Merge pull request #19 from UChicago-pol-methods/vignettes

Update estimates graphs layout. Add brief description to data[[1]]

Author: mollyow

DESCRIPTION

@@ -2,10 +2,11 @@ Package: banditsCI
 Title: Bandit-Based Experiments and Policy Evaluation
 Version: 1.0.0
 Authors@R: c(
-    person("Molly", "Offer-Westort", email = "mollyow@uchicago.edu", role = c("aut", "cre", "cph")),
-    person("Yinghui", "Zhou", email = "yz4690@columbia.edu", role = c("aut")))
+    person("Molly", "Offer-Westort", email = "mollyow@gmail.com", role = c("aut", "cre", "cph"), comment = c(ORCID = "0000-0003-2787-9919")),
+    person("Yinghui", "Zhou", email = "yz4690@columbia.edu", role = c("aut")),
+    person("Zhan", "Ruohan", email = "rhzhan@ust.hk", role = c("aut"), comment = c(ORCID = "0000-0002-3426-2784")))
 Description: Provides functions for conducting frequentist inference on adaptively generated data. The methods implemented are based on Zhan et al. (2021) <arXiv:2106.02029> and Hadad et al. (2021) <arXiv:1911.02768>. For illustration, several functions for simulating non-contextual and contextual adaptive experiments using Thompson sampling are also supplied. 
-License: GPL (>= 3)
+License: GPL (>= 3) + file LICENSE
 Encoding: UTF-8
 RoxygenNote: 7.2.3
 URL: https://github.com/UChicago-pol-methods/banditsCI, https://uchicago-pol-methods.github.io/banditsCI/

vignettes/banditsCI.Rmd

@@ -54,6 +54,15 @@ data from a pilot (this code is not evaluated here).
 # data <- generate_bandit_data(xs = xs, y = y, noise_std = 0.5)
 ```
 
+Components of data[[1]]:
+
+- `ys`: outcomes vector of shape [A];
+
+- `xs`: covariates of shape [A, p]. The value in xs [i, j] represents the j-th 
+  covariate of the i-th observation;
+  
+- `muxs`: true best arm for each context of shape [A, K]. The value in muxs [i, j] represents the predicted outcome or expected reward if the i-th observation is assigned to the j-th treatment arm.
+
 ## For the contextual case. 
 We run a contextual bandit experiment using our `run_experiment()` function. 
 The algorithm used here is a version of linear Thompson sampling. 
@@ -187,7 +196,7 @@ True mean values are represented by the dashed red line.
 ```{r contextual_means_plot, fig.width=7, fig.height=5, out.width = "45%"}
 op <- par()
 
-par(mar = c(5,11,4,2) + 0.1)
+par(mar = c(5,16,4,2) + 0.1)
 
 # set some plotting parameters across plots
 xmin <- min(unlist(lapply(out_full, `[`, TRUE, "estimate")), na.rm = TRUE) - 
@@ -262,7 +271,7 @@ supported in the original source. We can do non-contextual two-point estimation
 when calculating separate adaptive weights for contrasts, however we recommend 
 using other estimation procedures in contextual adaptive settings. 
 ```{r contextual_ATEs_plot1, fig.width=7, fig.height=5, out.width = "45%"}
-par(mar = c(5,11,4,2) + 0.1)
+par(mar = c(5,16,4,2) + 0.1)
 
 # set some plotting parameters across plots
 xmin <- min(unlist(lapply(out_full_te1, `[`, TRUE, "estimate")), na.rm = TRUE) - 
@@ -299,7 +308,7 @@ suppressWarnings(par(op))
 2. In the second approach, we implement adaptive weighting on treatment and control
 scores separately, and then take the difference. 
 ```{r contextual_ATEs_plot2, fig.width=7, fig.height=5, out.width = "45%"}
-par(mar = c(5,11,4,2) + 0.1)
+par(mar = c(5,16,4,2) + 0.1)
 
 # set some plotting parameters across plots
 xmin <- min(unlist(lapply(out_full_te2, `[`, TRUE, "estimate")), na.rm = TRUE) - 
@@ -396,7 +405,7 @@ True mean values are represented by the dashed red line.
 ```{r noncontextual_means_plot, fig.width=7, fig.height=5, out.width = "45%"}
 op <- par()
 
-par(mar = c(5,11,4,2) + 0.1)
+par(mar = c(5,16,4,2) + 0.1)
 
 # set some plotting parameters across plots
 xmin <- min(unlist(lapply(out_full, `[`, TRUE, "estimate"))) - 
@@ -463,7 +472,7 @@ out_full_te2 <- output_estimates(
 1. First take the difference in AIPW scores, and then conduct adaptive 
 weighting.
 ```{r noncontextual_ATEs_plot1, fig.width=7, fig.height=5, out.width = "45%"}
-par(mar = c(5,11,4,2) + 0.1)
+par(mar = c(5,16,4,2) + 0.1)
 
 # set some plotting parameters across plots
 xmin <- min(unlist(lapply(out_full_te1, `[`, TRUE, "estimate")), na.rm = TRUE) - 
@@ -500,7 +509,7 @@ suppressWarnings(par(op))
 2. Or implement adaptive weighting on treatment and control
 scores separately, and then take the difference. 
 ```{r noncontextual_ATEs_plot2, fig.width=7, fig.height=5, out.width = "45%"}
-par(mar = c(5,11,4,2) + 0.1)
+par(mar = c(5,16,4,2) + 0.1)
 
 # set some plotting parameters across plots
 xmin <- min(unlist(lapply(out_full_te2, `[`, TRUE, "estimate")), na.rm = TRUE) -