Support `nestedLogit`
vincentarelbundock opened this issue · 3 comments
vincentarelbundock commented
Would it be possible to support nestedLogit models in parameters?
Fit model:
library(nestedLogit)
data(Womenlf, package = "carData")
comparisons <- logits(
work = dichotomy("not.work", working = c("parttime", "fulltime")),
full = dichotomy("parttime", "fulltime"))
m <- nestedLogit(
partic ~ hincome + children,
dichotomies = comparisons,
data = Womenlf)The m object includes a list of glm objects stored in m$models. When we call summary(m), the nestedLogit package returns the summary of each glm object one after the other. This is equivalent to calling:
lapply(m$models, parameters::parameters)
# $work
# Parameter | Log-Odds | SE | 95% CI | z | p
# ----------------------------------------------------------------------
# (Intercept) | 1.34 | 0.38 | [ 0.61, 2.12] | 3.48 | < .001
# hincome | -0.04 | 0.02 | [-0.08, 0.00] | -2.14 | 0.032
# children [present] | -1.58 | 0.29 | [-2.16, -1.01] | -5.39 | < .001
#
# Uncertainty intervals (profile-likelihood) and p-values (two-tailed)
# computed using a Wald z-distribution approximation.
#
# The model has a log- or logit-link. Consider using `exponentiate =
# TRUE` to interpret coefficients as ratios.
#
# $full
# Parameter | Log-Odds | SE | 95% CI | z | p
# ----------------------------------------------------------------------
# (Intercept) | 3.48 | 0.77 | [ 2.11, 5.14] | 4.53 | < .001
# hincome | -0.11 | 0.04 | [-0.19, -0.04] | -2.74 | 0.006
# children [present] | -2.65 | 0.54 | [-3.80, -1.66] | -4.90 | < .001
#
# Uncertainty intervals (profile-likelihood) and p-values (two-tailed)
# computed using a Wald z-distribution approximation.Note that the m$models list of models has informative names:
names(m$models)
# [1] "work" "full"How are these types of models handled in parameters normally? Ideally, I think we would bind each parameters data frame into a single one, with a new column to identify the model names.
Is there an existing example of this approach somewhere?
strengejacke commented
I'm not sure, we have methods for lm(cbind(), ...) or crch:
https://github.com/easystats/parameters/blob/3bc6d1bd1a7d4109cbb067bedf82fa4fd224f653/R/methods_crch.R
Maybe in that direction?
strengejacke commented
library(nestedLogit)
library(parameters)
data(Womenlf, package = "carData")
comparisons <- logits(
work = dichotomy("not.work", working = c("parttime", "fulltime")),
full = dichotomy("parttime", "fulltime")
)
m <- nestedLogit(
partic ~ hincome + children,
dichotomies = comparisons,
data = Womenlf
)
model_parameters(m)
#> # {not.work} vs. {parttime, fulltime} response
#>
#> Parameter | Log-Odds | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 1.34 | 0.38 | [ 0.61, 2.12] | 3.48 | < .001
#> hincome | -0.04 | 0.02 | [-0.08, 0.00] | -2.14 | 0.032
#> children [present] | -1.58 | 0.29 | [-2.16, -1.01] | -5.39 | < .001
#>
#> # {parttime} vs. {fulltime} response
#>
#> Parameter | Log-Odds | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 3.48 | 0.77 | [ 2.11, 5.14] | 4.53 | < .001
#> hincome | -0.11 | 0.04 | [-0.19, -0.04] | -2.74 | 0.006
#> children [present] | -2.65 | 0.54 | [-3.80, -1.66] | -4.90 | < .001
#>
#> Uncertainty intervals (profile-likelihood) and p-values (two-tailed)
#> computed using a Wald z-distribution approximation.
#>
#> The model has a log- or logit-link. Consider using `exponentiate =
#> TRUE` to interpret coefficients as ratios.
model_parameters(m, ci_method = "wald")
#> # {not.work} vs. {parttime, fulltime} response
#>
#> Parameter | Log-Odds | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 1.34 | 0.38 | [ 0.58, 2.09] | 3.48 | < .001
#> hincome | -0.04 | 0.02 | [-0.08, 0.00] | -2.14 | 0.032
#> children [present] | -1.58 | 0.29 | [-2.15, -1.00] | -5.39 | < .001
#>
#> # {parttime} vs. {fulltime} response
#>
#> Parameter | Log-Odds | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 3.48 | 0.77 | [ 1.97, 4.98] | 4.53 | < .001
#> hincome | -0.11 | 0.04 | [-0.18, -0.03] | -2.74 | 0.006
#> children [present] | -2.65 | 0.54 | [-3.71, -1.59] | -4.90 | < .001
#>
#> Uncertainty intervals (equal-tailed) and p-values (two-tailed) computed
#> using a Wald z-distribution approximation.
model_parameters(m, vcov = "HC3")
#> # {not.work} vs. {parttime, fulltime} response
#>
#> Parameter | Log-Odds | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 1.34 | 0.42 | [ 0.52, 2.15] | 3.20 | 0.001
#> hincome | -0.04 | 0.02 | [-0.09, 0.00] | -1.88 | 0.061
#> children [present] | -1.58 | 0.30 | [-2.16, -1.00] | -5.33 | < .001
#>
#> # {parttime} vs. {fulltime} response
#>
#> Parameter | Log-Odds | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 3.48 | 0.76 | [ 1.98, 4.98] | 4.55 | < .001
#> hincome | -0.11 | 0.04 | [-0.18, -0.03] | -2.88 | 0.004
#> children [present] | -2.65 | 0.56 | [-3.75, -1.55] | -4.72 | < .001
#>
#> Uncertainty intervals (equal-tailed) and p-values (two-tailed) computed
#> using a Wald z-distribution approximation.
model_parameters(m, exponentiate = TRUE)
#> # {not.work} vs. {parttime, fulltime} response
#>
#> Parameter | Odds Ratio | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 3.80 | 1.46 | [1.83, 8.29] | 3.48 | < .001
#> hincome | 0.96 | 0.02 | [0.92, 1.00] | -2.14 | 0.032
#> children [present] | 0.21 | 0.06 | [0.12, 0.36] | -5.39 | < .001
#>
#> # {parttime} vs. {fulltime} response
#>
#> Parameter | Odds Ratio | SE | 95% CI | z | p
#> -------------------------------------------------------------------------
#> (Intercept) | 32.39 | 24.84 | [8.26, 171.34] | 4.53 | < .001
#> hincome | 0.90 | 0.04 | [0.83, 0.97] | -2.74 | 0.006
#> children [present] | 0.07 | 0.04 | [0.02, 0.19] | -4.90 | < .001
#>
#> Uncertainty intervals (profile-likelihood) and p-values (two-tailed)
#> computed using a Wald z-distribution approximation.
model_parameters(m) |> as.data.frame()
#> Parameter Coefficient SE CI CI_low CI_high
#> 1 (Intercept) 1.33582979 0.38376323 0.95 0.60590835 2.115336975
#> 2 hincome -0.04230843 0.01978012 0.95 -0.08225743 -0.004471769
#> 3 childrenpresent -1.57564843 0.29226284 0.95 -2.16143847 -1.012720928
#> 4 (Intercept) 3.47777346 0.76710910 0.95 2.11086529 5.143653414
#> 5 hincome -0.10726786 0.03915231 0.95 -0.18920584 -0.035088050
#> 6 childrenpresent -2.65145569 0.54107504 0.95 -3.80274064 -1.657253408
#> z df_error p Component Response
#> 1 3.480870 Inf 4.997885e-04 work {not.work} vs. {parttime, fulltime}
#> 2 -2.138937 Inf 3.244074e-02 work {not.work} vs. {parttime, fulltime}
#> 3 -5.391204 Inf 6.998736e-08 work {not.work} vs. {parttime, fulltime}
#> 4 4.533610 Inf 5.798406e-06 full {parttime} vs. {fulltime}
#> 5 -2.739758 Inf 6.148445e-03 full {parttime} vs. {fulltime}
#> 6 -4.900347 Inf 9.566732e-07 full {parttime} vs. {fulltime}
model_parameters(m, component = "work")
#> # Fixed Effects
#>
#> Parameter | Log-Odds | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 1.34 | 0.38 | [ 0.61, 2.12] | 3.48 | < .001
#> hincome | -0.04 | 0.02 | [-0.08, 0.00] | -2.14 | 0.032
#> children [present] | -1.58 | 0.29 | [-2.16, -1.01] | -5.39 | < .001
#>
#> Uncertainty intervals (profile-likelihood) and p-values (two-tailed)
#> computed using a Wald z-distribution approximation.
model_parameters(m, ci_method = "wald", component = "work")
#> # Fixed Effects
#>
#> Parameter | Log-Odds | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 1.34 | 0.38 | [ 0.58, 2.09] | 3.48 | < .001
#> hincome | -0.04 | 0.02 | [-0.08, 0.00] | -2.14 | 0.032
#> children [present] | -1.58 | 0.29 | [-2.15, -1.00] | -5.39 | < .001
#>
#> Uncertainty intervals (equal-tailed) and p-values (two-tailed) computed
#> using a Wald z-distribution approximation.
model_parameters(m, vcov = "HC3", component = "work")
#> # Fixed Effects
#>
#> Parameter | Log-Odds | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 1.34 | 0.42 | [ 0.52, 2.15] | 3.20 | 0.001
#> hincome | -0.04 | 0.02 | [-0.09, 0.00] | -1.88 | 0.061
#> children [present] | -1.58 | 0.30 | [-2.16, -1.00] | -5.33 | < .001
#>
#> Uncertainty intervals (equal-tailed) and p-values (two-tailed) computed
#> using a Wald z-distribution approximation.
model_parameters(m, exponentiate = TRUE, component = "work")
#> # Fixed Effects
#>
#> Parameter | Odds Ratio | SE | 95% CI | z | p
#> ----------------------------------------------------------------------
#> (Intercept) | 3.80 | 1.46 | [1.83, 8.29] | 3.48 | < .001
#> hincome | 0.96 | 0.02 | [0.92, 1.00] | -2.14 | 0.032
#> children [present] | 0.21 | 0.06 | [0.12, 0.36] | -5.39 | < .001
#>
#> Uncertainty intervals (profile-likelihood) and p-values (two-tailed)
#> computed using a Wald z-distribution approximation.
model_parameters(m, component = "work") |> as.data.frame()
#> Parameter Coefficient SE CI CI_low CI_high
#> 1 (Intercept) 1.33582979 0.38376323 0.95 0.60590835 2.115336975
#> 2 hincome -0.04230843 0.01978012 0.95 -0.08225743 -0.004471769
#> 3 childrenpresent -1.57564843 0.29226284 0.95 -2.16143847 -1.012720928
#> z df_error p Response
#> 1 3.480870 Inf 4.997885e-04 {not.work} vs. {parttime, fulltime}
#> 2 -2.138937 Inf 3.244074e-02 {not.work} vs. {parttime, fulltime}
#> 3 -5.391204 Inf 6.998736e-08 {not.work} vs. {parttime, fulltime}Created on 2023-05-22 with reprex v2.0.2
vincentarelbundock commented
This is awesome! Thanks so much @strengejacke