The fundamental model (what we sometimes call sentiment
model) is an importance piece of our methodology. Our first attempt was to
construct a standard model using the typical explanatory factors. In the literature many
variables have been identified as predictive of an election outcome. The proportion of votes for a particular party depends on
national/regional/provincial variables which suggest that an appropriate
method is a hierarchical model. For instance, at the national level many models
use the share of votes for the party in the previous election, if it is the
incumbent or not, the change in GDP (or personal income) during the last year
(and its interaction with incumbency), the change in unemployment, etc. On a side note, a quite
successful fundamental model for Presidential elections in the US is the so
called “bread and peace” model of Hibbs. It uses only two variables: the growth
rate of per capita real disposable income and the US military fatalities per million
population.
Sometimes fundamental models also include variables from
polls, e.g. the rate of approval of the Presidential candidate. We are going to
avoid including results from polls to the fundamental model. At the provincial/state level further variables to consider include the most recent electoral result for the party, the presidential or vice-presidential home
state advantage (for instance in the case of Rajoy the provincial advantage
will be assigned to A Coruña), change in state’s economic growth (and its interaction with incumbency), and even a partisanship or ideology indicator of the state.
Our first attempt for the fundamental model was along the
lines described above. We run a multilevel (national/provincial) Bayesian model
(lag share, incumbent, unemployment, change of per capita GDP at the national level,
change of per capita GDP at the regional level, etc.)
Although the performance of the model in previous Spanish
elections was quite good, the past is not very informative for the future in
the case of the 2015 Congressional Election. It seems clear that, for instance,
the effect of the recent macroeconomic improvement weights in favor of the
incumbent (in this case the Popular Party) but when you have two new parties it
is not reasonable to assume that the weight will be similar to the one estimated
when the Spanish political field was dominated by two large parties. Something similar happens with incumbency:
the historical effect of being the party in power before an election is not
going to be a good indication for predicting the effect of incumbency in the
2015 Congressional election when new parties enter the game. Summarizing, any
model that relied on historical data can explain well the past but cannot predict
the future, which is our objective. The search continued…