For those who follow this tale, the chart below shows the hitherto of vote intentions among the viable candidates as reported in various polls. Data collected over 2012 were simply disregarded and those from 2013 enter in the model as prior values (covariance) for the 2014 estimates. The big dots at the end of the graphs are my best guess where a particular candidate will end. The "plus" signs represent polling data estimates.

Neves (PSDB) shows a drift reaction in the last polls, in association with Marina's falling the gap between them narrows. Although this trend is not clear across all polls, one major national polling house – Ibope has picked it up with intense this week.
Perhaps most importantly, while Marina turns down, the polls suggest the government is recovering and has now the same share as in April this year. It's this trend I'd like to pay attention over the coming two weeks. If the big move in sentiment towards Marina (PSB) captured by the polls over the last weeks is genuine or ephemeral.

Finally, the undecideds are still significant in number: 8%, which means about 11 million of fellows. So, this stock of voters may play a role in the final spin of the campaign. Actually, who decide now, tend to be consistent and avoid decision costs bringing her decision to the runoff if her candidate succeed the primary.

DILMA ROUSSEFF (PT)

MARINA SILVA (PSB)

AÉCIO NEVES (PSDB)

OTHERS

Yesterday, more Scots than ever since universal suffrage was introduced cast a ballot on the matter of independence. The turnout was itself phenomenal and that implicating a series of questions for the government authorities and citizens, but for the time being the sole question was: would this benefit one side or the other? The verdict favored the "NO"--better together--by a margin little higher than indicated by major polling houses over the last week. And this returns to the point I raised roughly a month ago: the evidence from political science literature suggests that voters who failed to promptly decide are more likely to vote for the status quo in referendums like this.

Just few hours before Scots head to the polls, there is not an overwhelming advantage of the anti-independence vote. Actually, the margin is shorter than last time I looked at it, but despite such a growing trend in favor of the "Yes" campaign in the last weeks, the "NO" side has an edge still. To frame this in terms of probabilities that $\theta_{No}$ exceeds $\theta_{Yes}$, I write a short function (replicated here) that will use simulation from the Dirichlet distributions to compute the posterior probability that "No" exceeds "Yes" shown in the lovely chart below.

The data used here to draw the distributions were gathered from a series of polls and available at the wikipedia. The polls employ different methodologies and phrase questions differently. For instance, some surveys ask respondents how they would vote if this referendum were held today, others ask them how they intend to vote on 18th September. By aggregating them, any swing could be the by-product of the random variation to which all polls are subject.

New poll, new predictions. Dilma is moving towards the red dot as does Marina to the green according to the latest poll by Vox Populi (15th Sept). With about 20 days to the election day, nothing is guaranteed, but the runoff between Dilma and Marina. If the trend continues in this pace, and if Dilma holds the plurality among late deciders, she may get close to a 75% probability of avoiding a runoff by 5th October. Today, there is a 97% probability of runoff, not bad, it was 98% last week.

The following predictions are based on state and national polls. In total, there are more than 100 polls. Those fielded in 2013 enters the model as a prior covariance matrix while the rest are used as evidence and are shown in the charts. In this model, uncertain voters (Swing) are a bit higher than in the model based on national samples only. Also, Marina (PSB) and Aécio (PSDB) tend to get slightly lower numbers, whereas Dilma (PT) does virtually the same in both models. It happens because pollsters fare well on fielding Dilma's support, but they do diverge significantly about the other candidates across the states.

Each plus sign (+) is a poll measure for a candidate at that date and the trend-line is the median of the Gibbs sampling realizations. The blurred band around the median is 95% confidence region. So the predictions are based on that region.

DILMA ROUSSEFF (PT)

MARINA SILVA (PSB)

AÉCIO NEVES (PSDB)

OTHERS

WASTING

While waiting for the next election brings me some data, I'm playing with some old numbers. The following graphs show the rearrangement in the Lower Chamber of Brazil for the last 3 elections. It is astonishing how fast changes happen, specially for small parties. For some of them, having one seat means literally surviving as some just disappeared after having lost its seat. But the most important thing these figures suggest is the increasing fragmentation of the lower chamber in Brazil: In 2002 the number of effective party was 8.47, in 2010 it was 10.37, how like it will be in 2015?