v3.0.0

Import {#import .unnumbered .unnumbered}

New data types were implemented

• Noldus Observer compatible
• Eyelink

Untested data types

CNT data import has not been not tested -- please contact the developers
with sample data files.

Filtering for SCR models

Previous versions of PsPM have used a bi-directional high pass filter of
0.0159 for all SCR analyses. We have recently shown a better predictive
validity for GLM with a unidirectional filter of 0.05 Hz [@Bach:2013aa].
This also implies that different filters are used for different models.
These are now set as defaults, and the way the default settings are
implemented has changed. It is now possible to alter the filter settings
in the model definition, although we discourage this.

New SF method

A matching pursuit algorithm is now implemented to approximate the
number of spontaneous fluctuations (SF) in skin conductance
[@Bach:2015aa].

General linear modelling (GLM)

Parametric modulators

Parametric modulators (pmods) are z-normalised before being entered into
the design matrix. This is to account for possibly very large or very
small numbers -- a badly scaled design matrix can cause induced
instability in the inversion. The parameter estimates of the pmods were
not transformed back in previous versions, i. e. the parameter estimates
of the pmods were independent of the scaling of the pmods. This is
appropriate as long as they are the same for all datasets, or if
analysis is done strictly on a within-subject level. Some researchers
have reported designs in which inference was to be drawn on parameter
estimates of pmods on a between-group level, and where the pmods
systematically differed between these groups. To account for this
possibility, the parameter estimates are now transformed back, such that
they refer to the pmods in their original scaling/units.

Parametric confounds

Previous versions of PsPM contained an option to include a parametric
modulator across all event types, to account for confounds across all
conditions. For example, in an experiment with 5 conditions, one could
have included 5 regressors, plus one reaction time confound across all
events, without including an associated regressor that contains the
event onsets for all these events. This option was removed.

Design matrix filtering

Previous versions of PsPM filtered the design matrix after
orthogonalisation of basis sets. This can introduce unwanted
dependencies between regressors. PsPM 3.0 filters the regressors first,
then orthogonalises the basis sets.

File format

Some minor changes have been made to the data format. In particular,
marker channels from previous versions can not be read with the current
version - such data files have to be re-imported. Model files have
changed drastically, and model files from previous versions can not be
read with the current version of the software.

VB inversion

The VBA toolbox (http://mbb-team.github.io/VBA-toolbox/) was updated
in October 2014 [@Daunizeau:2014aa]. This update incorporates bugfixes
in this toolbox and slightly changed the model estimates in our test
models. In terms of predictive validity, we noted that there was no
consistent benefit of the old or new version of this code (Figure
[fig:VBA]{reference-type="ref" reference="fig:VBA"}).

<insert_image_here>
Model comparison between old and new versions of the VBA toolbox,
based on two delay fear conditioning datasets. The log Bayes factor
quantifies the difference between negative log likelihood (nLL) of
parameter estimates obtained from model inversion using the old and new
version of VBA. A difference in nLL above 3 indicates significant
differences in model evidence which is not exceeded for either data set.
Analysis and figure contributed by Matthias Staib.