| Title: | Gaussian Process Projection |
| Version: | 0.1 |
| Description: | Estimates a counterfactual using Gaussian process projection. It takes a dataframe, creates missingness in the desired outcome variable and estimates counterfactual values based on all information in the dataframe. The package writes Stan code, checks it for convergence and adds artificial noise to prevent overfitting and returns a plot of actual values and estimated counterfactual values using r-base plot. |
| Depends: | R (≥ 3.5.0), methods, rstan, parallel |
| LazyData: | true |
| Encoding: | UTF-8 |
| License: | GPL-2 | GPL-3 [expanded from: GPL (≥ 2)] |
| RoxygenNote: | 7.1.1 |
| NeedsCompilation: | no |
| Packaged: | 2020-11-25 14:58:45 UTC; david |
| Author: | Devin P. Brown [aut], David Carlson [aut, cre] |
| Maintainer: | David Carlson <carlson.david@wustl.edu> |
| Repository: | CRAN |
| Date/Publication: | 2020-11-27 10:20:06 UTC |
1960-2003 GDP dataset
Description
An example dataset for using GPP to estimate the counterfactual GDP of West Germany assuming no reunification.
Usage
GDPdata
Format
A data frame with 748 rows and 14 columns. For detailed explanations of the exact measures, see https://www.dropbox.com/s/n1bvqb54xrw8vyj/GPSynth.pdf?dl=0:
- index
- country
- year
- gdp
- infrate
- trade
- schooling
- invest60
- invest70
- invest80
- industry
- invest
- school
- ind
See Also
GPP plotGPPfit writeMod runMod autoConverge
Estimates a counterfactual with uncertainty using Gaussian process projection
Description
Returns a list of a plot object (after making the plot) of estimated counterfactual values after checking for model convergence and adjusting the noise level, and returns the fitted model.
Usage
GPP(
df,
controlVars,
nUntreated,
obvColName,
obvName,
outcomeName,
starttime,
timeColName,
ncores = NULL,
epsilon = 0.02,
noise = 0.1,
printMod = FALSE,
shift = 0.05,
iter = 25000,
filepath = NULL,
legendLoc = "topleft",
xlabel = NULL,
ylabel = NULL,
actualdatacol = "black",
preddatacol = "red",
...
)
Arguments
df |
The dataframe used for the model. |
controlVars |
String of column names for control variables. |
nUntreated |
The number of untreated units in the model. |
obvColName |
The column name that includes the observation subject to the counterfactual. |
obvName |
The name of the observation subject to the counterfactual. |
outcomeName |
The outcome variable of interest. |
starttime |
The start year of the counterfactual estimation. |
timeColName |
The name of the column that includes the time variable. |
ncores |
The number of cores to be used to run the model. See details. |
epsilon |
The desired level of convergence. |
noise |
The baseline level of noise to be added to the model to prevent overfit. Updates as the model runs. |
printMod |
Boolean. Defaults FALSE. If TRUE, prints each model block to the console. See details. |
shift |
The magnitude of adjustment for the noise level per iteration. Defaults to 0.05. |
iter |
The number of iterations you would like to run. Defaults to 25,000. See details. |
filepath |
Your preferred place to save the fit data. See Details. |
legendLoc |
The preferred location of the legend in the final graph. Defaults to "topleft". |
xlabel |
The label of the x-axis in the final graph. Defaults to input for 'timeColName'. |
ylabel |
The preferred label of the y-axis in the final graph. Defaults to input for 'outcomeName'. |
actualdatacol |
The preferred color for plotted line for actual data. Defaults to black. |
preddatacol |
The preferred color for plotted line for predicted counterfactual data. Defaults to red. |
... |
Further parameters passed to the plot function. |
Details
We recommend using all cores on your machine to speed up model run time. If you are unsure about the number of cores in your machine, see parallel::detectCores().
We recommend keeping printMod as FALSE, otherwise, the function will write the model to the console for every model run on the convergence.
For iterations, check that your model converged (we recommend all r-hats close to 1 and examining traceplots).
We recommend creating a new folder for the file path since the Stan fit creates a large number of files at runtime.
Value
A plot of the actual values and the estimated counterfactual values of the model, and the final model fit.
Author(s)
Devin P. Brown devinpbrown96@gmail.com and David Carlson carlson.david@wustl.edu
See Also
plotGPPfit writeMod runMod autoConverge
Examples
data(GDPdata)
out = GPP(df = GDPdata,
controlVars = c('invest', 'school', 'ind'),
nUntreated = length(unique(GDPdata$country))-1,
obvColName = 'country', obvName = 'West Germany',
outcomeName = 'gdp', starttime = 1989,
timeColName = 'year',
ncores = 2)
Checks Stan model for convergence, then runs model on actual data.
Description
Return a converged Stan model fit and the recommended noise level.
Usage
autoConverge(
df,
controlVars,
nUntreated,
obvColName,
obvName,
outcomeName,
starttime,
timeColName,
filepath = NULL,
ncores = NULL,
iter = 25000,
epsilon = 0.02,
noise = 0.1,
printMod = FALSE,
shift = 0.05
)
Arguments
df |
The dataframe used for the model. |
controlVars |
String of column names for control variables. |
nUntreated |
The number of untreated units in the model. |
obvColName |
The column name that includes the observation subject to the counterfactual. |
obvName |
The name of the observation subject to the counterfactual. |
outcomeName |
The outcome variable of interest. |
starttime |
The start time of the counterfactual estimation. |
timeColName |
The name of the column that includes the time variable. |
filepath |
Your preferred place to save the fit data. See Details. |
ncores |
The number of cores to be used to run the model. Default of NULL will utilize all cores. |
iter |
Preferred number of iterations. See details. |
epsilon |
The desired level of convergence, i.e. how close to the 0.95 coverage is acceptable. |
noise |
The baseline level of noise to be added to the model to prevent overfit. Updates as the model runs. |
printMod |
Boolean. Defaults FALSE. If TRUE, prints the model block for the run to the console. See details. |
shift |
The magnitude of adjustment for the noise level per iteration. Defaults to 0.05. |
Details
We recommend creating a new folder for the file path since the Stan fit creates a large number of files at runtime.
For iterations, check that your model converged (we recommend all r-hats close to 1 and examining traceplots).
We recommend keeping printMod as FALSE, otherwise, the function will write the model to the console for every model run on the convergence.
We also recommend using all cores on your machine to speed up model run time. If you are unsure about the number of cores in your machine, see doParallel::detectCores().
Value
The recommended noise level after convergence.
Author(s)
Devin P. Brown devinpbrown96@gmail.com and David Carlson carlson.david@wustl.edu
See Also
plotGPPfit runMod GPP writeMod
Plots results of a (converged) model, with true and projected values.
Description
Takes the results of a Gaussian Process Projection fit and generates a linear plot of the actual and predicted counterfactual values
Usage
plotGPPfit(
fit,
df,
obvColName,
obvName,
outcomeName,
starttime,
timeColName,
legendLoc = "topleft",
xlabel = NULL,
ylabel = NULL,
actualdatacol = "black",
preddatacol = "red",
...
)
Arguments
fit |
The fit results of the GPP stan model. |
df |
The dataframe used in your model. |
obvColName |
The column name that includes your observation of interest. Must be a string. |
obvName |
The name of the specific observation of interest. Must be a string. |
outcomeName |
The explanatory variable that is subjected to the counterfactual claim. |
starttime |
The start time of the treatment effect. |
timeColName |
The name of the column that includes your time variable. |
legendLoc |
The preferred location of the legend in the final graph. Defaults to "topleft". |
xlabel |
The label of the x-axis in the final graph. Defaults to input for 'timeColName'. |
ylabel |
The preferred label of the y-axis in the final graph. Defaults to input for 'outcomeName'. |
actualdatacol |
The preferred color for plotted line for actual data. Defaults to black. |
preddatacol |
The preferred color for plotted line for predicted counterfactual data. Defaults to red. |
... |
Further graphical parameters. |
Value
A plot built in r-base
Author(s)
Devin P. Brown devinpbrown96@gmail.com and David Carlson carlson.david@wustl.edu
See Also
autoConverge GPP runMod writeMod
Runs the model, given the data and treated case (may be a placebo).
Description
Returns a fit of the Stan model for all observations.
Usage
runMod(modText, dataBloc, unit, iter = 25000, filepath = NULL)
Arguments
modText |
This is the string that contains your Stan code. Can be written with |
dataBloc |
This is the data that you pass to the Stan code. It is automatically generated when you run |
unit |
The unit of observation to project. |
iter |
The number of iterations you would like to run. Defaults to 25,000. |
filepath |
Your preferred place to save the fit data. See Details. |
Details
For iterations, check that your model converged (we recommend all r-hats close to 1 and examining traceplots).
We recommend creating a new folder for the file path since the Stan fit creates a large number of files at runtime.
Value
The fit for the GPP counterfactual Stan model.
Author(s)
Devin P. Brown devinpbrown96@gmail.com and David Carlson carlson.david@wustl.edu
See Also
plotGPPfit writeMod GPP autoConverge
Writes Stan code for GPP model
Description
Returns string of Stan code that can be run to estimate the GPP.
Usage
writeMod(noise, ncov, printMod = FALSE)
Arguments
noise |
The desired amount of artificial noise to add to the model. |
ncov |
The number of covariates to include in the model. |
printMod |
Boolean. Defaults FALSE. If TRUE, prints each model block to the console. See details. |
Details
We recommend keeping printMod as FALSE, otherwise, the function will write the model to the console for every model run on the convergence.
Value
A string of Stan code that can be run with runMod
Author(s)
Devin P. Brown devinpbrown96@gmail.com and David Carlson carlson.david@wustl.edu
See Also
plotGPPfit runMod GPP autoConverge
Examples
writeMod(noise = 0.25, ncov = 2)