initial_value should be dimension K by J

[kuriwaki]

not J - 1, I think?

Here is a reprex with the existing version that shows that J does not work.

library(emlogit)
data(japan, package = "MNP")

# example ---
Y <- japan %>% select(LDP, NFP, SKG, JCP) %>% data.matrix()
X <- japan %>% select(gender, age) %>% data.matrix()

set.seed(1234)
fit <- emlogit(Y = Y, X = X)
dim(fit$coef)
#> [1] 3 4

# try with initial values ----
# 2 covariates + 1 intercept (rows) and 4 choices (columns)
init_mat <- matrix(c(0, 0.5, -0.2, 0.1,
                     0, 0.2,  0.2, 0.2,
                     0, 0.3,  0.3, 0.4),
                   byrow = TRUE, nrow = 3)

# these match. J - 1 does not seems necessary.
identical(dim(init_mat), dim(fit$coef))
#> [1] TRUE

# try it out. 
fit_2 <- emlogit(Y = Y, X = X, 
                 control = list(initial_value = init_mat))
#> Error in emlogit(Y = Y, X = X, control = list(initial_value = init_mat)): Dimension of initial value does not match.

^{Created on 2020-05-04 by the reprex package (v0.3.0)}

but with the change in bfc88fe, this works.

library(emlogit)
data(japan, package = "MNP")

# example ---
Y <- japan %>% select(LDP, NFP, SKG, JCP) %>% data.matrix()
X <- japan %>% select(gender, age) %>% data.matrix()

set.seed(1234)
fit <- emlogit(Y = Y, X = X)
dim(fit$coef)
#> [1] 3 4

# try with initial values ----
# 2 covariates + 1 intercept (rows) and 4 choices (columns)
init_mat <- matrix(c(0, 0.5, -0.2, 0.1,
                     0, 0.2,  0.2, 0.2,
                     0, 0.3,  0.3, 0.4),
                   byrow = TRUE, nrow = 3)

# try it out. 
fit_2 <- emlogit(Y = Y, X = X, 
                 control = list(initial_value = init_mat))
fit_2$coef
#>      [,1]         [,2]         [,3]         [,4]
#> [1,]    0  0.505046601 -0.037569740 -0.503925933
#> [2,]    0 -0.171375636 -0.051898286  0.088159440
#> [3,]    0 -0.005205218 -0.003313376 -0.009005319

^{Created on 2020-05-04 by the reprex package (v0.3.0)}

[sou412]

Thanks. Merging this. Do you think it's better to add 0's internally rather than asking users to add them because for now, only the first column can be the baseline?

[kuriwaki]

I think making the baseline explicit throughout is good, though either is ok. To make sure there are no errors is to write in a check could be to ensure all first columns are equal to zero.