Leveraging Complexity Science to understand Human Behavior

Final Mile
7 min readDec 21, 2021

--

Part 4 | Nudge Bias: Problem with centralized interventionist approach

By: Alok Gangaramany and Anushka Ashok

In a recent article, Cass Sunstein points to the distributional effects of nudging. The article suggests that while some nudges may have an overall positive outcome they may end up hurting a sub-group. It goes on to point the need to predict the consequences in advance to determine the groups that are more likely to benefit or less likely to be harmed. And use this understanding towards targeted and personalized nudges. This reflection is a welcome change to the mass nudge-based approaches that we see in global health and other population level interventions. From our own experience and reflections over the years, we believe the nudge approach needs to address some fundamental issues. In this post, we will attempt to discuss these issues particularly those related to group-level generalization, individual focus, prediction and centralized nudge approach.

The Ergodicity Problem

The focus on average effects of an intervention assumes that it captures the directional change of a significant (if not all) majority of population over a period of time. However, this particular assumption can be a serious issue. In the paper “Lack of group-to-individual generalizability is a threat to human subjects research”, discuss this issue in detail. They refer to the “Ergodicity” aspect of the system that is frequently missed. A system would be considered Ergodic if the effect of a situation or an intervention is homogenous amongst the population and more importantly stable over a period of time. Let’s use an example articulated in Nassim Taleb’s book Fooled by Randomness to understand the concept of Ergodicity.

Consider six people who are playing Russian Roulette (deadly game where an individual pulls the trigger of a revolver with a single round and wins if they survive) for a million dollars. When each member of the group pulls the trigger one at time (sample of n = 6), the odds of success (that is surviving and winning million dollars) of the overall group is fairly high. Essentially, all those who survive will become rich. However, if the same game is played by an individual six times (to maintain the same sample of n = 6) then their survival odds diminish significantly. The individual will most certainly die prior to completing their six attempts.

That is the game is non-ergodic where the average rate of survival of the group and individual over a period of time is not consistent. The inconsistency over a period of time is key. Intuitively and something discussed extensively in behavioral science literature, we know that people are not consistent or stable over a period of time. An intervention may increase saving rates for a subset of population today. But next month the increased saving rates could be achieved by a different subset of population leading to the same average effect. However, at an individual level the impact may be minimal. The heterogeneity and time inconsistency amongst humans means that effect of any intervention (nudge or otherwise) will be distributed.

Another issue with grouping heterogeneous populations is that the patterns we observe depends extensively on how we group the population. This is sometimes termed as Simpson’s paradox. For example one of the argument (or misinformation) that is being used against COVID vaccination is that the unvaccinated groups have a higher survival rate compared to thee vaccinated groups. As explained in this video here, The key issue here is vaccinated group is disproportionately in the older age group category. The older group will always have a lower survival rate. Once we break down the groups into different age brackets (e.g. 20–30, 30–40 year old’s etc. ), we will see that the story is completely opposite.

Predicting the Unpredictable

Our earlier article has already alluded to the issue around predicting complex phenomena that are inherently unpredictable. While in that article we discussed the issue from the perspective of developing a theory of change for complex interventions, the argument is also applicable for predicting impact of nudges. In his article, Sunstein suggests that before implementing a nudge, we should try to identify the sub-groups that are likely to benefit from it in advance. However, the process of identification is a key issue. When we narrowly focus on the effect of intervention on a particular action (e.g. saving rates), we frequently miss second order effects. For example, did increased 401K savings also lead to increased high-cost debts due to reduced cash flows? The challenge is that we may not be equipped to predict these second order effects. Another process issue arises from the inconsistency of individuals over a period of time which makes any prediction short lived. This means that our prediction may need to be continuously updated especially for behaviors and nudges that repeat over a period of time. So if we can’t evaluate a nudge with all its second order effects, how can we forecast its benefits?

Centralization & Scale

While the design of nudges has evolved to start taking into account the context, there still exists a large gap in creating an understanding of the existing system (as a whole and parts, interactions among the system’s parts, linkages, relationships and feedback loops) to design interventions. Furthermore, a crucial aspect that is typically missed is who is responsible for the design of nudges and at what scale.

As we have discussed earlier, nudges like increase savings can have second order effects for individuals. Imagine if the nudge was centrally designed by a “Nudge Unit” at a nation’s capital and rolled out at a national level. By doing that we would have essentially increased the scale of both the immediate impact (that may or may not be positive) but also second order and long-term impacts that may cause significant harm. In situations with high level of uncertainty around possible outcomes especially due to heterogenity within the population, we are likely to deliver more harm than good. A decentralized approach, where interventions are designed and executed at local community levels in a relatively independent manner, may be a far better approach to mitigate such issues.

Focus on Individual Decision-making

Finally, we know that our behaviors and decisions are significantly influenced by physical and virtual interactions with others. Societal norms can have significant influence on our everyday decisions including health (e.g. addressing pandemics ), employment (what kind of jobs are valued), economic (investments, savings and debts behaviors), environmental (reuse, recycling, organic produce etc. ) , and other cultural factors (marriage age, birth rates etc. ). However, most nudges tend to operate at an individual level and the whole community or system-level change is conceived simply as a matter of ‘‘aggregating up’. From experience, we know that such a linear problem solving approach can be problematic in non-linear settings with varied contexts. As a result, we frequently experience failure of popular nudges that may have been successful in one place but fail miserably when they are scaled up or applied in other contexts.

Implications

The above challenges suggest asking a few important questions:

  • Are nudges an effective way to address the complex social and behavioral issues in question? For example, if lack of trust in health care experts is a key issue when it comes to adoption of vaccines, can it be overcome by nudges like lottery based incentives?
  • Are we equipped to design and implement nudges while considering the local context? Do we have the buy-in from the community that we are working with?
  • How can we better deal with addressing second order effects? Do we have a process to eliminate societal level risks, receive rapid feedback and do the necessary course correction?
  • How might we evaluate the effectiveness of an intervention at both individual and group levels especially when the effects may not be consistent over a period of time? How might we capture the context evaluation aspect?
  • How might we allow for longer time frames for follow up to appreciate that changes in complex systems happens non-linearly where we may have long periods with no changes followed by sudden phase transitions?

We recognize that Nudge based approaches will continue to be popular due to multiple reasons. This is apparent from the number of Behavioral Insight Units (BIUs) that have come up across the globe influencing government policies and actions and in some cases successfully. There is a need to incorporate the learning from a decade of application of nudges and evolving our ideas on designing for systemic changes. And it is increasingly important to continue to reflect on the short-term and long-term impacts of nudges and consider implementation models (e.g. decentralized BIUs) that are capable of acting at community levels.

References:

Final Mile brings unique and proven capabilities in addressing complex behavioral challenges. As one of the first Behavioral Science & Design consultancies, Final Mile has had the opportunity to bring these to practice in a wide variety of sectors and contexts. We have executed highly complex behavior change projects across a wide variety of areas covering Global Health (HIV, TB, Maternal Health, WASH), Financial Inclusion, Safety across Africa, Asia, Europe, and the US.

Final Mile is also building a pandemic playbook that can be used as a potential toolkit by policymakers and implementors in mitigating Covid19 and future such pandemics.

Reach out to us at contact@thefinalmile.com.

--

--

Final Mile
Final Mile

Written by Final Mile

Final Mile is a research and consulting firm solving tough and relevant behavioral problems across the globe

Responses (1)