Clustered variance-covariance matrices and T statistic d.o.f.

Description

Variance-covariance matrices with robust clustered standard errors and degrees-of-freedom for T statistics, for tests and examples specifying vcov (d.o.f. defined as g - 1, where g is the number of clusters). Generated with margex data in this package.

Usage

cvcov

Format

A list of three lists, from an OLS model, logit model, and OLS with a polynomial interaction with missing data, each containing

clust

3-by-3 variance-covariance matrix

dof

integer, degrees of freedom for the T statistic

Details

See data-raw/make_cluster_vcov.R for details.

Source

http://cameron.econ.ucdavis.edu/research/Cameron_Miller_JHR_2015_February.pdf

Estimating predictive margins on a model

Description

This function estimates the predictive effects and levels for variables within a model using the delta method.

Usage

marg(
  mod,
  var_interest,
  data = NULL,
  weights = NULL,
  vcov_mat = NULL,
  dof = NULL,
  type = "levels",
  base_rn = 1,
  at_var_interest = NULL,
  at = NULL,
  cofint = 0.95,
  ...
)

Arguments

Details

The variable for the predictive margin is specified by var_interest. If margins are only needed at particular values of var_interest, at_var_interest should be used. If margins of var_interest are needed at across the levels of a different variable in the model, at should be used.

If higher-order polynomial terms (e.g. y ~ x + x^2) are added using the R function poly, the raw = TRUE argument should be used to include the basic polynomial terms instead of orthogonal polynomial terms. If orthogonal polynomials are used, marg will fail when the user specifies at for a small set of values for the variable in question (e.g. at = list(x = 10)), since poly needs more data to calculate orthogonal polynomials (e.g. poly(10, 2) fails, but poly(c(10, 8, 3), 2) will run).

P values are calculated with T tests for gaussian families, and Z tests otherwise. If a new variance-covariance matrix is provided (e.g. for clustering standard errors), the degrees of freedom for the T test / p-value calculation may need to be specified using dof.

This function currently only supports glm and ivreg objects. If you would like to use lm objects, consider running a glm with family gaussian.

When calculating predicted levels and effects for models built using weights, marg returns weighted averages for levels and effects by default. Users can remove this option by setting weights = NULL.

Value

list of dataframes with predicted margins/effects, standard errors, p-values, and confidence interval bounds

Predicted Margins for 'glm' objects

Description

Obtains predicted margins and standard errors of those predictions from a fitted generalized linear model object.

Usage

## S3 method for class 'glm'
marg(
  mod,
  var_interest,
  data = mod$data[names(mod$prior.weights), ],
  weights = mod$prior.weights,
  ...
)

Arguments

Examples

data(mtcars)
mod <- glm(vs ~ as.factor(gear) + mpg, data = mtcars, family = 'binomial')

# Get the level of the outcome variable at different values of `gear`
marg(mod, var_interest = 'gear', type = 'levels')
# Get the effect of `gear` on the outcome value, holding values of `mpg`
# constant
marg(mod, var_interest = 'gear', type = 'effects',
     at = list(mpg = c(15, 21)))

data(margex)
mod <- glm(outcome ~ as.factor(treatment) + distance,
       data = margex, family = 'binomial')
# Get the level of the outcome variable at different values of `treatment`
marg(mod, var_interest = 'treatment', type = 'levels', at = NULL)
# Get the effect of `treatment` on the outcome variable
marg(mod, var_interest = 'treatment', type = 'effects', at = NULL)
# Get the level of the outcome variable at different values of `distance`
marg(mod, var_interest = 'distance', type = 'levels',
          at = NULL, at_var_interest = c(10, 20, 30))

# Using a custom variance-covariance matrix for clustered standard errors
# (also requires custom degrees of freedom for T statistic with OLS model),
# clustering on the "arm" variable

data(margex)
data(cvcov)
# ?cvcov
v <- cvcov$ols$clust
d <- cvcov$ols$stata_dof
mod <- glm(outcome ~ treatment + distance,
           data = margex, family = 'binomial')
marg(mod, var_interest = 'treatment', type = 'levels',
          vcov_mat = v, dof = d)

# Using weights
data(margex)
mm <- glm(y ~ as.factor(treatment) + age, data = margex, family = 'gaussian',
          weights = distance)
z1 <- marg(mod = mm, var_interest = 'treatment', type = 'levels')[[1]]
z2 <- marg(mod = mm, var_interest = 'treatment', type = 'effects')[[1]]

Predicted Margins for 'ivreg' objects from the AER package

Description

Obtains predicted margins and standard errors of those predictions from a fitted ivreg model object.

Usage

## S3 method for class 'ivreg'
marg(mod, var_interest, data, weights = NULL, ...)

Arguments

Examples

# From ?AER::ivreg

# data
data("CigarettesSW", package = "AER")
CigarettesSW$rprice <- with(CigarettesSW, price/cpi)
CigarettesSW$rincome <- with(CigarettesSW, income/population/cpi)
CigarettesSW$tdiff <- with(CigarettesSW, (taxs - tax)/cpi)

# model
fm <- AER::ivreg(log(packs) ~ log(rprice) + log(rincome) |
                   log(rincome) + tdiff + I(tax/cpi),
                 data = CigarettesSW, subset = year == "1995")

# Get margins for different levels of price/cpi
rprice_levs <- round(quantile(CigarettesSW$rprice))

marg(fm, data = subset(CigarettesSW, year == "1995"),
     var_interest = 'rprice', at_var_interest = rprice_levs)

Artificial data for margins

Description

A fictitious dataset outcome, treatment, and demographic variables for 3000 observations.

Usage

margex

Format

A data frame with 3000 rows and 11 variables:

y

numeric

outcome

integer, 0 or 1

sex

character: "female" or "male"

group

integer

age

integer

distance

numeric

ycn

numeric

yc

numeric, 0 or 1

treatment

integer

agegroup

character: "20-29", "30-39", or "40+"

arm

integer

Source

https://www.stata-press.com/data/r14/margex.dta

Main wrapper function to calculate margins and standard errors

Description

For one set of transformed covariates (not including the variable of interest), calculate the predicted level and standard error for the variable of interest.

Usage

pred_se(
  df_levels,
  model,
  type,
  base_rn,
  vcov_mat,
  weights,
  deriv_func,
  link_func
)

Arguments