Add Transformations

The next step in defining a model is to transform the variables you’ve selected. This step is optional – for many models you may not need to make any modifications to the input variables.

To add a transformation, click Add Transformation and select an operation from the drop-down list.

The transformations currently supported by Neuroscout are a subset of the complete set of transformations detailed in the BIDS StatsModel specification (in active development). As Neuroscout matures, the number of supported transformations will grow.

Transformation	Description
Scale	Standardize the value of one or more variables. Can independently choose to demean and/or rescale.
Orthogonalize	Orthogonalizes one or more input columns with respect to one or more other columns.

Select input

For all transformations, you must select which features this transformation will operate on. Most operations will operate on each column independently, but specifying multiple columns will save you from having to specify the same operation for multiple predictors.

Transformation-specific options

Most transformations additionally have specific options which you can specify.

Scale

• Demean - If True, subtracts the mean from each input column (i.e., applies mean-centering).

• Rescale - If True, divides each column by its standard deviation.

• ReplaceNA - Whether/when to replace missing values with 0. If “Don’t replace”: no replacement is performed. If ‘before’: missing values are replaced with 0’s before scaling. If ‘after’: missing values are replaced with 0 after scaling.

Note

If using only DummyContrasts, it is not necessary to scale variables, as the statistical maps we interpret are scale invariant. Scaling be necessary if computing contrasts between variables (which we don’t typically reccomend with naturalistic features).

However, it may be useful to use the scale transformation if there are missing values (NAs) in the variable. Replacing missing values with 0s after scaling effectively sets the values of NAs to the mean of that variable (for that run).

Orthogonalize

• You must select the inputs to orthogonalize with respect to. The transformed variable will be uncorrelated to these variables.

Threshold

• Threshold - The value to binarize around (values above will be assigned 1, values below will be assigned 0)

• Binarize - If True, binarizes all non-zero values (i.e., every non-zero value will be set to 1).

• Above - Specifies which values to retain with respect to the cut-off. If True, all values above the threshold will be kept; if False, all values below the threshold will be kept.

• Signed - Specifies whether to treat the threshold as signed (default) or unsigned. For example, when passing above=True and threshold=3, if signed=True, all and only values above +3 would be retained. If signed=False, all absolute values > 3 would be retained (i.e.,values in the range -3 < X < 3 would be set to 0).

Editing and ordering transformations

It is important to note that transformations are applied sequentially, so the order of the transformation matters. To re-order transformations you can drag and drop them in the list.

You can also remove transformations you’ve created using the trash icon, and edit existing transformations with the blue edit icon.