API

effects!(reference_grid::DataFrame, model::RegressionModel;
         eff_col=nothing, err_col=:err, typical=mean)

Compute the effects as specified in formula.

Effects are the model predictions made using values given via the reference grid. For terms present in the model, but not in the reference grid, then the typical value of those predictors is used. (In other words, effects are conditional on the typical value.) The function for computing typical values is specified via typical. Note that this is also applied to categorical contrasts, thus yielding an average of the contrast, weighted by the balance of levels in the data set used to fit the model.

By default, the column corresponding to the response variable in the formula is overwritten with the effects, but an alternative column for the effects can be specified by eff_col. If this column does not exist, it is created. Pointwise standard errors are written into the column specified by err_col.

The reference grid must contain columns for all predictors in the formula. (Interactions are computed automatically.) Contrasts must match the contrasts used to fit the model; using other contrasts will lead to undefined behavior.

Interaction terms are computed in the same way for any regression model: as the product of the lower-order terms. Typical values of lower terms thus propagate up into the interaction term in the same way that any value would.

The use of typical values for excluded effects differs from other approaches such as "partial effects" used in R packages like remef. The R package effects is similar in approach, but due to differing languages and licenses, no source code was inspected and there is no attempt at API compatibility or even similarity.

The approach for computing effect is based on the effects plots described here:

Fox, John (2003). Effect Displays in R for Generalised Linear Models. Journal of Statistical Software. Vol. 8, No. 15

effects(design::AbstractDict, model::RegressionModel;
        eff_col=nothing, err_col=:err, typical=mean,
        lower_col=:lower, upper_col=:upper)

Compute the effects as specified by the design.

This is a convenience wrapper for effects!. Instead of specifying a reference grid, a dictionary containing the levels/values of each predictor is specified. This is then expanded into a reference grid representing a fully-crossed design. Additionally, two extra columns are created representing the lower and upper edge of the error band (i.e. [resp-err, resp+err]).