Movatterモバイル変換

CRAN Task View: Econometrics

Base R ships with a lot of functionality useful for (computational) econometrics, in particular in the stats package. This functionality is complemented by many packages on CRAN, a brief overview is given below. There is also a certain overlap between the tools for econometrics in this view and those in the task views onFinance,TimeSeries, andCausalInference.

The packages in this view can be roughly structured into the following topics. If you think that some package is missing from the list, please file an issue in the GitHub repository or contact the maintainer.

Basic linear regression

• Estimation and standard inference: Ordinary least squares (OLS) estimation for linear models is provided bylm() (from stats) and standard tests for model comparisons are available in various methods such assummary() andanova().
• Further inference and nested model comparisons: Functions analogous to the basicsummary() andanova() methods that also support asymptotic tests (z instead oft tests, and Chi-squared instead ofF tests) and plug-in of other covariance matrices arecoeftest() andwaldtest() inlmtest. (Non)linear hypothesis testing for a wide range of R packages can implemented through thedeltamethod() function ofmarginaleffects. This expands on older (non)linear hypothesis test functions likelinearHypothesis() anddeltaMethod() fromcar.
• Robust standard errors: HC, HAC, clustered, and bootstrap covariance matrices are available insandwich and can be plugged into the inference functions mentioned above.
• Nonnested model comparisons: Various tests for comparing non-nested linear models are available inlmtest (encompassing test, J test, Cox test). The Vuong test for comparing other non-nested models is provided bynonnest2 (and specifically for count data regression inpscl).
• Diagnostic checking: The packagescar andlmtest provide a large collection of regression diagnostics and diagnostic tests.
• Miscellaneous: Much of the above functionality is bundled together infixest, which provides a number of in-built convenience features that users may find attractive. This includes robust standard error specification, multi-model estimation, custom hypothesis testing, etc.

Microeconometrics

• Generalized linear models (GLMs): Many standard microeconometric models belong to the family of generalized linear models and can be fitted byglm() from package stats. This includes in particular logit and probit models for modeling choice data and Poisson models for count data.
• Effects and marginal effects: Effects for typical values of regressors in GLMs and various other probabilistic regression models can be obtained and visualized usingeffects. Marginal effect tables and corresponding visualizations for a wide range of models can be be produced withmarginaleffects. Other implementations of marginal effects for certain models are inmargins andmfx. Interactive visualizations of both effects and marginal effects are possible inLinRegInteractive.
• Binary responses: The standard logit and probit models (among many others) for binary responses are GLMs that can be estimated byglm() withfamily = binomial. Bias-reduced GLMs that are robust to complete and quasi-complete separation are provided bybrglm. Discrete choice models estimated by simulated maximum likelihood are implemented inRchoice.bife provides binary choice models with fixed effects. Heteroscedastic probit models (and other heteroscedastic GLMs) are implemented inglmx along with parametric link functions and goodness-of-link tests for GLMs.
• Count responses: The basic Poisson regression is a GLM that can be estimated byglm() withfamily = poisson as explained above. Negative binomial GLMs are available viaglm.nb() in packageMASS. Another implementation of negative binomial models is provided byaod, which also contains other models for overdispersed data. Zero-inflated and hurdle count models are provided in packagepscl. A reimplementation by the same authors is currently under development incountreg on R-Forge which also encompasses separate functions for zero-truncated regression, finite mixture models etc.
• Multinomial responses: Multinomial models with individual-specific covariates only are available inmultinom() from packagennet. An implementation with both individual- and choice-specific variables ismlogit. Generalized multinomial logit models (e.g., with random effects etc.) are ingmnl. A flexible framework of various customizable choice models (including multinomial logit and nested logit among many others) is implemented in theapollo package. The newerlogitr package combines many of the features from these preceding packages and also offers some meaningful performance improvements for fast estimation of multinomial and mixed logit models. Simulated maximum likelihood estimation of mixed logit models, especially for large data sets, is available inmixl. Generalized additive models (GAMs) for multinomial responses can be fitted with theVGAM package. A Bayesian approach to multinomial probit models is provided byMNP. Various Bayesian multinomial models (including logit and probit) are available inbayesm. The packageRSGHB fits various hierarchical Bayesian specifications based on direct specification of the likelihood function. Furthermore, theRprobitB package implements latent class mixed multinomial probit models for approximations of the true underlying mixing distribution.
• Ordered responses: Proportional-odds regression for ordered responses is implemented inpolr() from packageMASS. The packageordinal provides cumulative link models for ordered data which encompasses proportional odds models but also includes more general specifications. Bayesian ordered probit models are provided bybayesm andRprobitB.
• Censored responses: Basic censored regression models (e.g., tobit models) can be fitted bysurvreg() insurvival, a convenience interfacetobit() is in packageAER. Further censored regression models, including models for panel data, are provided incensReg. Censored regression models with conditional heteroscedasticity are incrch. Furthermore, hurdle models for left-censored data at zero can be estimated withmhurdle. Models for sample selection are available insampleSelection andssmrob using classical and robust inference, respectively. PackagematchingMarkets corrects for selection bias when the sample is the result of a stable matching process (e.g., a group formation or college admissions problem).
• Truncated responses:crch for truncated (and potentially heteroscedastic) Gaussian, logistic, and t responses. Homoscedastic Gaussian responses are also available intruncreg.
• Fraction and proportion responses: Beta regression for responses in (0, 1) is inbetareg andgamlss.
• Duration responses: Many classical duration models can be fitted withsurvival, e.g., Cox proportional hazard models withcoxph() or Weibull models withsurvreg(). Many more refined models can be found in theSurvival task view.
• High-dimensional fixed effects: Linear and generalized linear models with potentially high-dimensional fixed effects, also for multiple groups, can be fitted withfixest, using optimized parallel C++ code. Other implementations of high-dimensional fixed effects are inlfe andalpaca for linear and generalized linear models, respectively.
• Miscellaneous: Further more refined tools for microeconometrics are provided in themicEcon family of packages: Analysis with Cobb-Douglas, translog, and quadratic functions is inmicEcon; the constant elasticity of scale (CES) function is inmicEconCES; the symmetric normalized quadratic profit (SNQP) function is inmicEconSNQP. The almost ideal demand system (AIDS) is inmicEconAids. Stochastic frontier analysis (SFA) is infrontier. Semiparametric SFA in is available insemsfa and spatial SFA inssfa. The packagebayesm implements a Bayesian approach to microeconometrics and marketing. Inference for relative distributions is contained in packagereldist.

Common research designs for causal inference

We review packages related to some common research designs for causal inference below. This section is necessarily brief and should be paired with theCausalInference task view, since is there a high degree of overlap.

Difference-in-differences and synthetic control

• Basic difference-in-differences (DiD): The canonical 2x2 DiD model (two units, two periods) can be estimated as a simple interaction between two factor variables inlm() orglm(), etc. Similarly, the equivalent two-way fixed effects (TWFE) design can be obtained using factors to control for unit and time fixed effects. However, for high-dimensional datasets TWFE is more conveniently estimated using a dedicated panel data package likefixest orplm. The former even provides a conveniencei() operator for constructing and interacting factors in TWFE settings.
• Advanced DiD and TWFE corrections: Despite its long-standing popularity, recent research has uncovered various problems with (naive) TWFE; for example, severe bias in the presence of staggered treatment rollout. Acottage industry of workarounds and alternative estimators now exists to address these problems. R package implementations include:bacondecomp,did,did2s,DRDID,etwfe,fixest (via thesunab() function), andgsynth.
• Synthetic control: The original synthetic control (SC) implementation is available throughSynth, whiletidysynth offers a newer SC implementation with various enhancements (speed, inspection, etc.) Similarly,gsynth generalizes the original SC implementation to multiple treated units and variable treatment periods, and also supports additional estimation methods like the EM algorithm and matrix completion.

Instrumental variables

• Basic instrumental variables (IV) regression: Two-stage least squares (2SLS) is provided byivreg, which separates out the dedicated 2SLS routines previously found inAER. Another implementation is available astsls() in packagesem.
• Binary responses: TheLARF package estimates local average response functions for binary treatments and binary instruments.
• Panel data: Several panel data model packages (see below) provide their own dedicated IV routines for efficient estimation in the presence of high-dimensional data. These includefixest andlfe for fixed effects, andplm for first-difference, between, and multiple random effects methods.
• Miscellaneous:REndo fits linear models with endogenous regressor using various latent instrumental variable approaches.SteinIV provides semi-parametric IV estimators, including JIVE and SPS.

Regression discontinuity design

• Regression discontinuity design (RDD) methods are implemented inrdrobust (offering robust confidence interval construction and bandwidth selection),rddensity (density discontinuity testing (also known as manipulation testing)),rdlocrand (inference under local randomization), andrdmulti (analysis with multiple cutoffs or scores).
• Tools to perform power, sample size and minimum detectable effects (MDE) calculations are available inrdpower, whileRATest provides a collection of randomization tests, including a permutation test for the continuity assumption of the baseline covariates in the sharp RDD.

Panel data models

• Panel standard errors: A simple approach for panel data is to fit the pooling (or independence) model (e.g., vialm() orglm()) and only correct the standard errors. Different types of clustered, panel, and panel-corrected standard errors are available insandwich (incorporating prior work frommultiwayvcov),clusterSEs,pcse,clubSandwich,plm, andgeepack, respectively. The latter two require estimation of the pooling/independence models viaplm() andgeeglm() from the respective packages (which also provide other types of models, see below).
• Linear panel models:fixest provides very efficient fixed-effect routines that scale to high-dimensional data and multiple fixed-effects.plm, providing a wide range of within, between, and random-effect methods (among others) along with corrected standard errors, tests, etc. Various dynamic panel models are available inplm, with estimation based on moment conditions inpdynmc, and dynamic panel models with fixed effects inOrthoPanels.feisr provides fixed effects individual slope (FEIS) models. Panel vector autoregressions with homogeneous slope parameters are implemented inpanelvar and with heterogeneous parameters inpvars.
• GLMs and generalized estimation equations. The aformentionedfixest supports a variety of GLM-like models in addition to linear panel models. This includes efficient fixed-effect estimation of logit, probit, Poisson, and negative binomial models. Similar functionality is provided byalpaca (which also accounts for incidental parameter problems) andpglm.penppml further extends the high-dimensional case through penalized Poisson Pseudo Maximum Likelihood (PPML) regressions, using lasso or ridge penalties. GEE models for panel data (or longitudinal data in statistical jargon) are available in ingeepack.
• Mixed effects models: Linear and nonlinear models for panel data (and more general multi-level data) are available inlme4 andnlme.
• Instrumental variables:fixest. See also above.
• Miscellaneous: Threshold regression and unit root tests are inpdR. Dedicated fast data preprocessing for panel data econometrics is provided bycollapse.

Further regression models

• Nonlinear least squares modeling:nls() in package stats.
• Quantile regression:quantreg (including linear, nonlinear, censored, locally polynomial and additive quantile regressions).
• Generalized method of moments (GMM) and generalized empirical likelihood (GEL):gmm.
• Spatial econometric models: TheSpatial view gives details about handling spatial data, along with information about (regression) modeling. In particular, spatial regression models can be fitted usingspatialreg andsphet (the latter using a GMM approach).splm is a package for spatial panel models. Spatial probit models are available inspatialprobit and spatial seemingly unrelated regression (SUR) models inspsur.
• Bayesian model averaging (BMA): A comprehensive toolbox for BMA is provided byBMS including flexible prior selection, sampling, etc. A different implementation is inBMA for linear models, generalizable linear models and survival models (Cox regression).
• Linear structural equation models:lavaan andsem. See also thePsychometrics task view for more details.
• Machine learning: There are several packages that combine machine learning techniques with econometric inference (especially for identifying causal effects). These includegrf for causal random forests and estimation of heterogeneous treatment effects,DoubleML for double machine learning of a wide range of models from the mlr3 family, andhdm for selected high-dimensional econometric models. For a more general overview see theMachineLearning task view.
• Simultaneous equation estimation:systemfit.
• Nonparametric methods:np using kernel smoothing andNNS using partial moments.
• Linear and nonlinear mixed-effect models:nlme andlme4.
• Generalized additive models (GAMs):mgcv,gam,gamlss andVGAM.
• Design-based inference:estimatr contains fast procedures for several design-appropriate estimators with robust standard errors and confidence intervals including linear regression, instrumental variables regression, difference-in-means, among others.
• Extreme bounds analysis:ExtremeBounds.
• Miscellaneous: The packagesVGAM,rms andHmisc provide several tools for extended handling of (generalized) linear regression models.

Time series data and models

• TheTimeSeries task view provides much more detailed information about both basic time series infrastructure and time series models. Here, only the most important aspects relating to econometrics are briefly mentioned. Time series models for financial econometrics (e.g., GARCH, stochastic volatility models, or stochastic differential equations, etc.) are described in theFinance task view.
• Infrastructure for regularly spaced time series: The class"ts" in package stats is R’s standard class for regularly spaced time series (especially annual, quarterly, and monthly data). It can be coerced back and forth without loss of information to"zooreg" from packagezoo.
• Infrastructure for irregularly spaced time series:zoo provides infrastructure for both regularly and irregularly spaced time series (the latter via the class"zoo") where the time information can be of arbitrary class. This includes daily series (typically with"Date" time index) or intra-day series (e.g., with"POSIXct" time index). An extension based onzoo geared towards time series with different kinds of time index isxts. Further packages aimed particularly at finance applications are discussed in theFinance task view.
• Classical time series models: Simple autoregressive models can be estimated withar() and ARIMA modeling and Box-Jenkins-type analysis can be carried out witharima() (both in the stats package). An enhanced version ofarima() is inforecast.
• Linear regression models: A convenience interface tolm() for estimating OLS and 2SLS models based on time series data isdynlm. Linear regression models with AR error terms via GLS is possible usinggls() fromnlme.
• Structural time series models: Standard models can be fitted withStructTS() in stats. Further packages are discussed in theTimeSeries task view.
• Filtering and decomposition:decompose() andHoltWinters() in stats. The basic function for computing filters (both rolling and autoregressive) isfilter() in stats. Many extensions to these methods, in particular for forecasting and model selection, are provided in theforecast package.
• Vector autoregression: Simple models can be fitted byar() in stats, more elaborate models are provided in packagevars along with suitable diagnostics, visualizations etc. Structural smooth transition vector autoregressive models are available insstvars. Panel vector autoregressions are implemented inpanelvar and inpvars.
• Unit root and cointegration tests:urca,tseries,CADFtest. See alsoplm for panel unit root tests andpvars for system-based panel cointegration tests.
• Miscellaneous:
  • tsDyn - Threshold and smooth transition models.
  • midasr -MIDAS regression and other econometric methods for mixed frequency time series data analysis.
  • gets - GEneral-To-Specific (GETS) model selection for either ARX models with log-ARCH-X errors, or a log-ARCH-X model of the log variance.
  • bimets - Econometric modeling of time series data using flexible specifications of simultaneous equation models.
  • dlsem - Distributed-lag linear structural equation models.
  • lpirfs - Local projections impulse response functions.
  • apt - Asymmetric price transmission models.

Data sets

• Textbooks and journals: PackagesAER,Ecdat, andwooldridge contain a comprehensive collections of data sets from various standard econometric textbooks (including Greene, Stock & Watson, Wooldridge, Baltagi, among others) as well as several data sets from the Journal of Applied Econometrics and the Journal of Business & Economic Statistics data archives.AER andwooldridge additionally provide extensive sets of examples reproducing analyses from the textbooks/papers, illustrating various econometric methods. Inpder a wide collection of data sets for “Panel Data Econometrics with R” (Croissant & Millo 2018) is available. ThePoEdata package on GitHub provides the data sets from “Principles of Econometrics” (4th ed, by Hill, Griffiths, and Lim 2011).
• Penn World Table:pwt provides versions 5.6, 6.x, 7.x. Version 8.x and 9.x data are available inpwt8 andpwt9, respectively.
• Time series and forecasting data: The packagesexpsmooth,fma, andMcomp are data packages with time series data from the books “Forecasting with Exponential Smoothing: The State Space Approach” (Hyndman, Koehler, Ord, Snyder, 2008, Springer) and “Forecasting: Methods and Applications” (Makridakis, Wheelwright, Hyndman, 3rd ed., 1998, Wiley) and the M-competitions, respectively.
• Empirical Research in Economics: Packageerer contains functions and datasets for the book of “Empirical Research in Economics: Growing up with R” (Sun 2015).
• Panel Study of Income Dynamics (PSID):psidR can build panel data sets from the Panel Study of Income Dynamics (PSID).
• World Bank data and statistics: Thewbstats package provides programmatic access to the World Bank API.

Miscellaneous

• Model tables: A flexible implementation of side-by-side summary tables for a wide range of statistical models along with corresponding visualizations and data summary tables is provided inmodelsummary. Other implementations as well as further utilities for integrating econometric and statistical results in scientific papers etc. are discussed in theReproducibleResearch task view.
• Matrix manipulations: As a vector- and matrix-based language, base R ships with many powerful tools for doing matrix manipulations, which are complemented by the packagesMatrix andSparseM.
• Optimization and mathematical programming: R and many of its contributed packages provide many specialized functions for solving particular optimization problems, e.g., in regression as discussed above. Further functionality for solving more general optimization problems, e.g., likelihood maximization, is discussed in the theOptimization task view.
• Bootstrap: In addition to the recommendedboot package, there are some other general bootstrapping techniques available inbootstrap orsimpleboot as well some bootstrap techniques designed for time-series data, such as the maximum entropy bootstrap inmeboot or thetsbootstrap() fromtseries. ThevcovBS() andvcovJK() functions insandwich provide object-oriented implementations of (clustered) bootstrap and jackknife covariance matrices that work for various model objects. Thefwildclusterboot package provides a fast wild cluster bootstrap implementation for linear regression models, especially when the number of clusters is low.
• Inequality: For measuring inequality, concentration and poverty the packageineq provides some basic tools such as Lorenz curves, Pen’s parade, the Gini coefficient, Herfindahl-Hirschman index and many more.wINEQ provides these and other inequality measures for weighted data along with bootstrapping methods.
• Structural change: R is particularly strong when dealing with structural changes and changepoints in parametric models, seestrucchange andsegmented.
• Exchange rate regimes: Methods for inference about exchange rate regimes, in particular in a structural change setting, are provided byfxregime.
• Global value chains: Tools and decompositions for global value chains are ingvc anddecompr.
• Regression discontinuity design: A variety of methods are provided in therdrobust andrdlocrand packages. Therdpower package offers power calculations for regression discontinuity designs. Andrdmulti implements analysis with multiple cutoffs or scores.
• Gravity models: Estimation of log-log and multiplicative gravity models is available ingravity.
• z-Tree:zTree can import data from the z-Tree software for developing and carrying out economic experiments.
• Numerical standard errors:nse(archived) implements various numerical standard errors for time series data, especially in simulation experiments with correlated outcome sequences.

CRAN packages

Related links

• Articles:Special Volume on “Econometrics in R” in JSS (2008)
• Book:Applied Econometrics with R (Kleiber & Zeileis; 2008)
• Book:Introduction to Econometrics with R (Hanck, Arnold, Gerber, & Schmelzer; 2021)
• Book:Introduction to Econometrics with R (Oswald, Robin, & Viers; 2020)
• Book:Causal Inference: The Mixtape (Cunningham; 2021)
• Book:Hands-On Intermediate Econometrics Using R (Vinod; 2008)
• Book:Learning Microeconometrics with R (Adams; 2021)
• Book:Panel Data Econometrics with R (Croissant & Millo; 2018)
• Book:Principles of Econometrics with R (Colonescu; 2016)
• Book:Spatial Econometrics (Kelejian & Piras; 2017)
• Book:Statistical Inference via Data Science (Ismay & Kim; 2022)
• Book:The Effect (Huntington-Klein; 2022)
• Book:Using R for Introductory Econometrics (Heiss; 2019)
• Book:Quantitative Economics with R (Dayal; 2020)
• Course:Applied Empirical Methods (Goldsmith-Pinkham; 2021)
• Course:Data Science for Economists (McDermott; 2021)
• Course:Econometrics In-Class Labs (Ransom; 2021)
• Course:Introduction to Econometrics (Rubin; 2021)
• Course:PhD Econometrics (Rubin; 2022)
• Course:Program Evaluation for Public Service (Heiss, 2022)
• Course:Statistical Rethinking (McElreath; 2022)
• Website:Stata2R

Other resources

• CRAN Task View:CausalInference
• CRAN Task View:Finance
• CRAN Task View:MachineLearning
• CRAN Task View:Optimization
• CRAN Task View:Psychometrics
• CRAN Task View:ReproducibleResearch
• CRAN Task View:Spatial
• CRAN Task View:Survival
• CRAN Task View:TimeSeries
• R-Forge Project:countreg
• GitHub Project:fwildclusterboot
• GitHub Project:PoEdata