main.Rmd

---
title: "Stilized Facts about NBER Working Papers: a fully reproducible analysis"
author: "Fernando Hoces de la Guardia"
date: "1/29/2018"
output: html_document
editor_options: 
  chunk_output_type: console
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
Sys.setenv("DATAVERSE_KEY" = "examplekey12345")
Sys.setenv("DATAVERSE_SERVER" = "dataverse.harvard.edu")


#tidyverse, dataverse
```

## Web Scrapping (optional)

First: we obtain the data from the NBER website. This task is achieve through the following steps:
 1 - Identify the last paper pubished  
 2 - Run a loop that iterates over the webpage of each paper. For each paper: 
  2.1 - Get the authors and store it the variable `authors`. When multiple authors, separate them with the "-*-" symbol  
  2.2 - Get the title, date of publication, abstract and identify if the working paper was ever published. Use this information to create the variables `title`, `date`, `abstract`, `published`. 
  2.3 - Get the NBER categories of study and store it the variable `NBER_cat`. When multiple authors, separate them with the "-*-" symbol.  
 3 - Save the scraped data set and upload it to repository on dataverse [IN PROGRESS]. 

 


```{r web  scrapping1}

# Loading required libraries
list.of.packages <- c("tidyverse", "gender", "arm")

new.packages <- list.of.packages[!(list.of.packages %in% installed.packages()[,"Package"])]
if(length(new.packages)) install.packages(new.packages, repos= "http://cran.cnr.berkeley.edu/")

lapply(list.of.packages, require, character.only = TRUE)
# Setting working directory
library(here)

# Setting up workflow
for (folder in c("rawdata", "data", "documentation", "output", "paper", "scripts")) {
  if ( !(folder %in% dir()) ) {
    dir.create(folder)    
  }
}

#readme_file <- file.path("README.md")
writeLines("# Descriptive statistics of NBER Working Paper Series", "README.md")


# Define function to the last working paper
last.wp.f <- function() {
  # Download the NBER news web page
  readLines("https://www.nber.org/new.html") %>% 
    # then pull al the strings of the form "w[0-9]{5}"
    str_extract("w[0-9]{5}") %>%
    # then keep only the numeric part of those strings. 
    str_extract("[0-9]{5}") %>% 
    # transform to numeric
    as.numeric() %>% 
    # find the maximum
    max(na.rm = TRUE)
}

# Get the last working paper
last.wp <- last.wp.f()
#last.wp <- 10
# Define a vector that enumerates all possible papers
papers <- 1:last.wp

## Define function toget the extracted lines that match "pattern", and 
## keep only the second group according to "group_pat"
get.line.f <- function(pattern, group_pat) {
  matched.lines <- grep(pattern, raw_lines, value = TRUE)
  sub(group_pat,"\\2", matched.lines) 
}

scrape.nber.f <- function(papers.var){
  # Build a empty data set that will contain all the information of each paper (row) 
  df <- data.frame(authors = rep(NA, length(papers.var)), 
           title  = rep(NA, length(papers.var)), 
           date = rep(NA, length(papers.var)), 
           abstract = rep(NA, length(papers.var)), 
           published = rep(NA, length(papers.var)),
           NBER_cat = rep(NA, length(papers.var)))
  # "j" tracks the data that stores the results  
  j <- 0
  for (i in papers.var) {
    j <- j + 1
    ## read in the website that contains links to all the pages where we want to download data
    raw_lines <- tryCatch(
                readLines( paste("https://www.nber.org/papers/w", i, sep = "") ), 
                error = function(e) NULL)
    # If there is no such paper, jump to next one
    while (is.null(raw_lines)) {
      i <- i + 1
      j <- j + 1
      raw_lines <- tryCatch(
        readLines(paste("https://www.nber.org/papers/w", i, sep = "")), 
        error = function(e) NULL)
      }
  
    df$authors[j] <- get.line.f('<meta name="citation_author" content=.*', 
               '(<meta name=\"citation_author\" content=\")(.*)(\">)') %>% 
        paste(collapse = "-*-" )
  
    df$title[j] <- get.line.f('<meta name="citation_title" content=.*' , 
               '(<meta name="citation_title" content=\")(.*)(\">)')
        
    df$date[j] <- get.line.f('<meta name="DC.Date" content=.*', 
               '(<meta name="DC.Date" content=\")(.*)(\">)')
    
    df$abstract[j]<- raw_lines[which.max(nchar(raw_lines))]
    
    df$published[j] <- ifelse(length(grep("<p id='published_line'>.*"
                                          , raw_lines, value = TRUE)) == 0,
                              "not published", "published")
    # Get lines with NBER categories and clean them
    clean.categ <- grep("<b>NBER Program.*", raw_lines, value = TRUE) %>% 
      str_extract_all("([A-Z]{1,4})\\.html") %>% 
      str_extract_all("[A-Z]{1,4}") %>% 
      unlist() %>% paste(collapse = "-*-")
    
    df$NBER_cat[j] <- ifelse(length(clean.categ) == 0,
                              "No Category", clean.categ)
    # Track progress
    if (j%%50 == 0) {
      print(paste(round(j/length(papers.var), 3) * 100,"% done", sep = ""))
    } 
  }
  return(df)
}

# run the scrapping function
start.time <- Sys.time()
df <- scrape.nber.f(papers.var = papers)
print(Sys.time() - start.time)

# save, export to dataverse, pull from dv, set random check, do plot, 




#################################################################################
### Attempt to updload data to datavers (failed so far)
if (FALSE) {
# Identify the data: (apply some hash function)

# upload to dataverse/osf
Sys.setenv("DATAVERSE_SERVER" = "dataverse.harvard.edu")
Sys.setenv("DATAVERSE_KEY" = "b4d2c057-79a6-42ad-8749-9f9c2b711499")

# create a dataverse
dat <- create_dataverse("mydataverse")


# create a list of metadata
metadat <- list(title = "My Study",
                creator = "Doe, John",
                description = "An example study")

# create a list of metadata
metadat <- list(title = "My Study",
                creator = "Doe, John",
                description = "An example study")

# create the dataset
dat <- initiate_dataset("dataverse.harvard.edu", body = metadat)

# 
# create the dataset
ds <- create_dataset("mydataverse")

# add files
tmp <- tempfile()
write.csv(iris, file = tmp)
f <- add_dataset_file(file = tmp, dataset = ds)

# publish dataset
publish_dataset(ds)

# dataset will now be published
get_dataverse("mydataverse")

main <- create_project(title = "NBER Trends")
sub <- create_component(id = main, title = "Data")
x <- rnorm(1000, 2, 10)
write.csv(x, 'test.csv')
up_df <- upload_files(id = sub, path = "test.csv")
download.file("https://ndownloader.figshare.com/files/2292169",
              "data/portal_data_joined.csv")
#You are now ready to load the data:

surveys <- read.csv("portal_data_joined.csv")
}
```




## Read data

```{r}
load("~/rawdata/all_NBER_Papers_v1.RData")
```



## Analysis  

### Number of authors

```{r authors}
df$num_authors <- sapply(strsplit(df$authors, "-*-", fixed = TRUE), length)

df %>% mutate(time = format(as.Date(df$date), "%Y")) %>% group_by(time) %>% summarise("ave_author_num"= mean(num_authors)) %>%   ggplot(aes(x=time, y = ave_author_num,  group = 1)) + 
    geom_line() +
    labs(title = "Average Number of Authors Per NBER Working Paper")
   
```

### Analysis by gender and NBER categories   
```{r}
# Extract NBER categories from html text
aux1 <- with(df,  regmatches(NBER_cat, 
                             gregexpr("([A-Z]{1,4})\\.html", NBER_cat)))
# Paste multiple categories according to the following format "CAT1-*-CAT2-*-..."
df$NBER_cat <- sapply(aux1, function(y) paste( sub(x =y, ".html", "") , collapse = "-*-" ) )

# delete if no NBER category (306 papers)
df1 <- df[df$NBER_cat!='', ]
# For each paper: separate each cateary. 
asd1 <- strsplit(df1$NBER_cat[], "-*-", fixed = TRUE)

# Get the number of categories per paper
n1 <- sapply(asd1, length)

# Generate numeric var for published and repeat for each repetition (category) of the paper
published <- rep(1*(df1$published!="Not published"), times = n1)
# Get year of working paper and expand 
date1 <- format(as.Date( rep(df1$date, times = n1) ), "%Y")
# Same for categories
categories <- unlist(asd1)

# Expand authors var
authors <- rep((df1$authors), times = n1)

# Create a data frame with new vars
df2 <- data.frame(authors,published, categories, date1, stringsAsFactors = FALSE)

# Same as with categories, but for authors. Final data set has 79163 obs.
asd2 <- strsplit(df2$authors, "-*-", fixed = TRUE)
n2 <- sapply(asd2, length)
published <- rep(df2$published, times = n2)
categories <- rep(df2$categories, times = n2)
date1 <- rep(df2$date1, times = n2)

# Get the first name of each author.
first.name <- gsub("^(.*?),\\s(\\w+).*", "\\2",  unlist(asd2))

# Impute gender to each name
temp1 <- gender::gender(first.name)
# Keep only one obs per name
temp1 <- temp1[!duplicated(temp1),]

# Large size data with name, date and categories
temp2 <- data.frame("name"=first.name, date1, categories, 
                    "full_name" =unlist(asd2), published)

# Merge with gender
temp3 <- right_join(temp1, temp2)
# Delete missing obs NEED TO UNDERSTAND THE MISMATCH (7471)
temp3 <- temp3[!is.na(temp3$gender),]

# Plot percentage of female authors over time: general and by categories
plot1 <- temp3 %>% 
  mutate(num_gen = 1*(gender != 'male')) %>% 
  group_by(date1, categories) %>% 
  summarise("perc_fem" = mean(num_gen,na.rm=TRUE), 
            "n_cat" = n()) %>% 
  filter(date1%in%1985:2015) %>% 
  filter(n_cat>30) %>% 
  ggplot(aes(x=date1, y = perc_fem, color = categories, group = categories)) + 
    geom_line() +
    labs(title = "Percentage of Female Authors in NBER Working Papers", 
         subtitle = "Overall (red) and by research programe") 

temp_df <- temp3 %>% 
  mutate(num_gen = 1*(gender != 'male')) %>% 
  group_by(date1) %>% 
  summarise("perc_fem" = mean(num_gen,na.rm=TRUE)) %>% 
  filter(date1%in%1985:2015)
 
plot1 + 
  geom_line(data = temp_df, 
            aes(x = date1, y = perc_fem, group=1), 
            color = "red", size = 2) 



######
df1 <- df[!is.na(df$NBER_cat), ]
asd1 <- strsplit(df1$NBER_cat, "-*-", fixed = TRUE)
length(unique(unlist(asd1)))
n <- sapply(asd1, length)
published <- rep(1*(df1$published!="Not published"), times = n)
date1 <- format(as.Date( rep(df1$date, times = n) ), "%Y")
categories <- unlist(asd1)
authors <- rep((df1$authors), times = n)
df3 <- data.frame(authors,published, categories, date1, stringsAsFactors = FALSE)

# Publication rates over time by NBER program & total 
plot2 <- temp3 %>% 
  group_by(date1, categories) %>% 
  summarise("pub_perc" = mean(published), n_papers = n()) %>% 
  filter(date1%in%1985:2010) %>% 
  filter(n_papers> 30) %>%   
  ggplot(aes(x=date1, y = pub_perc, 
             color = categories, group = categories)) + 
    geom_line() +
    labs(title = "Percentage Of NBER Working Papers Published In Journals", 
         subtitle = "Overall (red) and by research programe") 

temp_df <- temp3 %>% 
  group_by(date1) %>% 
  summarise("pub_perc" = mean(published,na.rm=TRUE)) %>% 
  filter(date1%in%1985:2010)
 
plot2 + 
  geom_line(data = temp_df, 
            aes(x = date1, y = pub_perc, group=1), 
            color = "red", size = 2) 


# Publication rates over time by total and by gender 
plot3 <- temp3 %>% 
  group_by(date1, gender) %>% 
  summarise("pub_perc" = mean(published), n_papers = n()) %>% 
  filter(date1%in%1985:2010) %>% 
  filter(n_papers> 30) %>%   
  ggplot(aes(x=date1, y = pub_perc, 
             color = gender, group = gender)) + 
    geom_line() +
    labs(title = "Percentage Of NBER Working Papers Published In Journals", 
         subtitle = "Overall (red) and by gender") 

temp_df <- temp3 %>% 
  group_by(date1) %>% 
  summarise("pub_perc" = mean(published,na.rm=TRUE)) %>% 
  filter(date1%in%1985:2010)
 
plot3 + 
  geom_line(data = temp_df, 
            aes(x = date1, y = pub_perc, group=1), 
            color = "red", size = 2) 

# Publication rates by research programe (needs work)

if (FALSE) {
  table1 <- table(date1, categories)
  df3 <- tapply(published, INDEX = list(date1, categories), function(x) mean(x, na.rm = TRUE)) 
  df3 <- df3[which(rownames(df3)%in%1980:2015),]
  
  df2 <- as.data.frame.matrix(prop.table(table1, 1))
  df2 <- df2[which(rownames(df2)%in%1980:2015),]
  
  quartz()
  plot(1980:2015,df2[,1], type ="l" ,lwd=0.1, ylim = c(0,.4))
  for (i in ( 2:(dim(df2)[2]) ) ) lines(1980:2015,df2[,i], type ="l" ,lwd=1)
  
  lines(1980:2015,df2[,"EFG"], type ="l" ,lwd=1, col="red")
  lines(1980:2015,df2[,"LS"], type ="l" ,lwd=1, col="blue")
  lines(1980:2015,df2[,"PE"], type ="l" ,lwd=1, col="green")
  
  df3[is.na(df3)] <- 0
  
  apply(df3,2, function(x) which(x==0))
  
  quartz()
  plot(1980:2015,df3[,1], type ="l" ,lwd=0.1, ylim = c(0,1))
  for (i in ( 2:(dim(df3)[2]) ) ) lines(1980:2015,df3[,i], type ="l" ,lwd=1)
  lines(1980:2015,df3[,"EFG"], type ="l" ,lwd=1, col="red")
  lines(1980:2015,df3[,"LS"], type ="l" ,lwd=1, col="blue")
  lines(1980:2015,df3[,"PE"], type ="l" ,lwd=1, col="green")

}

## Any effect of alphabetical order?
## Gender over time
# last name is not working well
library(arm)
library(gender)
library(dplyr)


asd1 <- strsplit(df$authors, "-*-", fixed = TRUE)
#Total number of authors
length(unique(unlist(asd1)))

n <- sapply(asd1, length)
published <- rep(df$published, times = n)

first.name <- gsub("^(.*?),\\s(\\w+).*", "\\2",  unlist(asd1))
last.name <- gsub("^(.*?),\\s(\\w+).*", "\\1",  unlist(asd1))
first.letter <- toupper(substr(last.name, 1, 1))
temp1 <- gender(first.name)
temp1 <- temp1[!duplicated(temp1),]


date1 <- format(as.Date( rep(df$date, times = n) ), "%Y")

temp2 <- data.frame("name"=first.name, "full_name" = unlist(asd1), date1, published, first.letter)

temp3  <- right_join(temp1, temp2)

#####Droping 1545 (of 11639) authors that do not a predicted gender
temp3 <- temp3[!is.na(temp3$gender),]

#Publishing rate by gender over time
df2 <- (tapply(temp3$published!="Not published", list(temp3$date1, temp3$gender), mean))
plot(x = 1973:2015, y = df2[,"female"], type="l", lty=3)
lines(x = 1973:2015, y = df2[,"male"], type="l", lty=1)
legend("topright",lty=c(3,1), legend = c("Female", "Male"))

published <- rep(df$published, times = n)
df1 <- data.frame(published, unlist(asd1))

#Overall count of WP by first letter
df2 <- tapply(df1$published, first.letter, length)
barplot(df2[(rownames(df2)%in%LETTERS)])

#Overall percentage published
tapply(df1$published!="Not published", as.factor(first.letter%in%LETTERS[1:13]), mean)

#Regression (almost done to include gender interaction)
#Delete letters with very few authors:
to.delete <- which(LETTERS%in%c("Q", "U", "X", "Y", "Z"))

df3 <- data.frame(y = (temp3$published!="Not published"), temp3$first.letter, temp3$gender)
df3 <- df3[df$first.letter%in%LETTERS[-to.delete],]
fit1 <- lm(y~first.letter-1, data = df3)
arm::coefplot(fit1)
abline(v = mean(df3$y))

# If last name is between A-H then mean publication rate is 2% higher. 
d1 <- 1*(df3$first.letter%in%LETTERS[1:8])
coef(summary(lm(df3$y~d1)))


# Don't remember what comes next

# Count of WP over time
df2 <- (tapply(df1$published!="Not published", list(first.letter,format(as.Date(rep(df$date, times = n)), "%Y")), length))
plot(1973:2015,df2[1,], type ="l" ,lwd=0.1, ylim = c(0,300), col="red")
for (i in ( 2:(dim(df2)[1]/2) ) ) lines(1973:2015,df2[i,], type ="l" ,lwd=1, col="red")
for (i in ( (dim(df2)[1]/2+1):(dim(df2)[1]) ) ) lines(1973:2015,df2[i,], type ="l" ,lwd=1, col="Blue")

# Percentage published over time
df2 <- (tapply(df1$published!="Not published", list(first.letter,format(as.Date(rep(df$date, times = n)), "%Y")), mean))
plot(1973:2015,df2[1,], type ="l" ,lwd=0.1, ylim = c(0,1), col="red")
for (i in ( 2:(dim(df2)[1]/2) ) ) lines(1973:2015,df2[i,], type ="l" ,lwd=1, col="red")
for (i in ( (dim(df2)[1]/2+1):(dim(df2)[1]) ) ) lines(1973:2015,df2[i,], type ="l" ,lwd=1, col="Blue")


###Need to get NBER fields

###Do word clouds

### Extract 2-3 word phrases



```





### Author order