1/replication_package/scripts/hbg_analysis.R

### Initialize workspace.
## Clear workspace.
rm(list = ls(all = TRUE))

## Confirm working directory.
setwd("~/Downloads/hbg_replication")

## Set seed.
set.seed(123)

## Set number of iterations for bootstrap replication.
n_iter <- 10000

## Load relevant packages.
library(sandwich)
library(car)

## Load relevant helper functions.
source("scripts/helper_functions.R")

## Load data.
# Load experimental data.
tpnw <- read.csv("data/tpnw_data.csv", row.names = 1, 
				 stringsAsFactors = FALSE)

# Load YouGov data.
aware <- read.csv("data/tpnw_aware.csv", row.names = 1,
				  stringsAsFactors = FALSE)

### Define relevant objects.
## Define objects specifying outcomes.
# Specify join_tpnw object, representing main outcome.
join_tpnw <- "join_tpnw"

# Specify tpnw_atts object, representing attitudinal outcomes.
tpnw_atts <- names(tpnw)[startsWith(names(tpnw), "tpnw_atts")]

# Specify all_outs object, concatenating main and attitudinal outcomes.
all_outs <- c(join_tpnw, tpnw_atts)

## Define objects specifying predictors.
# Define object specifying main treatments.
treats <- c("group_cue", "security_cue", "norms_cue", "institutions_cue")

# Define object specifying general demographics.
demos <- c("age", "female", "midwest", "west", "south", "income", "educ")

# Define object specifying politically relevant demographics.
pol_demos <- c("ideo", "pid3")

# Define list of conditioning sets (NULL corresponds to Model 1, whereas the use
# of demographic and political covariates corresponds to Model 2).
covars <- list(NULL, c(demos, pol_demos))

### Produce analysis.
## Produce balance table.
# Specify covariates to be used for balance table.
bal_covars <- c("age", "female", "northeast", "midwest", "west", 
				"south", "income", "educ", "ideo", "pid3")

# Produce balance table matrix output, looping over treatment group.
bal_mat <- lapply(0:4, function (i) {
				# For each treatment value ...
		   		apply(tpnw[bal_covars][tpnw$treatment == i,], 2, function (x) {
		   			
		   			# Calculate the mean of each covariate.
		   			mean_x <- mean(x)

		   			# Calculate SE estimates using 10,000 bootstrap replicates.
		   			sd_x <- sd(replicate(10000, {
		   						samp <- x[sample(length(x), replace = TRUE)]
		   						return(mean(samp))
		   				}))

		   			# Return a list containing both point estimates.
		   			return(list(mean = mean_x, sd = sd_x))
		   			})
		   		})

# Bind point estimates for each treatment group.
bal_mat <- lapply(bal_mat, function (treat) {
				  do.call("rbind", unlist(treat, recursive = FALSE))
		   })

# Convert list into a matrix, with columns representing treatment group.
bal_mat <- do.call("cbind", bal_mat)

# Round all estimates to within three decimal points and convert to character
# for the purposes of producing tabular output.
bal_tab <- apply(bal_mat, 2, function (x) format(round(x, 3), digits = 3))

# Specify rows containing mean point estimates.
mean_rows <- endsWith(rownames(bal_tab), ".mean")

# Specify rows containing SE point estimates.
se_rows <- endsWith(rownames(bal_tab), ".sd")

# Reformat SE estimates to be within parentheses.
bal_tab[se_rows,] <- apply(bal_tab[se_rows,], 2, function (x) {
						   paste0("(", x, ")")
					 })

# Remove row names for rows with SE estimates.
rownames(bal_tab)[se_rows] <- ""

# Remove ".mean" string in row names for rows with mean estimates.
rownames(bal_tab)[mean_rows] <- gsub(".mean", "", rownames(bal_tab)[mean_rows])

# Concatenate data to comport with LaTeX tabular markup.
bal_tab <- paste(paste(paste(
		capwords(rownames(bal_tab)), apply(bal_tab, 1, function (x) {
										   paste(x, collapse = " & ")
									 }), 
	   sep = " & "), collapse = " \\\\\n"), "\\\\\n")
bal_tab <- gsub("\\( ", "\\(", bal_tab)

# Produce tabular output.
sink("output/balance_tab.tex")
cat("\\begin{table}\n",
	  "\\caption{Covariate Balance Across Treatment Arms}\n",
	  "\\centering\\small\n",
	  "\\sisetup{\n",
      "\tdetect-all,\n",
      "\ttable-number-alignment = center,\n",
      "\ttable-figures-integer = 1,\n",
      "\ttable-figures-decimal = 3,\n",
      "\tinput-symbols = {()}\n",
  	  "}\n",
	  paste0("\\begin{tabular}{@{\\extracolsep{5pt}}L{2.75cm}*{5}",
	  		 "{S[table-number-alignment = center, table-column-width = 1.75cm]}}\n"),
	  "\\toprule\n",
	  "& \\multicolumn{5}{c}{Arm}\\\\\\cmidrule{2-6}\n",
	  "& {Control} & {Group} & {Security} & {Norms} & {Institutions} \\\\\\midrule\n",
	  bal_tab,
	  "\\bottomrule\n",
	  "\\end{tabular}\n",
	  "\\end{table}\n")
sink()

## Produce main results.
# Compute main results, looping over conditioning sets.
main_results <- lapply(covars, function (covar) {
					# For each conditioning set ...
					# Specify the relevant regression formula.
					form <- as.formula(paste(join_tpnw, paste(c(treats, covar), 
									   collapse = " + "), sep = " ~ "))

					# Fit the OLS model per the specification.
					fit <- lm(form, data = tpnw)

					# Compute HC2 robust standard errors.
					ses <- sqrt(diag(vcovHC(fit, type = "HC2")))

					# Bind coefficient and SE output.
					reg_out <- cbind(fit$coef[2:5], ses[2:5])

					# Name output matrix columns and rows.
					colnames(reg_out) <- c("coef", "se")
					rownames(reg_out) <- treats

					# Return output
					return(as.data.frame(reg_out))
				})

# Name results to distinguish between Model 1 and Model 2 estimates.
names(main_results) <- c("model_1", "model_2")

## Assess significance of effect estimates and differences.
# Estimate Bonferroni-Holm-adjusted p-values.
bf_ps <- lapply(main_results, function (x) {
				round(p.adjust(pnorm(x[, 1] / x[, 2], lower.tail = TRUE), 
							   method = "holm"), 3)
		 })

# Estimate FDR-adjusted p-values, as an added robustness check.
fdr_ps <- lapply(main_results, function (x) {
				 round(p.adjust(pnorm(x[, 1] / x[, 2], lower.tail = TRUE), 
				 	   method = "fdr"), 3)
		  })

# Redefine the main model (Model 2), and store full VCOV matrix.
main_model <- lm(join_tpnw ~ group_cue + security_cue + norms_cue + 
				 institutions_cue + age + female + midwest + 
				 west + south + income + educ + ideo + pid3, tpnw)
main_vcov <- vcovHC(main_model, "HC2")

# Specify diff_sig function for assessing significance between two effect
# estimates (defined here for the sake of clarity).
diff_sig <- function (eff_1, eff_2) {
				diff <- main_model$coef[eff_1] - main_model$coef[eff_2]
				se <- sqrt(main_vcov[eff_1, eff_1] + main_vcov[eff_2, eff_2] -
						   2 * main_vcov[eff_1, eff_2])
				p <- 2 * (1 - pnorm(abs(diff) / se))
				return (p)
}

# Assess the significance of the difference between institution and security cue
# effect estimates .
inst_sec_diff_p <- diff_sig("institutions_cue", "security_cue")

# Assess the significance of the difference between institution and group cue
# effect estimates
inst_grp_diff_p <- diff_sig("institutions_cue", "group_cue")

# Assess the significance of the difference between security and group cue
# effect estimates
sec_grp_diff_p <- diff_sig("security_cue", "group_cue")

# Assess the significance of the difference between security and norms cue
# effect estimates
sec_norms_diff_p <- diff_sig("security_cue", "norms_cue")

# Assess the significance of the difference between institution and group cue
# effect estimates
inst_norms_diff_p <- diff_sig("institutions_cue", "norms_cue")

# Assess the significance of the difference between institution and group cue
# effect estimates
grp_norms_diff_p <- diff_sig("group_cue", "norms_cue")

# The significance of differences between effect estimates was also assessed
# using 10,000 bootstrap replicates and two-tailed p-values; relevant code is 
# included below with the institutions and security cues, for posterity, but is 
# not run.

# Compute SE estimates.
# diffs <- replicate(10000, {
# 	samp <- tpnw[sample(nrow(tpnw), replace = TRUE),]
# 	model <- lm(join_tpnw ~ group_cue + security_cue + norms_cue + 
# 				institutions_cue + age + female + midwest + 
# 				west + south + income + educ + ideo + pid3, samp)
# 	model$coef[5] - model$coef[3]
# })
# diffs_se <- sd(diffs)
# 
# # Fit model.
# model <- lm(join_tpnw ~ group_cue + security_cue + norms_cue + 
# 				institutions_cue + age + female + midwest + 
# 				west + south + income + educ + ideo + pid3, tpnw)
# 
# # Compute two-tailed p-value.
# 2 * (1 - pnorm(abs((model$coef[5] - model$coef[3])/diffs_se)))

## Assess YouGov results.
# Tabulate responses.
aware_table <- table(aware$awareness, useNA = "ifany")
names(aware_table) <-  c("Yes, support", "Yes, oppose", 
						 "No, support", "No, oppose", "Skipped")

# Compute both weighted and unweighted means.
aware_results <- lapply(1:4, function (resp) {
				 	# Calculate weighted mean.
				 	wt_mean <- with(aware, weighted.mean(awareness == resp, 
				 										 w = weight, na.rm = TRUE))
				 	
				 	# Calculate raw mean.
				 	rw_mean <- with(aware, mean(awareness == resp, na.rm = TRUE))
				 	
				 	# Concatenate means and rename vector.
				 	means <- c(wt_mean, rw_mean)
				 	names(means) <- c("weighted_mean", "raw_mean")

				 	# Calculate SE estimates with 10,000 bootstrap replicates.
				 	ses <- replicate(10000, {
		   					      samp <- aware[sample(nrow(aware), 
		   					      				replace = TRUE),]
		   					      wt_mean <- with(samp, weighted.mean(awareness == resp, 
		   					      									  w = weight, na.rm = TRUE))
		   					      rw_mean <- with(samp, mean(awareness == resp, 
		   					      							 na.rm = TRUE))
		   					      return(c(wt_mean, rw_mean))
		   				   })
					ses <- apply(ses, 1, sd)
					names(ses) <- c("weighted_mean", "raw_mean")
		   			
		   			# Bind mean and SE estimates.
		   			outs <- rbind(means, ses)
		   			rownames(outs) <- paste(names(aware_table)[resp], 
		   									c("mean", "se"), sep = "_")
		   			return(outs)
				 })

# Name results to distinguish between responses.
names(aware_results) <- c("Yes, support", "Yes, oppose", 
						  "No, support", "No, oppose")

## Assess covariate means for experimental and YouGov data (used in Table A1).
# Indicate the list of covariates to be assessed.
demo_tab_vars <- c("age", "female", "northeast", "midwest", "west", "south")

# Compute covariate averages for experimental data.
tpnw_means <- apply(tpnw[demo_tab_vars], 2, mean, na.rm = TRUE)

# Compute covariate averages for YouGov data.
aware_means <- apply(aware[demo_tab_vars], 2, function (x) {
					 	weighted.mean(x, na.rm = TRUE, w = aware$weight)
					})

# Compute bootstrap standard errors for demographic means.
demo_ses <- replicate(10000, {
	# Sample the experimental data.
	samp_tpnw <- tpnw[sample(nrow(tpnw), replace = TRUE), demo_tab_vars]
	
	# Sample the YouGov data.
	samp_aware <- aware[sample(nrow(aware), replace = TRUE), 
						c(demo_tab_vars, "weight")]
	
	# Compute bootstrap means for experimental data.
	tpnw_means <- apply(samp_tpnw[demo_tab_vars], 2, mean, na.rm = TRUE)

	# Compute bootstrap means for YouGov data.
	aware_means <- apply(samp_aware[demo_tab_vars], 2, function (x) {
					 	weighted.mean(x, na.rm = TRUE, w = samp_aware$weight)
					})

	# Return the results as a list, and ensure that replicate() also returns a 
	# list.
	return(list(tpnw = tpnw_means, aware = aware_means))
	}, simplify = FALSE)

# Compute SE estimates for each set of demographics.
demo_ses <- lapply(c("tpnw", "aware"), function (dataset) {
	# Group all estimates from each dataset.
	sep_res <- lapply(demo_ses, function (iteration) {
					  	return(iteration[[dataset]])
			   })

	# Bind estimates.
	sep_res <- do.call("rbind", sep_res)

	# Compute SE estimates.
	sep_ses <- apply(sep_res, 2, sd)

	# Return SE estimates.
	return(sep_ses)
	})

## Assess responses to the attitudinal battery.
# Assess responses to the attitudinal battery, looping over treatment group. For
# each treatment value ...
att_results <- lapply(0:4, function (i) {
	# Calculate the average response to each attitudinal battery question.
	atts_mean <- apply(tpnw[tpnw$treatment == i, tpnw_atts], 2, function (x) {
					   mean(x, na.rm = TRUE)
				 })

	# Calculate SE estimates using 10,000 bootstrap replicates.
	bl_atts_boot <- replicate(10000, {
		dat <- tpnw[tpnw$treatment == i, tpnw_atts]
		samp <- dat[sample(nrow(dat), replace = TRUE),]
		apply(samp, 2, function (x) mean(x, na.rm = TRUE))
	})
	bl_atts_ses <- apply(bl_atts_boot, 1, sd)
	
	# Combine mean and SE estimates and return results.
	return(cbind(atts_mean, bl_atts_ses))
})

# Compute treatment effects on responses to the attitudinal battery, looping 
# over conditioning sets.
att_effs <- lapply(covars, function (covar) {
			# For each conditioning set ...
			model_res <- lapply(tpnw_atts, function (out) {
								# Specify the relevant regression formula.
								form <- as.formula(paste(out, 
												   paste(c(treats, covar), 
												   collapse = " + "), 
												   sep = " ~ "))
								
								# Fit the OLS model per the specification.
								fit <- lm(form, data = tpnw)

								# Compute HC2 robust standard errors.
								ses <- sqrt(diag(vcovHC(fit, type = "HC2")))
								
								# Bind coefficient and SE output.
								reg_out <- cbind(fit$coef[2:5], ses[2:5])
								
								# Name output matrix columns and rows.
								colnames(reg_out) <- c("coef", "se")
								rownames(reg_out) <- treats

								# Return output.
								return(as.data.frame(reg_out))
						 })
			# Name results to distinguish between each attitudinal battery 
			# outcome and return results.
			names(model_res) <- tpnw_atts
			return(model_res)
	})

# Name results to distinguish between Model 1 and Model 2 estimates.
names(att_effs) <- c("model_1", "model_2")

## Perform subgroup analysis.
# Compute mean support by political party, looping over treatment group.
pid_results <- lapply(0:4, function (treat) {
					  # For each partisan group ...
					  out <- lapply(-1:1, function (i) {
					      			# Calculate average support.
					      			pid_mean <- with(tpnw, 
					      							 mean(join_tpnw[pid3 == i & 
					      							 	  treatment == treat], 
					      							 	  na.rm = TRUE))

					      			# Calculate SE estimates with 10,000 
					      			# bootstrap replicates.
					      			pid_boot <- replicate(10000, {
					      								 dat <- tpnw$join_tpnw[tpnw$pid3 == i & 
					      							   			tpnw$treatment == treat]
					      								 samp <- dat[sample(length(dat), 
					      											 		replace = TRUE)]
					      								 mean(samp, na.rm = TRUE)
					  		 					})

					      			# Concatenate and return mean and SE 
					      			# estimates.
					  	  	 		return(c(mean = pid_mean, se = sd(pid_boot)))
			   		 })

					 # Name results to distinguish estimates by political party,
					 # and return output.
					 names(out) <- c("dem", "ind", "rep")
					 return(as.data.frame(out))
			   })

# Name results to distinguish between treatment groups.
names(pid_results) <- c("Control", paste(c("Group", "Security", "Norms", 
										   "Institutions"), "Cue"))

# Assess significance between control-group means; for 10,000 bootstrap
# replicates ...
pid_diff_ses <- replicate(10000, {
						  # Sample with replacement.
		  				  samp <- tpnw[sample(nrow(tpnw), replace = TRUE),]
		  				  
		  				  # Compute the difference between Democrats' and
		  				  # Independents' support.
		  				  dem_ind_diff <- with(samp[samp$treatment == 0,], 
						  			 		   mean(join_tpnw[pid3 == -1], 
						  			 		   		na.rm = TRUE) - 
						  			 		   mean(join_tpnw[pid3 == 0], 
						  			 		   		na.rm = TRUE))
						  # Compute the difference between Democrats' and
		  				  # Republicans' support.
		  				  dem_rep_diff <- with(samp[samp$treatment == 0,], 
						  			 		   mean(join_tpnw[pid3 == -1], 
						  			 		   		na.rm = TRUE) - 
						  			 		   mean(join_tpnw[pid3 == 1], 
						  			 		   		na.rm = TRUE))
						  # Compute the difference between Independents' and
		  				  # Republicans' support.
		  				  ind_rep_diff <- with(samp[samp$treatment == 0,], 
						  			 		   mean(join_tpnw[pid3 == 1], 
						  			 		   		na.rm = TRUE) - 
						  			 		   mean(join_tpnw[pid3 == 0], 
						  			 		   		na.rm = TRUE))

		  				  # Concatenate and name results.
		  				  out <- c(dem_ind_diff, dem_rep_diff, ind_rep_diff)
		  				  names(out) <- c("dem_ind", "dem_rep", "ind_rep")
		  				  return(out)
				})

# Compute SE estimates for each difference.
pid_diff_ses <- apply(pid_diff_ses, 1, sd)

# Assess significance for each difference.
dem_ind_p <- 2 * (1 - pnorm(abs(pid_results$Control["mean", "dem"] - 
				pid_results$Control["mean", "ind"]) / pid_diff_ses["dem_ind"]))
dem_rep_p <- 2 * (1 - pnorm(abs(pid_results$Control["mean", "dem"] - 
				pid_results$Control["mean", "rep"]) / pid_diff_ses["dem_rep"]))
ind_rep_p <- 2 * (1 - pnorm(abs(pid_results$Control["mean", "ind"] - 
				pid_results$Control["mean", "rep"]) / pid_diff_ses["ind_rep"]))

# Compute mean support by political ideology, looping over treatment group.
tpnw$ideo <- recode(tpnw$ideo, "c(-2, -1) = 'liberal'; 
								0 = 'moderate'; 
								c(1, 2) = 'conservative'")
ideo_results <- lapply(0:4, function (treat) {
					  # For each ideological group ...
					  out <- lapply(c("liberal", "moderate", "conservative"), function (i) {
					      			# Calculate average support.
					      			pid_mean <- with(tpnw, 
					      							 mean(join_tpnw[ideo == i & 
					      							 	  treatment == treat], 
					      							 	  na.rm = TRUE))

					      			# Calculate SE estimates with 10,000 
					      			# bootstrap replicates.
					      			pid_boot <- replicate(10000, {
					      								 dat <- tpnw$join_tpnw[tpnw$ideo == i & 
					      							   			tpnw$treatment == treat]
					      								 samp <- dat[sample(length(dat), 
					      											 		replace = TRUE)]
					      								 mean(samp, na.rm = TRUE)
					  		 					})

					      			# Concatenate and return mean and SE 
					      			# estimates.
					  	  	 		return(c(mean = pid_mean, se = sd(pid_boot)))
			   		 })

					 # Name results to distinguish estimates by political ideology,
					 # and return output.
					 names(out) <- c("liberal", "moderate", "conservative")
					 return(as.data.frame(out))
			   })

# Name results to distinguish between treatment groups.
names(ideo_results) <- c("Control", paste(c("Group", "Security", "Norms", 
										   "Institutions"), "Cue"))

## Produce weighted main results.
# Compute weighted main results, looping over conditioning sets.
w_main_results <- lapply(covars, function (covar) {
				  	# For each conditioning set ...
				  	# Specify the relevant regression formula.
				  	form <- as.formula(paste(join_tpnw, paste(c(treats, covar), 
				  					   collapse = " + "), sep = " ~ "))

					# Fit the OLS model per the specification.
					fit <- lm(form, data = tpnw, weights = anesrake_weight)

					# Compute HC2 robust standard errors.
					ses <- sqrt(diag(vcovHC(fit, type = "HC2")))

					# Bind coefficient and SE output.
					reg_out <- cbind(fit$coef[2:5], ses[2:5])

					# Name output matrix columns and rows.
					colnames(reg_out) <- c("coef", "se")
					rownames(reg_out) <- treats

					# Return output
					return(as.data.frame(reg_out))
				})

# Name results to distinguish between Model 1 and Model 2 estimates.
names(w_main_results) <- c("model_1", "model_2")

### Produce plots and tables.
## Produce main results plot.
# Produce main results matrix for plotting.
main_mat <- do.call("rbind", lapply(1:2, function (model) {
									cbind(main_results[[model]], model)
							 }))

# Store values for constructing 90- and 95-percent CIs.
z_90 <- qnorm(.95)
z_95 <- qnorm(.975)

# Open new pdf device.
setEPS()
postscript("output/fg1.eps", width = 8, height = 5.5)

# Define custom graphical parameters.
par(mar = c(8, 7, 2, 2))

# Open new, empty plot.
plot(0, type = "n", axes = FALSE, ann = FALSE, 
	 xlim = c(-.3, .05), ylim = c(.8, 4))

# Produce guidelines to go behind point estimates and error bars.
abline(v = seq(-.3, .05, .05)[-7], col = "lightgrey", lty = 3)

# Add Model 1 point estimates.
par(new = TRUE)
plot(x = main_mat$coef[main_mat$model == 1], y = 1:4 + .05, 
	 xlim = c(-.3, .05), ylim = c(.8, 4), pch = 16, col = "steelblue2", 
	 xlab = "", ylab = "", axes = FALSE)

# Add Model 2 point estimates.
par(new = TRUE)
plot(x = main_mat$coef[main_mat$model == 2], y = 1:4 - .05, 
	 xlim = c(-.3, .05), ylim = c(.8, 4), pch = 16, col = "#FF8F37", main = "", 
	 xlab = "", ylab = "", axes = FALSE)

# Add horizontal axis indicating effect estimate size.
axis(side = 1, at = round(seq(-.3, 0, .05), 2), labels = FALSE)
mtext(side = 1, at = seq(-.3, .1, .1), text = c("-30", "-20", "-10", "0"), 
	  cex = .9, line = .75)
axis(side = 1, at = round(seq(-.25, .05, .05), 2), tck = -.01, labels = FALSE)

# Add vertical axis specifying treatment names corresponding to point estimates.
axis(side = 2, at = 1:4, labels = FALSE)
mtext(side = 2, line = .75, at = 1:4, 
	  text = paste(c("Group", "Security", "Norms", "Institutions"), "Cue"), 
	  las = 1, padj = .35, cex = .9)

# Add axis labels.
mtext(side = 2, line = 2.3, at = 4.2, text = "Treatment", 
	  font = 2, las = 1, xpd = TRUE)
mtext(side = 1, text = "Estimated Effect Size", line = 2.5, at = -.15, font = 2)

# Add a dashed line at zero.
abline(v = 0.00, lty = 2)

# Add two-sided, 90-percent CIs.
with(main_mat[main_mat$model == 1,], 
	 segments(x0 = coef - z_90 * se, y0 = 1:4 + .05, x1 = coef + z_90 * se, 
	 	      y1 = 1:4 + .05, col = "steelblue2", lwd = 3))
with(main_mat[main_mat$model == 2,], 
	 segments(x0 = coef - z_90 * se, y0 = 1:4 - .05, x1 = coef + z_90 * se, 
	 		  y1 = 1:4 - .05, col = "#FF8F37", lwd = 3))

# Add two-sided 95-percent CIs.
with(main_mat[main_mat$model == 1,], 
	 segments(x0 = coef - z_95 *se, y0 = 1:4 + .05, x1 = coef + z_95 *se, 
	 		  y1 = 1:4 + .05, col = "steelblue2", lwd = 1))
with(main_mat[main_mat$model == 2,], 
	 segments(x0 = coef - z_95 *se, y0 = 1:4 - .05, x1 = coef + z_95 *se, 
	 		  y1 = 1:4 - .05, col = "#FF8F37", lwd = 1))

# Add legend.
legend(legend = paste("Model", 1:2), x = -.15, y = -.275, horiz = TRUE, 
	   pch = 16, col = c("steelblue2", "#FF8F37"), xjust = .5, xpd = TRUE, 
	   text.width = .05, cex = .9)

# Draw a box around the plot.
box()

# Close the grpahical device.
dev.off()

## Create tabular output for main results.
# Define matrix object of main results.
tab_dat <- do.call("cbind", main_results)

# Compute control-group means, with SE estimates; define OLS formula.
ctrl_form <- as.formula(paste(join_tpnw, paste(treats, 
							  collapse = " + "), sep = " ~ "))

# Fit the OLS model per the specification and recover the control mean.
ctrl_fit <- lm(ctrl_form, data = tpnw)

# Recover the control-group mean.
ctrl_mean <- ctrl_fit$coef["(Intercept)"]

# Compute control SE.
ctrl_se <- sqrt(diag(vcovHC(ctrl_fit, "HC2")))["(Intercept)"]

# Concatenate mean and SE output with blank values for Model 2.
ctrl_results <- c(format(round(c(ctrl_mean, ctrl_se), 3) * 100, digits = 2), 
			  	  "|", "|")

# Reformat data to include a decimal point.
tab_dat <- apply(tab_dat, 2, function (y) format(round(y, 3) * 100, digits = 2))

# Bind control-group means with main results data.
tab <- rbind(ctrl_results, tab_dat)

# Rename row containing control-group means.
rownames(tab)[which(rownames(tab) == "1")] <- "control_mean"

# Relabel coefficient columns.
coef_cols <- grep("coef$", colnames(tab))

# Relabel SE columns.
se_cols <- grep("se$", colnames(tab))

# Reformat SE estimates to be within parentheses.
tab[,se_cols] <- apply(tab[, se_cols], 2, function (y) paste0("(", y, ")"))

# Concatenate data to comport with LaTeX tabular markup.
tab <- paste(paste(paste(capwords(gsub("_", " ", rownames(tab))), 
			 			 apply(tab, 1, function (x) {
			 			 	   paste(x, collapse = " & ")
				 		 }), sep = " & "), collapse = " \\\\\n"), "\\\\\n")

# Produce tabular output.
sink("output/main_results_tab.tex")
cat("\\begin{table}\n",
	  "\\caption{Estimated Treatment Effects on Support for TPNW}\n",
	  "\\begin{adjustbox}{width = \\textwidth, center}\n",
	  "\\sisetup{\n",
      "\tdetect-all,\n",
      "\ttable-number-alignment = center,\n",
      "\ttable-figures-integer = 1,\n",
      "\ttable-figures-decimal = 3,\n",
      "\ttable-space-text-post = *,\n",
      "\tinput-symbols = {()}\n",
  	  "}\n",
	  paste0("\\begin{tabular}{@{\\extracolsep{5pt}}L{3.5cm}*{4}",
	  		 "{S[table-number-alignment = right, table-column-width=1.25cm]}}\n"),
	  "\\toprule\n",
	  "& \\multicolumn{4}{c}{Model}\\\\\\cmidrule{2-5}\n",
	  "& \\multicolumn{2}{c}{{(1)}} & \\multicolumn{2}{c}{{(2)}} \\\\\\midrule\n",
	  tab,
	  "\\bottomrule\n",
	  "\\end{tabular}\n",
	  "\\end{adjustbox}\n",
	  "\\end{table}\n")
sink()

## Create tabular output for YouGov results.
# Restructure data as a matrix.
aware_tab <- rbind(do.call("rbind", aware_results))

# Reformat data to include three decimal points.
aware_tab <- apply(aware_tab, 2, function (y) format(round(y, 3) * 100, 
				   digits = 3))

# Relabel mean rows.
mean_rows <- endsWith(rownames(aware_tab), "mean")

# Relabel SE rows.
se_rows <- endsWith(rownames(aware_tab), "se")

# Reformat SE estimates to be within parentheses.
aware_tab[se_rows,] <- paste0("(", aware_tab[se_rows,], ")")

# Remove row names for rows with SE estimates.
rownames(aware_tab)[se_rows] <- ""

# Remove "_mean" indication in mean_rows.
rownames(aware_tab)[mean_rows] <- gsub("_mean", "", 
										rownames(aware_tab)[mean_rows])

# Add an empty row, where excluded calculations of responses among skips are
# noted in the table, and rename the relevant row.
aware_tab <- rbind(aware_tab, c("|", "|"))
rownames(aware_tab)[nrow(aware_tab)] <- "Skipped"

# Add an empty column to the table, and insert the count column at the relevant
# indices.
aware_tab[which(rownames(aware_tab) %in% names(aware_table)),]
aware_tab <- cbind(aware_tab, "")
colnames(aware_tab)[ncol(aware_tab)] <- "N"
aware_tab[which(rownames(aware_tab) %in% names(aware_table)), "N"] <- aware_table

# Concatenate data to comport with LaTeX tabular markup.
aware_tab <- paste(paste(paste(capwords(gsub("_", " ", rownames(aware_tab))), 
				 apply(aware_tab, 1, function (x) {
				 	   paste(x, collapse = " & ")
				 }), 
	   			 sep = " & "), collapse = " \\\\\n"), "\\\\\n")

# Produce tabular output.
sink("output/yougov_tab.tex")
cat("\\begin{table}\n",
	  "\\caption{YouGov Survey Responses}\n",
	  "\\centering\\small\n",
	  "\\sisetup{\n",
      "\tdetect-all,\n",
      "\ttable-number-alignment = center,\n",
      "\ttable-figures-integer = 1,\n",
      "\ttable-figures-decimal = 3,\n",      
      "\tinput-symbols = {()}\n",
  	  "}\n",
	  paste0("\\begin{tabular}{@{\\extracolsep{5pt}}L{3.5cm}*{5}", 
	  		 "{S[table-number-alignment = right, table-column-width=1.25cm]}}\n"),
	  "\\toprule\n",
	  "& \\multicolumn{5}{c}{Arm}\\\\\\cmidrule{2-6}\n",
	  "& {Control} & {Group} & {Security} & {Norms} & {Institutions} \\\\\\midrule\n",
	  aware_tab,
	  "\\bottomrule\n",
	  "\\end{tabular}\n",
	  "\\end{table}\n")
sink()

## Create tabular output for attitudinal results.
# Define matrix object of main results.
tab_dat <- do.call("cbind", att_results)

# Reformat matrix to alternate mean and SE estimates.
tab <- sapply(seq(0, 8, 2), function (i) {
	matrix(c(t(tab_dat[,1:2 + i])), 14, 1)
})

# Reformat data to include three decimal points.
tab <- apply(tab, 2, function (y) format(round(y, 3), digits = 3))

# Rename rows to indicate mean and SE estimates.
rownames(tab) <- paste(rep(rownames(tab_dat), each = 2), 
					   c("mean", "se"), sep = "_")

# Relabel mean rows.
mean_rows <- grep("_mean", rownames(tab))

# Relabel SE rows
se_rows <- grep("_se", rownames(tab))

# Reformat SE estimates to be within parentheses.
tab[se_rows,] <- apply(tab[se_rows,], 1, function (y) {
					   paste0("(", gsub(" ", "", y), ")")
				 })

# Rename rows to improve tabular labels; remove "tpnw_atts, "mean," and "se" row
# name strings.
rownames(tab) <- gsub("tpnw_atts|mean$|se$", "", rownames(tab))

# Remove leading and tailing underscores.
rownames(tab) <- gsub("^_|_$", "", rownames(tab))

# Remove row names for rows with SE estimates.
rownames(tab)[se_rows] <- ""

# Concatenate data to comport with LaTeX tabular markup.
tab <- paste(paste(paste(capwords(gsub("_", " ", rownames(tab))), 
						 apply(tab, 1, function (x) {
						 	   paste(x, collapse = " & ")
						 }), 
	   sep = " & "), collapse = " \\\\\n"), "\\\\\n")

# Produce tabular output.
sink("output/atts_tab.tex")
cat("\\begin{table}\n",
	  "\\caption{Attitudes Toward Nuclear Weapons by Arm}\n",
	  "\\centering\\small\n",
	  "\\sisetup{\n",
      "\tdetect-all,\n",
      "\ttable-number-alignment = center,\n",
      "\ttable-figures-integer = 1,\n",
      "\ttable-figures-decimal = 3,\n",
      "\ttable-space-text-post = *,\n",
      "\tinput-symbols = {()}\n",
  	  "}\n",
	  paste0("\\begin{tabular}{@{\\extracolsep{5pt}}L{3.5cm}*{5}", 
	  		 "{S[table-number-alignment = center, table-column-width=1.25cm]}}\n"),
	  "\\toprule\n",
	  "& \\multicolumn{5}{c}{Arm}\\\\\\cmidrule{2-6}\n",
	  "& {Control} & {Group} & {Security} & {Norms} & {Institutions} \\\\\\midrule\n",
	  tab,
	  "\\bottomrule\n",
	  "\\end{tabular}\n",
	  "\\end{table}\n")
sink()

## Create tabular output for results by political party.
# Restructure data such that mean and SE estimates are alternating rows in a
# 1 x 6 matrix, in each of five list elements, corresponding to each treatment
# group; and bind the results for each treatment group.
pid_tab <- lapply(pid_results, function (x) {
				  matrix(unlist(x), nrow = 6, ncol = 1)
		   })
pid_tab <- do.call("cbind", pid_tab)

# Assign row names to distinguish results for each partisan group, and mean and
# SE estimates.
rownames(pid_tab) <- paste(rep(c("democrat", "independent", "republican"), 
							   each = 2), c("mean", "se"))

# Relabel mean rows.
mean_rows <- endsWith(rownames(pid_tab), "mean")

# Relabel SE rows.
se_rows <- endsWith(rownames(pid_tab), "se")

# Label columns per treatment, for the computation of ATEs.
colnames(pid_tab) <- c("control", treats)

# Compute ATEs, with control as baseline, and update tabular data.
pid_tab[mean_rows, treats] <- pid_tab[mean_rows, treats] - 
							  pid_tab[mean_rows, "control"]

# Reformat data to include three decimal points.
pid_tab <- apply(pid_tab, 2, function (y) format(round(y, 3) * 100, digits = 3))

# Remove extraneous spacing.
pid_tab <- gsub(" ", "", pid_tab)

# Reformat SE estimates to be within parentheses.
pid_tab[se_rows,] <- paste0("(", pid_tab[se_rows,], ")")

# Remove row names for rows with SE estimates.
rownames(pid_tab)[se_rows] <- ""

# Concatenate data to comport with LaTeX tabular markup.
pid_tab <- paste(paste(paste(capwords(gsub("_", " ", rownames(pid_tab))), 
				 apply(pid_tab, 1, function (x) {
				 	   paste(x, collapse = " & ")
				 }), 
	   			 sep = " & "), collapse = " \\\\\n"), "\\\\\n")

# Produce tabular output.
sink("output/pid_support.tex")
cat("\\begin{table}\n",
	  "\\caption{Support for Joining TPNW by Party ID}\n",
	  "\\centering\\small\n",
	  "\\sisetup{\n",
      "\tdetect-all,\n",
      "\ttable-number-alignment = center,\n",
      "\ttable-figures-integer = 1,\n",
      "\ttable-figures-decimal = 3,\n",      
      "\tinput-symbols = {()}\n",
  	  "}\n",
	  paste0("\\begin{tabular}{@{\\extracolsep{5pt}}L{3.5cm}*{5}", 
	  		 "{S[table-number-alignment = right, table-column-width=1.25cm]}}\n"),
	  "\\toprule\n",
	  "& \\multicolumn{5}{c}{Arm}\\\\\\cmidrule{2-6}\n",
	  "& {Control} & {Group} & {Security} & {Norms} & {Institutions} \\\\\\midrule\n",
	  pid_tab,
	  "\\bottomrule\n",
	  "\\end{tabular}\n",
	  "\\end{table}\n")
sink()

## Create tabular output for results by political ideology.
# Restructure data such that mean and SE estimates are alternating rows in a
# 1 x 6 matrix, in each of five list elements, corresponding to each treatment
# group; and bind the results for each treatment group.
ideo_tab <- lapply(ideo_results, function (x) {
				   matrix(unlist(x), nrow = 6, ncol = 1)
			})
ideo_tab <- do.call("cbind", ideo_tab)

# Assign row names to distinguish results for each idelogical group, and mean 
# and SE estimates.
rownames(ideo_tab) <- paste(rep(c("liberal", "moderate", "conservative"), 
							   each = 2), c("mean", "se"))

# Reformat data to include three decimal points.
ideo_tab <- apply(ideo_tab, 2, function (y) format(round(y, 3) * 100,
				  digits = 3))

# Relabel mean rows.
mean_rows <- endsWith(rownames(ideo_tab), "mean")

# Relabel SE rows.
se_rows <- endsWith(rownames(ideo_tab), "se")

# Reformat SE estimates to be within parentheses.
ideo_tab[se_rows,] <- paste0("(", ideo_tab[se_rows,], ")")

# Remove row names for rows with SE estimates.
rownames(ideo_tab)[se_rows] <- ""

# Concatenate data to comport with LaTeX tabular markup.
ideo_tab <- paste(paste(paste(capwords(gsub("_", " ", rownames(ideo_tab))), 
				 apply(ideo_tab, 1, function (x) {
				 	   paste(x, collapse = " & ")
				 }), 
	   			 sep = " & "), collapse = " \\\\\n"), "\\\\\n")

# Produce tabular output.
sink("output/ideo_support_tab.tex")
cat("\\begin{table}\n",
	  "\\caption{Support for Joining TPNW by Ideology}\n",
	  "\\centering\\small\n",
	  "\\sisetup{\n",
      "\tdetect-all,\n",
      "\ttable-number-alignment = center,\n",
      "\ttable-figures-integer = 1,\n",
      "\ttable-figures-decimal = 3,\n",      
      "\tinput-symbols = {()}\n",
  	  "}\n",
	  paste0("\\begin{tabular}{@{\\extracolsep{5pt}}L{3.5cm}*{5}",
	  		 "{S[table-number-alignment = right, table-column-width=1.25cm]}}\n"),
	  "\\toprule\n",
	  "& \\multicolumn{5}{c}{Arm}\\\\\\cmidrule{2-6}\n",
	  "& {Control} & {Group} & {Security} & {Norms} & {Institutions} \\\\\\midrule\n",
	  ideo_tab,
	  "\\bottomrule\n",
	  "\\end{tabular}\n",
	  "\\end{table}\n")
sink()

## Create tabular output for weighted main results.
# Define matrix object of weighted main results.
w_tab_dat <- do.call("cbind", w_main_results)

# Compute weighted control-group means, with SE estimates; define OLS formula.
w_ctrl_form <- as.formula(paste(join_tpnw, paste(treats, 
							  collapse = " + "), sep = " ~ "))

# Fit the OLS model per the specification and recover the control mean.
w_ctrl_fit <- lm(w_ctrl_form, data = tpnw, 
				  weights = anesrake_weight)

# Recover the control-group mean.
w_ctrl_mean <- w_ctrl_fit$coef["(Intercept)"]

# Compute control SE.
w_ctrl_se <- sqrt(diag(vcovHC(w_ctrl_fit, "HC2")))["(Intercept)"]


# Concatenate mean and SE output with blank values for Model 2.
w_ctrl_results <- c(format(round(c(w_ctrl_mean, w_ctrl_se), 3) * 100, 
						   digits = 2), "|", "|")

# Reformat data to include a decimal point.
w_tab_dat <- apply(w_tab_dat, 2, function (y) format(round(y, 3) * 100, 
				   digits = 2))

# Bind control-group means with main results data.
w_tab <- rbind(w_ctrl_results, w_tab_dat)

# Rename row containing control-group means.
rownames(w_tab)[which(rownames(w_tab) == "1")] <- "control_mean"

# Relabel coefficient columns.
coef_cols <- grep("coef$", colnames(w_tab))

# Relabel SE columns.
se_cols <- grep("se$", colnames(w_tab))

# Reformat SE estimates to be within parentheses.
w_tab[,se_cols] <- apply(w_tab[, se_cols], 2, function (y) paste0("(", y, ")"))

# Concatenate data to comport with LaTeX tabular markup.
w_tab <- paste(paste(paste(capwords(gsub("_", " ", rownames(w_tab))), 
			 			 apply(w_tab, 1, function (x) {
			 			 	   paste(x, collapse = " & ")
				 		 }), sep = " & "), collapse = " \\\\\n"), "\\\\\n")

# Produce tabular output.
sink("output/weighted_main_results_tab.tex")
cat("\\begin{table}\n",
	  "\\caption{Estimated Treatment Effects on Support for TPNW (Weighted)}\n",
	  "\\begin{adjustbox}{width = \\textwidth, center}\n",
	  "\\sisetup{\n",
      "\tdetect-all,\n",
      "\ttable-number-alignment = center,\n",
      "\ttable-figures-integer = 1,\n",
      "\ttable-figures-decimal = 3,\n",
      "\ttable-space-text-post = *,\n",
      "\tinput-symbols = {()}\n",
  	  "}\n",
	  paste0("\\begin{tabular}{@{\\extracolsep{5pt}}L{3.5cm}*{4}", 
	  		 "{S[table-number-alignment = right, table-column-width=1.25cm]}}\n"),
	  "\\toprule\n",
	  "& \\multicolumn{4}{c}{Model}\\\\\\cmidrule{2-5}\n",
	  "& \\multicolumn{2}{c}{{(1)}} & \\multicolumn{2}{c}{{(2)}} \\\\\\midrule\n",
	  w_tab,
	  "\\bottomrule\n",
	  "\\end{tabular}\n",
	  "\\end{adjustbox}\n",
	  "\\end{table}\n")
sink()

### Save image containing all objects.
save.image(file = "output/hbg_replication_out.RData")