Variability of pollster during time

Selected only one pollster, to observe the variability of the time effect:

one_pollster <- polls_us_election_2016 %>%
  filter(pollster == "Ipsos" & state == "U.S.") %>%
  mutate(spread = rawpoll_clinton/100 - rawpoll_trump/100)

Calculation of the observed and theoretical standard error over all polls of this pollster:

se <- one_pollster %>%
  summarize(empirical = sd(spread),
            theoretical = 2*sqrt(mean(spread)*(1-mean(spread))/min(samplesize)))
se

A data.frame: 1 × 2

empirical	theoretical
<dbl>	<dbl>
0.04025194	0.03256719

The distribution of the data does not follow a normal distribution:

one_pollster %>% ggplot(aes(spread)) +
  geom_histogram(binwidth = 0.01, color = "black") +
  theme(axis.text = element_text(size=14))

The variability is accounted by the time variation due to party conventions, on both sides.
Graphics of the trend across time for one pollster and several pollsters:

one_pollster %>%
  filter(state == "U.S.") %>%
  mutate(spread = rawpoll_clinton/100 - rawpoll_trump/100) %>%
  ggplot(aes(enddate, spread)) +
  geom_smooth(formula = y ~ x, method = "loess", span = 0.1) +
  geom_point(aes(color = pollster), show.legend = FALSE, alpha = 0.6) +
  theme(axis.text = element_text(size=14))

polls_us_election_2016 %>%
  filter(state == "U.S." & enddate >= "2016-07-01") %>%
  group_by(pollster) %>%
  filter(n() >= 10) %>%
  ungroup() %>%
  mutate(spread = rawpoll_clinton/100 - rawpoll_trump/100) %>%
  ggplot(aes(enddate, spread)) +
  geom_smooth(formula = y ~ x, method = "loess", span = 0.1) +
  geom_point(aes(color = pollster), show.legend = FALSE, alpha = 0.6) +
  theme(axis.text = element_text(size=14))

This shows that there is an extra variable that needs to be modelled to represent the time variability. The sample of election can be rewritten as:
$Xi,j,t = d + b + hj + bt + f(t) + \varepsilon i,j,t$
    $bt$: bias term of time effect
    $f(t)$: trend of time effect