Posterior Mean Panel Predictor
Dynamic panel modelling framework based on an empirical-Bayes approach.
Contains tools for computing point forecasts and bootstrapping prediction intervals.
Reference: Liu et al. (2016) <doi:10.2139/ssrn.2889000>.
What is the Posterior Mean Panel Predictor?
Accurate predictions with micro-panels
Micro-panels are longitudinal data sets that contain observations on multiple units at only a few points in time. Examples include the performance of start-up companies, developmental skills of small children or revenues and leverage of banks after significant regulatory changes. When working with micro-panels, it is challenging to build accurate predictive models, as the time series are too short to contain enough information on their own.
Posterior Mean Panel Predictor (PMPP) takes an empirical-Bayes approach to computing forecasts with micro-panels. It uses cross-sectional information in the data to approximate the posterior mean of heterogeneous coefficients under a correlated random effects distribution. It has been shown to provide predictions of higher accuracy compared to the state-of-the-art methods for dynamic panel modelling. For more details, see the references in
pmpp() function manual.
The package allows for the following:
- Estimate the parameters of the PMPP model,
- Use the model to compute point forecasts,
- Compute prediction intervals with the Random-Window Block Bootstrap.
Additionally, the package exports a number of functions that can be used outside of the scope of PMPP modelling:
kde()for computing a robust kernel density estimate;
kde2D()for computing a robust 2-dimensional kernel density estimate;
create_fframe()for adding time periods to a panel-structured data frame;
ssys_gmm(), the suboptimal multi-step System-GMM estimator for AR(1) panel data model.
How to use
The central function in the package is
pmpp(). It estimates the model's coefficients and outputs an object of class
pmpp. This class has the
summary methods, with the former plotting the distribution of individual-specific effects and the latter allowing to inspect model's coeffcients and fit measures.
To compute predictions with the PMPP model, one needs to construct the forecast frame with
create_fframe(). The forecast frame and the corresponding model object can be passed along to the
predict method to obtain forecasts.
In order to calculate prediction intervals, the
pmpp_predinterval() function can be used. This function, similarly to the
predict method, takes the model object and the forecast frame as inputs. Be warned: bootstrapping of prediction interval might take time!
# Get data data(EmplUK, package = "plm") EmplUK <- dplyr::filter(EmplUK, year %in% c(1978, 1979, 1980, 1981, 1982)) # Run the model predicting employment pmpp_model <- pmpp(dep_var = "emp", data = EmplUK) summary(pmpp_model) # Compute predictions for following three years my_fframe <- create_fframe(EmplUK, 1983:1985) prediction <- predict(pmpp_model, my_fframe) # Compute prediction intervals intervals <- pmpp_predinterval(pmpp_model, my_fframe, bootReps = 20, confidence = 0.95)
Functions in pmpp
|kde2D||Compute a two-dimensional kernel density estimate|
|get_kernel||Obtain 2D kernel density estimates given sufficient statistics for lambdas and the initial data Y0|
|pmpp||Posterior Mean Panel Predictor for dynamic panel modelling|
|get_lambda0||Produce sufficient statistics (lambda0) given the common coefficients (rho0)|
|pmpp_data||Transform a single variable in the matrix format into the long panel format|
|loglikelihood_QMLE||Produce (negative) log marginal likelihood for QMLE with correlated random coefficients|
|plot.pmpp||Plot method for objects of class pmpp.|
|predict.pmpp||Compute forecasts with a PMPP model|
|create_fframe||Add empty rows with time stamps to each cress-sectional unit in the panel|
|get_sigma2||Produce variance of the shocks estimated using GMM residues (sigma2_0) given the common coefficients (rho0)|
|ssys_gmm||Suboptimal multi-step System-GMM estimator for AR(1) panel data model|
|kde||One-dimensional kernel density estimate|
|pmpp_predinterval||Random-Window Block Bootstrap for prediction intervals for PMPP model|
|GMM_parametric||Produce posterior means of lambda's for the parametric GMM implementation given autoregressive coefficient (rho)|
|post_mean_lambda_par||Provide posterior means of lambda_i's based on the Parametric Posterior Mean estimator with correlated random coefficients|
|summary.pmpp||Summary method for objects of class pmpp.|
|loglikelihood_GMM||Produce negative log-likelihood in the GMM case|
Last month downloads
|License||GPL (>= 2)|
|Packaged||2018-10-01 21:09:02 UTC; Asus|
|Date/Publication||2018-10-12 13:50:09 UTC|
Include our badge in your README