iiinova — where your digital self keeps evolving

Sebastian Hellmann

@sehellmann.bsky.social

150 followers 222 following 16 posts

PostDoc working at TU Munich. Interested in on computational modelling, decision-making, and confidence. Cat owner, Ireland lover and brass music fan

Posts Replies Media

Sebastian Hellmann @sehellmann.bsky.social · 25/08/2026

Pre-stimulus Beta activity mediated the effect of previous choices on drift rate only in a stable environment. In the volatile environment, previous choice was not credibly mediated by prior activation. Here, we see a cross-over of effects from cue, through Beta power, to drift rate.

Path diagrams showing how prior stimulus probability and previous choice predicted drift-diffusion model parameters in two studies. Left side shows the stable environment; the right side shows the volatile environment. In both, cued stimulus probability primarily affected the starting point (z), while previous choice primarily affected the drift rate (v). Only in the stable environment, the paths going from previous choice over Beta-power to the drift rate are boldt, i.e. credible.

010

Sebastian Hellmann @sehellmann.bsky.social · 25/08/2026

We used a joint model, integrating pre-stimulus Beta power over sensory areas into a hierarchical drift-diffusion model to see how the biases play out in the decision process and what role activation prior to stimulus onset plays in implementing these biases.

Graphical representation of the integrative, hierarchical drift-diffusion model used in the study and the effect of biases on drift rate and starting point on response times and response probabilities.

110

Sebastian Hellmann @sehellmann.bsky.social · 25/08/2026

We investigated two sources of bias in near-threshold perceptual decisions: stimulus probabilities (communicated with explicit probability cues) and previous choices in two different environments: stable vs. volatile, depending on whether stimulus probabilities change on a trial-level.

Schematic visualization of the experimental task. Participants first saw information about the probability of a stimulus, followed by an intertrial interval and a stimulus-onset cue. A near-threshold stimulus was then presented briefly with a known probability of either 25% or 75%. Participants indicated whether they detected it and reported their confidence. Performance feedback was provided after each trial.
The upper panel illustrates the stable environment, in which the probability cue was only shown once per block. The lower panel illustrates the volatile enviornment, in which cues were shown on a trial level.

101

Sebastian Hellmann @sehellmann.bsky.social · 10/09/2025

(e.g. the standard normal CDF) and normal distributions to fit the group-level distribution. But we cannot simply apply the same transformation to the mean of the real-valued normal distribution to derive the group-level mean on the parameter scale! This ignores individual variability.

A Figure with two panels, demonstrating the transformation of a normally distributed random variable resulting from transforming the values using the standard normal CDF to the interval 0 to 1. The figure shows the discrepancy between the mean of the transformed values and the result from simply transforming the mean of the normal distribution using the standard normal CDF.
The left side shows a normal distribution with low variance, such that the transformed values also show small variance and a symmetrical shape. The mean of the transformed values and the value resulting from transforming the mean directly coincide.
The right side shows a distribution with high variance. Due to the non-linearity of the normal CDF, the distribution of transformed values is non-symmetrical and highly skewed. The mean of the distribution is therefore pulled to the center of the range 0 to 1, compared to the result from directly transforming the mean of the normal distribution.

100

Sebastian Hellmann @sehellmann.bsky.social · 10/09/2025

If you’re estimating group-level means of constraint parameters, which are fitted with nonlinear transformations, beware that a common approach can produce biased estimates—especially with high individual variability. For constraint parameters, we often use nonlinear transformations...

Line graphs comparing correct and biased computations for exponential and probit (Φ) transformations in statistical modeling. Colored lines show how the bias changes with different group means and standard deviations.

210