-
Notifications
You must be signed in to change notification settings - Fork 1
/
fit.R
163 lines (129 loc) · 5.5 KB
/
fit.R
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
library(rstan)
library(dplyr)
CONSTS <- config::get()
options(mc.cores=parallel::detectCores())
options(readr.show_col_types = FALSE)
# Load -------------------------------------------------------------------------
polls_national <- readr::read_csv("data/polls_national.csv")
results_national <- readr::read_csv("data/results_national.csv")
lina <- readr::read_csv("data/lina.csv")
ps <- readr::read_csv("data/poststratification_table.csv")
results_by_division <- readr::read_csv("data/tpp_by_division.csv")
# Reconcile --------------------------------------------------------------------
# We are going to focus on the 2019 election, because that's what we have
# record-level data on. So we'll drop off the rest of the data where applicable
results_by_division <- results_by_division |>
tidyr::pivot_wider(names_from = year, values_from = tpp) |>
transmute(division, tpp = `2019`, tpp_prev = `2016`)
polls_national <- polls_national |>
filter(date >= CONSTS$walk_start, date <= CONSTS$walk_end) |>
filter(party=="tpp")
results_national <- results_national |>
filter(year %in% c(2016, 2019)) |>
filter(party=="tpp")
# Can only model divisions that appear in the data
ps <- ps |> filter(division %in% lina$division)
results_by_division <- results_by_division |> filter(division %in% lina$division)
# Prepare data for stan ------------------------------------------------------------
# Convert to factor and ensuring that levels match:
lina <- lina |>
mutate(
division = factor(division),
gender = factor(gender),
age_group = factor(age_group, c("ages18to29", "ages30to44", "ages45to59", "ages60plus")),
education = factor(education, c("highSchoolorCertIorIIorLess", "gradDipDipCertIIIandIV", "bachelorDegree", "postgraduateDegree"))
)
for (var in c('division', 'education', 'gender', 'age_group')) {
ps[[var]] <- factor(ps[[var]], levels(lina[[var]]))
}
# Collapse pollsters with < 10 obs over the entire period
polls_national <- polls_national |>
group_by(pollster) |>
mutate(pollster = ifelse(n() < 10, "Other", pollster))
# Merge results with polls, effectively treating the election as a special poll
df_national <- results_national |>
filter(year==2016) |>
mutate(pollster = "Election") |>
rename(obs=vote) |>
select(-year) |>
bind_rows(polls_national |> rename(obs=poll)) |>
select(date, pollster, obs) |>
arrange(date)
df_national$pollster <- factor(df_national$pollster)
df_national$pollster <- forcats::fct_relevel(df_national$pollster, "Election", after=Inf)
df_national$week <- ceiling(as.numeric((df_national$date - min(df_national$date))/7))+1
# Create dummies:
X <- model.matrix(~ 0 + gender + age_group + education, data=lina)
X <- X[,-1] # Don't need two columns for gender
# Test -------------------------------------------------------------------------
# Before we fit the main model, we test fitting the components:
walkdat <- list(
n_timesteps = max(df_national$week),
n_polls = nrow(df_national),
n_pollsters = nlevels(df_national$pollster)-1, # Treating election as special pollster,
walk0 = results_national$vote[1],
tpp_nat_prev = results_national$vote[nrow(results_national)-1],
polls = df_national$obs,
poll_timestep = df_national$week,
poll_pollster = as.integer(df_national$pollster)
)
# fitwalk <- stan("models/walk.stan", data = walkdat, seed = 2023-08-28)
mrpdat <- list(
n_records = nrow(X),
n_covariates = ncol(X),
n_divisions = nlevels(lina$division),
record_division = as.numeric(lina$division),
age_record = as.integer(lina$age_group),
sex_record = as.integer(lina$gender),
educ_record = as.integer(lina$education),
tpp_div_prev = results_by_division$tpp_prev,
tpp_record = (lina$tpp_imputed=="ALP")*1
)
# fitmrp <- stan(
# "models/mrp.stan",
# data = mrpdat,
# seed = 2023-08-28,
# control=list(adapt_delta=0.90)
# )
# Fit --------------------------------------------------------------------------
# Now we fit the main model
standat <- c(walkdat, mrpdat)
fit <- stan(
"models/model.stan",
data = standat,
iter = 5000,
seed = 2023-08-28,
control=list(adapt_delta=0.90)
)
rstan::check_hmc_diagnostics(fit)
rstan::stan_ess(fit)
rstan::stan_rhat(fit)
# Extract results --------------------------------------------------------------
tpp_walk <- rstan::extract(fit, "walk")[[1]]
poll_bias <- rstan::extract(fit, "poll_bias")[[1]]
colnames(poll_bias) <- levels(df_national$pollster)[1:ncol(poll_bias)]
# Extract coefficients
params <- list(
age_group = rstan::extract(fit, "b_age")[[1]] |> `colnames<-`(levels(lina$age_group)),
gender = rstan::extract(fit, "b_sex")[[1]] |> `colnames<-`(levels(lina$gender)),
education = rstan::extract(fit, "b_educ")[[1]] |> `colnames<-`(levels(lina$education)),
division = rstan::extract(fit, "tpp_div_curr")[[1]] |> `colnames<-`(levels(lina$division))
)
boxplot(params$age_group)
boxplot(params$education)
# Post-stratify
ps_est <- purrr::imap(params, function(coef, nm){
t(coef)[as.numeric(ps[[nm]]), ]
})
ps_est <- purrr::reduce(ps_est, `+`)
colnames(ps_est) <- paste0("est[", 1:ncol(ps_est), "]")
ps_est <- cbind(ps, ps_est) |>
group_by(division) |>
summarise(across(starts_with("est["), ~sum(.*number)/sum(number))) |>
tidyr::pivot_longer(starts_with("est["), names_to = "rep", values_to = "est") |>
mutate(rep = stringr::str_extract(rep, "\\d+"))
# Save -------------------------------------------------------------------------
# Results
saveRDS(tpp_walk, "outputs/tpp_walk.Rds")
saveRDS(poll_bias, "outputs/poll_bias.Rds")
saveRDS(ps_est, "outputs/estimates.Rds")