Package Management
setwd("D:/OneDrive - HKUST Connect/SOSC/paper_with_jean/sentiment_analysis/connected_case")
require("pacman")
p_load(tidyverse, tidyfst, purrr, furrr, textclean, tictoc, progress, lubridate, sentimentr, rvest, purrr, furrr, sf, terra, tmap, DT, reticulate, tidytext, lme4, modelsummary, sf, leaflet, countrycode)Fix the mistake of labeling non-democracy using v2x_libdem and v2x_regime simultaneously.
Create a new variable nondemocracy_robust which only use v2x_regime as reference.
Make some adjustments to the covariates and fix other mistakes.
Please find all the raw data here.
read_rds("D:\\OneDrive - HKUST Connect\\SOSC\\paper_with_jean\\V-Dem-CY-Core-v12.rds") %>%
tibble() %>%
filter(year %in% c(2007, 2008, 2011, 2018)) %>%
dplyr::select(v2x_libdem, v2x_regime, country_name, country_text_id, year) %>%
mutate(nondemocracy = if_else(v2x_libdem < median(v2x_libdem), 1, 0),
nondemocracy_robust = if_else(v2x_regime %in% 2:3, 0, 1)) %>%
rename(country_code = country_text_id) %>%
dplyr::select(country_name, nondemocracy, country_code, year, nondemocracy_robust) -> democracyrio::import("D:/OneDrive - HKUST Connect/桌面/Ra_Hong/China_aid_v2.xlsx", setclass = "tibble") %>%
dplyr::select(recipient, year, `Amount (Constant USD2017)`) %>%
rename(amount = `Amount (Constant USD2017)`, country_name = recipient) %>%
mutate(country_name = case_when(country_name == "Viet Nam" ~ "Vietnam",
country_name == "Lao People's Democratic Republic" ~ "Laos",
country_name == "Democratic People's Republic of Korea" ~ "North Korea",
country_name == "Congo" ~ "Republic of the Congo",
country_name == "Brunei Darussalam" ~ "Brunei",
country_name == "Cabo Verde" ~ "Cape Verde",
country_name == "Kyrgyz Republic" ~ "Kyrgyzstan",
country_name == "Micronesia" ~ "Federated States of Micronesia",
country_name == "Sint Maarten (Dutch part)" ~ "Sint Maarten",
country_name == "Timor-Leste" ~ "East Timor",
country_name == "Sint Maarten (Dutch part)" ~ "Sint Maarten",
country_name == "Syrian Arab Republic" ~ "Syria",
country_name == "Cote d'Ivoire" ~ "Ivory Coast",
country_name == "Myanmar" ~ "Burma",
country_name == "West Bank and Gaza Strip" ~ "Palestine/Gaza",
TRUE ~ country_name)) %>% # repair name
left_join(standard_code) %>%
filter(!str_detect(country_name, "egion")) %>%
group_by(country_code, year) %>%
summarise(amount = sum(amount, na.rm = T)) %>%
arrange(year) %>%
pivot_wider(names_from = year, values_from = amount) %>%
mutate(sum_2008 = rowSums(across(`2000`:`2008`), na.rm = T), # cumulatively
sum_2009 = rowSums(across(`2000`:`2009`), na.rm = T),
sum_2012 = rowSums(across(`2000`:`2011`), na.rm = T),
sum_2019 = rowSums(across(`2000`:`2017`), na.rm = T)) %>%
dplyr::select(country_code, sum_2008:sum_2019) %>%
pivot_longer(cols = sum_2008:sum_2019, names_to = "year") %>%
mutate(year = map_chr(year, ~ str_extract(., "[[:digit:]]{4}")) |> as.numeric()) %>%
rename(cum_amount = value) %>%
ungroup() -> aiddemocracy %>%
mutate(year = year + 1) %>%
left_join(aid) %>%
left_join(BRI) %>%
select(4, 1, 3, 2, 6, 7, 5) %>%
replace_na(list(BRI = 0, cum_amount = 0)) %>%
mutate(aid = if_else(cum_amount > 0, 1, 0)) -> all_variables
read_sf("D:/OneDrive - HKUST Connect/桌面/Ra_Hong/China/map/ne_50m_admin_0_countries.shp") %>%
rename(country_code = ADM0_A3) %>%
left_join(filter(all_variables, year == 2019)) %>%
mutate(nondemocracy = if_else(nondemocracy == 1, "Not Democracy", "Democracy")) -> world_shape
colorFactor(palette = c("#be64ac", "#5ac8c8"), domain = world_shape$nondemocracy) -> pal # color palette
read_rds("geo_point_world.Rdata") %>%
left_join(all_variables) -> point_geo # lat-long
point_geo %>%
filter(BRI == 1) %>%
mutate(BRI = "BRI") %>%
filter(year == 2008) -> BRI_geo # BRI point
point_geo %>%
filter(cum_amount > 0 & year == 2019) -> aid_geo # aid point
leaflet() %>%
addProviderTiles(providers$CartoDB.Positron,
options = tileOptions(tileSize = 256, minZoom = 2, maxZoom = 20)) %>%
addPolygons(data = sf::st_as_sf(world_shape),
fillColor = ~ pal(nondemocracy),
stroke = T,
color = "#8e92c1",
weight = 2,
smoothFactor = 1,
opacity = 1,
fillOpacity = 0.2,
label = ~ country_name,
highlightOptions = highlightOptions(color = "green",
weight = 2,
bringToFront = TRUE),
labelOptions = labelOptions(style = list("font-weight" = "normal", padding = "3px 8px"),
textsize = "12px",
direction = "auto")) %>%
addMarkers(data = BRI_geo,
lng = ~ longitude - 0.02,
lat = ~ latitude - 0.02,
label = "BRI Country",
clusterOptions = markerClusterOptions()) %>%
addLegend(na.label = "Missing",
data = sf::st_as_sf(world_shape),
pal = pal,
values = ~ nondemocracy,
opacity = 0.3,
title = NULL,
position = "bottomright") %>%
addCircles(opacity = 1,
color = "#5a64ac",
data = aid_geo,
lng = ~ longitude,
lat = ~ latitude,
weight = 6.6,
radius = ~ log(cum_amount), label = "Aid Recipient", popup = ~ str_c(format(round(cum_amount / 1e6, 1), trim = TRUE), "Million"))Sample data (n = 2000, N = 43,780) is as follows.
options(digits = 3)
read_rds("sentence_level_result.Rdata") %>%
left_join(read_rds("basic_info.Rdata")) %>%
mutate(year = year + 1) %>%
left_join(all_variables) %>%
group_by(country_code, events) %>%
mutate(caseid = cur_group_id() |> as_factor(),
treatment = if_else(day >= 0, 1, 0),
day_f = as_factor(day)) %>%
filter(day < 10 & day >= -10) %>%
ungroup() -> regression_data
regression_data %>%
dplyr::select(-c(link, sentiment_sr)) %>%
sample_n(2000) %>%
datatable(filter = "top", options = list(pageLength = 5, autoWidth = TRUE)) regression_data %>%
pivot_longer(cols = compound:pos, names_to = "sentiment_name") %>%
group_nest(sentiment_name) %>%
mutate(models = map(data, ~ lmer(value ~ 1 + treatment * (aid + nondemocracy + BRI) + (1 | caseid) + (1 | day_f), data = .))) -> models
modelsummary(list("neu" = models$models[[3]],
"pos" = models$models[[4]],
"neg" = models$models[[2]],
"compound" = models$models[[1]]),
stars = c("*" = 0.10, "**" = 0.05, "***" = 0.01),
coef_map = c("(Intercept)" = "Constant",
"treatment" = "Treatment",
"aid" = "Aid recipient",
"nondemocracy" = "Not democracy",
"BRI" = "BRI member",
"treatmentaid" = "Treatment * Aid recipient",
"treatmentnondemocracy" = "Treatment * Not democracy",
"treatmentBRI" = "Treatment * BRI member"))| neu | pos | neg | compound | |
|---|---|---|---|---|
| Constant | 0.566*** | 0.072*** | 0.053*** | 0.076*** |
| (0.024) | (0.002) | (0.003) | (0.021) | |
| Treatment | 0.004 | -0.004*** | 0.002* | -0.015** |
| (0.003) | (0.001) | (0.001) | (0.007) | |
| Aid recipient | 0.068 | -0.0008 | -0.013** | 0.107*** |
| (0.044) | (0.005) | (0.006) | (0.039) | |
| Not democracy | -0.125** | -0.004 | 0.029*** | -0.226*** |
| (0.062) | (0.008) | (0.008) | (0.055) | |
| BRI member | 0.091* | 0.002 | -0.026*** | 0.193*** |
| (0.049) | (0.007) | (0.007) | (0.045) | |
| Treatment * Aid recipient | 0.010 | 0.007* | -0.004 | 0.033** |
| (0.012) | (0.004) | (0.003) | (0.016) | |
| Treatment * Not democracy | 0.002 | -0.007 | -0.009 | 0.051 |
| (0.026) | (0.009) | (0.006) | (0.035) | |
| Treatment * BRI member | -0.022 | 0.005 | 0.011* | -0.066** |
| (0.024) | (0.008) | (0.006) | (0.032) | |
| Num.Obs. | 43780 | 43780 | 43780 | 43780 |
| R2 Marg. | 0.008 | 0.0009 | 0.015 | 0.025 |
| R2 Cond. | 0.199 | 0.018 | 0.074 | 0.107 |
| AIC | -33564.8 | -129158.5 | -158019.2 | -7041.7 |
| BIC | -33469.2 | -129062.9 | -157923.6 | -6946.1 |
| ICC | 0.2 | 0.02 | 0.06 | 0.08 |
| RMSE | 0.16 | 0.06 | 0.04 | 0.22 |
| * p < 0.1, ** p < 0.05, *** p < 0.01 |
regression_data %>%
pivot_longer(cols = compound:pos, names_to = "sentiment_name") %>%
group_nest(sentiment_name) %>%
mutate(models = map(data, ~ lmer(value ~ 1 + treatment * (aid + nondemocracy_robust + BRI) + (1 | caseid) + (1 | day_f), data = .))) -> models_robust
modelsummary(list("neu" = models_robust$models[[3]],
"pos" = models_robust$models[[4]],
"neg" = models_robust$models[[2]],
"compound" = models_robust$models[[1]]),
stars = c("*" = 0.10, "**" = 0.05, "***" = 0.01),
coef_map = c("(Intercept)" = "Constant",
"treatment" = "Treatment",
"aid" = "Aid recipient",
"nondemocracy_robust" = "Not democracy",
"BRI" = "BRI member",
"treatmentaid" = "Treatment * Aid recipient",
"treatmentnondemocracy_robust" = "Treatment * Not democracy",
"treatmentBRI" = "Treatment * BRI member"))| neu | pos | neg | compound | |
|---|---|---|---|---|
| Constant | 0.570*** | 0.072*** | 0.052*** | 0.086*** |
| (0.024) | (0.002) | (0.003) | (0.022) | |
| Treatment | 0.004 | -0.003*** | 0.002* | -0.015** |
| (0.003) | (0.001) | (0.001) | (0.007) | |
| Aid recipient | 0.051 | -0.002 | -0.007 | 0.067* |
| (0.040) | (0.005) | (0.006) | (0.037) | |
| Not democracy | -0.151** | -0.006 | 0.024** | -0.231*** |
| (0.062) | (0.009) | (0.010) | (0.060) | |
| BRI member | 0.133** | 0.006 | -0.028*** | 0.242*** |
| (0.056) | (0.009) | (0.009) | (0.056) | |
| Treatment * Aid recipient | 0.007 | 0.005 | -0.005* | 0.034** |
| (0.011) | (0.004) | (0.003) | (0.015) | |
| Treatment * Not democracy | 0.056 | 0.004 | -0.015* | 0.134** |
| (0.039) | (0.013) | (0.009) | (0.052) | |
| Treatment * BRI member | -0.073* | -0.005 | 0.018* | -0.151*** |
| (0.038) | (0.013) | (0.009) | (0.052) | |
| Num.Obs. | 43780 | 43780 | 43780 | 43780 |
| R2 Marg. | 0.006 | 0.0008 | 0.014 | 0.022 |
| R2 Cond. | 0.193 | 0.017 | 0.085 | 0.115 |
| AIC | -33568.4 | -129158.4 | -158016.8 | -7044.3 |
| BIC | -33472.9 | -129062.9 | -157921.2 | -6948.8 |
| ICC | 0.2 | 0.02 | 0.07 | 0.09 |
| RMSE | 0.16 | 0.06 | 0.04 | 0.22 |
| * p < 0.1, ** p < 0.05, *** p < 0.01 |
Both question 1 and question 2 are caused by different treatment coding (i.e., when does the treatment effect begin). Since I tried two ways before and after which brought ambiguities and inconsistencies. I am very sorry. Please code the treatment based on the day.
full <- read_rds("sentence_level_result.Rdata") %>%
left_join(read_rds("basic_info.Rdata")) %>%
mutate(year = year + 1) %>% # since we use previous year's democracy information
left_join(all_variables) %>%
group_by(country_code, events) %>%
mutate(caseid = cur_group_id() |> as_factor(),
day_f = as_factor(day)) %>%
ungroup() # all the data
before <- full %>%
mutate(treatment = if_else(day > 0, 1, 0)) %>%
filter(day > -10 & day <= 10) # before
after <- full %>%
mutate(treatment = if_else(day >= 0, 1, 0)) %>%
filter(day >= -10 & day < 10) # after
question_1_country <- c("Kazakhstan", "Malaysia", "Turkmenistan", "United States of America") %>%
imap(~ countrycode(., origin = "country.name", destination = "iso3c")) %>%
unlist()
before_n <- before %>%
filter(country_code %in% question_1_country) %>%
count(events, country_code, treatment, name = "before_n")
after_n <- after %>%
filter(country_code %in% question_1_country) %>%
count(events, country_code, treatment, name = "after_n")
before_n %>%
left_join(after_n) %>%
mutate(diff = after_n - before_n) %>%
datatable()question_2_country <- c("Brazil", "Cambodia", "Croatia", "Malaysia", "Nepal", "New Zealand", "Senegal", "Vietnam") %>%
imap(~ countrycode(., origin = "country.name", destination = "iso3c")) %>%
unlist()
before_n_2 <- before %>%
ungroup() %>%
filter(country_code %in% question_2_country & events == "chenguangcheng") %>%
count(events, country_code, treatment, name = "before_n")
after_n_2 <- after %>%
ungroup() %>%
filter(country_code %in% question_2_country & events == "chenguangcheng") %>%
count(events, country_code, treatment, name = "after_n")
before_n_2 %>%
left_join(after_n_2) %>%
mutate(diff = after_n - before_n) %>%
datatable()As I have mentioned, only sentimentr in R will commit errors when it processes a very long sentence without proper punctuation. Since we choose the data calculated by the asent, so no articles will be dropped.
Asent in Python:
0% 25% 50% 75% 100%
-0.9951 -0.0154 0.1088 0.2401 0.9988 Sentimentr in R: