2-Poder360Fixer.Rmd

---
title: "Poder360 Polls Fixer"
author: "Daniel T. Ferreira"

output: html_document
---

<!--
This file is (c) 2020 CincoNoveSeis Jornalismo Ltda.
It is licensed under the GNU General Public License, version 3.
-->

Let's fix Poder360 polls! First, let's import some useful libraries.

```{r}
library(purrr)
library(tidyr)
library(dplyr)
library(readr)
library(stringi)
library(stringr)
library(lubridate)
library(mgsub)
library(sqldf)

source('polling_utils.R')
source('polls_registry.R')
```

And define an useful function:

```{r}
concatenate_id = function(x, y) {
  paste0(ifelse(x == 'TSE', 'BR', ifelse(
    nchar(x) == 6, substring(x, first=5, last=6), NA
  )), y)
}
```

Then, we import the TSE Registry data:

```{r}
estatisticos_ids = read_csv('data/manual-data/estatisticos_ids.csv')

df = load_poll_registry_data('./data/tse', estatisticos_ids)
df_for_merge = get_poll_registry_for_merge(df)
```

We start by importing the dataset, removing unwanted polls (e.g. second round
simulations before the first round of the election takes place), and performing
basic normalizations.

```{r}
all_p3 = read_delim('data/poder360/all.csv', ';', escape_double = F, trim_ws = T)
p360_empresa_correspondence <- read_csv("data/manual-data/p360_empresa_correspondence.csv")

dfp3 = all_p3 %>%
  filter(year(data_pesquisa) == ano) %>%
  mutate(num_registro = normalize_simple(num_registro)) %>%
  mutate(raw_nr = str_replace_all(num_registro, '[\\s\\-\\/]', '')) %>%
  mutate(raw_nr = recode(raw_nr,
    `RN00032014` = 'RN000032014',
    `PI00782014` = 'PI000782014',
    `PI00852014` = 'PI000852014',
    `PI00932014` = 'PI000932014',
    `DF00043214` = 'DF000432014',
    `BR10372014` = 'BR010372014',
    `GO14952016` = 'GO014952016',
    `C055462018` = 'AC055462018',
    `0001362012` = 'MG001362012'
  )) %>%
  mutate(raw_nr = ifelse(nchar(raw_nr) == 9, concatenate_id(orgao_registro, raw_nr), raw_nr)) %>%
  mutate(cmp_ue = normalize_simple(ifelse(!is.na(cidade), cidade, unidade_federativa_nome))) %>%
  mutate(has_problematic_id = is.na(raw_nr) | (ambito != 'BR' & ambito != 'RE' & ambito != str_sub(raw_nr, end = 2))) %>%
  left_join(election_dates, by = c('ano' = 'year')) %>%
  mutate(turno_realizacao = ifelse(data_pesquisa <= first_round_date, 1, 2)) %>%
  mutate(position = recode(cargos_id,
    `2` = 'p',
    `1` = 'g',
    `4` = 's',
    `3` = 'pr'
  )) %>%
  mutate(CD_CARGO = recode(position,
    `p` = 11,
    `pr` = 1,
    `g` = 3,
    `s` = 5
  )) %>%
  mutate(estimulada = tipo_id == 2) %>%
  mutate(raw_cand = normalize_cand(candidato)) %>%
  mutate(candidate_without_title = normalize_cand_rm_titles(raw_cand)) %>%
  filter(turno_realizacao == turno) %>%
  mutate(norm_cenario_desc = normalize_simple(cenario_descricao)) %>%
  filter(tipo_id != 3 & !grepl('REJEICAO', norm_cenario_desc)) %>%
  filter(ambito != 'RE') %>%
  filter(percentual != 0) %>%
  filter(condicao == 0 & !grepl('EM BRANCO|NULO|NENHUM|BASE|TOTAL|OUTROS|OUTRAS|NAO SABE|NAO RESPOND|RECUSA|NS\\/NR|CITOU OUTRO', raw_cand))
```

The Poder360 database comes with a number of ID-less polls, or polls with
invalid IDs. To address this, we try matching these polls with the TSE poll
database by pollster and date.

Note that there is some room for improvement here! There are some companies
that should be easy enough to match if we put in more effort.

```{r}
without_id = dfp3 %>%
  filter(has_problematic_id) %>%
  select(pesquisa_id, data_pesquisa, instituto, unidade_federativa_nome, cidade, cmp_ue) %>%
  distinct(data_pesquisa, instituto, unidade_federativa_nome, cidade, .keep_all = T) %>%
  inner_join(p360_empresa_correspondence, by = c('instituto' = 'instituto')) %>%
  mutate(joinid = row_number())

matches = sqldf('
SELECT * FROM without_id INNER JOIN df_for_merge ON
without_id.cmp_ue = df_for_merge.cmp_ue AND
without_id.cnpj = df_for_merge.NR_CNPJ_EMPRESA AND
ABS(julianday(without_id.data_pesquisa) - julianday(df_for_merge.DT_FIM_PESQUISA)) < 5') %>%
  subset(select = -c(cmp_ue)) %>%
  group_by(joinid) %>%
  filter(n() == 1) %>%
  ungroup()

dfp3_2 = dfp3 %>%
  left_join(matches %>% select(pesquisa_id, NR_IDENTIFICACAO_PESQUISA),
            by = c('pesquisa_id' = 'pesquisa_id')) %>%
  mutate(raw_nr = ifelse(has_problematic_id, NR_IDENTIFICACAO_PESQUISA, raw_nr)) %>%
  filter(!is.na(raw_nr) & nchar(raw_nr) == 11) %>%
  select(-NR_IDENTIFICACAO_PESQUISA)
```

Now, we basically repeat the algorithm we applied to our manual polls to solve
the "BR problem". Fortunately, the post-filtered Poder360 database doesn't
have presidential polls outside BR polls; all we have to deal with are non-BR
presidential polls:

```{r}
br_polls = df_for_merge %>%
  filter(startsWith(NR_IDENTIFICACAO_PESQUISA, 'BR')) %>%
  select(DT_INICIO_PESQUISA, DT_FIM_PESQUISA, NR_CNPJ_EMPRESA, QT_ENTREVISTADOS, f_id, NR_IDENTIFICACAO_PESQUISA, DS_DADO_MUNICIPIO) %>%
  rename(br_poll_id = f_id) %>%
  rename(br_poll_id_raw = NR_IDENTIFICACAO_PESQUISA)

state_polls = df_for_merge %>%
  filter(!startsWith(NR_IDENTIFICACAO_PESQUISA, 'BR')) %>%
  select(DT_INICIO_PESQUISA, DT_FIM_PESQUISA, NR_CNPJ_EMPRESA, QT_ENTREVISTADOS, f_id) %>%
  rename(state_poll_id = f_id)

p3merged = dfp3_2 %>%
  filter(position == 'pr' & !startsWith(raw_nr, 'BR')) %>%
  distinct(raw_nr, cmp_ue, instituto, position) %>%
  mutate(joinid = row_number()) %>%
  inner_join(df_for_merge, by = c('raw_nr' = 'NR_IDENTIFICACAO_PESQUISA')) %>%
  group_by(joinid) %>%
  filter(n() == 1) %>%
  ungroup() %>%
  mutate(joinid = row_number())
  
p3j1 = p3merged %>%
  inner_join(br_polls %>% select(-DS_DADO_MUNICIPIO), by = c(
    'DT_INICIO_PESQUISA' = 'DT_INICIO_PESQUISA',
    'DT_FIM_PESQUISA' = 'DT_FIM_PESQUISA',
    'QT_ENTREVISTADOS' = 'QT_ENTREVISTADOS',
    'NR_CNPJ_EMPRESA' = 'NR_CNPJ_EMPRESA'
  )) %>%
  group_by(joinid) %>%
  filter(n() == 1) %>%
  ungroup()

p3j2 = p3merged %>%
  filter(!(joinid %in% p3j1$joinid)) %>%
  inner_join(br_polls, by = c(
    'DT_INICIO_PESQUISA' = 'DT_INICIO_PESQUISA',
    'DT_FIM_PESQUISA' = 'DT_FIM_PESQUISA',
    'QT_ENTREVISTADOS' = 'QT_ENTREVISTADOS',
    'NR_CNPJ_EMPRESA' = 'NR_CNPJ_EMPRESA',
    'DS_DADO_MUNICIPIO' = 'DS_DADO_MUNICIPIO'
  )) %>%
  group_by(joinid) %>%
  filter(n() == 1) %>%
  ungroup()

p3j = bind_rows(p3j1, p3j2) %>%
  select(raw_nr, position, br_poll_id_raw)

dfp3_3 = left_join(dfp3_2, p3j, by = c('raw_nr' = 'raw_nr', 'position' = 'position')) %>%
  mutate(raw_nr = ifelse(!is.na(br_poll_id_raw), br_poll_id_raw, raw_nr)) %>%
  filter(!(position == 'pr' & !startsWith(raw_nr, 'BR'))) %>%
  mutate(cmp_ue = ifelse(position == 'pr', 'BRASIL', cmp_ue))
```

Let's remove excessive scenarios:

```{r}
dfp3_4 = dfp3_3 %>%
  group_by(raw_nr, estimulada, position, voto_tipo, instituto, cmp_ue, cenario_id) %>%
  mutate(scenario_count = n()) %>%
  mutate(scenario_id = cur_group_id()) %>%
  ungroup() %>%
  distinct(scenario_id, raw_cand, percentual, .keep_all = T) %>%
  group_by(scenario_id) %>%
  filter(n_distinct(raw_cand) == n()) %>%
  ungroup() %>%
  group_by(raw_nr, estimulada, position, voto_tipo, instituto, cmp_ue) %>%
  mutate(multiple_scenarios = n_distinct(scenario_id) > 1) %>%
  filter(scenario_count == max(scenario_count)) %>%
  ungroup() %>%
  select(-scenario_count)
```

Finally, we merge the Poder360 database with the TSE polls database, and
subsequently match the candidates:

```{r}
p360_joined = dfp3_4 %>%
  mutate(joinid = row_number()) %>%
  inner_join(df_for_merge, by = c(
    'raw_nr' = 'NR_IDENTIFICACAO_PESQUISA',
    'cmp_ue' = 'cmp_ue'
  )) %>%
  group_by(joinid) %>%
  filter(n() == 1) %>%
  ungroup() %>%
  mutate(norm_voto_tipo = normalize_simple(voto_tipo)) %>%
  mutate(vv = norm_voto_tipo == 'VOTOS VALIDOS' | grepl('VALIDOS', norm_cenario_desc)) %>%
  group_by(raw_nr, SG_UE, NR_CNPJ_EMPRESA) %>%
  filter(n_distinct(ambito) == 1) %>%
  ungroup()
```

And let's save the result:

```{r}
p360_joined %>%
  mutate(main_source = 'Poder360', source = NA) %>%
  rename(NR_IDENTIFICACAO_PESQUISA = raw_nr, year = ano, candidate = raw_cand, result = percentual) %>%
  select(year, NR_IDENTIFICACAO_PESQUISA, SG_UE, NM_UE, CD_CARGO, estimulada, scenario_id, vv,
         candidate, result, DT_INICIO_PESQUISA, DT_FIM_PESQUISA, DT_REGISTRO,
         NR_CNPJ_EMPRESA, QT_ENTREVISTADOS, cmp_ue, norm_est, est_id, is_fluxo, is_phone, self_hired, partisan,
         first_round_date, second_round_date, turno, main_source, source, candidate_without_title,
         multiple_scenarios) %>%
  write.csv('output/pindograma_poder360_polls.csv', row.names = F)
```

<!--
# TODO:
# Resolver n_distinct(ambito) > 1 dentro da mesma pesquisa
# Tentar reduzir as without_id.
# use party to cross cands and p360 as a backup!
-->