Here’s a snapshot of some of our recent work into politics and voter behaviour

probabilistic and path-analytic methods show intensity of interest in elections predicts perceptions that society has changed for the worse

Testing causal pathways is a critical and often difficult analytical exercise on non-experimental data. However, these causal relationships can be teased out using a range of statistical methods. Using data from the 2019 Australian Election Study, Orbisant sought to build a structural equation model (SEM) to see if intensity of election interest predicts perceptions that society has changed.

The key trick for this analysis, is that societal change perceptions and intensity of interest in elections are not “variables” that you can easily model. Indeed, these are actually higher order “factors” or constructs that are comprised of numerous underlying actual variables. This is where SEM and path analysis comes in (see graph above).

The model revealed that election interest significantly predicts perceptions that society has changed, such that higher election intensity drives more negative societal change perceptions. However, despite its significance, the magnitude of this effect is still somewhat weak at -0.11 (on a -1 to 1 scale), but this does not lessen its practical significance.

Also importantly, all the lower-level variables significantly predicted their respective higher-order factors. Perceptions of health service quality change was the strongest driver of overall societal change perceptions, whilst showing public support for a political party and going to rallies were the strongest drivers of overall election interest intensity.

However, one may begin to hypothesise that these lower-level variables themselves may hold directional or causal relationships. Let's take the election interest variable for example. How can we tease out the causal flow of these variables within themselves before they are aggregated to the higher-order election interest factor? Enter Bayesian Networks (see graph below).

What Bayesian Networks lack in mathematical simplicity they make up for in semantic simplicity and transparency. These networks leverage Bayesian Statistics' immense ability to handle messy and missing data, and react dynamically to model changes. This is due to their use of probabilistic relationships. For example, if an election expert was to say that one of the paths in the above graph was in fact non-existent, we could easily remove that connection and the entire model re-arranges due to its conditional probabilistic nature. Importantly, these connections do not innately represent causality. That is an intellectual leap that we as the interpreters must make.

In this example, we can see that discussing politics in-person appears to drive persuasion of others to vote for a particular party, which in-turn increases the potential for that person to show public support for the party and so on. Of some interest is the relationship between persuading others and discussing politics online. Rather than the showing the same direction of discussing politics in person driving persuasion, we see it go the opposite way for online discussion, whereby persuading others leads people to take their discussion further to the online environment.

Furthermore, Bayesian Networks can also be used for inferential purposes. Specifically, we can use the network to tell us the probability of an event occurring (a score on one of the variables) given the existence of certain scores on others. For example, let's say we want to know the probability that someone will contribute a lot of money (highest possible survey score on that dimension) to a party given that they discuss politics in-person very frequently (highest possible survey score on that dimension). The network returns a very high probability of 0.92 out of 1.

Together, these models highlight just some of the complexity of factors at play when considering voter behaviour and party preferences. But when modelled using appropriate methods, these factors can reveal some interesting and important relationships that may go a long way in understanding particular behaviours.

Voter preference of Liberal or ALP can be predicted from demographics using an artificial neural network

Understanding the drivers of voter choice is a lofty goal sought by most political parties. Using data from the 2019 Australian Election Study, Orbisant sought to build a machine learning model to predict reported voting choice for those voting either Liberal or Labor (ALP). The output of this model returns probabilities that a person (or persons) with given input data would vote for either party.

The final model chosen was an artificial neural network with just 1 hidden layer. Model inputs were some commonly-studied demographic characteristics. These included age, gender, highest post-school qualification, and home ownership status. While the data set included many other characteristics, the goal for this analysis was to produce the best model with the least inputs possible. This goal stemmed from the real-world implication that potential input data for future prediction applications may be somewhat limited.

The model was “successfully” able to predict, to an acceptable degree of accuracy and suitable minimisation of loss, reported party choice. To create a use case beyond the regular training-test data model evaluations, Orbisant took the analysis one step further and used the model to compute the probability that the “average” person in the survey data set would vote for either party. This average person was defined as a combination of the modes (most frequent value) of all the input demographic characteristics. It turns out the average person would likely report voting Labor.

To improve this model, future designs may seek to build in other, non-demographic inputs, and also seek to leverage the longitudinal version of the survey to predict change in party vote selection rather than the party vote preference at one time point. In addition, further work can be done to tweak the algorithm’s hyperparameters to improve accuracy and convergence performance. Further development may also seek to change the model outputs to include other major parties, such as the National Party and the Greens.