Two related biases can lead us to believe that we’re getting a good answer to a good question, when what we’re really getting is an arbitrary answer to an irrelevant question.
Both biases relate to metrics – the numerical attributes associated with our data records.
- Existing metrics must matter to the questions I want to ask (sunk-cost)
- A computer-generated numerical result is more objective than human judgment (related to automation bias)
These biases might have no impact, if our available metrics were always pertinent to our most important questions. However, while the metrics we naturally have on hand are convenient to calculate – things like counts, sizes, costs, prices, and rankings – what we often require are metrics describing value. Value metrics are much harder to develop – just ask the sabermetrics guys.
That presents a real problem, but there are valid responses. We might invest in developing the metrics we need – sometimes a very approximate value metric still allows conclusions and decisions. Or, we can trade-in the original value-based question for another one answerable within the confines of our existing data.
But driven by bias and imprecise question formulation, what I’ve often encountered is this: mentally, we substitute the metrics we have for the value metrics we want. Essentially, we start to believe that by storing a metric that allows us to say A > B, we can also say that A is more valuable than B.
I’ve seen: employees evaluated by cost (particularly when the issue is layoffs), products evaluated on gross sales, databases evaluated on record count, new-product ideas evaluated on potential while ignoring development difficulties, proposals evaluated on average reviewer scores, and promotions and hirings based on evaluator scores.
Unfortunately swapping metrics isn’t valid. It pushes the hard work of metric development under the analytics rug. Although our process might feel objective and ordered due to automation bias, in my experience it has usually become arbitrary. For example, an employee is evaluated on cost but this can be overruled when an employee is considered essential – so a kind of subjective value metric (really a constraint) has been inserted into the process. Or, we have a set of scores for promotion candidates, and simply average them – thereby creating a de-facto but unconsidered value metric. (And that leaves aside the fact that most scoring systems I’ve seen are “gamed” by the evaluators.)
We can avoid this trap through a commitment to framing precise data questions, and making the results available to all stakeholders. This can feel uncomfortable, because we’re committing to recognizing data limitations, and to what feels like a loss of objectivity. But the limitations were always real, and the objectivity probably wasn’t. Regardless, a well-defined question with clear assumptions is a good way to know. I’ll post a process I like for formulating data-based questions shortly.