Is Cognitive Bias Impacting Your AI Outcomes? Decoding The Decoy Effect

AI models have been more successful in a world of machines than in the world of humans. For example, algorithms can predict a machine failure better than predicting if a customer would buy the recommended product. The main reason for this is the various cognitive biases and psychological aspects that impact human decision-making.

Even the most seasoned data scientists are susceptible to these cognitive biases; understanding and identifying them can help improve the data science models. They need to be able to consider these biases (when existing) and the context of the behavior in the models for better prediction.

There are some common cognitive biases, such as anchoring bias, confirmation bias, logical fallacy bias, cognitive dissonance, cognitive framing effect, etc., that can interfere with your insights. In this blog, we will talk about the decoy effect, which is part of the framing effect. The decoy effect became prevalent, and many top companies use decoy effect pricing strategies to improve their revenues.

What is Decoy Effect?

First, let’s have a look into what a decoy effect is with a well-known example mentioned by Dan Ariely, Professor of Psychology and Behavioral Economics at Duke University, in his book Predictably Irrational. Dan spotted something peculiar on The Economist’s membership website a few years ago. An online subscription, a print + web subscription, and only a print subscription were all available from the magazine. What was interesting was the print only and print+web options pricing.

Ariely gave these options to MIT’s Sloan School of Management students. They opted as follows:

Web Only: 16% students
Print Only: 0% students
Print+Web: 84% students

They saw the advantage of the print and web offers over the print-only offer. Ariely wanted to check if they were influenced by print only (Let’s call it decoy) option. Then he gave two options to choose from by removing the decoy.

Surprisingly, students didn’t react in the same way as earlier. They opted as follows:

Web Only: 68% students
Print + Web: 32% students

What changed their minds? Nothing rational. The presence and absence of a decoy made them choose differently. Including a similar but inferior third option can tip people’s preference towards the slightly better option. Consumers tend to be drawn to things that seem better than other options in their selection (a decoy) and to things that are perceived as a middle ground between other possibilities. We also tend to focus on comparing things that are easily comparable and avoid comparing things that cannot be compared easily. In the above example, we always compare print-only options with print+web.

Decoy pricing is quite popular with many organizations and marketing teams. Though all companies don’t use this strategy, many do. Now as a data scientist, how do you consider cognitive biases like the decoy effect while building marketing or customer behavior models? How do we bring the context of consumer behavior to the model?

There is ongoing research happening in behavioral economics to model decoy effects, and we borrow some of those methods for marketing. For example, Random Regret Modeling (RRM) can be used in conjoint analysis to predict the choice. RRM model assumes consumers make choices to avoid the potential regret of missing something rather than choosing the option with the best combination of product characteristics that suits them.

To consider decoy effects in the models, we need to collect relevant data. Check if the marketing team has any such targeted campaigns with a decoy. In the case of agent-driven sales (for example insurance), check if the team is capturing all the options in the CRM system that has before presented to the consumer during the sales conversation. Capturing this information can help in either removing the bias completely from data with the decoy effect or introducing new features in your model to understand how different segments and different consumers will react to decoy, hence modeling the behavior.