Stimulus
In this study, we used a Random Dots Motion paradigm (RDM) in which participants were asked to decide the net motion direction of a set of dots moving left or right with coherence varying between trials (0, 3.2, 6.4, 12.8, 25.6 or 51.2%). The more coherent the dot motion, the higher the probability of giving a correct answer. Detailed information on RDM properties were as follows: dot density of 16.7 dots/deg2/s, dot size of 3 × 3 pixels (0.075°), velocity of 5°/s, and aperture size of 5°. The stimulus consisted of white dots moving against a black background, located in one of three locations: the center, above (top), or below (bottom) the fixation point, at the eccentricity of 5°. In each block of 120 trials, having fixed the aperture location, each set of dots was displayed on the screen for one frame (16.7ms) whose position got updated every three frames, i.e. the position of dots in the first frame was updated in the fourth frame and so on, as previously has been done 7,63.
Data analysis
Having the binary nature of the outcomes (either true or false), in order to assess the impact of different parameters of the stimulus in this experiment, we used multiple logistic regression models. Logit[P] stands for log\(\:\left(\frac{p}{1-p}\right)\) and βi implies fitted coefficients. The fitting method was maximum likelihood under a binomial error model (i.e., a GLM).
The probability of a correct choice for single pulse trials was estimated as:
\(\:\text{L}\text{o}\text{g}\text{i}\text{t}\:\left[{P}_{correct}\right]={{\beta\:}}_{0}+{{\beta\:}}_{1}\text{C}\) (Eq. 1)
where C was the motion coherence.
A modified version of Eq. 1 was applied to check for right/left bias:
\(\:\text{L}\text{o}\text{g}\text{i}\text{t}\:\left[{\text{P}}_{\text{r}\text{i}\text{g}\text{h}\text{t}}\right]={{\beta\:}}_{0}+{{\beta\:}}_{1}{\text{C}}_{\pm\:}\) (Eq. 2)
Where C+ corresponds to rightward motion and C− to leftward motion.
To find the impact of time gap between the two pulses in double-pulse trials we used:
\(\:\text{L}\text{o}\text{g}\text{i}\text{t}\:\left[{\text{P}}_{\text{c}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t}}\right]={{\beta\:}}_{0}+{{\beta\:}}_{1}{\text{C}}_{1}+{{\beta\:}}_{2}{\text{C}}_{2}+{{{\beta\:}}_{3}\text{T}+{\beta\:}}_{4}{\text{C}}_{1}\text{T}+{{\beta\:}}_{5}{\text{C}}_{2}\text{T}\) (Eq. 3)
where T indicates the gap duration, and C1 and C2 indicate coherence in the first and second pulse, respectively.
To see whether location of the dots patch had any contribution to the probability of giving a correct answer, we defined another equation, where L indicates the location of the dots patch including three categories of ‘center’ (L = 0), ‘same’ (L = 1) and ‘different’ (L = -1):
\(\:\text{L}\text{o}\text{g}\text{i}\text{t}\:\left[{\text{P}}_{\text{c}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t}}\right]={{\beta\:}}_{0}+{{\beta\:}}_{1}{\text{C}}_{1}+{{\beta\:}}_{2}{\text{C}}_{2}+{{{\beta\:}}_{3}\text{T}+{\beta\:}}_{4}{\text{C}}_{1}\text{T}+{{\beta\:}}_{5}{\text{C}}_{2}\text{T}+{{\beta\:}}_{6}\text{L}+{{\beta\:}}_{7}\text{L}\text{T}\) (Eq. 4)
In the next step, we fitted a logistic regression model to compare the expected accuracy of a perfect integrator with the observed accuracy:
\(\:\text{L}\text{o}\text{g}\text{i}\text{t}\:\left[{\text{P}}_{\text{c}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t}}\right]=\text{L}\text{o}\text{g}\text{i}\text{t}\left[{\text{P}}_{\text{e}}\right]+{\beta\:}\) (Eq. 5)
where Pe is the expected probability derived from a perfect integrator. If β ends up with a positive value, then observed accuracy is higher than expected. Pe was calculated as:
\(\:{\text{P}}_{\text{e}}=1-\phi\:(0,{e}_{1}+{e}_{2},\sqrt{2})\) (Eq. 6)
where \(\:\phi\:\) is the normal cumulative distribution function, calculated as:
\(\:\phi\:\left(s,\mu\:,\sigma\:\right)=\:{\int\:}_{-\infty\:}^{s}N\left(v,\:\mu\:,\sigma\:\right)dv\) (Eq. 7)
where \(\:N\left(v,\:\mu\:,\sigma\:\right)\) is the normal probability density function with mean of (µ) and standard deviation of (σ).
Also, e1 and e2 are pieces of evidence from the two pulses. The distribution of the evidence is calculated from the probability of correct answers from single-pulse trials:
\(\:{e}_{i}={\phi\:}^{-1}\left({P}_{i},0,\:1\right)\:,\:i=1,\:2\) (Eq. 8)
where Pi is the probability of correct response of single-pulse trials based on Eq. 1, and \(\:{\phi\:}^{-1}\) is the inverse of the normal cumulative distribution function from Eq. 7.
To see the effect of pulse sequences on the probability of correct responses we used:
\(\:\text{L}\text{o}\text{g}\text{i}\text{t}\:\left[{P}_{correct}\right]={{\beta\:}}_{0}+{\:{\beta\:}}_{1}\left[{C}_{1}+{C}_{2}\right]+{\:{\beta\:}}_{2}\left[{C}_{2}-{C}_{1}\right]\) (Eq. 9)
where C1 and C2 are motion strength of the first and second pulse, respectively.
Furthermore, to see the interaction between two pulses (whether the stronger pulse 1 increases or decreases the effect of the second pulse), we used:
\(\:\text{L}\text{o}\text{g}\text{i}\text{t}\:\left[{\text{P}}_{\text{c}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t}}\right]={{\beta\:}}_{0}+{\:{\beta\:}}_{1}{\text{C}}_{1}+{\:{\beta\:}}_{2}{\text{C}}_{2}+{\:{\beta\:}}_{3}{\text{C}}_{1}{\text{C}}_{2}\) (Eq. 10)
Motion energy analysis
Since there are fluctuations in the RDM stimulus and direction of dots varies in each generated frame, no two pulses carry the same amount of information regarding a direction of motion. To overcome this issue, we calculated motion energy for each pulse of every trial. To analyze the amount of motion energy in favor of the intended direction on each trial, we used two pairs of quadrature spatiotemporal filters 41,65,66. Each pair was selective to either right or left. The filters were convolved with the three-dimensional pattern of RDM stimulus and then squared and summed. The result was summated across space to give the motion energy estimate as a function of time. Finally, the net motion energy in each direction was calculated as the difference between motion energies of opponent directions.
To test for the larger effect of the second pulse on the performance, we fit the following linear regression model:
\(\:\text{M}={{\beta\:}}_{0}+{\:{\beta\:}}_{1}\text{C}+{\:{\beta\:}}_{2}\text{E}+{\:{\beta\:}}_{3}\text{E}\text{S}\) (Eq. 11)
$$\:\text{E}=\left\{\begin{array}{c}0,\:\:correct\:response\\\:1,\:\:\:\:\:\:error\:response\end{array}\right.\:,\:\text{S}=\left\{\begin{array}{c}0,\:\:\:\:\:\:\:\:\:\:first\:pulse\\\:1,\:\:\:\:\:\:second\:pulse\end{array}\right.$$
Where C is the coherence of the pulse, E indicates if the response of subject was correct or wrong, S stands for the pulse sequence and M is the motion energy of that pulse. M is calculated as the summation of motion energy across a window of 200ms starting 50ms after the onset of stimulus. Since the motion energy for zero gap intervals overlapped with each other, they were omitted.
In order to compare the effect of motion energy of each of the pulses on the performance, a logistic regression model was used:
\(\:\text{L}\text{o}\text{g}\text{i}\text{t}\:\left[{\text{P}}_{\text{c}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t}}\right]={{\beta\:}}_{0}+{\:{\beta\:}}_{1}{\text{C}}_{1}+{\:{\beta\:}}_{2}{\text{C}}_{2}+{\:{\beta\:}}_{3}({M}_{1}+{M}_{2})+{\:{\beta\:}}_{4}{\text{M}}_{2}\) (Eq. 12),
in which \(\:{M}_{1}\)and \(\:{M}_{2}\) are the motion energy of the first and second pulses, respectively. For trials with equal coherence, the term \(\:{\:{\beta\:}}_{2}{\text{C}}_{2}\) was dropped. The null hypothesis is that both pulses have the same effect (\(\:{{\beta\:}}_{4}=0\)).