add custom models to be used in the transition functions

[asiripanich]

• ModelCustom: provides the base class for building custom models.

• ModelLinear: for creating a linear combination model.

• ModelBinaryChoice: for creating a logistic regression model.

• ModelMultinomialLogit: for creating a multinomial logistic regression model. This requires on the mlogit package for formula and structuring of prediction data.

• choice simulation (eventually will replace the MatchingMarket classes) (move to market simulation #85 )

• To construct choiceset for mlogit see https://stackoverflow.com/questions/51458223/multinomial-logit-estimation-on-a-subset-of-alternatives-in-r

• To manually calculate the probs see https://monashdatafluency.github.io/r-linear/topics/linear_models.html

[asiripanich]

The multinomial predict function of ModelCustom works as intended when there is no alternative specific constants or variables in the formula. However, if those were to be added then it would give a wrong result. This should be generalised to work with all mlogit formulations.

library(mlogit)
#> Loading required package: Formula
#> Loading required package: zoo
#> 
#> Attaching package: 'zoo'
#> The following objects are masked from 'package:base':
#> 
#>     as.Date, as.Date.numeric
#> Loading required package: lmtest
library(dymiumCore) # branch: mlogit
#> ── * dymium's options * ───────────────────────────────────────────────────────────────────────────────────────────────────
#> ● dymium.input_dir: /var/folders/0d/9srpj_750lxbkfs2_8nwkcpw0000gn/T//Rtmp3LJUfZ/scenario/inputs
#> ● dymium.output_dir: /var/folders/0d/9srpj_750lxbkfs2_8nwkcpw0000gn/T//Rtmp3LJUfZ/scenario/outputs
#> ● dymium.scenario_dir: /var/folders/0d/9srpj_750lxbkfs2_8nwkcpw0000gn/T//Rtmp3LJUfZ/scenario
#> ● dymium.simulation_clock: 0
#> ● dymium.simulation_scale: 1
data("Fishing", package = "mlogit")
Fish <- mlogit.data(Fishing, varying = c(2:9), shape = "wide", choice = "mode")

## a pure "conditional" model
fm <- mFormula(mode ~ price + catch | 0)
mm <- model.matrix(fm, Fish)

mod <- mlogit(mode ~ price + catch | 0, data = Fish)
md <- ModelCustom$new(params = mod$coefficients, formula = fm, type = "multinomial")

mlogit_prediction <- predict(mod, Fish)
my_prediction <- md$predict(Fish, chooser_id_col = "chid", choice_id_col = "alt")

head(my_prediction)
#>    chooser_id choice_id linear_comb      prob
#> 1:          1     beach -3.16923742 0.2424937
#> 2:          1      boat -2.98595663 0.2912717
#> 3:          1   charter -3.23195532 0.2277521
#> 4:          1      pier -3.18591664 0.2384826
#> 5:          2     beach -0.20950218 0.2470243
#> 6:          2      boat -0.06568204 0.2852332
head(mlogit_prediction)
#>        beach      boat   charter       pier
#> 1 0.24249366 0.2912717 0.2277521 0.23848260
#> 2 0.24702432 0.2852332 0.2344037 0.23333881
#> 3 0.03617426 0.4577779 0.4725643 0.03348346
#> 4 0.34947274 0.1661746 0.1311631 0.35318957
#> 5 0.11609212 0.4605316 0.3092044 0.11417185
#> 6 0.02923272 0.5866468 0.3570623 0.02705826

^{Created on 2020-05-28 by the reprex package (v0.3.0)}

[asiripanich]

I retract my earlier comment. Apparently the model.matrix object that gets created from mFormula (from mlogit) does have dummy columns for alternative specific constants. This works in with alternative specific variables as well.

library(mlogit)
#> Loading required package: Formula
#> Loading required package: zoo
#> 
#> Attaching package: 'zoo'
#> The following objects are masked from 'package:base':
#> 
#>     as.Date, as.Date.numeric
#> Loading required package: lmtest
library(dymiumCore) # branch: mlogit
#> ── * dymium's options * ───────────────────────────────────────────────────────────────────────────────────────────────────
#> ● dymium.input_dir: /var/folders/0d/9srpj_750lxbkfs2_8nwkcpw0000gn/T//Rtmp3LJUfZ/scenario/inputs
#> ● dymium.output_dir: /var/folders/0d/9srpj_750lxbkfs2_8nwkcpw0000gn/T//Rtmp3LJUfZ/scenario/outputs
#> ● dymium.scenario_dir: /var/folders/0d/9srpj_750lxbkfs2_8nwkcpw0000gn/T//Rtmp3LJUfZ/scenario
#> ● dymium.simulation_clock: 0
#> ● dymium.simulation_scale: 1
data("Fishing", package = "mlogit")
Fish <- mlogit.data(Fishing, varying = c(2:9), shape = "wide", choice = "mode")

## a pure "conditional" model
fm <- mFormula(mode ~ price + catch)
mm <- model.matrix(fm, Fish)

mod <- mlogit(fm, data = Fish)
md <- ModelCustom$new(params = mod$coefficients, formula = fm, type = "multinomial")

mlogit_prediction <- predict(mod, Fish)
my_prediction <- md$predict(Fish, chooser_id_col = "chid", choice_id_col = "alt")

head(my_prediction, 4)
#>    chooser_id choice_id linear_comb      prob
#> 1:          1     beach   -3.889442 0.1282391
#> 2:          1      boat   -2.945537 0.3295740
#> 3:          1   charter   -2.832532 0.3690034
#> 4:          1      pier   -3.588987 0.1731834
head(mlogit_prediction, 1)
#>       beach     boat   charter      pier
#> 1 0.1282391 0.329574 0.3690034 0.1731834

^{Created on 2020-05-28 by the reprex package (v0.3.0)}