Code Monkey home page Code Monkey logo

gk2011's Introduction

Gaines and Kuklinski (2011) Estimators for Hybrid Experiments

GK2011 is a package for providing implementations of the treatment effect estimators for hybrid (self-selection) experiments, as developed by Gaines and Kuklinski (2011).

These functions estimate local average treatment effects for unobserved population subgroups inclined and disinclined to be treated, as revealed by a three-condition (two-arm) "hybrid" experimental design. In the design, participants are randomly assigned to one of three conditions: 1) treatment (T), 2) control (C), or 3) self-selection (S) of treatment or control. The design enables the estimation of four treatment effects:

  1. The sample average treatment effect,
  2. The effect for those inclined to choose treatment,
  3. The effect for those disinclined to choose treatment (or, equivalent, to inclined to choose control), and
  4. The naive difference in the outcome between those who chose treatment rather than control.

Code Examples

The only function is estimate(), which provides a simple interface for estimating all effects:

library("GK2011")

# create fake data
set.seed(12345)
d <- 
data.frame(rand = c(rep(1, 200), rep(0, 100)),
           tr = c(rep(0, 100), rep(1, 100), rep(0, 37), rep(1, 63)),
           y = c(rnorm(100), rnorm(100) + 1, rnorm(37), rnorm(63) + 3))

# estimate effects
estimate(rand = d$rand, tr = d$tr, y = d$y)
##   Effect   Estimate        SE         t            p
## 1      t  0.8000359 0.2076146  3.853467 1.570643e-04
## 2    t_s  2.5374464 0.3327255  7.626245 9.735505e-13
## 3    t_n -2.1582576 0.5848236 -3.690442 2.889546e-04
## 4  naive  2.9034906 0.3065426  9.471736 1.554308e-15

The response structure is a data.frame containing the name of each effect, the effect estimate, and bootstrapped standard error.

Gaines and Kuklinski's data

A small example dataset is included that contains treatment and outcome data from Gaines and Kuklinski's (2011) study. The following code replicate the main descriptive results from Table 2 of their paper:

data(ajps)
pmean <- function(x) sprintf("%0.1f", mean(x))
cbind(
  # Democrats
  aggregate(cbind(therm.mccain, therm.obama) ~ tr, data = ajps[ajps$pid == 1, ], FUN = pmean)[, 1:3],
  n_dem = aggregate(therm.obama ~ tr, data = ajps[ajps$pid == 1, ], FUN = length)[, 2],
  # Republicans
  aggregate(cbind(therm.mccain, therm.obama) ~ tr, data = ajps[ajps$pid == -1, ], FUN = pmean)[, 2:3],
  n_rep = aggregate(therm.obama ~ tr, data = ajps[ajps$pid == -1, ], FUN = length)[, 2]
)
##   tr therm.mccain therm.obama n_dem therm.mccain therm.obama n_rep
## 1  1         27.1        80.9    53         79.2        22.4    56
## 2  2         30.6        76.7    65         81.3        28.0    54
## 3  3         27.9        77.9    47         76.0        24.8    34
## 4  4         25.6        82.0    68         81.4        24.6    63

The paper reports four main sets of experimental results, separating the experiment by outcome (feeling thermometers for Obama and McCain) and by respondents' party identification (Democrats and Republicans).

Here are effects for McCain among Democrats:

with(ajps[ajps$pid == 1, ], {
  estimate(rand = tr %in% 1:2, tr = tr %in% c(1,3), y = therm.mccain)
})
##   Effect    Estimate       SE          t         p
## 1      t  -3.5175617 4.735196 -0.7428544 0.4590587
## 2    t_s -10.0114566 9.744498 -1.0273958 0.3056205
## 3    t_n   0.9708657 7.138275  0.1360084 0.8919784
## 4  naive   2.2546934 4.826405  0.4671579 0.6412786

Here are effects for McCain among Republicans:

with(ajps[ajps$pid == -1, ], {
  estimate(rand = tr %in% 1:2, tr = tr %in% c(1,3), y = therm.mccain)
})
##   Effect   Estimate       SE          t         p
## 1      t -2.1177249 3.419383 -0.6193296 0.5369916
## 2    t_s -5.0217865 8.877852 -0.5656533 0.5724746
## 3    t_n -0.5504535 4.673361 -0.1177853 0.9063932
## 4  naive -5.4444444 3.845384 -1.4158389 0.1600589

Here are effects for Obama among Democrats:

with(ajps[ajps$pid == 1, ], {
  estimate(rand = tr %in% 1:2, tr = tr %in% c(1,3), y = therm.obama)
})
##   Effect  Estimate       SE          t         p
## 1      t  4.201451 4.458794  0.9422843 0.3479887
## 2    t_s  8.762684 9.338501  0.9383394 0.3493344
## 3    t_n  1.048835 6.703290  0.1564657 0.8758550
## 4  naive -4.040989 4.675241 -0.8643381 0.3892172

Here are effects for Obama among Republicans:

with(ajps[ajps$pid == -1, ], {
  estimate(rand = tr %in% 1:2, tr = tr %in% c(1,3), y = therm.obama)
})
##   Effect   Estimate        SE           t         p
## 1      t -5.5522487  5.085318 -1.09181929 0.2773200
## 2    t_s -9.4531590 13.110036 -0.72106281 0.4719932
## 3    t_n -3.4469955  7.068427 -0.48766092 0.6264927
## 4  naive  0.1774043  5.689454  0.03118125 0.9751897

Installation

CRAN Build Status codecov.io

This package is not yet on CRAN. To install the latest development version you can pull directly from GitHub:

if(!require("remotes")){
    install.packages("remotes")
}
remotes::install_github("leeper/GK2011")

References

Brian J. Gaines and James H. Kuklinski, (2011), "Experimental Estimation of Heterogeneous Treatment Effects Related to Self-Selection," American Journal of Political Science 55(3): 724-736, doi:10.1111/j.1540-5907.2011.00518.x.

gk2011's People

Contributors

leeper avatar

Stargazers

 avatar  avatar

Watchers

 avatar  avatar

Forkers

shuai114

gk2011's Issues

Standard errors are probably too small

The current implementation doesn't acknowledge that alpha is itself an estimate and therefore deflates SEs. Fixing this might require an internal rewrite to some extent, even though it shouldn't change the interface.

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.