Post No.: 0681
Furrywisepuppy says:
We’ve talked heck aplenty about biases in this blog. But there’s another collection of problems that affects human judgements that’s equally as important to recognise…
Whereas bias is analogous to having all of your shots land consistently off-target in the same direction – ‘noise’ is like having these shots land all over the place. A system can exhibit bias and/or noise, or no bias or noise at all.
As an example – given the same facts, one judge might hand a dognapper a sentence of 1 year, another judge 2, and yet another decides 5. Another example is – given the same facts, one employee at an insurance firm quotes you £534, another employee from the same firm quotes £655, and yet another £701.
Therefore the same evidence can lead to different interpretations or verdicts by different individuals who have different backgrounds, personalities and biases, and/or who experience different moods at different times. Noise is basically the unwanted or unwarranted variability in judgements. Subjectivity, arbitrariness and disagreements lead to noise. It results in a gross lack of consistency or fairness and thus a huge dose of luck depending on whether your case is handled by Brian or Briony.
‘Noise’ means random noise or interference, as contrasted to a ‘signal’ or a true positive of what we’re trying to detect. Here, in this context, you can think of noise as the unwanted scatter that interferes with the signal or the correct, or at least consistent, answer we’re trying to find.
Variability isn’t always unwanted or unwarranted e.g. different people like different coffees, movies, songs or interior designs because they have different tastes – and that isn’t a problem. They’re highly individual judgements. But we expect the same crimes to be punished in a consistent manner. So tackling noise in professional judgements should be just as critical as tackling bias. Noise is overall as impacting as bias in professional judgement contexts. Biases (or alternatively incompetence) aren’t always to blame for errors or injustices. The cognitive heuristics that lead to bias can also lead to noise.
If we’re all, on average, biased in similar ways then there’s systematic bias. If we’re biased in different ways then there’s system noise.
There’s level noise and pattern noise, which together make ‘system noise’, or the total noise in the system.
So you can have the same case that’s judged differently by different people (because e.g. some judges are simply on average harsher than others, or one judge weights intention more greatly than result while with another it’s the opposite) and the dispersion in the average judgements made by different individuals is called the ‘level noise’.
‘Pattern noise’ is when different people regard different kinds of cases as deserving different kinds of judgements, regardless of their own average judgements. There are two subcomponents to pattern noise.
You can have the same person judging particular cases differently because of the stable biases, values or personality traits of that person (e.g. one person is more lenient on those who are of the same colour as him/her yet harsher on those who aren’t, or one person is more unusually softer on women yet unusually heavier on drug dealers). This is ‘stable pattern noise’, which is the largest source of pattern noise.
And you can have the same person judging similar cases differently (because of e.g. his/her particular mood that day, how stressed or fatigued he/she felt or other changeable situational factors at the time). This is unstable or ‘occasion noise’.
The unwanted variation in decisions from the same person (same judge, different similar cases) isn’t as great as the unwanted variation in decisions between different people (different judges, same cases), but it still shows how we can even disagree with ourselves from one case or day to the next(!)
We can measure both bias and noise via statistical calculations. Bias is worked out by calculating the average of all errors. Noise is worked out by calculating the standard deviation of all measurements.
To recognise bias, you’ll need to somehow know what the correct answer should be (where the bullseye is supposed to be) – but all you need to recognise that noise exists is the presence of scatter; although this requires being able to see multiple results in aggregate, which we can’t always e.g. because we only hear the sentencing decision of one judge per case and not what other judges would’ve decided if they had handled that case instead.
Noise is typically neglected despite its perverse effect on fairness in both public and private sector contexts. It occurs amongst sports referee verdicts, job interview outcomes, deciding the length of jail sentences, DNA, fingerprint and handwriting analyses, welfare benefit decisions, teachers’ grades, university admissions, doctors’ diagnoses, market forecasts, insurance premiums, asylum decisions, child custody decisions, patent award decisions… wherever there’s human judgement there’s noise, and far more of it than you may realise. And the consequences are often unfair and/or costly. We underestimate how much noise exists – in part because we accept humans as being capricious decision-makers.
Evaluative judgements include a doctor’s diagnosis. Predictive judgements include a doctor’s prognosis. Forecasts that come with lower predictability imply greater noise.
Noise is less salient than bias but is just as problematic. We tend to attribute errors as down to biases rather than noise because it’s easier to generate a causal story about them e.g. ‘he/she made a decision that I disagree with because of his/her politics’. We assume our own judgements are objectively fair and if others differ from us then they are biased or incompetent at their job, not us! We also assume we’re always consistent with ourselves and never swayed by irrelevant factors like how hungry we feel or whether it’s sunny at a particular moment. Hence we don’t suspect noise. Yet people do often recognise situations like ‘I hope I get this driving examiner instead of that other one’ – it shouldn’t matter who judges our cases, yet in our experience we know it can.
In TV dance competitions or sports like boxing or weightlifting, the exact same dance, round or lift can lead to different scores/lights from different judges. They appear to be judged by a panel of independent judges and it’s the combined or average of these scores that’s used though, which greatly reduces noise and improves scoring accuracy (although not always perfectly because even entire panels can occasionally miss things). But a student doesn’t receive an average grade on his/her coursework based on what a panel of teachers think, thus depends on being lucky that his/her particular teacher isn’t prejudiced against him/her or marks the coursework on a bad-mood day. We can shop around for the best insurance quote – but we can’t shop around for the most lenient trial judge!
Football fans are acutely aware of the incredible amount of frustrating inconsistency in referee decisions! But most fans just haven’t called this problem what it is – noise. Fans and pundits notice inconsistencies and are livid about its unfairness every time they spot similar kinds of tackles being judged differently depending on whether they were committed near the halfway line or in the penalty box, or when the same kinds of tackles are being judged differently on different occasions even when they’re all in the box. Whenever different people watching the exact same events come to different verdicts on whether something was a foul or not, or whether it deserved a yellow or red card, they’re highlighting the noise in such decisions. Woof!
Although it may have reduced some unwarranted spread in judgements, even the introduction of VAR (video assistant referee – which, in its current form, is basically a bunch of camera replays and some off-site referees advising the match referee to check his/her decision for certain decisions) has, so far, hardly eliminated fuzzy noise. This is because capricious, inconstant humans, rather than algorithms, still make the decisions at all stages of the process. Some rules are subjective like determining what’s a ‘clear and obvious’ error. Others like the offside rule are completely objective but the technology isn’t available yet to make accurate enough assessments, and humans still ultimately draw the offside lines and make the final verdicts. Allowing more time for referees to make decisions hasn’t completely solved the problem either.
Some fans argue that disagreements create talking points – but inconsistent decisions are manifestly unfair, and that’s arguably a far worse problem than taking away the discretion of human decision-makers. It does however precisely reveal, in our everyday lives, how much of what we each personally think are objective judgements are in fact subjective opinions based on our own perceptions e.g. when it comes to how we each decide and interpret what’s sufficient ‘contact’, ‘force’ or ‘intent to harm’ in an attempted tackle. It’s not always a matter of individual bias because even neutrals have differences in perception and interpretation. And our subjective biases, perceptions and interpretations speak more about us than the world we’re judging.
Noise requires taking a broader picture to notice the range of various judgements made on essentially identical issues. It requires statistical thinking rather than single-case thinking – like taking into consideration how different judges differ in their political biases and how it depends on luck which judge you get for your particular case.
Noise exists in singular judgements too however – like when your country was considering its first ever COVID-19 lockdown. The trick is to just imagine them as only the first judgement of many.
So non-recurrent judgements can still exhibit noise but we just can’t see the alternative judgements that could’ve been made had a different decision-maker judged it. It’s like shooting once from a gun that may be inconsistently inaccurate but we won’t ever know unless we take some more shots to identify the scatter. Such decisions include a particular President deciding to go to war – we won’t know if different Presidents would’ve done the same? We’ll never know how noisy this decision was. We need to imagine ‘could this judgement have been different?’ And if the answer is affirmative then there’s noise.
Non-verifiable predictive judgements are like guessing that there’s an 8% chance of losing the war. Only predicting a 0% or 100% chance can really be falsified. Okay you guessed there’d be an 8% chance of us losing a particular war. We lost that war. Does that mean your probability was incorrect? 8% chance events do happen. One-in-a-trillion events do happen for that matter. We can’t repeatedly turn back the clock to re-run that war several times to check the accuracy of your probability.
In these cases, we can instead assess the quality of the judgement process. The problem is, most organisations try to assess judgements based solely on their results rather than their processes. We can test various processes to find out which one produces the most accurate judgements over an ensemble of cases. We’ll explore in more detail how we can cur-tail noise in a forthcoming post.
…Human decision-makers really love and want to preserve their ability to exercise personal discretion – whereas clear and rigid rules would allow everyone to know where they stand. (Many will argue that the way the race director exercised his discretion when applying rules controversially decided the 2021 Formula 1 Championship.) More autonomy in making decisions leads to greater job satisfaction – but should the happiness of decision-makers come at the expense of the fairness for those affected by their decisions? People may want to be treated as individuals rather than as part of an indistinguishable mass, but this increases noise and arbitrary idiosyncrasy – similarly situated individuals need to be treated similarly!
Woof. If you’re starting to think more deeply about the magnitude of unfairness that results from fickle, inconsistent judgements, and believe that it’s a problem that should be eliminated as much as possible instead of accepted as a part of life, then you can share your thoughts, and maybe own experiences, via the Twitter comment button just below.
Comment on this post by replying to this tweet:
