The package permits the covariate effects of trinomial regression models to be represented graphically by means of a ternary plot. The aim of the plots is helping the interpretation of regression coefficients in terms of the effects that a change in regressors’ values has on the probability distribution of the dependent variable. Such changes may involve either a single regressor, or a group of them (composite changes), and the package permits both cases to be handled in a user-friendly way. Methodological details are illustrated and discussed in Santi, Dickson, and Espa (2019).

The package can read the results of **both categorical and ordinal trinomial logit** regression fitted by various functions (see the next section) and creates a `field3logit`

object which may be represented by means of functions `gg3logit`

and `stat_field3logit`

.

The `plot3logit`

package inherits graphical classes and methods from the package `ggtern`

(Hamilton and Ferry 2018) which, in turn, is based on the package `ggplot2`

(Wickham 2016).

Graphical representation based on **standard graphics** is made available through the package `Ternary`

(Smith 2017) by functions `plot3logit`

and `TernaryField`

, and by the `plot`

method of `field3logit`

objects.

See the help of `field3logit`

for representing composite effects and `multifield3logit`

for drawing multiple fields.

Function `field3logit`

of package `plot3logit`

can read trinomial regression estimates from the output of the following functions:

`clm`

and`clm2`

of package`ordinal`

(ordinal logit regression);`mlogit`

of package`mlogit`

(logit regression);`multinom`

of package`nnet`

(logit regression);`polr`

of package`MASS`

(ordinal logit regression);`vgam`

and`vglm`

of package`VGAM`

(logit regression).

Moreover, explicit estimates can be passed to `field3logit()`

. See the help of the package (type `? 'plot3logit-package'`

) and the help of functions `field3logit()`

and `extract3logit()`

for further details.

Fit a trilogit model by means of package `nnet`

where the student’s employment situation is analysed with respect to all variables in the dataset `cross_1year`

:

```
library(plot3logit)
data(cross_1year)
library(nnet)
mod0 <- multinom(employment_sit ~ ., data = cross_1year)
#> # weights: 42 (26 variable)
#> initial value 3605.645531
#> iter 10 value 2167.042903
#> iter 20 value 2136.782685
#> iter 30 value 2134.363158
#> final value 2134.352162
#> converged
```

The gender effect is analysed by means of a ternary plot which is generated in two steps.

Firstly, the vector field is computed:

Secondly, the field is represented on a ternary plot, using either `gg`

-graphics:

or standard graphics:

Ternary plots represent the effect of a change in covariate values on the probability distribution of the dependent variable. The function `field3logit`

permits such change to be specified in two different ways, which are illustrated below.

As an example, the following subsections refer to this dataset:

```
data(cross_1year)
str(cross_1year)
#> 'data.frame': 3282 obs. of 7 variables:
#> $ employment_sit: Factor w/ 3 levels "Employed","Unemployed",..: 1 1 1 3 1 1 1 1 1 2 ...
#> $ gender : Factor w/ 2 levels "Male","Female": 2 1 2 2 2 1 2 1 1 2 ...
#> $ finalgrade : Factor w/ 3 levels "Average","Low",..: 3 3 1 1 2 3 1 1 2 3 ...
#> $ duration : Factor w/ 3 levels "Average","Short",..: 3 3 3 3 3 3 1 3 3 3 ...
#> $ social_class : Factor w/ 5 levels "Working class",..: 1 2 4 4 3 3 1 4 1 2 ...
#> $ irregularity : Factor w/ 3 levels "Average","Low",..: 1 1 3 3 3 3 1 3 3 3 ...
#> $ hsscore : num 100 95 82 64 69 ...
head(cross_1year)
#> employment_sit gender finalgrade duration social_class irregularity
#> 1 Employed Female High Long Working class Average
#> 2 Employed Male High Long White-collar workers Average
#> 3 Employed Female Average Long Upper middle class High
#> 4 Trainee Female Average Long Upper middle class High
#> 5 Employed Female Low Long Lower middle class High
#> 6 Employed Male High Long Lower middle class High
#> hsscore
#> 1 100.00000
#> 2 95.00000
#> 3 82.00000
#> 4 64.00000
#> 5 69.00000
#> 6 86.66667
```

and this trinomial logistic regression model:

```
mod0 <- nnet::multinom(employment_sit ~ finalgrade + irregularity + hsscore, cross_1year)
#> # weights: 21 (12 variable)
#> initial value 3605.645531
#> iter 10 value 2187.709284
#> iter 20 value 2157.087955
#> final value 2157.087854
#> converged
mod0
#> Call:
#> nnet::multinom(formula = employment_sit ~ finalgrade + irregularity +
#> hsscore, data = cross_1year)
#>
#> Coefficients:
#> (Intercept) finalgradeLow finalgradeHigh irregularityLow
#> Unemployed -0.4481761 0.05551765 -0.07810893 -0.01874354
#> Trainee -1.3751140 0.14456683 -0.26849829 0.05764144
#> irregularityHigh hsscore
#> Unemployed 0.15691595 -0.016619227
#> Trainee -0.03477569 -0.009964381
#>
#> Residual Deviance: 4314.176
#> AIC: 4338.176
```

This method for specifying the change in the covariate values requires the vector \(\Delta x\) to be explicitly defined, thus it may be suitable when \(\Delta x\) results from some calculations. On the other hand, it is less user-friendly than implicit syntax, as it depends on the order of regressors in the design matrix.

If the effect of a *high final grade* has to be assessed, the vector of changes \(\Delta x\) can be set according to the position of the dummy variable `finalgradeHigh`

in the matrix of coefficients of the model `mod0`

:

```
coef(mod0)
#> (Intercept) finalgradeLow finalgradeHigh irregularityLow
#> Unemployed -0.4481761 0.05551765 -0.07810893 -0.01874354
#> Trainee -1.3751140 0.14456683 -0.26849829 0.05764144
#> irregularityHigh hsscore
#> Unemployed 0.15691595 -0.016619227
#> Trainee -0.03477569 -0.009964381
```

in this case, we have that \[
\Delta x=[0, 0, 1, 0, 0, 0]'
\] since `finalgradeHigh`

is the fourth coefficient (including the intercept) of the matrix of coefficients.

It follows that the function `field3logit`

can be invoked as it follows:

```
field0 <- field3logit(mod0, c(0, 0, 1, 0, 0, 0))
field0
#> Object of class "field3logit"
#> -------------------------------
#> Label : <empty>
#> Possible outcomes : Employed; Unemployed; Trainee
#> Reference level : Employed
#> Type of model : categorical
#> Effect : 0 0 1 0 0 0
#> Model has been read from : nnet::multinom
#> Number of stream lines : 8
#> Number of arrows : 182
#> Covariance matrix : available
#> Confidence regions : not available
```

It is also possible to set \(\Delta x\) so as to consider changes involving more than one regressor, as well as fractional changes. In such cases, \(\Delta x\) will consist in a vector where there are several non-zero elements which may take any positive or negative value.

Assume, for example, that we want to study the effect of a decrease by 10 in the high school final score, associated to an high final grade. In such a case, we have that: \[ \Delta x =[0, 0, 1, 0, 0, -10]'\,, \] thus:

```
field0 <- field3logit(mod0, c(0, 0, 1, 0, 0, -10))
field0
#> Object of class "field3logit"
#> -------------------------------
#> Label : <empty>
#> Possible outcomes : Employed; Unemployed; Trainee
#> Reference level : Employed
#> Type of model : categorical
#> Effect : 0 0 1 0 0 -10
#> Model has been read from : nnet::multinom
#> Number of stream lines : 8
#> Number of arrows : 166
#> Covariance matrix : available
#> Confidence regions : not available
```

Unlike the explicit method, this syntax allows the user to initialise the vector \(\Delta x\) by writing the name of the covariate which should vary: the function `field3logit`

will build \(\Delta x\) up associated to a unitary change in the specified covariate.

If more complex changes in covariate values have to be considered, implicit syntax allows the user to express them in terms of `R`

expressions involving the covariates.

If the effect of a *high final grade* has to be assessed, the implicit syntax which allow to assess the effect of a unitary change of `finalgradeHigh`

is the following:

```
field0 <- field3logit(mod0, 'finalgradeHigh')
field0
#> Object of class "field3logit"
#> -------------------------------
#> Label : <empty>
#> Possible outcomes : Employed; Unemployed; Trainee
#> Reference level : Employed
#> Type of model : categorical
#> Effect : finalgradeHigh
#> Explicit effect : 0 0 1 0 0 0
#> Model has been read from : nnet::multinom
#> Number of stream lines : 8
#> Number of arrows : 182
#> Covariance matrix : available
#> Confidence regions : not available
```

Note that the console output produced by printing `field0`

shows both the implicit effect (line `Effect`

) and the associated vector \(\Delta x\) (line `Explicit effect`

).

If we want to study the effect of a decrease by 10 in the high school final score, associated to an high final grade, the implicit syntax is:

```
field0 <- field3logit(mod0, 'finalgradeHigh - 10 * hsscore')
field0
#> Object of class "field3logit"
#> -------------------------------
#> Label : <empty>
#> Possible outcomes : Employed; Unemployed; Trainee
#> Reference level : Employed
#> Type of model : categorical
#> Effect : finalgradeHigh - 10 * hsscore
#> Explicit effect : 0 0 1 0 0 -10
#> Model has been read from : nnet::multinom
#> Number of stream lines : 8
#> Number of arrows : 166
#> Covariance matrix : available
#> Confidence regions : not available
```

Compare the line `Explicit effect`

of this output to the line `Effect`

of the same example in the previous section: as expected, they are the same.

When effects of multiple changes have to be compared at a time, multiple fields should be computed and represented on the same plot. This task can be easily done by creating a `multifield3logit`

object and directly representing it.

Since objects `multifield3logit`

result by putting together two or more `field3logit`

objects, the package `plot3logit`

allows the user to create a `multifield3logit`

object by adding up two or more `filed3logit`

or `multifield3logit`

objects using standard sum operator `+`

.

Here it is an example. The following command fit a trilogit model where all available variables are used as regressors. Then four `fields3logit`

objects are computed for assessing the effects of a some combined changes in the duration of studies and in students’ final degree score.

Note that each field is computed just with respect to a single probability distribution (`refpoint`

) of the dependent variable, and only one arrow is computed. The reason of this is that we have to represent four fields on the same plot, thus olny a small number of arrows can be drawn in order to preserve the clarity of the graph.

```
data(cross_1year)
mod0 <- nnet::multinom(employment_sit ~ ., data = cross_1year)
refpoint <- list(c(0.7, 0.15, 0.15))
field_Sdur <- field3logit(mod0, 'durationShort', label = 'Short duration', p0 = refpoint, narrows = 1)
field_Ldur <- field3logit(mod0, 'durationLong', label = 'Long duration', p0 = refpoint, narrows = 1)
field_Hfgr <- field3logit(mod0, 'finalgradeHigh', label = 'High final grade', p0 = refpoint, narrows = 1)
field_Lfgr <- field3logit(mod0, 'finalgradeLow', label = 'Low final grade', p0 = refpoint, narrows = 1)
```

Now the `multifield3logit`

object can be created by adding all the `field3logit`

objects up together:

```
mfields <- field_Sdur + field_Ldur + field_Lfgr + field_Hfgr
mfields
#> Object of class "multifield3logit"
#> ------------------------------------
#> Number of fields : 4
#> Labels
#> 1. Short duration (dX: durationShort)
#> 2. Long duration (dX: durationLong)
#> 3. Low final grade (dX: finalgradeLow)
#> 4. High final grade (dX: finalgradeHigh)
```

and the `multifield3logit`

object `mfield`

can be represented in a graph:

The code needed for generating the object `mfields`

may be conveniently made shorter in this way (see the help of `field3logit`

for details on syntax):

```
depo <- list(
list(delta = 'durationShort', label = 'Short duration'),
list(delta = 'durationLong', label = 'Long duration'),
list(delta = 'finalgradeHigh', label = 'High final grade'),
list(delta = 'finalgradeLow', label = 'Low final grade')
)
mfields <- field3logit(mod0, delta = depo, p0 = refpoint, narrows = 1)
mfields
#> Object of class "multifield3logit"
#> ------------------------------------
#> Number of fields : 4
#> Labels
#> 1. Short duration (dX: durationShort)
#> 2. Long duration (dX: durationLong)
#> 3. High final grade (dX: finalgradeHigh)
#> 4. Low final grade (dX: finalgradeLow)
```

It is also possible to rely on syntax based on delimiters `<<...>>`

(see the help of `field3logit`

, section **Details**) in order to create a `multifield3logit`

object consisting of fields based on the effect of each dummy variable generated by the same `factor`

regressor:

The package `plot3logit`

allows also to draw the confidence regions associated to each effect, both in case of `field3logit`

and `multifield3logit`

objects.

The confidence regions can be computed when the function `field3logit`

is called by setting the argument `conf`

. Otherwise, they can be added later through the function `add_confregions`

as it follows:

```
field0 <- add_confregions(field0, conf = 0.95)
field0
#> Object of class "field3logit"
#> -------------------------------
#> Label : <empty>
#> Possible outcomes : Employed; Unemployed; Trainee
#> Reference level : Employed
#> Type of model : categorical
#> Effect : finalgradeHigh - 10 * hsscore
#> Explicit effect : 0 0 1 0 0 -10
#> Model has been read from : nnet::multinom
#> Number of stream lines : 8
#> Number of arrows : 166
#> Covariance matrix : available
#> Confidence regions : 95%
```

and through the same syntax in case of `multifield3logit`

objects:

The statistic `stat_conf3logit`

permits confidence regions to be drawn, if available:

and

Hamilton, N. E., and M. Ferry. 2018. “ggtern: Ternary Diagrams Using ggplot2.” *Journal of Statistical Software, Code Snippets* 87 (3): 1–17. doi:10.18637/jss.v087.c03.

Santi, F., M. M. Dickson, and G. Espa. 2019. “A Graphical Tool for Interpreting Regression Coefficients of Trinomial Logit Models.” *The American Statistician* 73 (2): 200–207. doi:10.1080/00031305.2018.1442368.

Smith, M. R. 2017. “Ternary: An R Package for Creating Ternary Plots.” *Zenodo*.

Wickham, H. 2016. *ggplot2: Elegant Graphics for Data Analysis*. New York: Springer-Verlag.