Rittman Analytics

View Original

Behavioural Analytics — Querying fast and slow

This article aims to explore a layman’s approach to behavioural analytics, using concepts within behaviouralism and behavioural science to unlock new ways of using analytics to help us achieve organisational goals.

These are complex disciplines. Behaviouralism is philosophic and behavioural science is the scientific¹. Like most philosophical and scientific disciplines it requires a lifetime of study to become truly informed.

However, gaining an understanding of the foundations of these two subjects has become increasingly accessible and I believe that with this accessibility, it’s possible to gain a basic level of understanding that can open up a way of thinking about analytics that can dramatically improve analytics functions within many organisations. This way of thinking puts understanding human behaviour at the heart of our decision-making, leading us to make decisions and create interventions that nudge end users to make more desirable choices. It’s this way of thinking I want to explore.

So What Is Behaviourism?

The first explicit behaviouralist was John B. Watson “who in 1913 issued a kind of manifesto called Psychology As the Behaviouralist Views It”.¹ What Watson was arguing is that psychology should be redefined as the study of behaviours. This was radically different to the thinking at the time, where most psychologists were predominantly concerned with consciousness and mental processes. This disruption was compounded when in 1971 B.F Skinner published Beyond Freedom and Dignity, which marked a landmark within the paradigm of radical behaviouralism, something much more exciting than plain old behaviouralism… This new radical behaviourism saw a future where the often self-destructive and unpredictable nature of humans could be corrected through behavioural conditioning. Skinner writes;

“What we need is a technology of behaviour, We could solve our problems quickly enough if we could adjust the growth of the world’s population as precisely as we adjust the course of a spaceship…”²

Skinner saw behaviouralism as a way of saving ourselves from ourselves and perhaps the only way to save ourselves. This radical view speaks to his belief in the power of human behaviour not only in our lives as individuals but as a species.

Paired with this radical new philosophy, in the coming years another powerful force emerged; the collection and processing of data on a mass scale. Behavioural scientists and commercial interest over this period became increasingly aware of how these two powerful disciplines can be used together.

What does this mean to the modern data scientist or analytics engineer? Well, the behaviouralists would probably argue, we are in a position of great power.

The basics of behavioural science in application to data analysis

So far, we have covered a very brief overview of the emergence of behaviouralism, the concepts of which we will apply to our analytics using methodologies that sit within behavioural science. Let’s look into how analytics fits into all of this.

Beyond freedom and dignity AND descriptive analytics…

Broadly speaking, there are three types of analysis:

  1. Descriptive

  2. Predictive

  3. Causal³

Developing a BI platform that provides reliable and fast descriptive analytics is no mean feat and can be very powerful. It allows you to understand things that have happened and perhaps why they happened. It is however limited in its ability as it only looks to describe past events. Predictive analytics helps you look forward, forecasting future trends to better help decision-making in the present. Causal analytics on the other hand offers the opportunity to carve out the path for future activity by modifying the behaviours of end-users. We can use Skinner to provide us with an example of this, he writes:

“We tend to say, often rashly, that if one thing follows another, it was probably caused by it — following the ancient principle of post hoc, ergo propter hoc (after this, therefore because of this). An example of this is if we speak sharply with a friend and are asked why a reasonable response would be ‘I felt angry’”¹. Anger here is the cognition or emotion and the reaction is the behaviour it invokes. It’s this relationship that casual analytics looks to understand and utilise, which in turn brings us to the question, what can we utilise?

Wendel’s summary of research on this topic gives us some useful pointers:

  • We’re limited beings in attention, time, willpower, etc.

  • We’re of two minds: our actions depend on both conscious thought and non-conscious reactions, like habits.

  • Because of these limitations, our minds use shortcuts to economise and make quick decisions.

  • Our decisions and our behaviour are deeply affected by the context we’re in, worsening or ameliorating our biases and our intention–action gap.

  • One can cleverly and thoughtfully design a context to improve people’s decision-making and lessen the intention–action gap.³

Understanding these behavioural qualities allows us to better approach causal analytics. The acknowledgement that we are limited beings and do not operate on a purely rational basis can change the treatment we give end-users. The fact that we are of two minds, the conscious and the non-conscious can prompt us to think about how our neurology impacts how we structure our business behaviours. Taking into account the mental shortcuts people use to make decisions and how context can also affect these decisions enables us to better align the actions we take to the people we are focused towards.

What does this mean to behavioural analysis? If you can develop models and methodologies that take advantage of how individuals are likely to behave, you can nudge them in the direction you want them to go in. If your business has the potential, you can constantly nudge them into the direction you want them to go in and become habit-forming through cues, routines, and rewards.³

With these considerations in mind, you can see how using our knowledge of behavioural science can help us encourage end-users to do certain things for their or our betterment.

The process for developing behavioural models — A case study

To bring theory to life, I have used basic understandings and processes to approach a business question. It will hopefully provide practical context for the concepts and processes we have and will be covering.

For our case study, we will be framing it within the bounds of a behavioural story. For our behavioural story, we are very lucky as there is a double coincidence of wants. We want employees to gain certification and employees want to be certified. From a commercial point of view, it strengthens our ability to deliver and brings us closer to our partners. For the individual, it increases their equity, satisfies their desire to learn and enables them to succeed in their personal development.

Despite this fortunate convergence of wants, putting in time for certification in a growing consultancy can be a challenge. The business and individual want to do it but billable client work often gets prioritised when utilisation starts to run hot. This leads to the intention-action gap where people want to put time aside for certification but it gets pushed aside for seemingly more important activities. This manifests into a behaviour where engineers default to prioritising client work over training.

This behavioural story is a sensitive one. Client work pays the bills and strategic messaging and company culture can move in opposite directions, a case of “Culture eats strategy for breakfast”. This being said, there’s nothing wrong with hoping for better and behavioural analytics can be a way of reaching higher levels of self-awareness through empirical evidence.

To tackle this behavioural story, we are eventually going to be looking at developing a theory of change (ToC) to intervene with this behaviour but first, we need to work through some procedures to equip ourselves to approach it. The process we will adopt is the one outlined in Behavioural Data Analysis with R and Python

by Florent Buisson which is an excellent resource.

However, before we can start working through a method for behavioural analysis, we need to figure out what are our variables.

What data am I using?

We collect timesheet data on our staff from Harvest, it’s best used when anonymised and tells us an interesting story about how people spend their time across internal and external activities. We also collect behaviour and sentiment data through a platform called OfficeVibe. These two data sets together allow us to gain insights into how workload, types of work and project concurrency impact people’s emotions and well-being within the workplace. This provides great potential for analysing factors that impact utilisation and workplace sentiment.

The approach

Please note, this section is very much a digest of Behavioral Data Analysis with R & Python by Florent Buisson. I have made some minor modifications to what I interpret their workflow to be but I take no credit for the methods discussed.

So far we have explored some of the conceptual foundations and how they can apply to analytics, which seems brilliant! This however leads us to question how we approach it… Luckily Buisson has put forward a great method which looks something like the workflow below:

Using this method we can approach analysis in a methodological and process-driven fashion.

  1. Our first step is the frame the business goal. We’re going to be looking into the goal to “Increase the level of certification within our business”

  2. We then need to identify our key metric to measure this. In our instance, it is “rate of certification against target”

  3. We then need to create our causal diagram (CD). There’s a method to doing this. The approach I have taken took three iterations. I first of all mapped out all of the variables I am aware we have data points for.

I then looked at correlations between these variables.

4. Once we have an understanding of the correlations, we can refer back to our CD.

By matching our variables with the correlation matrix, we can start to prioritise which variables are likely to allow us the influence our key metric (rate of certification against target) and then rationalise our CD.

5. Now that we have iterated through our CD and formed some opinions about how to best achieve our goal we can plan our intervention. I found a tight correlation between overall feelings of satisfaction, delivery headcount and the number of hours spent on certification (surprise, surprise). These are levers that I can now pull to try and improve our key metric. To avoid compounding variables, I’m only going to make one change which is to ensure we up headcount. Having a control and intervention group with employees feels a bit mean so we won’t deploy that in this instance…

6. We now need to monitor the results of our intervention over time. An example of this would be an uptake in certification against a target or a tighter correlation.

Steps 1 through 5 have also helped us to establish our theory of change:

Using the process and methods above, we can systemically build behavioural models that allow us to better influence the behaviours of end-users to achieve our organisation’s goals

What did I find out?

I found out that there’s a close correlation between feelings of satisfaction and time spent training. It does raise the question, does workplace satisfaction encourage learning or does learning lead to workplace satisfaction or do both lead to each other? Either way, further promoting this emotion is a good lever to pull. It also brought into focus the importance of adequate staffing, yet another lever to focus on. It also got me thinking about other variables I could include in our behavioural logic such as talking about it in our daily stand-ups, something that’s harder to measure.

HOW THIS APPLIES TO THE MODERN DATA STACK

So far, we have established a top-level view of behavioural science and a layman’s approach to it. What is interesting to me is how these concepts and processes could fit within a data warehouse.

We are provisioning for the causal layer of the analytics hierarchy by having what we call the methodological layer as part of the warehouse architecture. The architecture above looks to leverage the semantic layer to call out behavioural variables for analysis bringing you higher up the analytic hierarchy. The aim of this would be for causal analytics to fit within a standard warehouse architecture, making use of the ontological layer and semantics to feed behavioural variables into the shape of our causal diagrams. This methodological layer should not be restricted to behavioural data and the aim would be to extend it into multiple disciplines all of which impose methods on business entities, moving us away from over-indexing on descriptive analytics.

Future articles may look into using https://github.com/LewisCharlesBaker/droughty to provide an interpreter for behavioural grammar.

The counterargument and the tempering of behaviouralism

In recent years, behaviouralism has fallen under scrutiny as behaviouralists commonly used animals in experiments under the assumption that what was learned using animal models could, to some degree, be applied to human behaviour. These assumptions fell when the original research that supposedly validated the arguments of behavioural had consistently different results when repeated. Critics of behaviouralism use this data to validate an important argument against the impact that nudging individuals by methods such as the cue, routine and reward system has when applied. This puts forward the notion that behaviouralism is perhaps scientistic rather than scientific. Furthermore, what tempers the power of behaviouralism is the ability for individuals to make their own choices through a conscious effort which is something the behaviouralists have been seen to overlook.

In conclusion…

The concepts and research within both behaviouralism and behavioural science offer valuable lenses we can use to model, test and act on our data. Even having a rudimentary understanding of these disciplines provides a way of thinking about data that can unlock potential from data that we would otherwise be unable to access. The case study of certification within our company, it’s enabled me to think about how business behaviours impact our employee behaviours and how we can modify these variables in a methodological and controlled way to better achieve our business goals (and most importantly, to help them learn!).

¹ About Behaviorism — Skinner, B.F.

²Beyond Freedom and Dignity — Skinner, B. F.

³ Behavioral Data Analysis with R & Python — Florent Buisson

⁴ How the Mind Works — Steven Pinker

⁵ Designing for Behavior Change: Applying Psychology and Behavioral Economics — Wendel