78–163) and 65 long lag trials (SD = 9; range = 39–87) per 20-minute block.

knitr::include_graphics("figures/figure_3_task.png", auto_pdf = TRUE)

Figure 2.2: Attentional blink task. Participants viewed rapid serial visual presentation streams of 15 letters, all of which were distractors (gray letters) except for T1 and T2. T1 was presented in red at position 5; T2 was presented in green and followed T1 after 2 distractors (lag 3, inside the AB window) or 7 distractors (lag 8, outside the AB window). At the end of the trial, participants reported the identity of T1 and then T2 (self-paced).

2.4 tDCS

Transcranial direct current stimulation was delivered online (i.e. during performance of the attentional blink task) using a DC-STIMULATOR PLUS (NeuroCare Group GmbH). The current was ramped up to 1 mA in 1 minute, stayed at 1 mA for 20 minutes, and was ramped down again in 1 minute.

One electrode was placed at F3 (international 10-20 system) to target the lDLPFC; the other was placed over the right forehead, centered above the eye (approximately corresponding to position Fp2). Both electrodes were 5 x 7 cm in size (35 cm²), leading to a current density of 0.029 mA/cm². The electrode montage and tDCS parameters are identical to London and Slagter (2015), with two exceptions. First, we also measured EEG (see the sEBR section), so the EEG electrodes and headcap were applied on top of the tDCS electrodes. Second, we used Ten20 paste (Weaver and Company) as the conductive medium, whereas London and Slagter (2015) used sponges soaked in saline.

Participants received either anodal tDCS (anode on F3, cathode on right forehead) or cathodal tDCS (cathode on F3, anode on right forehead) in separate sessions. The procedure was double-blinded: both the participant and the experimenters were unaware which polarity was applied in a given session. The experimenter loaded a stimulation setting on the tDCS device (programmed by someone not involved in data collection), without knowing whether it was mapped to anodal or cathodal tDCS. In the 2nd session, the experimenter loaded a second setting mapped to the opposite polarity (half the dataset), or simply connected the terminals of the device to the electrodes in the opposite way.

At the start of the experiment, participants received a brief trial stimulation, based on which they decided whether or not they wanted to continue with the rest of the session. The experimenter offered to terminate the experiment in case tDCS was experienced as too uncomfortable, but none of the participants opted to do so. For the trial stimulation, the current was ramped up to 1 mA in 45 seconds, stayed at 1 mA for 15 seconds, and was ramped down again in 45 seconds.

2.5 sEBR

Movement of the eyelids across the eyes affects the electrical potential that naturally exists across the eyeball (Matsuo, Peters, and Reilly 1975). Blinks can thus be recorded with electrodes placed on the face and/or scalp (Luck 2005).

We used a BioSemi ActiveTwo system with 64 Ag/AgCl active electrodes, placed according to the (10-10 subdivision of the) international 10-20 system. Two pairs of additional external electrodes were placed to record the electro-oculogram (EOG): above and below the left eye, and next to the left and right outer canthi. Finally, another pair of electrodes on the left and right earlobes served as the reference. This full setup was used because we also recorded the EEG during task performance. This dataset is available elsewhere (Reteig et al. 2019a).

# Session intervals in hours
hs <- df_sub_info %>%
  select(subject,session,time) %>%
  spread(session,time, sep = "_") %>% # times for session_1 and session_2
  # pull the hour of day and subtract
  mutate(d_h = abs(hour(session_2) - hour(session_1))) %>%
  count(d_h, sort = TRUE)

sEBR was recorded for 6 minutes after setting up the EEG in each session. Participants were asked to sit still and look straight ahead at a white wall (about 1 meter away). Participants were told they were allowed to move their eyes, but the experimenter made no mention of eye blinks. The “cover story” was that we needed to monitor the quality of the EEG signal before being able to start the recordings. Because blink rate can increase in the evening (Barbato et al. 2000), but is stable during the day time (Barbato et al. 2000; Doughty and Naase 2006), we made sure all recordings were completed before 5 PM. Most participants started their first and second sessions at the exact same time of day (34 participants; 4 participants started their second session 5 hours earlier/later than their first, 2 started 3 hours earlier/later).

The raw data were preprocessed using the EEGLAB toolbox (Delorme and Makeig 2004) in MATLAB (MathWorks, Inc.). First, data were re-referenced offline to the average of the earlobe electrodes. Next, horizontal and vertical EOG channels were created by subtracting the signals from each member of a horizontal/vertical electrode pair. A high-pass filter with a cut-off of .5 Hz was then applied. Finally, we ran an independent component analysis (ICA) to capture the eye blink events in a single time series. For each recording, we visually inspected the independent components and selected one that appeared to contain the eye blink signals, based on the waveform (large amplitude, positive deflections) and scalp distribution of the ICA weights (loading on frontal and EOG electrodes).

Eye blinks in this component were then detected using a semi-automatic procedure (cf. Slagter and Georgopoulou 2013; Kruis et al. 2016). First, a voltage threshold was set (initialized to the standard deviation of the signal) which captured most eye blink peaks. This threshold was moved up or down by the analyst if necessary. The sample with the maximum voltage between two threshold crossings was marked as an eye blink, with the restriction that two eye blinks must be at least 400 ms apart. We picked 400 ms as an upper estimate of the duration of a single blink (Caffier, Erdmann, and Ullsperger 2003). The analyst then inspected the output, and removed or added eye blinks in the case of clear false positives (e.g., a muscle contraction) or false negatives (e.g., an eye blink waveform did not exceed the threshold, or was followed by another clear eye blink within 400 ms).

2.6 Statistical analysis

# make text for general r-citations;
# withold certain packages to cite individually
r_citations <- cite_r(file = "r-references.bib", 
                      pkgs = c("broom","cowplot","here","tidyverse","papaja","knitr","lubridate"), 
                      withhold = FALSE,
                      footnote = TRUE)

Data were analyzed using R (Version 3.5.1; R Core Team 2018)² from within RStudio (Version 1.1.463; RStudio Team 2016). Our analyses were focused on three depedendent variables. sEBR was expressed as the number of eye blinks per minute. For the AB, we examined T2|T1 accuracy, i.e. the percentage of trials where T2 was reported correctly, out of the subset of trials in which T1 was reported correctly. The size of the attentional blink (AB magnitude) was quantified by subtracting T2|T1 accuracy at lag 3 from T2|T1 accuracy at lag 8, for each session in each block. Lastly, we created AB magnitude change scores for each session by subtracting AB magnitude in the “baseline” block from the “tDCS” and the “post” blocks, respectively.

2.7 Data, materials, and code availability

All of the data, materials, and code for this study are available on the Open Science Framework³. The raw task data and sEBR recordings can be downloaded from this page [FIXME: Insert data citation]. The code to preprocess the sEBR data and perform the semi-automatic eye blink detection (cf. Slagter and Georgopoulou 2013; Kruis et al. 2016) is supplied in MATLAB scripts. We provide the statistical analysis code in the form of an R notebook, detailing all the analyses that we ran for this project, along with the results. We also include an Rmarkdown (Xie, Allaire, and Grolemund 2018) source file for this paper that can be run to reproduce the pdf version of the text, along with all the figures and statistics.

3.1 Reliability

# Intra-class correlation for AB magnitude
ICC_ABmag <- df_ABmag %>%
  filter(block == "pre") %>%
  select(subject, session, AB_magnitude) %>%
  spread(session, AB_magnitude, sep = "_") %>% # create 2 sEBR columns: session 1 and session 2
  select(-subject) %>% # drop subjects column
  ICC(lmer=FALSE) # intra-class correlation

# Intra-class correlation for sEBR
ICC_sEBR <- df_sEBR %>%
  select(subject, session, sEBR) %>%
  spread(session, sEBR, sep = "_") %>% # create 2 sEBR columns: session 1 and session 2
  select(-subject) %>% # drop subjects column
  ICC(lmer=FALSE) # intra-class correlation, using AOV instead of LMER (no missing data)

pearson_ABmag <- df_ABmag %>%
  filter(block == "pre") %>%
  select(subject, session, AB_magnitude) %>%
  spread(session, AB_magnitude, sep = "_") %>%
  cor.test(data = ., .$session_1, .$session_2, 
                               method = "pearson", alternative = "two.sided")

pearson_sEBR <- df_sEBR %>%
  select(subject, session, sEBR) %>%
  spread(session, sEBR, sep = "_") %>% # create 2 sEBR columns: session 1 and session 2
  cor.test(data = ., .$session_1, .$session_2, 
                               method = "pearson", alternative = "two.sided")

# Plot AB over sessions
df_mean_AB <- df_AB %>%
  group_by(session = factor(session), lag = factor(lag)) %>%
  summarise(M = mean(T2_given_T1))

plot_lag3 <- df_AB %>%
  left_join(df_ABmag) %>%
  select(-T1,-T2,-trials) %>%
  filter(block == "pre") %>%
  spread(lag, T2_given_T1, sep = "_") %>% 
  mutate(session = factor(session),
         subject = fct_reorder(subject, AB_magnitude)) %>%
  
  ggplot(aes(subject,lag_8)) +
  #coord_flip() +
  geom_hline(data = df_mean_AB, aes(yintercept = M, colour = lag, linetype = session)) +
  geom_linerange(aes(x = subject, group = subject, ymin = lag_3, ymax = lag_8, linetype = session),
                 color = "grey", show.legend = FALSE, position = position_dodge2(.75)) +
  geom_point(shape = 21, aes(x = subject, y = lag_3), fill = "#E69F00", 
             colour = "white", stroke = 1, position = position_dodge2(.75)) +
  geom_point(shape = 21, aes(x = subject, y = lag_8), fill = "#D55E00", 
             colour = "white", stroke = 1, position = position_dodge2(.75)) +
  #geom_text(aes(label = round(AB_magnitude,2), x = T2_given_T1), size = 2) +
  theme_minimal_hgrid(font_size = base_font_size, font_family = base_font_family) +
  theme(axis.text.x=element_blank(), axis.ticks.x = element_blank(), axis.line.x = element_blank()) +
  xlab("participant") +
  scale_y_continuous("accuracy (T2|T1)", labels = scales::percent_format()) +
  scale_colour_manual(values = c("#E69F00", "#D55E00")) +
  NULL

# function for test-retest plots
paired_plot <- function(df, lims, break_int, labs = waiver()) {

  df_axes <- data.frame(
  x = c(lims[1], -Inf), 
  xend = c(lims[2], -Inf),
  y = c(-Inf, lims[1]),
  yend = c(-Inf, lims[2])
)
  
  ggplot(df, aes(session_1,session_2)) +
  geom_abline(color = "grey") +
  geom_point(shape = 21, fill = "gray25", size = 2, color = "white", stroke = .3) +
  geom_segment(
    data = df_axes,
    aes(x = x, xend = xend, y = y, yend = yend),
    size = 0.5,
    inherit.aes = FALSE
  ) +
  scale_x_continuous(breaks = seq(lims[1],lims[2],break_int),
                     labels = labs,
                     expand = expand_scale(mult = c(.1, .05))) +
  scale_y_continuous(breaks = seq(lims[1],lims[2],break_int), 
                     labels = labs,
                     expand = expand_scale(mult = c(.1, .05))) +
  coord_fixed(clip ="off") +
  theme_half_open(font_size = base_font_size, font_family = base_font_family) +
  theme(axis.line = element_blank()) +
  NULL
    
}

# plot AB test-retest
ICC_label_ABmag = paste0("ICC(2,1) = ", 
                         printnum(ICC_ABmag$results["Single_random_raters","ICC"], gt1 = F),
                         ",\n",
                         "95% CI[",
                         printnum(ICC_ABmag$results["Single_random_raters","lower bound"], gt1 = F), 
                         ",", 
                         printnum(ICC_ABmag$results["Single_random_raters","upper bound"], gt1 = F)
                         ,"]")

plot_AB <- df_ABmag %>%
  filter(block == "pre") %>%
  select(subject, session, AB_magnitude) %>%
  spread(session, AB_magnitude, sep = "_") %>%
  
  paired_plot(c(.1,.7),.1, scales::percent_format(accuracy = 1)) +
  annotate("text", label = ICC_label_ABmag, size = geom_text_size,
           x = .1, y = .6, hjust = "left", vjust = "center") +
  labs(title = "Attentional Blink magnitude", x = "session 1", y = "session 2") +
  NULL

We first examined task performance in the baseline block of both sessions, i.e. before tDCS onset. All participants showed the characteristic AB effect in both sessions: T2|T1 accuracy was higher for lag 8 trials than for lag 3 (Figure 3.1A).

There was also a significant difference in AB magnitude (lag 8 minus lag 3 T2|T1 accuracy) between the sessions: (F(1, 39) = 16.53, p < .001). For most participants, AB magnitude was smaller in the 2nd session than the first (Figure 3.1B). The average difference in AB magnitude over sessions seemed to be driven by lag 3 performance only (Figure 3.1A), meaning the AB genuinely decreased with practice.

To be able to uncover relationships between sEBR and AB magnitude (see the subsequent sections), it is crucial that the test-retest reliability of both measures is adequate, i.e. that there is substantial agreement between the scores in session 1 and 2.

The intraclass correlation for AB magnitude (Figure 3.1B) is .60, indicating “moderate” test-retest reliability (Koo and Li 2016), with a 95% confidence interval of .25 (poor) – .79 (good). The standard (interclass) Pearson correlation for AB magnitude between sessions is comparable (r(38) = .68, CI_95% [.47, .82]) and comparable to previous reports (Dale, Dux, and Arnell 2013).

# plot sEBR test-retest
ICC_label_sEBR = paste0("ICC(2,1) = ", 
                        printnum(ICC_sEBR$results["Single_random_raters","ICC"], gt1 = F),
                        ",\n",
                        "95% CI[", 
                        printnum(ICC_sEBR$results["Single_random_raters","lower bound"], gt1 = F), 
                        ",", 
                        printnum(ICC_sEBR$results["Single_random_raters","upper bound"], gt1 = F),
                        "]")

plot_sEBR <- df_sEBR %>%
  select(subject, session, sEBR) %>%
  spread(session, sEBR, sep = "_") %>%
  
  paired_plot(c(0,60),10) +
  annotate("text", label = ICC_label_sEBR, size = geom_text_size,
           x = 0, y = 50, hjust = "left", vjust = "center") +
  labs(title = "spontaneous Eye Blink Rate", 
       x = "session 1", 
       y = "session 2")

# combine plots
fig_retest_bottom <- plot_grid(plot_AB, plot_sEBR, labels = c("B","C"), align = 'h')
fig_retest <- plot_grid(plot_lag3, fig_retest_bottom, labels = c("A",""), 
                        ncol = 1, rel_heights = c(1, 1.33))
fig_retest

Figure 3.1: Reliability of the attentional blink and spontaneous eye blink rate. (A) All participants showed an attentional blink in both sessions: higher T2|T1 accuracy (% T2 correct in trials where T1 was also correct) for lag 8 (orange) than lag 3 (yellow). Horizontal lines show group-average T2|T1 accuracy. The attentional blink magnitude (lag 8 - lag 3) is slightly smaller in the 2nd session (dotted lines) than the first session (solid lines), due to better lag 3 performance on average. (B) AB magnitude for each participant in session 1 vs. 2. The intraclass correlation indicates moderate test-retest reliability, though the 95% confidence interval ranges all the way from poor to good. AB magnitude in (A) and (B) was calculated on the baseline block only, before tDCS onset. (C) sEBR values for each participant in session 1 vs. 2. The intraclass correlation indicates that the test-retest reliability for sEBR is good.

sEBR_EEG <- df_sEBR %>%
  left_join(df_sub_info, by = c("subject", "session", "polarity")) %>%
  group_by(EEG_experience,session) %>%
  summarise(n = n_distinct(subject), median_sEBR = median(sEBR))

median_sEBR2 <- median(df_sEBR$sEBR[df_sEBR$session == 2])
median_sEBR1 <- median(df_sEBR$sEBR[df_sEBR$session == 1])
sEBR_S27 <- df_sEBR$sEBR[df_sEBR$subject == "S27" & df_sEBR$session == 2]
sEBR_S27_diff <- diff(df_sEBR$sEBR[df_sEBR$subject == "S27"])

median_sEBR1_EEG <- sEBR_EEG$median_sEBR[sEBR_EEG$EEG_experience == "yes" & sEBR_EEG$session == 1]
median_sEBR2_EEG <- sEBR_EEG$median_sEBR[sEBR_EEG$EEG_experience == "yes" & sEBR_EEG$session == 2]
n_EEG <- sum(df_sub_info$EEG_experience == "yes" & df_sub_info$session == 1)
n_smoker <- sum(df_sub_info$smoker == "yes" & df_sub_info$session == 1)
n_contacts <- sum(df_sub_info$contacts_glasses == "contacts" & df_sub_info$session == 1)

In contrast to AB magnitude, sEBR (Figure 3.1C), did not differ significantly between sessions (F(1, 39) = 0.149, p = 0.701). We had some concerns that participants would blink less in the 2nd session, because they had been instructed (after the sEBR measurement in session 1) that blinking can cause artifacts in the EEG signal (recorded during task performance). Yet, if anything, median sEBR was slightly higher in the 2nd session (12.6) than the first (11.7). However, we also asked participants whether they had been in an EEG experiment before. Participants that had done so exhibited a greatly reduced median sEBR in both sessions (6.3 in session 1; 4.1 in session 2). Because these were only 6 cases in an already small sample, we are unsure whether this finding is robust, but it is a cautionary note to others aiming to measure sEBR using (a full setup of) EEG electrodes. Because smoking has been reported to increase sEBR (Klein, Andresen, and Thom 1993), we also asked whether participants self-identified as a smoker (n = 5). These individuals were not clear outliers in the distribution, neither were those wearing contact lenses (n = 4), which also generally should increase blink rate.

Most importantly, the test-retest reliability for sEBR was “good” (Koo and Li 2016), indicated by an intraclass correlation of .85 CI_95% [.73, .92]. The Pearson correlation was r(38) = .84, CI_95% [.72, .91] (cf. Dang et al. 2017); Cronbach’s alpha was .91 CI_95% [.84, .95] (cf. Kruis et al. 2016).

One participant’s sEBR in session 2 seems remarkably high (57 blinks per minute), and is quite a bit higher in session 2 than in session 1 (difference of 12). However, we confirmed this was not a data quality issue, and rerunning the analyses without this participant did not qualitatively change the results in the subsequent sections. Hence, this participant was included in all analyses.

3.2 Relation between sEBR and baseline AB magnitude

lCorr_AB <- df_sEBR %>%
  left_join(df_ABmag) %>%
  filter(block == "pre") %>% # only baseline block
  nest(-session) %>% # for each session
  mutate(stats = map(data, ~cor.test(.$sEBR, .$AB_magnitude, 
                                     alternative = "two.sided", 
                                     method = "pearson"))) %>% # correlate sEBR to AB magnitude
  unnest(map(stats,tidy), .drop = TRUE) # make output into data frame, drop rest

uCorr_pre_session1 <- df_sEBR %>%
  left_join(df_ABmag) %>%
  filter(block == "pre", session == 1) %>%
  twolines(AB_magnitude~sEBR, graph = 1, data = .)

uCorr_pre_session2 <- df_sEBR %>%
  left_join(df_ABmag) %>%
  filter(block == "pre", session == 2) %>%
  twolines(AB_magnitude~sEBR, graph = 1, data = .)

# Make df with all results of model fits
x_1 <- pull(filter(df_sEBR, session == 1),sEBR)
x_2 <- pull(filter(df_sEBR, session == 2),sEBR)

df_u <- bind_cols(tribble(
  ~session,    ~breakpoint,           ~b_1,                  ~b_2, 
  "session 1", uCorr_pre_session1$xc, uCorr_pre_session1$b1, uCorr_pre_session1$b2,
  "session 2", uCorr_pre_session2$xc, uCorr_pre_session2$b1, uCorr_pre_session2$b2,
), tribble(
  ~intercept,                               ~d, 
  uCorr_pre_session1$glm1$coefficients[1],  uCorr_pre_session1$glm1$coefficients[4], 
  uCorr_pre_session2$glm1$coefficients[1],  uCorr_pre_session2$glm1$coefficients[4],
), tribble(
  ~fit,                                        ~x,               ~xmin,     ~xmax,
  uCorr_pre_session1$yhat.smooth[order(x_1)],  x_1[order(x_1)],  min(x_1),  max(x_1),
   uCorr_pre_session2$yhat.smooth[order(x_2)], x_2[order(x_2)],  min(x_2),  max(x_2)
))

Based on the results of Colzato et al. (2008), we should expect to find a negative correlation between sEBR and AB magnitude (in the baseline block). However, the correlation here was not significant in either session 1 (r(38) = .08, CI_95% [-.24, .38], p = .637), or session 2 (r(38) = .003, CI_95% [-.31, .31], p = .987). The direction of the effect is close to zero or slightly positive (Figure 3.2), which is not in accord with Colzato et al. (2008).

BF_default <- df_sEBR %>%
  left_join(df_ABmag) %>%
  filter(block == "pre") %>% # only baseline block
  nest(-session) %>% # for each session
  mutate(BF_null = 1 / map_dbl(data, ~extractBF(correlationBF(.$sEBR, .$AB_magnitude), 
                     onlybf = TRUE))) # correlate sEBR to AB magnitude

BF_one_sided <- df_sEBR %>%
  left_join(df_ABmag) %>%
  filter(block == "pre") %>% # only baseline block
  nest(-session) %>% # for each session
  mutate(BF_one_sided = map(data, ~rownames_to_column(extractBF(
    correlationBF(.$sEBR, .$AB_magnitude, nullInterval = c(-1,0)))))) %>% # correlate sEBR to AB magnitude
  unnest(BF_one_sided, .drop = T) %>%
  mutate(BF_null = 1/ bf) %>% # transform to evidence for null
  filter(!str_detect(rowname, "!")) # filter out complement to interval

bf0R_1 <- 1 / repBfR0(nOri = 20, rOri = -.530, nRep = lCorr_AB$parameter[1]+2, rRep = lCorr_AB$estimate[1])
bf0R_2 <- 1 / repBfR0(nOri = 20, rOri = -.530, nRep = lCorr_AB$parameter[2]+2, rRep = lCorr_AB$estimate[2])

Bayesian correlations show that the data are BF₀₁ = 2.57 times (session 1) and BF₀₁ = 2.84 times (session 2) more likely under the null hypothesis, using the default prior. This constitutes some evidence for absence of a correlation, though not more than anecdotal. If we evaluate the prior over negative effect sizes only (based on the negative correlation in Colzato et al. 2008), the support for the null increases slightly and becomes moderate (session 1: BF_0- = 3.90, session 2: BF_0- = 2.87). Finally, if we use the correlation as in Colzato et al. (2008) (r(18) = -.53) as a prior (Wagenmakers, Verhagen, and Ly 2016), the support for the null hypothesis becomes strong (session 1: BF_0r = 15.80, session 2: BF_0r = 10.43).

We also probed for an Inverted-U-shaped relationship between sEBR and AB magnitude, using the “two lines test” (Simonsohn 2018). In session 1, a cubic spline-fit indeed suggests an Inverted-U-shaped relationship (Figure 3.2). The linear regressions do as well, since the first slope is positive (b = .012, p = .058) and the second is negative (b = -.006, p = .602). However, neither slope is significant. Furthermore, this pattern is absent in session 2 (line 1: b = .001, p = .931; line 2: b = .003, p = .545), with the spline fit also showing a more erratic pattern.

df_sEBR %>%
  left_join(df_ABmag) %>%
  mutate(session = paste("session", session)) %>% # add "session" to facet labels
  left_join(select(df_u, session, breakpoint)) %>% # add breakpoint info
  filter(block == "pre") %>% # AB mag from baseline only
  mutate(reg = factor(ifelse(sEBR <  breakpoint, 1, 2))) %>% # colors based on breakpoint
  
  ggplot(aes(sEBR, AB_magnitude)) +
  facet_wrap(~session, ncol = 1) + # for each session
  geom_segment(data = unnest(df_u), # plot the first linear fit
               aes(x = xmin, xend = breakpoint,
                   y = intercept + b_1*(xmin - breakpoint), # solve interrupted regression
                   yend = intercept),
               colour = "#009E73") +
  geom_segment(data = unnest(df_u), # plot the 2nd linear fit
               aes(x = breakpoint, xend = xmax, 
                   y = intercept + d, 
                   yend = intercept + b_2*(xmax - breakpoint) + d),  # solve for ending-y
               colour = "#CC79A7") +
  geom_smooth(method = "lm", se = FALSE, colour = "gray50") + # plot the overall linear fit
  geom_point(shape = 21, fill = "gray25", size = 2, aes(colour = reg), stroke = .3, show.legend = FALSE) +
  geom_path(data = unnest(df_u), aes(x = x, y = fit), # plot the spline fit
            lty = "dashed", colour = "gray50", size = 1) +
  scale_x_continuous("spontaneous Eye Blink Rate (per min)", limits = c(0,60),
                     breaks = seq(0,60,10), expand = c(0,0)) +
  scale_y_continuous("Attentional blink magnitude", limits = c(.1,.7), 
                     breaks = seq(.1,.7,.1), labels = scales::percent_format(accuracy = 1)) +
  scale_colour_manual(values = c("#009E73", "#CC79A7")) +
  theme_minimal_grid(font_size = base_font_size, font_family = base_font_family) +
  theme(strip.background = element_rect(fill = "grey90")) +
  NULL

Figure 3.2: No significant relationships between sEBR and AB magnitude in the block before tDCS onset. Grey solid lines show a linear fit over all data points (individual participants), with no clear relationship in both sessions. Grey dashed lines show a cubic spline fit over all data points. Colored lines show two separate linear fits, delimited by the break point in the cubic spline, as estimated with the “two-lines test” (Simonsohn 2018). Both the spline fit and the two linear slopes suggest an Inverted-U shaped relationship in session 1, but neither slope is significant, and this pattern is not present in session 2.

3.3 Relation between sEBR and the effect of tDCS on AB magnitude

r_ABmag_sEBR <- df_sEBR %>%
  left_join(df_ABmag, by = c("subject", "session", "polarity")) %>%
  nest(-polarity,-block) %>% # for each session
  mutate(stats = map(data, ~cor.test(.$sEBR, .$AB_magnitude, 
                                     alternative = "two.sided", 
                                     method = "pearson"))) %>% # correlate sEBR to AB magnitude
  unnest(map(stats,tidy), .drop = TRUE)

sd_ABmag <- df_ABmag %>%
  group_by(polarity,block) %>%
  summarise(SD = sd(AB_magnitude))

r_ABmag <- df_ABmag %>%
  spread(block, AB_magnitude) %>% # separate column of scores for each block
  rename(baseline_AB_magnitude = pre) %>% # so we can keep "pre" as separate column
  gather(block, AB_magnitude, tDCS, post) %>% # collapse tdcs and post columns
  nest(-block,-polarity) %>% # for each tdcs/post block and anodal/cathodal
  mutate(stats = map(data, ~cor.test(.$baseline_AB_magnitude, .$AB_magnitude, 
                                     alternative = "two.sided", 
                                     method = "pearson"))) %>% # correlate blocks
  unnest(map(stats,tidy), .drop = TRUE)

df_ABmag_change <- df_ABmag %>%
  group_by(subject,session,polarity) %>%
   summarise(`tDCS - baseline` = AB_magnitude[block == "tDCS"] - AB_magnitude[block == "pre"],
             `post - baseline` = AB_magnitude[block == "post"] - AB_magnitude[block == "pre"]) %>%
   gather(change, AB_magnitude, `tDCS - baseline`, `post - baseline`) %>%
   mutate(change = fct_relevel(change, "tDCS - baseline")) # reorder so this comes first

lCorr_tDCS <- df_sEBR %>%
  left_join(df_ABmag_change) %>%
  nest(-polarity,-change) %>% # for anodal/cathodal & during/after
  mutate(stats = map(data, ~cor.test(.$sEBR, .$AB_magnitude, 
                                     alternative = "two.sided", 
                                     method = "pearson"))) %>% # correlate sEBR to AB magnitude
  unnest(map(stats,tidy), .drop = TRUE) # make output into data frame, drop rest

lCorr_tDCS_BF <- df_sEBR %>%
  left_join(df_ABmag_change) %>%
  nest(-polarity,-change) %>% # for anodal/cathodal & during/after
  mutate(BF_null = 1 / map_dbl(data, ~extractBF(correlationBF(.$sEBR, .$AB_magnitude), 
                     onlybf = TRUE))) %>% # correlate sEBR to AB magnitude
  select(-data)

corr_co <- r_ABmag_sEBR %>% # reshape data frame for joining later
  select(polarity, block, estimate, parameter) %>% # only keep correlations, and key columns
  spread(block, estimate) %>%
  gather(block, estimate, tDCS, post) %>% # separate column for subsequent blocks
  left_join(r_ABmag, by = c("polarity", "block")) %>% # join with reliability df
  rename(r_reliability = estimate.y, r_subsequent = estimate.x, r_baseline = pre) %>%
  nest(-polarity, -block) %>% # for each polarity and block
  mutate(stats = map(data, ~cocor.dep.groups.overlap(.$r_baseline, .$r_subsequent, .$r_reliability,
                                     n = n_total, return.htest = TRUE))) %>% # compare correlations
  # This yields a nested list of 10 different statistical test, for each 4 comparsions
  unnest(stats, .drop = TRUE) %>% # peel off first layer
  mutate(method = as.character(map(stats, "method"))) %>% # add column identifying lists with this method
  filter(method == "Pearson and Filon's z (1898)") %>% # keep only them
  select(-method) %>%
  unnest(map(stats,tidy), .drop = TRUE) # unnest lists into table

# plot_tDCS_corr_labels <- lCorr_tDCS %>%
#   left_join(lCorr_tDCS_BF) %>%
#   mutate(label = paste0("r(", parameter ,")", " = ", printnum(estimate,gt1=F),
#                         ", p = ", printp(p.value),
#                         ", BF = ", round(BF_null,1)))

# separate df so arrows only show on one facet
AB_arrows_y <- tibble(x = c(-16, -16),
                    xend = c(-16, -16),
                    y = c(-.05, .05),
                    yend =  c(-.32, .32),
                    change = factor("tDCS - baseline", levels = levels(df_ABmag_change$change)))

AB_labs_y <- tibble(x = c(-20, -20),
                    y = c(-.32, .32),
                    label = c("Smaller AB","Larger AB"),
                    hjust = c("left","right"),
                    change = factor("tDCS - baseline", levels = levels(df_ABmag_change$change)))

df_sEBR %>%
  left_join(df_ABmag_change) %>%
  
  ggplot(aes(sEBR, AB_magnitude)) +
  facet_grid(polarity~change) +
  geom_hline(yintercept = 0, linetype = "dashed", color = "gray") +
  geom_smooth(aes(color = polarity, linetype = change), method = "lm", show.legend = FALSE) +
  geom_point(shape = 21, fill = "gray25", size = 2, color = "white", stroke = .3) +
  # geom_label(data = plot_tDCS_corr_labels, aes(label = label), x = 62, y = .3, 
  #            size = geom_text_size, label.size = 0, hjust = "right", vjust = "top") +
  scale_color_manual(values = c("#EF7881", "#ABCCE7")) +
  scale_x_continuous("spontaneous Eye Blink Rate (per min)", breaks = seq(0,60,10)) +
  scale_y_continuous("Change in AB magnitude", breaks = seq(-.3,.3,.1), 
                     labels = scales::percent_format(accuracy = 1)) +
  coord_cartesian(xlim = c(0, 60), ylim = c(-.3, .3), clip = "off") +
  geom_segment(data = AB_arrows_y, # y-axis arrows
               aes(x = x, xend = xend, y = y, yend = yend), 
               arrow = arrow(angle = 20, length = unit(4,"mm"))) +
  geom_text(data = AB_labs_y, # y-axis arrow text
            aes(x = x, y = y, label = label, hjust = hjust), size = geom_text_size, angle = 90) + 
  theme_minimal_grid(font_size = base_font_size, font_family = base_font_family) +
  theme(strip.background = element_rect(fill = "grey90"), 
        panel.spacing = unit(1,"lines"), # add spacing between facets
        # move axis labels away from annotatations
        axis.title.y = element_text(margin = margin(t = 0, r = 23, b = 0, l = 0))) +
  NULL

Figure 3.3: No significant relationships between sEBR and AB magnitude change scores. Each plot shows spontaneous eye blink rates on the x-axis, and the change in AB magnitude on the y-axis (difference scores of the tDCS block - the baseline block, or the post block - baseline) in the anodal and cathodal stimulation conditions.

Although there was no relationship between sEBR and AB magnitude itself, sEBR could potentially still be associated with tDCS-induced changes in AB magnitude. We therefore computed AB magnitude change scores (tDCS - baseline, post - baseline) in each stimulation condition (anodal, cathodal), and correlated these to sEBR. However, this analysis did not uncover any significant relationships (Figure 3.3 and columns 2–4 in Table 3.2). We also separately computed the correlations between sEBR and AB magnitude in the baseline block and in the subsequent blocks (Table 3.1). A Pearson-Filon test revealed that these correlations did not differ signifcantly from each other (columns 5–6 in Table 3.1), again suggesting there was no relationship between sEBR and AB magnitude change scores.

ABmag_tbl <- r_ABmag_sEBR %>%
  left_join(sd_ABmag) %>%
  select(polarity,block,estimate,SD) %>%
  mutate(block = factor(block, levels = c("pre","tDCS","post")),
         block = fct_recode(block, baseline = "pre")) %>% # reorder so this comes first) 
  arrange(polarity,block) %>%
  select(block,estimate,SD)

ABmag_tbl %>%
  kable(col.names = c("Block", 
                      paste0("$r$ sEBR", footnote_marker_alphabet(1)),
                      paste0("SD", footnote_marker_alphabet(2))),
        caption = "Variability of attentional blink magnitude scores and correlation with sEBR, per stimulation condition and block.",
        booktabs = TRUE,
        digits = 3,
        escape = FALSE) %>%
  kable_styling() %>%  
  pack_rows(index = c("anodal session" = 3, "cathodal session" = 3)) %>%
  column_spec(1, width = "10em") %>%
  footnote(escape = FALSE, threeparttable = TRUE, alphabet = c(
    "Pearson's correlation between AB magnitude and spontaneous eye blink rate",
    "Standard deviation of AB magnitude"))

corr_co_tbl <- corr_co %>%
  select(polarity, change = "block", Z = "statistic", p = "p.value", r_reliability = "estimate3") %>%
  mutate(change = factor(change), # make equal to other data frame
         change = fct_recode(change,  `tDCS - baseline` = "tDCS", `post - baseline` = "post"),
         change = fct_relevel(change, "tDCS - baseline")) %>%
  left_join(lCorr_tDCS, by = c("polarity", "change")) %>% # Pearson correlations difference scores
  left_join(lCorr_tDCS_BF, by = c("polarity", "change")) %>% # Bayesian correlations difference scores
  arrange(polarity) %>% # order rows
  select(change, estimate, p.value, BF_null, Z, p, r_reliability) # order columns

corr_co_tbl %>%
  kable(col.names = c("contrast","$r$","$p$","$BF_{01}$","$Z$","$p$","$r$"),
        digits = 3,
        caption = "Attentional blink magnitude and spontaneous eye blink rate correlations.",
        booktabs = TRUE,
        escape = FALSE) %>%
  kable_styling() %>%
  pack_rows(index = c("anodal session" = 2, "cathodal session" = 2)) %>%
  add_header_above(setNames(c(1,3,2,1), c(" ",
              paste0("Correlation", footnote_marker_alphabet(1, double_escape = TRUE)),
              paste0("Pearson-Filon test", footnote_marker_alphabet(2, double_escape = TRUE)),
              paste0("Reliability", footnote_marker_alphabet(3, double_escape = TRUE)))),
                   escape = FALSE) %>%
  footnote(escape = FALSE, threeparttable = TRUE, alphabet = c(
    "Linear correlation (Pearson and Bayesian) between sEBR and AB magnitude change scores",
    "Test for significant differences between the sEBR vs. AB magnitude correlations (reported in Table \\3.1) in the baseline vs. tDCS or post blocks",
    "Pearson correlation between AB magnitude scores in the baseline vs. tDCS or post blocks"))

4.1 sEBR and AB magnitude have good to moderate test-retest reliability

The test-retest reliability of sEBR across two testing sessions (separated by about one week) was “good”, bordering on “excellent” (Koo and Li 2016). Only two other studies examined the reliability of sEBR to date. Dang et al. (2017) (n = 18) found a Pearson correlation of .86 between sEBR under administration of bromocriptine (a dopamine agonist) or placebo (separated by 4 hours). Here, Pearson’s correlation was comparable (.84). In Kruis et al. (2016), Cronbach’s alpha for sEBR ranged from .85 (three measurements of 27 long-term meditators) to .79 (2–3 measurements of 114 meditation-naive participants). In our data, Cronbach’s alpha was even slightly higher (.91). Together, all three studies suggest that sEBR scores are trait-like and stable over time, even for short measurements of only several minutes (6 minutes here and in Kruis et al. (2016); 5 minutes in Dang et al. (2017)). Although the reliability estimates are comparable, we measured sEBR twice under baseline conditions, whereas Dang et al. (2017) administered a dopamine agonist in one measurement (vs. placebo), and Kruis et al. (2016) measured sEBR following meditation practice (vs. a regular day). We did find that participants with prior EEG experience exhibited a two- to threefold lower sEBR. Future studies that measure sEBR with EEG electrodes might want to take this into account. However, this observation was based on only 6 participants, and sEBR did not significantly differ between sessions, despite the fact that all participants had EEG experience after the first session.

In contrast to sEBR, AB magnitude was significantly smaller in the second session. Previous studies have reported that performance on an AB task can improve across sessions (Dale, Dux, and Arnell 2013; Slagter et al. 2007), but for targets at all lags, whereas here we observed a specific increase in lag 3 T2|T1 performance (inside the attentional blink window). Test-retest reliability of AB magnitude was lower than sEBR: the point-estimate indicated only “moderate” reliability, though the 95% confidence interval was also consistent with “poor” to “good” reliability. However, this is comparable to previous reports (see Table 4.1). The Pearson correlation in the present study (r = .68) is even on the higher end of the range reported in previous studies (though note that regular correlations can overestimate “true” test-retest reliability (Bland and Altman 1986)).

The moderate reliability for AB magnitude limits the correlation that can be obtained between AB magnitude and sEBR. The AB phenomenon might suffer from the “individual differences paradox” (Hedge, Powell, and Sumner 2018): precisely because it is robust at the group level (almost everyone has an AB), it might not exhibit sufficient between-subject variability to be reliable. On the other hand, the AB seems to have a larger range of individual differences than other tasks (Willems and Martens 2016), and even a moderate reliability should provide “enough room” to detect correlations between sEBR and AB magnitude in a plausible range.

4.2 No significant relationships between sEBR and baseline AB magnitude

ABmag_mean <- mean(df_ABmag$AB_magnitude[df_ABmag$block == "pre"])
ABmag_best <- min(df_ABmag$AB_magnitude[df_ABmag$block == "pre"])

Colzato et al. (2008) (n = 20) found a negative correlation between sEBR and AB magnitude. Here, this correlation was not significant in either session, with an effect size around 0 (session 2) or even slightly positive (session 1). The main difference between both studies seems to be the distribution of AB magnitude scores. The AB was shallow in Colzato et al. (2008), as average AB magnitude was less than 10%, with 5 out of 20 participants showing either a “negative” AB (higher T2|T1 accuracy at the shortest lag) or an AB magnitude around 0. Here, the blink was much more pronounced (40% on average)—even the best performing participant had an AB magnitude (11%) that exceeded the average in Colzato et al. (2008).

Our AB task also involves an attentional set switch, while the task used by Colzato et al. (2008) did not. In our task, T1 and T2 had different features (T1 was red; T2 was green), so participants had to update their attentional template and detect the second target on the basis of a different feature (color) than T1. In Colzato et al. (2008), T1 and T2 were both digits, and thus belonged to the same set. Kelly and Dux (2011) suggest that such a set switch introduces an additional bottle neck that is dissociable from the AB (Potter et al. 1998). However, follow-up studies have shown that there is still a sizable correlation between non-switch and switch-versions of AB tasks (Dale, Dux, and Arnell 2013; Dale and Arnell 2013). Although a set switch may introduce additional variance, it seems unlikely this would be sufficient to completely abolish a correlation between sEBR and AB magnitude. However, it could have reduced the size of the effect to a degree where it could no longer be reliably detected. The correlation between sEBR and AB magnitude in Colzato et al. (2008) (r = -.53) appears to exceed the average reliability of AB magnitude itself (Table 4.1), suggesting it might be an overestimate of the true correlation. Further evidence that the effect might be small comes from Unsworth, Robison, and Miller (2019), who conducted a large-scale study (n = 204) and only found a correlation of .17 between sEBR and attentional control (psychomotor vigilance task, antisaccade task), and no correlation with working memory (operation, symmetry, and reading span)—a core component in the AB (Dux and Marois 2009; Martens and Wyble 2010). Although our sample size is twice that of Colzato et al. (2008), an effect of this magnitude would not have been detectable. It should also be noted that we used almost the same task as Slagter and Georgopoulou (2013) (n = 39), who also did not find a significant correlation between sEBR and AB magnitude. So both our studies do not support a relationship between sEBR and the AB.

No study to date has investigated an Inverted-U-shaped relationship between sEBR and AB magnitude, even though the underlying theory strongly suggests it (Cools and D’Esposito 2011; Slagter et al. 2012). Here the data do conform to a weak Inverted-U-shaped relationship, but only in session 1, and neither the upward nor the downward slope were significantly different from 0. Because both would have to be significant (Simonsohn 2018), and each slope is estimated on a different subset of the sample, we likely did not have sufficient power to detect an Inverted-U-shaped relationship. To be able to detect an Inverted-U-shaped relationship, the participants in our sample should also cover the whole range of AB task performance and dopamine levels. If this assumption is not met, then a true Inverted-U-shaped relationship could also masquerade as a linear effect (if all participants are sub-optimal), or no effect (if all participants are in the optimal range).

Two recent studies have cast doubt on the validity of sEBR as proxy for striatal dopamine. Both used PET to measure dopamine non-invasively in humans, but found no relation between sEBR and striatal dopamine receptor availability (Dang et al. 2017) or dopamine synthesis capacity (Sescousse et al. 2018). So, even though we find no relation between sEBR and the AB, there could still be a true relationship between dopamine and the AB—sEBR might simply not have high validity to measure it accurately.

4.3 No significant relationship between sEBR and the effect of tDCS on AB magnitude

We suggested two factors that might underlie individual differences in the effects of tDCS on the AB (See Figure 1.1). First, baseline cortical excitability might partially determine whether tDCS leads to behavioral improvements or impairments (Krause, Márquez-Ruiz, and Cohen Kadosh 2013) (Figure 1.1, left panel). This seems to fit the results of London and Slagter (2015), who found that individuals that benefitted from anodal tDCS tended to worsen with cathodal tDCS (and vice versa). However, we did not replicate this negative correlation between the effects of anodal and cathodal tDCS (see Reteig et al. 2019b), so this finding may not be robust.

Second, baseline dopamine levels may also partially determine the behavioral effects of tDCS (Wiegand, Nieratschker, and Plewnia 2016) (Figure 1.1, right panel). Assuming that sEBR is a valid index of dopamine levels, we found no evidence for this result, though our Bayes Factor analyses also did not indicate strong evidence of absence. The model proposed by Wiegand, Nieratschker, and Plewnia (2016) is based on just two studies (Plewnia et al. 2013; Nieratschker et al. 2015) that show different effects of tDCS in different COMT-subtypes, who should vary in baseline dopamine levels. However, a more recent study found that offline effects of prefrontal tDCS on working memory did not interact with COMT genotype (Jongkees et al. 2018). Because tDCS may itself also boost dopamine release in the striatum (Tanaka et al. 2013; Fonteneau et al. 2018), the model is complicated even further. On a more basic level, it is also still disputed whether the COMT polymorphism can affect cognitive functioning (Barnett, Scoriels, and Munafò 2008), as most studies are likely severely underpowered (Border et al. 2019). The only study that looked at the relation between COMT and the AB also found no association (Colzato et al. 2011), although this was a very small study for genotyping standards. Thus, the overall evidence that the effect of tDCS on the AB depends on baseline dopamine levels is preliminary.

Finally, it is unclear how both axes of the model (Figure 1.1) relate to each other. Pairing tDCS with dopaminergic drugs produces complex and non-linear effects on cortical excitability (Monte-Silva et al. 2009; Fresnoza et al. 2014). Two studies that have combined tDCS and administration of tyrosine (a dopamine precursor) in cognitive tasks have indeed led to divergent results. Cathodal tDCS of the lDLPFC appears to counteract the beneficial effects of tyrosine, as their combination had no net behavioral effect (Jongkees et al. 2017; Dennison et al. 2018). In contrast, combining anodal tDCS with tyrosine led to impaired performance on an n-back task, where anodal tDCS alone would be expected to enhance performance (Jongkees et al. 2017). Future studies that manipulate both dopamine levels and cortical excitability are necessary to uncover the physiology that could lead to such interactions.

4.4 Conclusions

We found that spontaneous eye blink rate is a reliable measure, but that it does not relate to the attentional blink, or to changes in attentional blink size following tDCS. Because dopamine is central to healthy cognitive and brain function, it remains plausible that dopamine interacts with manipulations of brain function, such as tDCS, as well as cognitive tasks such as the AB. But there is no good prior for how large the influence of dopamine is. Considering the large individual variability in AB size (Willems and Martens 2016), and the many factors that play a role in tDCS outcome (Li, Uehara, and Hanakawa 2015; Krause and Cohen Kadosh 2014), dopamine might account for only a small portion of the total variability. In addition, sEBR only provides an indirect measure of dopamine function, and this has recently also been questioned (Dang et al. 2017; Sescousse et al. 2018). More large-scale studies that include more direct measurement (e.g., using ligand-PET) or manipulation (e.g., using pharmacology) of dopamine function will be needed to pinpoint the unique contribution of dopamine.