General Method of Moments (GMM) estimators for panel data with lagged and differenced instruments in R.
The function panelGMM estimates one-step and two-step GMM-IV linear models for panel data. Panel-robust standard errors are calculated using the algorithm laid out in Chapter 22 of Cameron and Trivedi (2005).
The package also provides convenience functions to lag/lead/difference variables (panelGMM::shiftCols) and to convert the panelGMM object (the estimated parameters and hypothesis tests) into a data.frame (panelGMM::tidy_panelGMM). This is similar to broom::tidy.
library("devtools")
install_github("lrdegeest/panelGMM")Recreate the base-case two-step GMM-IV model in Table 22.2 of Cameron and Trivedi (2005). The authors estimate wages for a panel of 532 workers from 1981 - 1988.
library(panelGMM)
library(tidyverse) # for the pipe (%>%) and ggplot2
# load the data
data("hours_wages")
# use panelGMM::shiftCols to create lags and differences with automatic names
hours_wages_gmm = hours_wages %>%
shiftCols(data = ., panel=id, type = "difference", shifts = 1:2, columns=c("lnhr", "lnwg", "kids", "age", "agesq", "disab")) %>%
shiftCols(data = ., panel=id, type = "lag", shifts = 1:2, columns=c("lnhr", "lnwg", "kids", "age", "agesq", "disab"))
# define the model
model = lnhr_diff_1 ~ lnwg_diff_1 + kids_diff_1 + age_diff_1 + agesq_diff_1 + disab_diff_1 | kids_lag_1 + kids_lag_2 + age_lag_1 + age_lag_2 + agesq_lag_1 + agesq_lag_2 + disab_lag_1 + disab_lag_2 + lnwg_lag_2
# estimate the parameters and view the summary
panelGMM(model, panel = id, time = year, twostep = TRUE, data = hours_wages_gmm) %>%
summary()
-------------------------------------------------------------------
CALL
panelGMM(formula = model, panel = id, time = year, twostep = TRUE,
data = hours_wages_gmm)
-------------------------------------------------------------------
MODEL DETAILS
Unique cross-sections (N): 532
Unique obs. per cross-section (T): 8
Total number of obs. (N x T): 4256
Number of instruments: 9
-------------------------------------------------------------------
COEFFICIENTS
Estimate Std.Error z-value Pr(>|z|)
lnwg_diff_1 0.5467961 0.3276239 1.669 0.0951 .
kids_diff_1 -0.0449042 0.0271441 -1.654 0.0981 .
age_diff_1 0.0274759 0.0129598 2.120 0.0340 *
agesq_diff_1 -0.0003591 0.0001594 -2.253 0.0243 *
disab_diff_1 -0.0468348 0.0623601 -0.751 0.4526
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
-------------------------------------------------------------------
MODEL SUMMARY
Root mean square error (RMSE): 0.307
Overidentification Restrictions Test (OIR)
J-statistic: 5.4504, p-value: 0.2441, DOF (r - k) = 4
panelGMM is not yet set up to work with broom or sjPLot. But you can use panelGMM::tidy_panelGMM to convert the estimated model to a data.frame and pipe the output to ggplot:
panelGMM(model, panel = id, time = year, twostep = TRUE, data = hours_wages_gmm) %>%
tidy_panelGMM() %>%
ggplot(., aes(x = terms, y = estimates)) +
geom_point() +
geom_pointrange(aes(ymin = estimates -standard.errors, ymax = estimates + standard.errors)) +
coord_flip()
Several core routines in panelGMM are written in C++ so estimation is pretty fast:
microbenchmark::microbenchmark(
"panelGMM" = panelGMM(model, panel = id, time = year, twostep = TRUE, data = hours_wages_gmm),
times = 1000) %>%
ggplot2::autoplot(.)
Cameron, A. Colin, and Douglas L. Miller. "A practitioner’s guide to cluster-robust inference." Journal of human resources 50, no. 2 (2015): 317-372.
Cameron, A. Colin, and Pravin K. Trivedi. Microeconometrics: Methods and Applications. Cambridge university press, 2005.