The loneliness of the penguin: data science and cognitive biases
I have only recently come across an interesting article on The Psychology of Data Science by Lisa Christina Winter. Her contribution deals with the fundamental issue of cognitive biases. In particular, she mentions the confirmation bias, “the human tendency to confirm, rather than disconfirm, pre-existing hypothesesâ€. Since it is one of the most common biases, it may be worth to investigate it deeper.
As Karl Popper insisted in his outline of the scientific method, hypotheses can be falsified (proved to be false) but not verified (proved to be true). This is the case since - no matter how much evidence is collected to corroborate (i.e. to support) a hypothesis - it can never be proved to be true, because we cannot rule out that - sooner or later - some counter-evidence will be found, even just a single observation, which is enough to prove that the hypothesis is false:
- Hypothesis: All birds fly.
- Evidence: doves, crows, pigeons (corroboration).
- Counter-evidence: a penguin (falsification).
Nonetheless, a scientist may unconsciously try to confirm a hypothesis in different ways (if this is performed consciously, then it becomes an ethical problem). The researcher may select a sample which is not properly randomized, and therefore does not represent the general population. If only flying birds are selected (for example because the sample is small), then the study presents a sampling bias. Similarly, in a review study, only papers favoring a hypothesis may be collected (selection bias). Sampling bias and selection bias are almost synonyms, but the former is usually referred to the choice of the data to be collected, while the latter to the choice of the results to be considered.
In other cases, scientists may hold on to the initial hypothesis - no matter how much counter-evidence is collected - because they rely too heavily on the first information they acquire. For example, they may have not been to Antarctica, and so far they have only observed flying birds. This bias is called anchoring, or insufficient adjustment.
Ad hoc hypotheses may be used to anchor to a prior hypothesis:
领英推è
- Hypothesis: All birds fly.
- Counter-evidence: a penguin.
- Ad hoc hypothesis: Penguins are not birds.
Note that hypotheses are usually general, while evidence is particular, as in classic deduction. Ad hoc hypotheses are not necessarily false, but their adoption in order to make a hypothesis more robust to falsification may imply a confirmation bias. To mitigate this risk, explanations should be kept as simple as possible, and the number of necessary hypotheses minimized (adopting the so-called Ockham's razor).
Similarly, judges may confirm the presumption of innocence, even if the collected evidence is against the suspect, according to the general assumption that in democratic countries a free criminal (false negative) is preferred to a convicted innocent (false positive). A famous argument used to confirm a hypothesis is the No True Scotsman, where an assumption is changed ad hoc.
The opposite of anchoring - the base rate fallacy - may happen in medicine. Doctors may hospitalize a patient who shows some symptoms of an illness, even if its prevalence (the base rate or relative frequency) is very low in the general population (i.e. the disease is rare). The base rate is more likely to be neglected if the suspected illness is life-threatening. In an article on Psychology Today, JI Krueger suggested that the two opposing biases (anchoring and base rate fallacy) could work together to mitigate each other.
Nonetheless, the confirmation bias seems to be intrinsic to classic deductive approach to inference, which relies on a priori hypotheses (and therefore on assumptions or prejudices). Thus, part of the problem (and of the solution) may be methodological, not only psychological. For example, exploratory data analysis (i.e. hypothesis finding) can be performed before confirmatory data analysis (i.e. hypothesis testing) in order to evaluate new hypotheses which are suggested by the data themselves.