2. Main Res-IN algorithim
2.1 Item binarization
2.2 Res-In adjacency matrix
2.3 Force-directed location estimation
2.4 Coordinate rotation (PCA)
ResIN R Workshop Bremen CSS 2023
2023-07-05
Welcome to the 2023 BIGSSS CSS 2023 workshop and tutorial on Response-Item Networks (ResIN) in R. For a theoretical overview of the method and its capabilities, and exciting applications, please refer to https://www.resinmethod.net/. Also, we strongly recommend our having the package documentation handy whilst working on this tutorial (https://cran.r-project.org/web/packages/ResIN/ResIN.pdf). To kick things off, we include a couple of teasers of different -graphs we will construct as part of this tutorial:
Workshop contents
Data preparation & cleaning
Main algorithm
qgraph and igraph integration & visualization with ggplot
Advanced features of the ResIN function + exercise
Workshop prerequisites
To get started, let’s import some political attitude data from the 2020 wave of the American National Election Studies (ANES). The ResIN package takes care of all computational legwork for the examples presented here; we need to load the remaining packages solely for the purpose of visualization.
## Loading required packages: Here with "pacman" manager
if(!require("pacman")) install.packages('pacman')
library(pacman)
p_load(ResIN)
p_load(knitr)
p_load(tidyverse)
p_load(visNetwork)
## devtools::install_github("nsgrantham/ggdark") ## ggdark (fancy plotting) requires devtools install the first time around
p_load(ggdark)
## 2020 Wave of the ANES as example data
anes_timeseries_2020 <- read.csv("anes_timeseries_2020.csv")1. Data preparation & cleaning
Select items for ResIN network and optional co-variates
Clean data
ResIN node variables: Character, factor, or numeric are accepted. Be sure to categorize continuous variables beforehand and to apply uniform coding within each variable. ResIn handles character lables for numeric values very well - this feature is useful for visualization
Co-variates should be numeric
NA’s: Apply a uniform code for all
Non-monotonic Likert scale items do not need to be recoded <3
1.1 Data preparation - before
| V200002 | V200003 | V200004 | V200005 | V200006 | V200007 | V200008 | V200009 |
|---|---|---|---|---|---|---|---|
| 3 | 2 | 3 | 0 | -2 | -2 | -2 | 0 |
| 3 | 2 | 3 | 0 | 4 | -1 | 3 | 0 |
| 3 | 2 | 3 | 0 | -2 | -2 | -2 | 0 |
| 3 | 2 | 3 | 0 | -2 | -2 | -2 | 0 |
| 3 | 2 | 3 | 1 | -2 | -2 | -2 | 0 |
| 3 | 2 | 3 | 0 | -2 | -2 | -2 | 0 |
| 3 | 2 | 3 | 0 | -2 | -2 | -2 | 0 |
| 3 | 2 | 3 | 0 | -2 | -2 | -2 | 0 |
| 3 | 2 | 3 | 0 | -2 | -2 | -2 | 0 |
| 3 | 2 | 3 | 1 | 4 | -1 | 3 | 0 |
| 3 | 2 | 3 | 0 | -2 | -2 | -2 | 0 |
| 3 | 2 | 3 | 0 | -2 | -2 | -2 | 0 |
| 3 | 2 | 3 | 1 | 4 | -1 | 3 | 0 |
1.2 Data preparation - recode missing
## Selecting Res-IN node variables:
anes_nodes <- dplyr::select(anes_timeseries_2020, c("V201336", "V202257",
"V201417", "V201312",
"V201416", "V201262",
"V202337", "V201258"))
## Recoding missing values (check ANES codebook!)
anes_nodes$V201336[anes_nodes$V201336 %in% c(-9, -8, -7, -6, -5, 99)] <- NA
anes_nodes$V202257[anes_nodes$V202257 %in% c(-9, -8, -7, -6, -5, 99)] <- NA
anes_nodes$V201417[anes_nodes$V201417 %in% c(-9, -8, -7, -6, -5, 99)] <- NA
anes_nodes$V201312[anes_nodes$V201312 %in% c(-9, -8, -7, -6, -5, 99)] <- NA
anes_nodes$V201416[anes_nodes$V201416 %in% c(-9, -8, -7, -6, -5, 99)] <- NA
anes_nodes$V201262[anes_nodes$V201262 %in% c(-9, -8, -7, -6, -5, 99)] <- NA
anes_nodes$V202337[anes_nodes$V202337 %in% c(-9, -8, -7, -6, -5, 99)] <- NA
anes_nodes$V201258[anes_nodes$V201258 %in% c(-9, -8, -7, -6, -5, 99)] <- NA
## Give more recognizable names
colnames(anes_nodes) <- c("abort", "income", "immigr", "welfare", "gay_mar",
"environm", "gun_own", "aid_black")1.4 Data preparation - adding lables
## Re-code response labels to make them more recognizable
anes_nodes <- anes_nodes %>% dplyr::mutate(
abort = dplyr::recode(abort, `1` = "never",
`2` = "spec_cases",
`3` = "in_need",
`4` = "always",
`5` = "other"),
income = dplyr::recode(income, `1` = "favor",`2` = "oppose", `3` = "neither"),
immigr = dplyr::recode(immigr, `1` = "send_back", `2` = "guest_prog",`3` = "neither"),
welfare = dplyr::recode(welfare, `1` = "increase",`2` = "decrease", `3` = "same",),
gay_mar = dplyr::recode(gay_mar, `1` = "recogn",`2` = "civ_union", `3` = "no_recogn"),
environm = dplyr::recode(environm, `1` = "regul+++",`2` = "regul++", `3` = "regul+", `4` = "neut",`5` = "regul-", `6` = "regul--", `7` = "regul---"),
gun_own = dplyr::recode(gun_own, `1` = "more_diffic",`2` = "easier", `3` = "same"),
aid_black = dplyr::recode(aid_black, `1` = "yes++",`2` = "yes+", `3` = "yes", `4` = "neut",`5` = "no", `6` = "no+", `7` = "no++"))1.5 Data preparation - after
| abort | income | immigr | welfare | gay_mar | environm | gun_own | aid_black |
|---|---|---|---|---|---|---|---|
| spec_cases | favor | send_back | decrease | no_recogn | regul— | same | no++ |
| always | neither | neither | decrease | recogn | NA | same | neut |
| always | favor | NA | increase | recogn | regul+++ | more_diffic | yes |
| never | favor | send_back | same | recogn | regul+++ | same | NA |
| in_need | neither | guest_prog | same | civ_union | neut | same | no+ |
| always | neither | neither | same | recogn | NA | more_diffic | yes |
| in_need | neither | guest_prog | increase | recogn | regul- | more_diffic | no++ |
| always | neither | send_back | same | no_recogn | NA | more_diffic | no++ |
| always | neither | send_back | same | recogn | NA | more_diffic | NA |
| always | neither | guest_prog | same | civ_union | NA | easier | neut |
| never | neither | neither | increase | no_recogn | regul+++ | more_diffic | yes++ |
| spec_cases | oppose | neither | same | recogn | regul+++ | more_diffic | no++ |
| always | neither | send_back | increase | civ_union | neut | more_diffic | no++ |
| always | favor | neither | same | recogn | regul+++ | more_diffic | yes |
| always | oppose | guest_prog | same | civ_union | NA | more_diffic | yes++ |
| in_need | neither | neither | same | recogn | NA | easier | neut |
| always | favor | neither | same | recogn | regul+ | same | neut |
| always | favor | neither | same | recogn | regul+ | more_diffic | yes |
2.1.1 Item binarization
Expand original data-set into dummy variables for each possible item response
Results in \(n\) by \(\sum_{l}k\) data-frame, where \(k\) denotes the number of columns in the original data-frame, and \(\sum_{l}k\) is the sum of all unique item response options \(l\) within columns \(k\).
For the ANES example, 8 item selection results in a dataframe of \(5+3+3+3+3+7+3+7=34\) binary column vectors.
2.1.2 Item binarization - example
## Using the ResIN function from the ResIN package:
df_dummies <- ResIN::ResIN(anes_nodes)$df_dummies
## Number of collumns in the binarized, item-response data-frame:
dim(df_dummies)[2]## [1] 34| abort_always | abort_in_need | abort_never | abort_other | abort_spec_cases | income_favor | income_neither |
|---|---|---|---|---|---|---|
| 0 | 0 | 0 | 0 | 1 | 1 | 0 |
| 1 | 0 | 0 | 0 | 0 | 0 | 1 |
| 1 | 0 | 0 | 0 | 0 | 1 | 0 |
| 0 | 0 | 1 | 0 | 0 | 1 | 0 |
| 0 | 1 | 0 | 0 | 0 | 0 | 1 |
| 1 | 0 | 0 | 0 | 0 | 0 | 1 |
| 0 | 1 | 0 | 0 | 0 | 0 | 1 |
| 1 | 0 | 0 | 0 | 0 | 0 | 1 |
2.2 Res-IN adjacency matrix
Compute pairwise correlation matrix
Set within-item correlations to zero
Set all negative correlations to zero (optional but recommended)
2.2.1 ResIN (sparse) correlation matrix (no within-item correlations)
## Generate a symmetric correlation matrix & apply item response lables
res_in_cor <- ResIN(anes_nodes, remove_negative = FALSE, cluster = FALSE)$adj_matrix| abort_always | abort_in_need | abort_never | abort_other | abort_spec_cases | income_favor | |
|---|---|---|---|---|---|---|
| abort_always | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.48 |
| abort_in_need | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | -0.12 |
| abort_never | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | -0.36 |
| abort_other | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | -0.05 |
| abort_spec_cases | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | -0.36 |
| income_favor | 0.48 | -0.12 | -0.36 | -0.05 | -0.36 | 0.00 |
| income_neither | -0.10 | 0.04 | 0.12 | 0.00 | 0.04 | 0.00 |
| income_oppose | -0.47 | 0.10 | 0.27 | 0.06 | 0.34 | 0.00 |
| immigr_guest_prog | -0.18 | 0.06 | 0.03 | 0.03 | 0.14 | -0.16 |
2.2.2 Zero-out all remaining negative correlations (default & strongly recommended)
res_in_cor <- ResIN(anes_nodes, remove_negative = TRUE)$adj_matrix| abort_always | abort_in_need | abort_never | abort_other | abort_spec_cases | income_favor | |
|---|---|---|---|---|---|---|
| abort_always | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.48 |
| abort_in_need | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| abort_never | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| abort_other | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| abort_spec_cases | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| income_favor | 0.48 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| income_neither | 0.00 | 0.04 | 0.12 | 0.00 | 0.04 | 0.00 |
| income_oppose | 0.00 | 0.10 | 0.27 | 0.06 | 0.34 | 0.00 |
| immigr_guest_prog | 0.00 | 0.06 | 0.03 | 0.03 | 0.14 | 0.00 |
2.3 Force-directed location estimation
We rely on a force-directed algorithm that models item response nodes as a simulated physical system
Positively correlated nodes attract one another; un-correlated (and negatively correlated) nodes repel one another
Many possible algorithms to choose from; we use the commonly used “Fruchterman-Reingold” method (main work-horse in igraph package)
Can be modeled in 2D or 3D; for simplicity we focus on 2D solution here
2.3.1 Force-directed location estimation - using igraph package
- We rely on the “layout_with_fr” function from the igraph package (Nepusz et.al. 2020) https://cran.r-project.org/web/packages/igraph/index.html
## Here we take advantage of the "layout_with_fr" function taken from the igraph package
res_in_layout <- ResIN(anes_nodes)$node_frame2.3.2 Force-directed location estimation - data-frame
| x | y | node_names |
|---|---|---|
| -8.50 | -0.21 | abort_always |
| 1.34 | -0.41 | abort_in_need |
| 5.11 | 1.20 | abort_never |
| 2.61 | -2.06 | abort_other |
| 4.94 | -0.35 | abort_spec_cases |
| -8.44 | 0.62 | income_favor |
| 1.81 | 1.25 | income_neither |
| 5.00 | 0.21 | income_oppose |
| 4.36 | -1.84 | immigr_guest_prog |
| -6.83 | 0.04 | immigr_neither |
| 6.57 | 1.22 | immigr_send_back |
| 5.51 | 0.15 | welfare_decrease |
| -9.33 | -0.45 | welfare_increase |
| -1.56 | 0.12 | welfare_same |
| 4.14 | -0.61 | gay_mar_civ_union |
2.3.3 Force-directed location estimation - plot
2.4 Coordinate rotation
We use principle components analysis (PCA) to make graphical result more interpretable
X-axis rotated s.t. it captures most of the variance within the attitude network
We use orthogonal (non-distorting) rotation but different oblique rotations would theoretically be possible as well
2.4.1 Coordinate rotation using PCA
## Performing PCA with the R-native "prcomp" function; Note that is step is no longer required as ResIN package automates this step for us
res_in_layout_pca <- prcomp(res_in_layout[,1:2])$x
res_in_layout[1:2] <- res_in_layout_pca2.4.1 Coordinate rotation using PCA - R example
3 Integration & visualization with qgraph and igraph packages
qgraph-package
igraph-package
Visualization with ggplot
3.1 qgraph-integration
## qgraph package (Epskamp and Costantini et.al.)
resin_qgraph <- ResIN_qgraph(anes_nodes, plot_graph = TRUE,
qgraph_arglist = list(layout = "spring", maximum = 1, vsize = 4,
DoNotPlot = TRUE, sampleSize = nrow(anes_nodes),
title = "ANES 2020 ResIN plot made with qgraph package",
mar = c(3,3,3,3), normalize = FALSE))
3.2 Visualization with igraph and ggplot2
ResIN function can transform edge lists generated by igraph output to ggplot friendly data.frame
Here, we use the geom_curve() layer in ggplot for network visualization
We can add node-level covariates or network statistics to enhance our plot
3.2.1 Visualization with ggplot2 - prep work
## The ResIN function by default provides a plotting-optimized, edge-list data-frame
edgelist_frame <- ResIN(anes_nodes)$edgelist_frame3.2.2 Visualization with ggplot2 example data-frame
| from | to | weight | from.x | from.y | to.x | to.y | x | y | node_names |
|---|---|---|---|---|---|---|---|---|---|
| abort_always | income_favor | 0.48 | -8.50 | -0.21 | -8.44 | 0.62 | -8.50 | -0.21 | abort_always |
| abort_always | immigr_neither | 0.30 | -8.50 | -0.21 | -6.83 | 0.04 | -8.50 | -0.21 | abort_always |
| abort_always | welfare_increase | 0.42 | -8.50 | -0.21 | -9.33 | -0.45 | -8.50 | -0.21 | abort_always |
| abort_always | gay_mar_recogn | 0.67 | -8.50 | -0.21 | -7.68 | -0.28 | -8.50 | -0.21 | abort_always |
| abort_always | environm_regul++ | 0.18 | -8.50 | -0.21 | -6.15 | 0.07 | -8.50 | -0.21 | abort_always |
| abort_always | environm_regul+++ | 0.55 | -8.50 | -0.21 | -9.41 | 0.18 | -8.50 | -0.21 | abort_always |
| abort_always | gun_own_more_diffic | 0.50 | -8.50 | -0.21 | -7.83 | 0.52 | -8.50 | -0.21 | abort_always |
| abort_always | aid_black_yes | 0.14 | -8.50 | -0.21 | -5.05 | 0.08 | -8.50 | -0.21 | abort_always |
| abort_always | aid_black_yes+ | 0.34 | -8.50 | -0.21 | -8.29 | -1.06 | -8.50 | -0.21 | abort_always |
| abort_always | aid_black_yes++ | 0.45 | -8.50 | -0.21 | -9.35 | 0.90 | -8.50 | -0.21 | abort_always |
| abort_in_need | income_neither | 0.04 | 1.34 | -0.41 | 1.81 | 1.25 | 1.34 | -0.41 | abort_in_need |
| abort_in_need | income_oppose | 0.10 | 1.34 | -0.41 | 5.00 | 0.21 | 1.34 | -0.41 | abort_in_need |
| abort_in_need | immigr_guest_prog | 0.06 | 1.34 | -0.41 | 4.36 | -1.84 | 1.34 | -0.41 | abort_in_need |
| abort_in_need | immigr_neither | 0.01 | 1.34 | -0.41 | -6.83 | 0.04 | 1.34 | -0.41 | abort_in_need |
| abort_in_need | welfare_decrease | 0.06 | 1.34 | -0.41 | 5.51 | 0.15 | 1.34 | -0.41 | abort_in_need |
3.2.3 Visualization with ggplot2 - basic plot
There are many options to plot networks with ggplot. Here, we show an option using the geom_curve function which works well with edge-list type dataframes (i.e. with from.x, from.y, to.x, to.y edge-specific information).
ANES_net_fig <- ggplot() +
geom_curve(data = edgelist_frame, aes(x = from.x, xend = to.x, y = from.y,
yend = to.y), curvature = 0, color = "grey") +
## geom_curve requires the x, xend, y and yend aesthetics to be able to plot the network edges
geom_text(data = edgelist_frame, aes(x = from.x, y = from.y, label = from), size = 3, color = "black") +
geom_text(data = edgelist_frame, aes(x = to.x, y = to.y, label = to), size = 3, color = "black") +
## Make sure to plot node labels both based on the to and from columns in the plotting frame!
ggtitle("Belief Structure among ANES 2020 Respondents")+
theme_bw()+
expand_limits(x = c(min(edgelist_frame$x-2), max(edgelist_frame$x+2)),
y = c(min(edgelist_frame$y-2), max(edgelist_frame$y+2))) +
theme(axis.text.x = element_blank(), axis.title.x = element_blank(),
axis.text.y = element_blank(), axis.title.y = element_blank(),
axis.ticks = element_blank(), panel.grid.major = element_blank(),
panel.grid.minor = element_blank(), legend.position = "bottom",
legend.text = element_blank(), plot.title = element_text(hjust = 0.5))3.2.4 Visualization with ggplot2 - basic plot
3.3.5 Visualization with ggplot2 - more fancy plot
Let’s add a few additional visual elements to enhance our plot. This step also prepares us to for multi-variate data-analysis leveraging each item-response node’s spacial position and various outside information. In this example, we show how to add information on Feeling-thermometer scores to map information on affective polarization between partisan groups onto the attitude space.
## Here, we use Democratic and & Republican feelings thermometers: V201156 and V201157
anes_timeseries_2020$V201156[anes_timeseries_2020$V201156 < 0] <- NA
anes_timeseries_2020$V201157[anes_timeseries_2020$V201157 < 0] <- NA
anes_nodes_plus <- cbind(anes_nodes, anes_timeseries_2020$V201157, anes_timeseries_2020$V201156)
colnames(anes_nodes_plus) <- c(colnames(anes_nodes), "V201156", "V201157")
## We would like to extract the mean of each thermometer variable for the subset of respondents who selected each item response:
ResIN_anes <- ResIN(anes_nodes_plus, node_vars = colnames(anes_nodes), node_covars = c("V201156", "V201157"),
node_costats = c("mean", "mean"), network_stats = TRUE)
edgelist_frame <- ResIN_anes$edgelist_frame
## We can simply calculate node-level affective polarization as the distance between Democrat and Republican FT's
edgelist_frame$aff_pola <- abs(edgelist_frame$V201156_mean - edgelist_frame$V201157_mean)3.3.7 Visualization with ggplot2 - data-frame
| from | to | weight | from.x | from.y | to.x | to.y | x | y | node | R_mean_FT | D_mean_FT | node_label | aff_pola | NA | NA | NA | NA |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| abort_always | income_favor | 0.480 | -8.498 | -0.208 | -8.443 | 0.617 | -8.498 | -0.208 | abort_always | 0 | 0.0018063 | 4.041088 | 4.041088 | 28.74123 | 59.17015 | abort_always | 30.43 |
| abort_always | immigr_neither | 0.302 | -8.498 | -0.208 | -6.832 | 0.038 | -8.498 | -0.208 | abort_always | 0 | 0.0018063 | 4.041088 | 4.041088 | 28.74123 | 59.17015 | abort_always | 30.43 |
| abort_always | welfare_increase | 0.425 | -8.498 | -0.208 | -9.325 | -0.446 | -8.498 | -0.208 | abort_always | 0 | 0.0018063 | 4.041088 | 4.041088 | 28.74123 | 59.17015 | abort_always | 30.43 |
| abort_always | gay_mar_recogn | 0.674 | -8.498 | -0.208 | -7.681 | -0.282 | -8.498 | -0.208 | abort_always | 0 | 0.0018063 | 4.041088 | 4.041088 | 28.74123 | 59.17015 | abort_always | 30.43 |
| abort_always | environm_regul++ | 0.183 | -8.498 | -0.208 | -6.152 | 0.071 | -8.498 | -0.208 | abort_always | 0 | 0.0018063 | 4.041088 | 4.041088 | 28.74123 | 59.17015 | abort_always | 30.43 |
| abort_always | environm_regul+++ | 0.545 | -8.498 | -0.208 | -9.406 | 0.175 | -8.498 | -0.208 | abort_always | 0 | 0.0018063 | 4.041088 | 4.041088 | 28.74123 | 59.17015 | abort_always | 30.43 |
| abort_always | gun_own_more_diffic | 0.498 | -8.498 | -0.208 | -7.834 | 0.517 | -8.498 | -0.208 | abort_always | 0 | 0.0018063 | 4.041088 | 4.041088 | 28.74123 | 59.17015 | abort_always | 30.43 |
| abort_always | aid_black_yes | 0.139 | -8.498 | -0.208 | -5.052 | 0.078 | -8.498 | -0.208 | abort_always | 0 | 0.0018063 | 4.041088 | 4.041088 | 28.74123 | 59.17015 | abort_always | 30.43 |
| abort_always | aid_black_yes+ | 0.345 | -8.498 | -0.208 | -8.289 | -1.063 | -8.498 | -0.208 | abort_always | 0 | 0.0018063 | 4.041088 | 4.041088 | 28.74123 | 59.17015 | abort_always | 30.43 |
| abort_always | aid_black_yes++ | 0.452 | -8.498 | -0.208 | -9.348 | 0.900 | -8.498 | -0.208 | abort_always | 0 | 0.0018063 | 4.041088 | 4.041088 | 28.74123 | 59.17015 | abort_always | 30.43 |
| abort_in_need | income_neither | 0.038 | 1.336 | -0.411 | 1.806 | 1.254 | 1.336 | -0.411 | abort_in_need | 0 | 0.0023380 | 1.424381 | 1.424381 | 48.37396 | 40.89372 | abort_in_need | 7.48 |
| abort_in_need | income_oppose | 0.098 | 1.336 | -0.411 | 5.004 | 0.207 | 1.336 | -0.411 | abort_in_need | 0 | 0.0023380 | 1.424381 | 1.424381 | 48.37396 | 40.89372 | abort_in_need | 7.48 |
| abort_in_need | immigr_guest_prog | 0.056 | 1.336 | -0.411 | 4.362 | -1.836 | 1.336 | -0.411 | abort_in_need | 0 | 0.0023380 | 1.424381 | 1.424381 | 48.37396 | 40.89372 | abort_in_need | 7.48 |
| abort_in_need | immigr_neither | 0.006 | 1.336 | -0.411 | -6.832 | 0.038 | 1.336 | -0.411 | abort_in_need | 0 | 0.0023380 | 1.424381 | 1.424381 | 48.37396 | 40.89372 | abort_in_need | 7.48 |
| abort_in_need | welfare_decrease | 0.065 | 1.336 | -0.411 | 5.507 | 0.154 | 1.336 | -0.411 | abort_in_need | 0 | 0.0023380 | 1.424381 | 1.424381 | 48.37396 | 40.89372 | abort_in_need | 7.48 |
3.3.8 Visualization with ggplot2 - more fancy plot
ANES_dark <- ggplot() +
geom_curve(data = edgelist_frame, aes(x = from.x, xend = to.x, y = from.y,
yend = to.y, size = weight^10), curvature = 0.15, color = "grey", alpha = 0.4) +
## Note that we use the size aesthetic to control the thickness of the graph edges
geom_text(data = edgelist_frame, aes(x = from.x, y = from.y, label = from, color = aff_pola), size = 6) +
geom_text(data = edgelist_frame, aes(x = to.x, y = to.y, label = to, color = aff_pola), size = 6) +
## Make sure to plot node labels both based on the to and from columns in the plotting frame
ggtitle("Belief Structure among ANES 2020 Respondents")+
expand_limits(x = c(min(edgelist_frame$x-1.5), max(edgelist_frame$x+1.5)),
y = c(min(edgelist_frame$y-1.5), max(edgelist_frame$y+1.5))) +
scale_color_continuous(name = "Intensity of Partisan Affective Polarization", high = "yellow", low = "blue")+
scale_size_continuous(guide = "none")+
dark_theme_gray()+
theme(axis.text.x = element_blank(), axis.title.x = element_blank(),
axis.text.y = element_blank(), axis.title.y = element_blank(),
axis.ticks = element_blank(), panel.grid.major = element_blank(),
panel.grid.minor = element_blank(), legend.position = "bottom",
legend.text = element_blank(), plot.title = element_text(hjust = 0.5, size=18))3.3.9 Visualization with ggplot2 - more fancy plot
3.4. Correlational analysis
Although statistical analysis with ResIN can be as complex as you want, we are simply correlating feelings thermometer scores with the spacial location of ResIN nodes. Also, we run a simple linear regression model predicting whether left-wing or right wing attitudes more easily connect to other idea elements in the belief system.
## Correlation between ResIN Left-right position and mean Republican feelings thermometer:
cor(ResIN_anes$node_frame$x, ResIN_anes$node_frame$V201156_mean)## [1] 0.9636615## Simple regression model predicting node centrality from squared x-position
ResIN_anes$node_frame$x_sq <- (ResIN_anes$node_frame$x-mean(ResIN_anes$node_frame$x))^2
summary(lm(ResIN_anes$node_frame$Closeness ~ ResIN_anes$node_frame$x +
ResIN_anes$node_frame$x_sq))##
## Call:
## lm(formula = ResIN_anes$node_frame$Closeness ~ ResIN_anes$node_frame$x +
## ResIN_anes$node_frame$x_sq)
##
## Residuals:
## Min 1Q Median 3Q Max
## -7.665e-04 -1.013e-04 1.499e-05 1.024e-04 4.021e-04
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 2.466e-03 8.376e-05 29.437 < 2e-16 ***
## ResIN_anes$node_frame$x 2.660e-05 9.336e-06 2.849 0.00773 **
## ResIN_anes$node_frame$x_sq -5.354e-06 2.123e-06 -2.522 0.01700 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.0002375 on 31 degrees of freedom
## Multiple R-squared: 0.5847, Adjusted R-squared: 0.5579
## F-statistic: 21.82 on 2 and 31 DF, p-value: 1.216e-064. ResIN function advanced features
4.1 Advanced ResIN features
- cor_method: Which correlation method should be
used? Possible are “pearson”, “kendall”, “spearman”, and “auto”. The
“auto” option calls the qgraph::corr_auto function which
provides a fast implementation of the polychoric correlation coefficient
for binary data. Defaults to “auto”.
- weights: Optionally specify a weights vector of length n to calculate weighted correlation coefficients using the wCorr method. Defalts to NULL.
- method_wCorr: If weights are used, which type of weighted correlation should be used. Options include “Pearson”, “Spearman”, “Polychoric”, and “Polyserial”. Defaults to “Polychoric”
- EBICglasso: Optionally further increase the sparseness of ResIn correlation matrix by performing Gaussian graphical lasso algorithim using the Extended Baysian Information Criterion. Method calls the qgraph::EBICglasso Useful for visualizing very large ResIn graphs although the algorithm might take a while to converge. Defaults to FALSE.
- EBICglasso_arglist: An arglist that can optionally be fed to the qgraph::EBICglasso function.
- node_covars: Optionally select a series of co-variates to obtain node-level statistics. For example the share of female or Republican respondents selecting a particular item response. Can provide useful contextual information for plotting. Defaults to NULL.
- node_costats: A character vector of the same length as node_covars detailing the summary statistic to be obtained from the co-variates on the basis of subsets of respondents who selected each item response. First element in node_costats corresponds to the summary statistic to be calculated from the first element in node_covars. Possible stats are “mean”, “median”, “mode”, “sd”, and “var”.
- network_stats: Should node level and network level structural descriptive statistics be obtained? If set to TRUE, the function obtains betweenness-, closeness-, and strength centrality, as well as expected influence values for each ResIn network node. At the graph level, the function calculates betweenness-, closeness-, strength- and expected influence centralization, as well as the graph average path length and the global clustering coefficient. Defaults to FALSE.
4.2.1 ResIN function returns
adj_matrix: The k*k ResIn adjacency matrix where k are all unique item responses across all items
same_items: A character vector of length k grouping the ResIn nodes by the original items they pertain to. Useful for grouped plotting of ResIn items
edgelist_frame: A ggplot friendly edge-list dataframe. The dataframe has as many rows as ResIn network edges. Features plotting coordinates for all graph edges and nodes as well as node-level network statistics if called for by the main function.
node_frame: A ggplot friendly node-list dataframe. The dataframe has as many rows as ResIn network nodes. Features plotting coordinates for all nodes as well as node-level network statistics if called for by the main function.
graph_structuration: Named vector containing graph-level structuration statistics.
graph_centralization: Named vector containing graph-level centralization statistics.
4.2.2 Let’s test things out using ESS data!
- Select the following issue attitudes from wave 9 of the ESS:
- “gincdif” (Government should reduce income differences)
- “freehms” (Gays and lesbians should live their lifes as they please)
- “impcntr” (Immigration is good for the country)
- “ipstrgv” (Need strong government that ensures safety)
- “impenv” (People should care for the environment)
- Filter by country Switzerland (cntry==“CH”)
- Select “lrscale” (people’s left-right placements) as a node_covars and specify “mean” as node_costats
- Plot the network using mean “lrscale” as a color aesthetic
## Importing ESS 2018 dataset and selecting all relevant variables
ESS_9 <- read.csv("ESS_9.csv")
ESS_9_n <- ESS_9 %>% dplyr::select(cntry, lrscale, gincdif, freehms, impcntr,
ipstrgv, ipeqopt, impenv)
## Sub-setting only Austrian respondents
ESS_9_ch <- ESS_9_n[ESS_9_n$cntry=="CH", ]
## Using the advanced function to generate a plotable data-frame
ResIN_gen_ch <- ResIN(ESS_9_ch, node_vars =
c("gincdif", "freehms", "impcntr",
"ipstrgv", "impenv"),
node_covars = "lrscale", node_costats = "mean", EBICglasso = TRUE)4.2.3 Advanced features - plotting European example
## ResIn_gen stores the ggplot-ready dataframe in an object called "ggplot_frame"
plotting_frame <- ResIN_gen_ch$edgelist_frame
## Now we can plot the network using geom_curve!
ESS_net_fig <- ggplot() +
geom_curve(data = plotting_frame, aes(x = from.x, xend = to.x, y = from.y,
yend = to.y, size = weight^3), curvature = 0.2, color = "grey") +
## Here, we specified edge size according to the correlation weight and color edges based on node strength centrality
geom_text(data = plotting_frame, aes(x = from.x, y = from.y, label = from, color = lrscale_mean), size = 5.5) +
geom_text(data = plotting_frame, aes(x = to.x, y = to.y, label = to, color = lrscale_mean), size = 5.5) +
ggtitle("Belief System Structure among Swiss ESS 2018 Respondents")+
scale_color_continuous(name = "Political ideology: Red = left, blue = right", high = "blue", low = "red")+
scale_size_continuous(guide = "none")+
theme(axis.text.x = element_blank(), axis.title.x = element_blank(),
axis.text.y = element_blank(), axis.title.y = element_blank(),
axis.ticks = element_blank(), panel.grid.major = element_blank(),
panel.grid.minor = element_blank(), legend.position = "bottom",
legend.text = element_blank())
saveRDS(ESS_net_fig, "ESS_net_fig.rds")4.2.4 Advanced features - plotting European example
4.3 Interactive visualization with visNetwork package
- Here, we rely on the visNetwork package to provide a java-script based, HTML implementation for visualizing the ResIn network.
- visNetwork models a force-directed algorithm in real time; the package allows for displaying connected edged and node-level statistics upon mouse hovering.
4.3.1 ANES 2020 Interactive Plot
## Very simple example employing the ResIn_gen function to ANES data:
ResIN_ANES <- ResIN(anes_nodes, EBICglasso = TRUE, network_stats = TRUE)
## Here we need to do just a little bit of prep-work for the visNetwork package
## At the node level, visNetwork minimally needs a unique id column.
## Additional columns such as "label", "group", and "title" are optional but very useful for interactive plotting
ResIN_ANES$node_frame$id <- ResIN_ANES$node_frame$node_names
ResIN_ANES$node_frame$label <- ResIN_ANES$node_frame$node_names
ResIN_ANES$node_frame$group <- ResIN_ANES$same_items
## The "title" column allows for a flexible HTML pop-up window as part of the integration.
## Here, we customize this window with network descriptive statistics:
ResIN_ANES$node_frame$title <- paste0("<center> <strong>", ResIN_ANES$node_frame$label, "</strong> </center> <br>",
"Strength centrality: ", round(ResIN_ANES$node_frame$Strength, 2), "<br>",
"Betweeness centrality: ", ResIN_ANES$node_frame$Betweenness, "<br>",
"Closeness centrality: ", round(ResIN_ANES$node_frame$Closeness, 3) , "<br>",
"Expected influence: ", round(ResIN_ANES$node_frame$ExpectedInfluence, 2), "<br>")
## At the edge-level, visNetwork minimally requires a "from" and "to" vector.
## Here we additionally specify the edge weights as "value" collumn.
ResIN_ANES$edgelist_frame$value <- ResIN_ANES$edgelist_frame$weight
## Note how we keep the visualization as sparse as possible by specifying a 0-degree nearest neighbor highlighting option:
ResIN_ANES_plot <- visNetwork(ResIN_ANES$node_frame, ResIN_ANES$edgelist_frame,
main = "Interactive ANES 2020 Belief System Visualization") %>%
visPhysics(solver = "forceAtlas2Based", forceAtlas2Based = list(stabilization = 10)) %>%
visOptions(highlightNearest = list(enabled = TRUE, degree = 0, hover = TRUE)) %>%
visNodes(size = 24) %>% visInteraction(hideEdgesOnDrag = TRUE)
saveRDS(ResIN_ANES_plot, "ResIN_ANES_plot.rds")