CONTEXT

A while ago I noticed that the website of Bosnia’s parliament also provides the voting records for each separate vote taken in both houses. These records also include the voting behavior for each individual MP. Here an example of one of these voting records:

While the dominant role of the ethnic cleavage in Bosnian politics has been a constant theme and continuously emphasized for the last 25 years (to an extent which frequently blanks out other salient issues), I hadn’t seen any empirical record on MP’s individual voting behavior which could corroborate the cleavage’s prevalence in the legislative arena (one insightful exception is Birgit Bahtić-Kunrath’s 2011 paper).

This post digs into the now available records for the 2014-2018 parliament and seeks to shed some light on the legislative process. The first part presents the results, the second highlights the most relevant steps in R to produce them. The entire code for the analysis is again available on my github account. Since the post got a bit longer than initially expected, this one will mainly focus on the lower chamber (House of Representatives). A later post will take up the upper chamber in detail (House of Peoples).

As always, if you see any glaring error, feel free to let me know (best via twitter DM). If you think the blog is interesting and feel like quoting it, please refer to it as Roland Schmidt, “Bosnia: Voting behavior of MPs in the House of Representatives (2014-2018)”, https://werk.statt.codes/post/bih-voting-behavior-of-mps-in-house-of-representatives/, 8 February 2020.

For those working in R, the blog post might be particularly interesting when it comes to retrieving the actual data. As so often with empirical research the challenges has been less the analysis of the data itself, but ‘liberating’ it from the format within which it is provided. In this case, the voting records for both the House of Peoples and the House of Representative are provided in pdfs, such as the one above or this one or this one. Hence, there’s been some hefty OCR-ing involved thanks to the powerful pdf_tools package by Jeroen Ooms.

RESULTS

Number of bills considered and outcome

Overall, during the 2014 to 2018 parliament session 163 bills were considered in both houses. Out of this total, only 37 % were eventually passed and published to become law. 38 % of all bills were rejected in at least one of both houses.¹

by house

Looking separately at each of the two houses provides a more detailed picture. In the lower chamber, the House of Representatives, 153 bills were considered in total out of which a bit more than 50 % were approved. In the upper chamber, the House of Peoples, which comprises five delegates of each constituent people (Bosniak, Croats, Serbs), 56.7 % of all considered bills were passed.

intersection of houses

If we intersect the developments for each law in each chamber, we get a clearer picture where bills were actually rejected/not passed. The table/heatmap below intends to show this. The x-axis shows a bill’s outcome in the House of Representative, the y-axis the outcome of the same bill in the House of Peoples (if a bill was considered only in one chamber, the respective status in the other chamber is ‘missing’).

Out of the 67 bills adopted in the House of Peoples and 77 in the House of Representatives, 60 were actually considered in both chambers and eventually passed. With 37 % of all bills it’s the largest category in the table. The second largest category comprises 27 bills (16.7 % of all bills) which were rejected in the House of Representatives and never made it to the House of Peoples. The third largest category contains 12 bills (7.4 % of all bills) which were adopted in the lower house but rejected in the upper chamber.

Legislative votes over time

How was the legislative activity distributed over time? The graph below provides the number of ballots in both houses per month. Note that the graph also includes votes which were held in both houses outside of the 2014-2018 legislative session. The reason is that bills considered in 2014-2018 also include bills which were initiated prior to 2014 and hence also ballots which were taken outside of the 2014-2018 legislative session.

If we focus only on the 2014-2018 session, it is rather clear that both chambers were most active in the first half of 2016 and the end of 2017. I am not sure exactly what caused this bout of activity, but it’s quite distinct if contrasted e.g. with 2015 or the early parts of 2017.

Voting behavior per MP

So far the analysis presented was limited to bills as the unit of analysis. Now let’s dig a bit deeper and see how individual MPs voted. The graph below provides the results for every voting decision (including abstentions, no vote, not present) of every MP, by party affiliation, during the 2014-2018 legislative period. Since a bill undergoes several readings before it can become a law/is forwarded to the other chamber, there are more votes than bills. Overall, MPs in the House of Representatives were asked to cast their votes 451 times in total. These votes pertain to 148 laws (out of the initial 162 laws above eleven have not entailed any vote by the end of the 2014-2018 legislative period and three entailed ballots prior to the 2014 elections).

In the graphs below each vertical line (small rectangle) represents an MP’s vote in one ballot (hence the x axis is 451 small rectangles long; ordered chronologically). According to the voting records, there are five possible options: an MP can cast 1) a ‘yes’ (za), or 2) an ‘against’ or ‘no’ vote (protiv); 3) be not present during the ballot (nije prisuta); 4) abstained (suzdržan), or 5) do not cast a vote (nije glasao). Note that the Rules of Procedure for the House of Representatives (as for the House of Peoples) stipulate that ‘in case a representative does not vote “in favour”, “against” or “abstained”, but he/she is present at the session during the voting, he/she is considered “abstained” (see here).’

Unfortunately, I had to remove names’ diacritics which added a major source of complication (at least for me).

Even without going into numeric details, I find the graph rather informative. While with most parties/MPs the yes votes (green) dominate, the SNSD is a glaring aberration. Milorad Dodik’s SNSD clearly stands out as the party which says ‘no’ (to paraphrase Ian Paisley) or simply doesn’t show up when asked to cast a vote.

Note that the representatives Vjekoslav Bevanda and Halid Genjac were only briefly members of the House of Representatives. If not mistaken, Bevanda subsequently became Chairperson of the Council of Ministers. From the graph it appears that his place was taken by Predrag Kožul. Similarly, SDA’s Genjac was only for the early sessions member of the HoR. His place seems to have been taken over by Hazim Rančić.

As a general caveat, note that I took MPs’ party affiliation as stated on the parliament’s website which may omit instances in which MPs have changed their party affiliation in the course of the 2014-2018 legislative period. I am not aware that this was actually the case, but if so, please let me know.²

MP ranking per vote category

To get a more detailed view on MPs’ individual voting behavior, the next graph presents each MP’s voting behavior per voting category as % of an MP’s number of total voting decisions (incl. abstentions etc). In other words, the graph reveals how often an MP votes for e.g. yes, no etc. as % of all instances in which she was tasked to cast a vote. Party affiliation is presented next to each MP’s name. Furthermore, to highlight discrepancies between representatives originating (in terms of mandate) from the Federation and the RS, the bars are colored differently. The variation is rather remarkable, particularly with regard to MPs’ ‘no’ votes and absence (not present and abstentions).

I leave it up to the reader to go through each category, but just to highlight one case: According to the published voting records, Nikola Špirić, one of SNSD’s main figure heads (and previously Chairperson of the Council of Ministers), was absent in 44.8 % of all ballots. Even with the benefit of doubt and acknowledging that an MP’s work entails more than participating in ballots (work in committees etc), this is a rather, let’s say, ‘remarkably’ low rate of presence. His party colleagues Markovic and Milovanovic participated only a few times more often. In those instances in which they were present, they voted predominantly with ‘no’. Constructive engagement in the legislative process looks different, at least to me.

Party differences

As a next step, let’s group MPs’ voting behavior by party and the respective (ascribed or self-declared) ethnic affiliation. Admittedly, the latter is not without problems, but I guess the categorization below reflects some kind of middle-ground (SDP could probably be moved to the ‘Bosniak’ category, but well…). The table provides the total number of votes casted by all representatives of each party and their respective share in each vote category. Note the high percentage rate of SDA in the ‘yes’ category and SNSD’s ratio in the ‘no’ or ‘not present’ category.

Below the same results, put presented as a graph.

Conclusion

Overall, the results are pretty much in line with what I expected, but the details nevertheless are quite intriguing. Delegates of SNSD are either absent or vote with ‘no’ in the House of Representatives. There is a clear gulf between representatives from the Federation and the RS when it comes to ‘no’ votes.

Since this has been already rather lengthy (not to speak of the actual coding), I leave it here and will continue the analysis with a new post in a bit. If you spotted anything peculiar or have any other constructive comment etc. feel free to let me know and I will try to follow up on it. In any case, thanks for reading.

STEPS IN R

Below the most important steps to obtain the results from above. The entire code for the (and more) is available at my github here.

As already mentioned above, getting the actual data from the voting records/pdfs has been the main challenge. Essentially, there were three steps required:

• First, get the links to each law
• Second, from there extract
  • data on the status and details of the law in each chamber
  • the links to each voting record(s)
  • use this links to download each voting results file
• Third, get
  • the raw text from the voting result files and
  • extract data of interest (e.g. MP and vote)
• Finally, start with the actual analysis ;-)

Get the data

Getting the links to each law

The parliament’s website provides a list with all bills considered during the 2014-2018 session. See here. Overall, there are 17 pages which contain short summaries on each law plus the links a site with more details on each law. Iterating over these 17 pages and extracting the pertaining links provides us with a dataframe containing the links to the 162 (proposed) laws which were considered in at least one of both houses. To do this, the purrr and rvest package are our tools of choice. The possibly option provides us with a NULL response should the link no yield any info (instead of causing an error).

library(tidyverse)
library(glue)
library(rvest)

#define pages to iterate over
seq_page<- seq(1:17)
seq_page_links <- glue("http://parlament.ba/oLaw/GetOLawsByStatus?page={seq_page}&MandateId=4&Status=-1") %>% 
  enframe(name=NULL)

#define function to extract short info + link of each law 
pb <- progress_estimated(nrow(seq_page_links))

fn_scrap <- function(x){
  
  pb$tick()$print()
  
  x %>% 
    read_html() %>% 
    html_nodes("a") %>% 
    html_attr(.,"href") %>% 
    enframe(name=NULL) %>% 
    filter(str_detect(value, "OLawDetails")) %>% 
    mutate(links_to_law_details=paste0("http://parlament.ba", value)) 
  
}

#apply function to each of the 17 pages
df_links_to_laws <- seq_page_links$value %>% 
  set_names() %>% 
  map_dfr(.,  possibly(fn_scrap, otherwise=NULL), .id="seq_page_links")

With these links we are now able to scrap the status and details of each single (proposed) law (see e.g. here). In short, the code produces the same result as if we were clicking through our initial list from above and copy-paste the information from each law’s details page. The rvest’s functions were again applied to each link with the purrr’s map_dfr.

pb <- dplyr::progress_estimated(nrow(df_links_to_laws)) #define length for progress bar

#scrap law details
fn_scrap_law_details <- function(x) { 
  
  pb$tick()$print()
  
  df.contents <- x %>% 
    read_html() %>% 
    html_nodes("td") %>% 
    html_text(.,trim=T) %>% 
    enframe(name=NULL, value="content") %>% 
    slice(c(1:13))
  
  df.names<- x %>% 
    read_html() %>% 
    html_nodes("th") %>% 
    html_text(.,trim=T) %>% 
    enframe(name=NULL, value="heading") %>% 
    slice(c(1:13))
  
  bind_cols(df.names, df.contents) %>% 
    mutate(house=case_when(row_number()>7 ~ "DP",    #House of Peoples
                           TRUE ~ "PS"))           #House of Representatives
  
}

#Scrap overview of law status 
fn_get_law_status <- function(x) {
  x %>%
    read_html() %>%
    html_nodes("body > section > div.container > div > div.col-md-8 > article > div > div:nth-child(2) > table") %>%
    html_table() %>%
    enframe(name = NULL) %>%
    map_df(., bind_rows) %>%
    rename(
      heading = X1,
      content = X2
    ) %>%
    mutate(heading = case_when(
      lag(heading) %>% str_detect(., "Status u DNPSBiH") ~
        paste("DNPSBiH -", heading),
      lag(heading) %>% str_detect(., "Status u PDPSBiH") ~
        paste("PDPSBiH -", heading),
      TRUE ~ as.character(heading)
    ))
}

#apply function to each link
df_details_all_status<- df_links_to_laws$links_to_law_details %>% 
  set_names() %>% 
  future_map_dfr(.,  fn_get_law_status, 
                 .progress=T,
                 .id="seq_page_links")

Get status of each law

As a first step, I want to get info on each law’s status by the end of the 2014-2018 legislative period in each house. We get this info by applying the below scrapping function to each law’s link (obtained above). Subsequently I reshape the data to a more convenient wider rectangular form (pivot_wider) and extract each field of interest with stringr and regex.

df_details_all_status_wide <- df_details_all_status %>% 
  mutate(content=str_trim(content, side=c("both"))) %>% 
  mutate(content=iconv(content, from="UTF-8", to="windows-1253")) %>% #removes diacrits; makes it easier to work with
  pivot_wider(id_cols = "seq_page_links",
              names_from = "heading",
              values_from = "content") %>% 
  mutate(law_id=str_extract(seq_page_links, "[0-9]+$")) %>% 
  rename(doc_name ="Naziv dokumenta",
         period = Saziv,
         HoR_draft_law_number_date  = "Broj i datum Prijedloga zakona u PDPSBiH",
         HoR_tabler = `Predlagač u PDPSBiH`,
         HoR_status = "Status u PDPSBiH",
         HoR_order_number_session_date_agenda_item = `PDPSBiH - Red. br. i datum sjednice - tačka dnevnog reda`,
         HoP_draft_law_number_date= `Broj i datum Prijedloga zakona u DNPSBiH`,
         HoP_tabler =`Predlagač u DNPSBiH`,
         HoP_status =`Status u DNPSBiH`,
         HoP_order_number_session_date_agenda_item =`DNPSBiH - Red. br. i datum sjednice - tačka dnevnog reda`,
         BiH_Parl_final_status=`Konačni status u PSBiH`,
         ended_tabler_session =`Utvrđen na sjednici predlagača`) %>% 
  separate(col=HoR_draft_law_number_date, 
           into=c("HoR_draft_law_number", "HoR_draft_law_date"),
           sep="od",
           remove=F) %>% 
  separate(col=HoP_draft_law_number_date, 
           into=c("HoP_draft_law_number", "HoP_draft_law_date"),
           sep="od",
           remove=F) %>% 
  mutate_at(vars(contains("draft_law_date")), ~str_remove(., ".$") %>% 
              lubridate::dmy(.)) %>% 
  mutate_at(vars(ends_with("draft_law_number")), ~str_trim(., side=c("both")) %>% 
            str_remove(., ",$") %>% str_replace(., ",","-")) %>% 
  mutate(HoP_order_number_session_date_agenda_item=ifelse(HoP_order_number_session_date_agenda_item=="",
                                                          NA, HoP_order_number_session_date_agenda_item)) %>% 
  mutate(HoP_order_number=stringr::str_extract_all(HoP_order_number_session_date_agenda_item,
                                                   "^[0-9]+") %>% as.numeric) %>% 
  mutate(HoP_session_date=stringr::str_extract_all(HoP_order_number_session_date_agenda_item,
                                                   "[:digit:]+\\.[:digit:]+\\.[:digit:]{4}") %>% 
           map_chr(., paste, collapse="; ")) %>% 
  mutate(HoP_agenda_item=stringr::str_extract(HoP_order_number_session_date_agenda_item,
                                              "Ad\\. [:digit:]+")) %>% 
#  select(-HoP_order_number_session_date_agenda_item) %>% 
  mutate(HoR_order_number_session_date_agenda_item=ifelse(HoR_order_number_session_date_agenda_item=="",
                                                          NA, HoR_order_number_session_date_agenda_item)) %>% 
  mutate(HoR_order_number=stringr::str_extract_all(HoR_order_number_session_date_agenda_item,
                                                   "^[0-9]+") %>% as.numeric) %>% 
  mutate(HoR_session_date=stringr::str_extract_all(HoR_order_number_session_date_agenda_item,
                                                   "[:digit:]+\\.[:digit:]+\\.[:digit:]{4}") %>% 
           map_chr(., paste, collapse="; ")) %>% 
  mutate(HoR_agenda_item=stringr::str_extract(HoR_order_number_session_date_agenda_item,
                                              "Ad\\. [:digit:]+")) %>% 
#  select(-HoR_order_number_session_date_agenda_item) %>% 
  mutate_at(vars(contains("status")),
            .funs=list(~case_when(
              .=="Usvojen"~"adopted",
              .=="Procedura - nije preuzet" ~ "procedure - not started",  #?
              .=="Procedura" ~ "procedure",
              str_detect(., regex("Povucen")) ~ "withdrawn",  #Povucen encoding issue
              .=="Donesen i objavljen" ~ "submitted and published",
              .=="Odbijen" ~ "rejected",
              .=="Obustavljen zakonodavni postupak" ~ "legislative process suspended",
              str_detect(., "Ceka na pokretanje procedure") ~ "waiting for the procedure to start", #encoding issue
              .=="Nije razmatran" ~ "not considered",
              is.na(.) ~ "missing",
              TRUE ~ .)))

Finally, to make things more accessible to clowns like me who haven’t managed to properly learn BCS after too many years, I translate each law’s title with google translate (checkout the vignette of the powerful googleLanguageR package and to how get the necessary authorization for the google translate engine).

df_details_all_status_wide<- df_details_all_status_wide %>% 
  mutate(doc_name_eng=googleLanguageR::gl_translate(doc_name, target="en", format=c("text"),
                                                    source="bs")$translatedText) %>%
  select(doc_name, doc_name_eng, everything())

So this all yields us with a table containing plenty of information on each bill, including the eventual status in each house, the date of the bills submission and the name of those who tabled the bill (in each house). Not too bad me thinks.

Get details of each law

In the next step I want to go beyond each bill’s eventual status, but get details on each bill’s development. The function below, again applied to each link we initially obtained provides us with this. See e.g. here. Note, to speed things up, I use the furrr package with the multiprocess option to apply the function in parallel and not in sequence to each link. Hence, rather than scrapping the details of each bill one after the other, they are downloaded in parallel and therefore much quicker.

library(furrr)
plan(multiprocess)

fn_get_law_status <- function(x) {
  x %>%
    read_html() %>%
    html_nodes("body > section > div.container > div > div.col-md-8 > article > div > div:nth-child(2) > table") %>%
    html_table() %>%
    enframe(name = NULL) %>%
    map_df(., bind_rows) %>%
    rename(
      heading = X1,
      content = X2
    ) %>%
    mutate(heading = case_when(
      lag(heading) %>% str_detect(., "Status u DNPSBiH") ~
        paste("DNPSBiH -", heading),
      lag(heading) %>% str_detect(., "Status u PDPSBiH") ~
        paste("PDPSBiH -", heading),
      TRUE ~ as.character(heading)
    ))
}


df_details_all_status<- df_links_to_laws$links_to_law_details %>% 
  set_names() %>% 
  future_map_dfr(.,  fn_get_law_status, 
                 .progress=F,
                 .id="seq_page_links")

Getting links to each voting record

Critically, the details page of each bill also contains links to the records for each vote taken pertaining to the bill in question. This can be only one vote in only one house, or multiple votes in each house since a bill goes through several readings and entails also votes on procedural issues (e.g. whether the bill should be fast-tracked). For an example see here and note that the links to the voting records are called ‘Listing glasanja o skracenom postupku’ or ‘Listing glasanja o zahtjevu za hitni postupa’. The critical point here is that links to voting records always contain the word ‘listing’. What is hence required at this stage is to extract from all bills’ detail pages every link which contains the word ‘listing’. Once these links are obtained, the voting records to which these links lead can be downloaded. The function below extracts these links of interest.

fn_scrap_links_to_voting_records <- function(x) {

    x %>% 
    read_html() %>% 
    html_nodes(xpath="//a[contains(text(), 'Listing')]") %>%  #filters links based on text/name of links
    html_attr('href') %>%  #extracts links
    enframe(name=NULL) %>% 
    mutate(link_to_voting_record=paste0("http://parlament.ba", value))
  
}

df_voting_results_links <- df_links_to_laws$links_to_law_details %>% 
  set_names() %>% 
  future_map_dfr(.,  possibly(fn_scrap_links_to_voting_records, otherwise=NULL), 
                 .progress = F, .id="seq_page_links") %>% 
  select(-value) %>% 
  mutate(law_id=str_extract(seq_page_links, "[0-9]+$"),
         record_id=str_extract(link_to_voting_record, "(?<=documentId=)[0-9]+")) #pos. look-behind