Post No.: 0702
Furrywisepuppy says:
We had learned from Fluffystealthkitten in Post No.: 0689 that inconsistent verdicts don’t cancel each other out. So what can we do to tackle the noise or unwanted variability in professional assessments or evaluations? Some noise is inevitable but we should still aim to minimise it.
The subjective values of the individual assessor usually get mixed in with the objective facts of the case being assessed. So keeping them separate is paramount – accuracy should be our only goal as a judge.
Organisations need to conduct regular noise audits to reveal and understand the extent of the unwarranted variability that plagues their systems. This’ll be more challenging for some organisations or institutions than others though e.g. justice compared to insurance because it’s easier to determine whether two insurance applications are the same versus two crimes.
We might never be able to eliminate our natural variability in mental performance from one day to another e.g. our memory just sucks some days but is dandy on others. It’s just like our basketball free throws don’t always sink despite seemingly the same objective and variables being at play, because our muscles don’t fire in the exact identical way each time. But we can strive to control what we can, like by going through the same routine before each throw or decision.
We might, in many cases, not know what the ‘correct’ answer should be, like the right length of sentence for a specific crime, because there is no objective answer. (There’s not even an objective answer to say how much variability is ‘acceptable’ e.g. 11%, 15%, 12.642%?) But we can recognise that inconsistency is unfair. So there might not be objectively right answers for a lot of questions – but we should aim for consistent answers whatever we (together) determine them to be.
Utilising the wisdom of a diverse crowd of independent minds, implementing rules, rubrics or guidelines (like Apgar scoring for assessing newborns), specifying standards, or using formulas or algorithms, are a few possible solutions.
However, a diverse crowd of independent-minded people might just bicker amongst themselves when it comes to trying to come to a consensus when an aggregate answer cannot be simply based on the average of their opinions. Badly-written algorithms or badly-trained machine-learning AIs can be (highly) systematically biased. And decision-makers often object to having their own discretionary power overridden too. Judges argue that they need to apply judicial discretion because every case is individual.
Yet even if every case is ‘individual’, like being more lenient on the sentence or fine dealt to someone who has dependents (whether this is itself fair or not) – we should still aim to deal consistent sentences or fines to all others who are similarly situated in every pertinent way.
But this doesn’t happen. Even when different judges evaluate the exact same defendant and exact same case in experiments – they can come to different verdicts! The exact same crime, and even the exact same precise case – but different judges coming to different verdicts or punishments; or even the exact same judge on a different day coming to a different decision depending on how they felt that particular day! It’s a verdict lottery, yet people will attempt to post-rationalise the variability as justified. (Rationalisations are ‘system two’ processes hence this highlights that system two can produce errors too.) Moreover, humans are frequently biased. AIs can indeed be biased too (although it’s often only because they’ve been taught from past human decisions and trained by datasets selected by humans) – but, in many ways, it’s much easier to tweak an AI to become less biased than to train humans to become less so.
Regarding ‘rules’ or guidelines – it’s noisier with qualitative than quantitative assessments, hence guidelines that still contain too many qualitative assessments can remain noisy. This is a common quandary in psychiatry – different schools of thought, training, clinical experiences and interview styles can result in a variation in qualitative assessments. The main reason is that the diagnostic criteria for some disorders are still vague and it’s tricky to concretely define something complex that’s not yet fully understood. More standardised diagnostic guidelines would still help though. Now even outside of mental health – variability in diagnoses from different specialists when they’re looking for signs of cancer from the exact same x-ray images exists too, for example. When different medical practitioners give different diagnoses in medical contexts – and this happens frequently enough – it highlights the amount of unwanted variability in the system.
You could try to quantify and use objective metrics to represent different qualitative attributes but the risk is poor proxies. It’s like using ‘the quantity of patients seen’ as the measure of a hospital’s ‘quality of service’. Incentivising the wrong metrics can also lead to problems where hospitals will game the figures to maximise the number of patients seen by, ironically, giving each patient a rushed and thus poor-quality service. Or it’s like not all customer orders or queries are as quick or easy to serve as others hence counting them all as ‘one customer served’ for the purpose of giving a bonus to the telephone clerk who serves the highest number of customers each day would be unfair, as well as incentivise a potentially rudely hurried customer service.
‘Standards’ more vaguely state a level of quality that one must attain, and they are often implemented as a practical matter because it’s not always possible to specify a comprehensive enough set of explicit and unambiguous commands or rules. Highly-specific laws are clearer but can potentially be more easily circumvented – whereas more vaguely-expressed laws can cover unexpected events to prevent loopholes. This is like when drafting patent claims – you want vaguer claims that cover as wide a scope for your invention as possible so that other people will have less chance of finding a way to circumvent what you’ve protected, if you can get those claims through. Clear rules are most ideal and we should strive for black-or-white rules and laws but it’s just not always practical. You can end up with using a combination of both standards and rules though. And standards can eventually evolve into rules, and vice-versa, to reduce misunderstandings or as needs must.
Regarding the ‘wisdom of a diverse crowd’ of independent minds – the result of different people holding different positive and negative, implicit and explicit, biases is noise. But averaging out multiple independent views reduces noise, and also bias as a bonus too. The average answer from a group of diverse judges (as long as their views are each independently gathered otherwise this diversity will be lost) will likely be better than the average answer from a bunch of judges who share similar views with each other. This is because the latter might produce a skewed or biased overall answer despite the variability between the members of the group being lower initially. For a simple illustration, let’s say the correct answer for some estimation task is 10. A diverse group produces guesses of 3, 6, 14 and 17 (high initial variability but the average is 10), whereas a less diverse group produces guesses of 11, 12, 12 and 13 (low initial variability but the average is 12). The average guess of a noisy group can thus be more accurate than a unanimous set of guesses.
But along with the diversity and independence of the contributors, and having a sufficiently large enough number of contributors – each contributor to the ‘wisdom of the crowd’ does need sufficient expertise too. Everyone with eyesight and a fleshy body has the life experience to roughly guess the mass of a cow of a certain volume – but if you’ve no idea what a quasar is then can you guess how many are in the Milky Way at present? The average answer from laypeople, even if independent-minded and from heterogeneous backgrounds, won’t be close to the true answer of about nought. It was kind of a trick question but it shows that sufficient expertise matters too. (This may call into question the reliance on public referenda, at least when a question concerns a complex social and economic subject and people don’t put in the effort to do their homework or they listen to fake news sources?)
Notwithstanding – diversity pays if it can be harmoniously managed. An initially diverse group, if we aggregate the results of every member, can lead to both less noise and less bias. On Wikipedia, according to the site’s own internal metrics, the most ideologically-polarised teams of editors produce higher-quality pages than homogenous teams, especially for political articles. Diversity boosts creativity too because more ideas bounce around. It’s more soothing when it seems like everybody agrees with us, hence we tend to gravitate towards others who already think the same way as us – but opposing voices can point out the errors of the thinking processes and conclusions of each other. Healthy competition pushes innovation and product quality in the marketplace. A risk is that diverse groups can lead to cacophony and discord, but if everybody shares a common goal then they should be able to find a way to put aside their fluffy differences to collaborate.
To a lesser extent, if we average out our own judgements made on different days, we’ll more likely get closer to the right or optimal answer too – especially if we actively consider reasons why our first estimate was wide of the mark (the ‘crowd within’). Don’t average your guesses if there’s a reason why your latter guess alone should be more accurate though e.g. because you’ve considered additional information during your second guess that wasn’t available when you made your first. Better though is to ask for an independent second opinion from someone else.
People who are relevantly skilled, qualified or are higher-IQ critical thinkers, who make less impulsive decisions, are actively open-minded (they look for disconfirming as well as confirming evidence), intellectually humble (changing one’s mind is not a sign of weakness), methodical and thoughtful (less reliant on baseless intuitions) make better judgements. So-called ‘superforecasters’ possess these kinds of qualities. They consider the statistical base rates for events, and don’t downplay them by presuming (without knowledge of more relevant base rates) that a particular case at paw is ‘a totally unique and special case’ that renders that base rate irrelevant. They update their estimates as they gather new information, and give all estimates in terms of a probability e.g. ‘there’s a 64% chance that x will happen tomorrow’. They know it’s more important for us to listen to those we disagree with than those we agree with. And they constantly strive for improvement – try, fail, analyse, adjust, try again, and repeat. The aggregate view of multiple skilled independent experts is even better, as above with the ‘wisdom of the crowd’.
Someone who is aware of and owns up to their own biases is able to somewhat compensate for them – in a similar way that one can adjust the dial on a set of scales that is known to consistently under-report the weight placed on it by 2kg, or the way that one can simply mentally add that 2kg to every reading manually afterwards.
If you are called to repeatedly make the same kinds of assessments, you could check your thinking against a checklist of biases that you might fall foul of. But keep this list simple otherwise it might get ignored.
You could utilise an independent decision observer too. Consider the Delphi method. And you could sleep on a decision then think about it again when in a different mood.
Woof! We will delve into how some other methods, including algorithms, can be used to minimise noise another time…
Comment on this post by replying to this tweet:
