The {logitr} package requires that data be structured in a
data.frame
and arranged in a “long” format [@Wickham2014] where each row contains data on a
single alternative from a choice observation. The choice observations do
not have to be symmetric, meaning they can have a “ragged” structure
where different choice observations have different numbers of
alternatives. The data must also include variables for each of the
following:
1
is chosen, 0
is not chosen). Only one alternative should have a 1
per
choice observation.obsID
variable would be
1, 1, 2, 2, 3, 3
.The {logitr} package contains several example data sets that
illustrate this data structure. For example, the yogurt
contains observations of yogurt purchases by a panel of 100 households
[@Jain1994]. Choice is identified by the
choice
column, the observation ID is identified by the
obsID
column, and the columns price
,
feat
, and brand
can be used as model
covariates:
library("logitr")
head(yogurt)
#> # A tibble: 6 × 7
#> id obsID alt choice price feat brand
#> <dbl> <int> <int> <dbl> <dbl> <dbl> <chr>
#> 1 1 1 1 0 8.1 0 dannon
#> 2 1 1 2 0 6.10 0 hiland
#> 3 1 1 3 1 7.90 0 weight
#> 4 1 1 4 0 10.8 0 yoplait
#> 5 1 2 1 1 9.80 0 dannon
#> 6 1 2 2 0 6.40 0 hiland
This data set also includes an alt
variable that
determines the alternatives included in the choice set of each
observation and an id
variable that determines the
individual as the data have a panel structure containing multiple choice
observations from each individual.
Variables are modeled as either continuous or discrete based on their
data type. Numeric variables are by default estimated with a
single “slope” coefficient. For example, consider a data frame that
contains a price
variable with the levels $10, $15, and
$20. Adding price
to the pars
argument in the
main logitr()
function would result in a single
price
coefficient for the “slope” of the change in
price.
In contrast, categorical variables (i.e. character
or
factor
type variables) are by default estimated with a
coefficient for all but the first level, which serves as the reference
level. The default reference level is determined alphabetically, but it
can also be set by modifying the factor levels for that variable. For
example, the default reference level for the brand
variable
is "dannon"
as it is alphabetically first. To set
"weight"
as the reference level, the factor levels can be
modified using the factor()
function:
<- yogurt
yogurt2
<- c("weight", "hiland", "yoplait", "dannon")
brands $brand <- factor(yogurt2$brand, levels = brands) yogurt2
If you wish to make dummy-coded variables yourself to use them in a
model, I recommend using the dummy_cols()
function from the
{fastDummies}
package. For example, in the code below, I create dummy-coded columns
for the brand
variable and then use those variables as
covariates in a model:
<- fastDummies::dummy_cols(yogurt2, "brand") yogurt2
The yogurt2
data frame now has new dummy-coded columns
for brand:
head(yogurt2)
#> # A tibble: 6 × 11
#> id obsID alt choice price feat brand brand_weight brand_hiland
#> <dbl> <int> <int> <dbl> <dbl> <dbl> <fct> <int> <int>
#> 1 1 1 1 0 8.1 0 dannon 0 0
#> 2 1 1 2 0 6.10 0 hiland 0 1
#> 3 1 1 3 1 7.90 0 weight 1 0
#> 4 1 1 4 0 10.8 0 yoplait 0 0
#> 5 1 2 1 1 9.80 0 dannon 0 0
#> 6 1 2 2 0 6.40 0 hiland 0 1
#> # … with 2 more variables: brand_yoplait <int>, brand_dannon <int>
Now I can use those columns as covariates:
<- logitr(
mnl_pref_dummies data = yogurt2,
outcome = 'choice',
obsID = 'obsID',
pars = c(
'price', 'feat', 'brand_yoplait', 'brand_dannon', 'brand_weight')
)
summary(mnl_pref_dummies)
#> =================================================
#>
#> Model estimated on: Thu Jun 16 17:36:40 2022
#>
#> Using logitr version: 0.7.0
#>
#> Call:
#> logitr(data = yogurt2, outcome = "choice", obsID = "obsID", pars = c("price",
#> "feat", "brand_yoplait", "brand_dannon", "brand_weight"))
#>
#> Frequencies of alternatives:
#> 1 2 3 4
#> 0.402156 0.029436 0.229270 0.339138
#>
#> Exit Status: 3, Optimization stopped because ftol_rel or ftol_abs was reached.
#>
#> Model Type: Multinomial Logit
#> Model Space: Preference
#> Model Run: 1 of 1
#> Iterations: 18
#> Elapsed Time: 0h:0m:0.03s
#> Algorithm: NLOPT_LD_LBFGS
#> Weights Used?: FALSE
#> Robust? FALSE
#>
#> Model Coefficients:
#> Estimate Std. Error z-value Pr(>|z|)
#> price -0.366581 0.024366 -15.045 < 2.2e-16 ***
#> feat 0.491412 0.120063 4.093 4.259e-05 ***
#> brand_yoplait 4.450197 0.187118 23.783 < 2.2e-16 ***
#> brand_dannon 3.715575 0.145419 25.551 < 2.2e-16 ***
#> brand_weight 3.074399 0.145384 21.147 < 2.2e-16 ***
#> ---
#> Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#>
#> Log-Likelihood: -2656.8878788
#> Null Log-Likelihood: -3343.7419990
#> AIC: 5323.7757575
#> BIC: 5352.7168000
#> McFadden R2: 0.2054148
#> Adj McFadden R2: 0.2039195
#> Number of Observations: 2412.0000000