In IQSS/RobustSE: Tests for model misspecification using GIM test and robust standard errors

knitr::opts_knit$set(
        stop_on_error = 2L
)
knitr::opts_chunk$set(
        fig.path="man/figures/"
)

RobustSE

The RobustSE package implements the generalized information matrix (GIM) test to detect model misspecification described by King & Roberts (2015).

When a researcher suspects a model may be misspecified, rather than attempting to correct by fitting robust standard errors, the GIM test should be utilized as a formal statistical test for model misspecification. If the GIM test rejects the null hypothesis, the researcher should re-specify the model, as it is possible estimators of the misspecified model will be biased.

The GIM() function in the RobustSE package can be used to derive a quick rule of thumb estimate for whether or not a model is misspecified by comparing the size of classic and robust standard errors, as well as perform the full GIM test for linear, Poisson, and Negative binomial regressions.

The GIM() function accepts glm(..., family = gaussian(link="identity")), glm(..., family=binomial) with logit and probit links, glm(..., family = "poisson"), and glm.nb(...) objects as input. It is suggested that the researcher first perform the quick rule of thumb evaluation for model misspecification on their desired model by setting full = FALSE. In this case, GIM() will return the estimated coefficients, classic standard errors, robust standard errors, z-values, p-values with respect to the robust standard errors, and the rule of thumb for misspecification. If the rule of thumb is greater than 1.5 (i.e. any robust standard error is 1.5 times larger than its classic standard error), the output suggests there may be misspecification and to perform the full GIM test on the fitted model.

The full GIM test can be performed by setting full = TRUE. In this scenario, the researcher must also set the number of bootstraps for each part of the double bootstrap (B and B2) that makes up the full GIM test. The full GIM test returns the same things as the quick test above, as well as the GIM test statistic and the GIM test p-value. If the GIM test null hypothesis is rejected, this is a signal of model misspecification and the reseacher should consider re-specifying the model.

If the researcher believes it is appropriate to use cluster-robust standard errors, GIM() can handle this as well for both the rule of thumb test and the full test by setting cluster = data$cluster and the same parameters described above. In this scenario, both the rule of thumb and the full test return the same things, except the reported robust standard errors are cluster-robust standard errors.

Installation

You can install RobustSE from Github with:

# install.packages("devtools")
devtools::install_github("IQSS/RobustSE")

Example

NOTE: In order for standard errors to be calcualted correctly GLM objects must be used.

OLS with and without clustering

# install.packages("Ecdat")
library(Ecdat)
library(RobustSE)
data(Fatality)

# ols modeling traffic fatality rate
ols <- glm(mrall ~ beertax + factor(year), data = Fatality)

# Quick rule of thumb for model misspecification
GIM(ols, full = FALSE)

# Full GIM test for model misspecification
GIM(ols, full = TRUE, B = 30, B2 = 25)

# Quick rule of thumb for model misspecification; data clustered by state
GIM(ols, full = FALSE, cluster = Fatality$state)

# Full GIM test for model misspecification; data clustered by state
GIM(ols, full = TRUE, B = 30, B2 = 25, cluster = Fatality$state)

Poisson regression with and without clustering

# install.packages("MASS")
library(MASS)
data(epil)

# Poisson regression modeling seizure counts for epileptics
pois <- glm(y ~ trt + base + age + V4 + lbase + lage,
            data = epil, family = "poisson")

# Quick rule of thumb for model misspecification
GIM(pois, full = FALSE)

# Full GIM test for model misspecification
GIM(pois, full = TRUE, B = 30, B2 = 25)

# Quick rule of thumb for model misspecification; data clustered by subject
GIM(pois, full = FALSE, cluster = epil$subject)

# Full GIM test for model misspecification; data clustered by subject
GIM(pois, full = TRUE, B = 30, B2 = 25, cluster = epil$subject)

IQSS/RobustSE documentation built on May 7, 2019, 6:02 a.m.