Psychological Safety in Data Science Teams (Part 1): The Overlooked Factor for Enhancing Machine Learning Models
Psychological safety in Data Science teams is the link between team dynamics, communication, decision-making and machine learning model performance. Data leaders must enable a psychologically safe environment to unlock the full potential of their teams and achieve exceptional business outcomes.
The field of Data Science, although — at least under this term — relatively young, has already seen significant growth and transformations. During the last years, we have engaged in extensive discussions about the needed skills of Data Scientists (Rose, 2019), the actual essence of Data Science (Ramanathan, 2016), its optimal integration within organizations (van der Heuvel, 2021), and the best practices for deploying machine learning (ML) models in production.
Now, the demand for Senior Data Scientists to assume leadership roles is on the rise. As leaders of tomorrow’s Data Science Teams, it is crucial for us to reflect on what our Data Scientists truly need to thrive and deliver value to our companies. One such critical factor is psychological safety, which refers to the collective belief that team members feel safe when they take risks.
The presence of psychological safety within a team can significantly influence its dynamics and overall performance (Edmondson, 1999). In fact, I would argue that the efficiency and effectiveness of productive ML systems relies not only on technical and methodological aspects, but also on the leader’s ability to cultivate psychological safety. And here is why.
Psychological Safety as Predictor for Team Performance
There have been several studies that suggest a positive relationship between psychological safety and team performance (for review see: Newman, et al., 2017). The most famous study was conducted by Google under the name “Project Aristoteles” (Google, 2017). They found that psychological safety was the most important factor in determining team performance. Surprisingly, factors like seniority, individual performance levels, or team size did not have the same influence.
Psychological safety describes a shared belief among team members that it is safe to take (interpersonal) risks, such as speaking up, asking questions, and sharing ideas, without fear of negative consequences. For example, in a psychological safe team, a team member proposes a new, unconventional idea, and the others respond with openness and curiosity instead of dismissing it right away. Psychological safety is somewhat related to trust but has a distinct focus. Trust describes the confidence of team members to rely in each other’s abilities and intentions, and involves the belief that team members will fulfill their commitments and support one another. For example, in a trusting team, if a team member is assigned a critical task, others feel certain that this task will be completed eagerly and on time. While psychological safety relates to the emotional comfort and freedom team members feel to share ideas, admit mistakes, and ask for help, trust focuses on the confidence team members have in each other’s actions.
But why should you care as Data (Science) Leader?
There may not be a direct causal relationship between psychological safety in data science teams and the performance of ML models in production, but the results of several studies suggest that there might be an indirect effect on model performance through its impact on team dynamics, communication, and decision-making. For example:
1) Collaboration and knowledge sharing: Psychological safety can lead to more effective collaboration and knowledge sharing among team members, which can potentially result in better model development, testing, and deployment practices, finally leading to improved model performance in production (cp. Woolley et al., 2010).
Imagine a team of Data Scientists where everyone keeps its insights and experiences about certain model configurations to themselves? How would the ML model perform in production? At its real potential?
2) Experimentation and innovation: Psychological safety can create an environment where team members feel comfortable taking calculated risks, trying out new approaches, and experimenting with different techniques and algorithms. This can encourage innovation and new ideas, which can lead to the development of more impactful machine learning services (cp. Stollberger et al., 2017).
We all experienced the situation, where we train again and again a model, and are not sure if its good enough to go productive. What if we could discuss in our team the potential risks to do it anyway? Maybe even a “low performing” machine learning model is much better than the status-quo ? (I would argue that in most companies it is like that.)
3) Error reporting and learning from failures: In a psychologically safe environment, team members are more likely to report errors and failures without fear of blame. This can facilitate a culture of learning, allowing the team to identify and address issues and iterate on model development processes, leading to more robust and reliable models in production (cp. Edmondson, 2003).
Imagine you made a “terrible” mistake during training you ML model. Let’s say the k-fold-randomization you tried, didn’t worked as expected (e.g., test data were included in the training data …). You fix it, but don’t tell your team…
Reaching the real potential
To be clear, I am not saying that psychological safety is the main factor for model performance. There are also other moderating factors, like team diversity or psychological empowerment. And of course, we should not forget other technical and methodological aspects, such as data quality, model architecture, and deployment practices. They play crucial roles. But experienced psychological safety might determine if you service is reaching its real potential in the form of business impact. Of course, more empirical studies are needed to better understand the specific relationship between psychological safety and ML model performance in production. However, looking at the data so far, I would not expect it to be the other way around.
As a leader in the Data Science domain, it is your responsibility to enhance this experience by showing inclusive behaviours, asking for feedback, and valuing input from team members and to create a team atmosphere where everyone can belong (Nembhard & Edmondson, 2006). And believe me, it sounds easier as it is.
In the following articles, I will share methods that you as leader can employ to create a psychologically safe environment for your team. But in the next article, we must talk about the biggest problem: the Gremlin of each Data Scientist , the imposter syndrome.
Stay tuned.