Author: bs king

Human Bias vs Machine Bias

/ bs king / 3 Comments

One of the most common running themes on this blog is discussions of human bias, a topic I clearly believe deserves quite a bit of attention. In recent years though, I have started thinking a lot more about machine bias, and how it is slowly influencing more of our lives. In thinking about machine bias though, I’ve noticed recently that many people (myself included) actually tend to anthropomorphize machine bias and attempt to evaluate it as though it was bias coming from a human with a particularly wide-spread influence. Since anthropomorphism is actually a cognitive bias itself, I thought I’d take a few minutes today to talk about things we should keep in mind when talking about  computers/big data algorithms/search engine results. Quite a few of my points here will be based off of the recent kerfluffle around Facebook offering to target your ads to “Jew haters”, the book Weapons of Math Destruction, and the big data portion of the Calling BS class. Ready? Here we go!

  1. Algorithm bias knows no “natural” limit. There’s an old joke where someone, normally a prankster uncle, tells a child that they’ll stop pouring their drink when they “say when”. When the child subsequently says “stop” or “enough” or someone other non-when word, the prankster keeps pouring and the glass overflows. Now, a normal prankster will pour a couple of extra tablespoons of milk on the table. An incredibly dedicated prankster might pour the rest of the container. An algorithm in this same scenario would not only finish the container but go run out to the store and buy more so they could continue pouring until you realized you were supposed to say when. Nearly every programming 101 class starts at some point with a professor saying “the nice part is, computer’s do what you tell them. The downside is, computers do what you tell them”. Thus, despite the fact that no sane person, even a fairly anti-Semitic one, would request an advertising group called “Jew haters”, a computer will return a result like this if it hits the right criteria.
  2. Thoroughness does not indicate maliciousness. Back in the 90s, there was a sitcom on called “Spin City” about a fictional group of people in the mayor’s office in New York City. At one point the lone African American in the office discovered that you could use their word processing software to find certain words and replace them with others, so in an attempt to make the office more PC, he sets them up to replace the word “black” with “African-American”. This of course promptly leads to the mayor’s office inviting some constituents to an “African-American tie dinner”, and canned laughter ensues. While the situation is fictional, this stuff happens all the time. When people talk about the widespread nature of an algorithm bias, there’s always a sense that some human had to put extra effort in to making the algorithm do absurd things, but it’s almost always the opposite. You have to think of all the absurd things the algorithm could do ahead of time in order to stop them. Facebook almost certainly got in to this mess by asking its marketing algorithm to find often-repeated words in people’s profiles and aggregate those for its ads. In doing so, it forgot that the algorithm would not filter for “clearly being an asshole” and exclude that from the results.
  3. While algorithms are automatic, fixing them is often manual. Much like your kid blurting out embarrassing things in public, finding out your algorithm has done something embarrassing almost certainly requires you to intervene. However, this can be like a game of whack-a-mole, as you still don’t know when these issues are going to pop up. Even if you exclude every ad group that goes after Jewish people, the chances that some other group has a similar problem is high. It’s now on Facebook to figure out who those other groups are and wipe the offending categories from the database one by one. The chances they’ll miss some iteration of this is high, and then it will hit the news again in a year. With a human, this would be a sign they didn’t really “learn their lesson” the first time, but with an algorithm it’s more a sign that no one foresaw the other ways it might screw up.
  4. It is not overly beneficial to companies to fix these things, EXCEPT to avoid embarrassment. Once they’re up and running, algorithms tend to be quite cheap to maintain, until someone starts complaining about them. As long as their algorithms are making money and no one is saying anything negative, most companies will assume everything is okay. Additionally, since most of these algorithms are proprietary, people outside the company almost never get insight in to their weaknesses until they see a bad result so obvious they realize what happened. In her book Weapons of Math Destruction , Cathy O’Neill tells an interesting story about one teachers attempt (and repeated failure) to get an explanation for why an algorithm called her deficient despite excellent reviews, and why so much faith was put in it that she was fired. She never got an answer, and ultimately got rehired by a (ironically better funded, more prestigious) district. One of O’Neills major take-aways is that people will put near unlimited trust in algorithms, while not realizing that the algorithms decision making process could be flawed. It would be nearly impossible for a human to wield that much power while leaving so little trace, as every individual act of discrimination or unfairness would leave a trail. With a machine, it’s just the same process applied over and over.
  5. Some groups have more power than others to get changes made, because some people who get discriminated against won’t be part of traditional groups. This one seems obvious, but hear me out here. Yes, if your computer program ends up tagging photos of black people as “gorillas”, you can expect the outcry to be swift. But with many algorithms, we don’t know if there are new groups we’ve never thought of that are being discriminated against. I wrote a piece a while ago about a company that set their default address for unknown websites to the middle of the country, and inadvertently caused a living nightmare for the elderly woman who happened to own the house closest to that location. This woman had no idea why angry people kept showing up at her door, and had no idea what questions to ask to find out why they got there. We’re used to traditional biases that cover broad groups, but what if a computer algorithm decided to exclude men who were exactly age 30? When would someone figure that out? We have no equivalent in human bias for more oddly specific groups, and probably won’t notice them. Additionally, groups with less computer savvy will be discriminated against, solely due to the “lack of resources to trouble-shoot the algorithm” issues. The poor. Older people. Those convicted of crimes.The list goes on.

Overall, things aren’t entirely hopeless. There are efforts underway to come up with software that can systematically test “black box” algorithms on a larger scale to help identify biased algorithms before they can cause problems. However, until something reliable can be found, people should be aware that the biases we’re looking for are not the ones you would normally see if humans were running the show. One of the reasons AI freaks me out so much is because we really do all default to anthropomorphizing the machines and only look out for the downsides that fit our pre-conceived notions of how humans screw up. While this comes naturally to most of us, I would argue it’s on of the more dangerous forms of underestimating a situation we have going today. So uh, happy Sunday!

On Predictions and Definitions (After the Fact)

/ bs king / 3 Comments

Twice recently I’ve seen minor characters on both sides of the political spectrum claim that they foresaw/predicted some recent event with “eerie precision”, on topics where their predictions had actually appeared (to me at least) only loosely connected to what actually happened.

While I was annoyed by these people, I was more annoyed by the fans of theirs who rushed to agree that it was clear that they had amazing foresight in making their calls. While obviously some of that is just in-group defensiveness, some of them really seem to believe that this person had done something amazing. While none of those fans are people who read my blog, I figured I’d blow off some steam by reminding everyone of two things:

  1. Redefining words makes improbably results quite common. In the John Ionnaidis paper “Why Most Published Research Findings Are False”, one of his 6 corollaries for published research is “Corollary 4: The greater the flexibility in designs, definitions, outcomes, and analytical modes in a scientific field, the less likely the research findings are to be true.” This is true in research, and more true in political pontificating/predictions. Allowing yourself any latitude at all in redefining words will give you nearly unlimited predictive power after the fact. This is not a “minor quibble”, that’s literally the whole trick.
  2. Making lots of predictions has no consequences for a pundits career. Making predictions as a pundit is like playing roulette with the house’s money. You can’t actually lose. In “The Signal and the Noise” (a book which I once owned and have now lost to lending), Nate Silver reviews how often some pundits known for their “amazing predictions” actually make predictions. Answer: a lot. Most of them are wrong. However, the career of the pundit takes almost no hit for those wrong answers, yet skyrockets every time they are right. Thus they have no motivation to stop making crazy predictions, as they lose nothing for a wrong one and gain everything for a right one. When I read TSATN initially, I made this matrix to illustrate:

So yes, next time you see someone make an “amazing prediction”, take a deep breath and ask yourself how much redefining is going on and how many predictions they had to throw out to get to that one that hit. Well, probably not you specifically dear reader, you’re probably fine. This is almost certainly one of those “if you’re thinking about it enough to read this post, you’re probably not the problem” things. Regardless, thanks for letting me get that off my chest.

In Praise of Confidence Intervals

/ bs king / 2 Comments

This past week I was having a discussion with my high school teacher brother about an experiment his class was running and appropriate statistical methods for analyzing the data. We were discussing using the chi square statistic to compare data from an in class calorimetry experiment to the expected/published values (this is the point where the rest of my family wandered off), and he asked what other statistical analysis his kids could do that might help them understand their results. I mentioned that I was a big fan of confidence intervals for understanding data like this, and started to rattle off my reasons. While testing that produces a p-value is more commonly used in scientific work, I think for most people confidence intervals are more intuitive to use and should be worked in to the mix. Since we were talking about all this at around 7am (prior to both the second cup of coffee AND a trip out to move the cows at his farm), I figured I’d use my blog post today to more thoroughly lay out a few reasons confidence intervals should be more widely used (particularly by teachers) and provide a few helpful links.

The foundations of my argument comes from a paper published a few years ago called “Why the P-value culture is bad and confidence intervals a better alternative“, which gets in to the weeds on the stats, but makes a good overall case for moving away from a reliance on p-values and towards a focus on confidence intervals. The major reasons are:

  1. Confidence intervals use values you already have p-values and confidence intervals are mathematically similar and take the same basic variables in to account: number of observations (n) and variation of those observations (standard error or SE). More observations and lower variability within those observations generally are considered good things. If you can get a p-value, you can get a confidence interval.
  2. Confidence intervals give you more information than the p-value Where a p-value tells you just the statistical significance of the difference, the confidence interval tells you something about the magnitude of the difference. It gives an upper and lower limit, so you get a sense of what you’re really see. For kids learning about variation, I also thinks this can give them a better sense of how each of their experimental values affects the overall outcome. For example, if you’re doing a calorimetry experiment and know that the expected outcome is 100, and your class of 30 kids gets an average of 98 with a standard deviation of 5, you would tell them that p=.03 and thus this a significant difference. Using the confidence interval however, you would give them the range 96.2 to 99.8. This gives a better sense of how different the difference really is, as opposed to just accepting or rejecting a binary “is there a difference” assumption.
  3. Confidence intervals are more visual. The paper I mentioned above has a great figure with it that illustrates what I’m talking about:On this graph you can draw lines to show not just “is it different” but also “when do we really care”. I think this is easier to show kids than just a p-value by itself, as there’s no equivalent visual to show p-values.
  4. It’s easier to see the effect of sample size with a confidence interval. For the calorimetry experiment mentioned above, let’s show what happens if the class is different sizes, all with the same result of 98 with a standard deviation of 5:
    n 95% Confidence interval p-value
    10 94.9-101.1 .2377
    15 95.5-100.5 .1436
    20 95.8-100.2 .0896
    25 96-100 .0569
    30 96.2-99.8 .0366

    I think watching the range shrink is clearer than watching a p-value drop, and again, this can easily be converted in to a graph. If you’re running the experiment with multiple classes, comparing their results can also help show kids a wider range of what the variation can look like.

  5. Confidence intervals reiterate that some variation is to be expected. One of the harder statistical concepts for people to grasp is how predictable a bit of unpredictability really is. For some things we totally get this (like our average commute times), but for other things we seem to stumble (like success of medical treatments) and let outliers color our thinking. In the calorimetry experiment, if 1 kid gets 105 as a value, confidence intervals make it much easier to see how that one outlier fits in with a bigger picture than a single p-value.

So there you go. Confidence intervals are a superior way of presenting effect size, significance of the finding, and are easy to visualize for those who have trouble with written numbers. While they don’t do away with all of the pitfalls of p-values, they really don’t add any new pitfalls to the mix, and they confer some serious benefits for classroom learning.  I used Graphpad to quickly calculate the confidence intervals I used here, and they have options for both summary and individual data.

How to Make Friends and Influence Doctors

/ bs king / 1 Comment

I got a little behind in my reading list this year, but I’m just finishing up Ben Goldacre’s Bad Pharma and it’s really good. Highly recommended if you want to know all sorts of excruciating detail about how we get the drugs we do, and lose faith in most everything.

The book introduced me to a paper from 2007 called “Following the Script: How Drug Reps Make Friends and Influence Doctors“, where a former pharma salesman lays out the different categories of doctors he encountered and how he worked to sell to each of them. This includes a whole table with doctor categories including “The friendly and outgoing doctor” and the “aloof and skeptical” doctor, along with the techniques used to sell to each.

Since Goldacre is the absolute epitome of “aloof and skeptical” he added his own explanation of the tactic they use on evidence based doctors:

“If they think you’re a crack, evidence-based medicine geek, they’ll only come to you when they have a strong case, and won’t bother promoting their weaker drugs at you. As a result, in the minds of bookish, sceptical evidence geeks, that rep will be remembered as a faithful witness to strong evidence; so when their friends ask about something the rep has said, they’re more likely to reply, ‘Well, to be fair, the reps from that company have always seemed pretty sound whenever they’ve brought new evidence to me…’ If, on the other hand, they think you’re a soft touch, then this too will be noted.”

Maybe it’s just because I’ve never been in sales, but it really had not occurred to me that was a sales technique a person could use. Sneaky.

Of course I then realized I’ve seen other, similar things in other situations. While most people know better than to come to me with shoddy stats during political debates, I’ve definitely seen people who told me that they personally agree certain numbers are shoddy later use those same numbers in Facebook arguments with others who aren’t quite as grouchy. It’s an offshoot of the old “be pleasant while the boss/parent is in the room, show your true colors later” thing. Like a data Eddie Haskell. I may have a new bias name here. Gotta work on this one.

What I’m Reading: September 2017

/ bs king / 1 Comment

If you’ve been seeing talk about the PURE study that recently was being reported under headlines like “Huge new study casts doubt on conventional wisdom about fat and carbs“? The study found that those with low fat diets were more likely to die by the end of the study than those with higher fat diets. However, Carbsane took a look and noticed some interesting things. First, the US wasn’t included, so we may need to be careful about generalizing the results there. They also included some countries that were suffering other health crises at the time, like Zimbabwe. Finally, the group they looked at was adults age 35 to 70, but they excluded anyone who had any pre-existing heart problems. This was the only disease they excluded, and it makes some of the “no correlation with heart disease”  conclusions a little harder to generalize. To draw an equivalency, it’s like trying to figure out if smoking leads to lung cancer by excluding everyone in your sample who has lung problems already. What you really want to see is both groups, together and separately.

For my language oriented friends: this article about how cultures without words for numbers get by was really interesting. They make the assumption that counting distinct quantities is an inherently an unnatural thing to do, but I have to wonder about that. Some people do seem more numbers oriented than others, so what happens to those folks? Do people who are good at numbers and quantities just get really depressed in these cultures? Do they find another outlet? As someone who starts counting things to deal with all kinds of emotions (boredom, stress, etc), I feel like not having words for numbers would have a serious impact on my well being.

There’s a lot of herbs and supplements out there being marketed with dubious health claims, but exactly how those claims are worded depends on who you are. This article on how the same products are marketed on InfoWars and Goop is an interesting read, and a good reminder about how much information we get can be colored by marketing spin.

On a political note, this Economist article about the concept of anti-trust laws in the data age was food for thought. 

Finally, I decided to do my capstone project for my degree on a topic I’ve become a little bit obsessed with: dietary variability. Specifically, I’m looking at those who identify that they are food-insecure (defined as not having the resources to obtain enough food to eat in the last 30 days) , and comparing their health habits to those who have enough. While I already have the data set, I’ve been looking for interesting context articles like this one, which explores the “food-insecurity paradox”. Apparently in the US, women who are food insecure are actually more likely to be obese than those who aren’t. Interesting stuff.

Baby, You Can’t Drive My Car: Part 2

/ bs king

I didn’t get back to all the comments on my previous “Baby, You Can’t Drive My Car” post, but there were some good ones. Between those and some emails I got, there was some interesting statistical/follow-up issues raised that I felt deserved their own post.

First, the general “not getting driver’s licenses” trend among young people seems to actually be reflective of some larger delayed adulthood trends among “iGen”, or those born in 1995-2005. This Wall Street Journal article by a researcher studying the newest up and coming generation says the following ” As I found in analyzing seven large national surveys of teens, today’s adolescents are less likely to drive, drink, work, date, go out and have sex than were teens just 10 years ago. Today’s 18-year-olds look like 15-year-olds used to.” This may not be a bad thing. Per the article, this group is less likely to get in car accidents than teens in previous years, which is obviously a good thing. On the other hand, on my previous post commenter Michael pondered if delaying getting your driver’s license is good for your long term driving skills. Will the group that failed to get a license at 16 be getting in more accidents at 30 as a result? TBD.

Second point, related to the first, is the success of teen driver’s license initiatives. The initial results on these look good, but most of those numbers are obtained using raw numbers. In other words, we don’t know if those licensing laws improved the driving of 16 and 17 year olds, or if the reductions in crashes were due to fewer of them being able to drive at all. Not a bad outcome either way, but possibly a misleading one if the numbers pick back up at a later date. While I can definitely see some types of impulsiveness associated with teen drivers ebbing with age, there are certain errors inexperienced drivers might make regardless. Again, TBD.

Third point I started to wonder about….are we seeing a corresponding uptick in non-driver’s license IDs? If not, how are people without licenses identifying themselves? I can see people not wanting the hassle of getting a driver’s license, but it seems awfully hard to function without an ID. Even buying alcohol would be a challenge, at least in my area.

Fourth, related to #3, is the low driving rate in some states being driven by a population that can’t get a license? I mentioned in the first post that states with higher populations of those under 18 would likely have lower rates, but someone also noted that it’s not clear where the population numbers came from. Since some population counts include all residents regardless of citizenship, and since states vary wildly in allowing those without documentation of citizenship to get licenses, it’s possible that accounts for some of the differences.

 

 

 

Measuring Weather Events: A Few Options

/ bs king

Like most people in the US this past week, I was glued to the news watching the terrible devastation Hurricane Harvey was inflicting on Houston. I have quite a few close friends and close work collaborators in the area, who thankfully all are okay and appear to have suffered minimal property damage themselves. Of course all are shaken, and they have been recommending good charities in the area to donate to.  As someone’s whose lived through a home flood/property destruction/disaster area/FEMA process at least once in my adult life, I know how helpless you feel watching your home get swallowed by water and how unbelievably long the recovery can be. Of course when I went through it I had a huge advantage….the damage was limited to a small number of people and we had easy access to lots of undamaged places. I can’t imagine going through this when the whole city around you has suffered the same thing, or how you begin to triage a situation like this.

However, as the total cost of Hurricane Harvey continues to be fully assessed, I want to make a comment on some numbers I’m seeing tossed around. It’s important to remember in the aftermath of weather events of any kind that there are actually a few different ways of measuring how “big” it is:

  1. Disaster Measurement:
    1. Lives lost
    2. Total financial cost
    3. Long term impacts to other infrastructure (i.e. ability to rebuild, gas pipelines,etc)
  2. Storm measurement
    1. Size of weather event (i.e. magnitude of hurricane/volume of rainfall)
    2. Secondary outcomes of weather event (i.e. flooding)

538 has a good breakdown of where Harvey ranks (so far) on all of these scales here.

Now none of this matters too much to those living through the storm’s aftermath, but in terms of overall patterns they can be important distinctions to keep in mind. For example, many people have been wondering how Harvey compares to Katrina. Because of the large loss of life with Katrina (almost 2,000 deaths) and the high cost ($160 billion) it’s clear that Katrina is the benchmark for disastrous storms. However, in terms of wind speed and power of the storm, Katrina was actually pretty similar to Hurricane Andrew in 1992 which resulted in 61 deaths and had a quarter of the cost. So why did Katrina have such a massive death toll? Well, as 538 explains:

Katrina’s devastation was a result of the failure of government flood protection systems, violent storm surges, a chaotic evacuation plan and an ill-prepared city government.

So the most disastrous storms are not always the biggest, and the biggest storms are not always the most disastrous. This is important because the number of weather events causing massive damage (as in > $1 billion) is going up:

Source of graph.
Source of disaster data.

However, the number of hurricanes hitting the US has not gone up in that time (more aggregate data ending in 2004 here, storm level data through 2016 here). This graph does not include 2017 or Hurricane Harvey.

Now all of these methods of measuring events are valid, depending on what you’re using them for. However, that doesn’t mean the measures are totally interchangeable. As with all data, the way you intend to use it matters. If you’re making a case about weather patterns and climate change, costly storm data doesn’t prove the weather itself is worsening. However if you’re trying to make a case for infrastructure spending, cost data may be exactly what you need.

Stay safe everyone, and if you can spare a bit of cash, our friends in Houston can use it.

Baby, You Can’t Drive My Car

/ bs king / 5 Comments

This week I was rather surprised to find out that a early-20s acquaintance of mine who lives in a rather rural area did not have a driver’s license. I know this person well enough to know there is no medical reason for this, and she is now pursuing getting one. Knowing the area she lives in I was pretty surprised to hear this, particularly considering she has a job and a small child.

Now living around a large city with lots of public transportation options, I do know people who are medically unable to get a license (and thus stick close to public transit) and those who simply dislike the thought of driving (because they grew up around public transit), but I hadn’t met many people in the (non-big city) area I grew up in who didn’t get a license.

As often happens when I’m surprised by something, I decided to go ahead and look up some numbers to see how warranted my surprise was. It turns out I was right to be surprised, but there may be a trend developing here.

According to the Federal Highway Administration, in 2009 87% of the over-16 population had a driver’s license. There’s a lot of state to state variation, but the states I’ve lived in do trend high when it comes to the number of drivers per 1000 people:

Note: that map is licenses per 1000 people of all ages, so some states with particularly young populations may get skewed.

This appeared to confirm my suspicions that not having a license was relatively unusual, particularly in the New England region. Most of the places that have lower rates of licenses are those that have big cities, where presumably people can still get around even if they don’t know how to drive. I am a little curious about what’s driving the lowish rates in Utah, Texas and Oklahoma, so if anyone from there wants to weigh in I’d appreciate it.

I thought that was the end of the story, until I did some more Googling and found this story from the Atlantic, about the decreasing number of people who are getting driver’s licenses. The data I cited above is from 2009, and apparently the number of licensed drivers has been falling ever since then. For example, this paper found that in 2008, 82% of 20-24 year olds had driver’s licenses, but by 2014 76.7% did. In contrast, in 1983 that number was 91.8%.

So what’s the reason for the decline? According to this survey, the top 5 reasons for those aged 18 to 39 are “(1) too busy or not enough time to get a driver’s license (37%), (2) owning and maintaining a vehicle is too expensive (32%), (3) able to get transportation from others (31%), (4) prefer to bike or walk (22%), (5) prefer to use public transportation (17%)”.

Like most surveys though, I don’t think this tells the whole story. For example, the top reason for not having a license is that people are “too busy” to get one, but the study authors noted that those without licenses are less likely to be employed and have less education than those with licenses. This suggests that it is not extended outside commitments that are preventing people from getting licenses. Additionally, anyone who has ever lost the use of their car knows it can take a lot more time to get a ride from someone else than it does just to hop in your own vehicle.

My personal theory is that getting a drivers license is something that requires a bit of activation energy to get going. Over the last decade or two, state legislatures have progressively enacted laws that put more restrictions on teen drivers, so the excitement of “got my license this morning and now I’m taking all my friends out for a ride tonight” no longer exists in many states. For example, in Massachusetts drivers under 18 can’t drive anyone else under 18 (except siblings) for the first 6 months. This is probably a good practice, but it almost certainly decreases the motivation of some kids to go through all the work of getting their license. After all, this is an age group pretty notorious for being driven by instant gratification.

Additionally, with high costs for insurance and vehicles, many parents may not be as excited for their kids to get their license. Without parental support, it can be really hard to navigate the whole process, and if a parent starts to think it may be easier to keep driving you than to pay for insurance, this could further decrease the numbers. With younger generations spending more time living at home, parental support is an increasing factor. Anyone attempting to get a license has a certain reliance on the willingness of others to teach them to drive and help them practice, so the “too busy” reason may actually be driven just as much by the business of those around you as your own business. You can be unemployed and have plenty of time to practice driving, but if no one around you has time to take you out, it won’t help.

Finally, there may be a small influence of new technology. With things like Uber making rides more available more quickly and Amazon delivering more things to your door, it may actually be easier to function without a license than it was 10 years ago. Even the general shift from “have to go out to do anything fun” to “can stay home and entertain myself on line” may account for a bit of the decreased enthusiasm for getting licensed to drive. For any one person it’s doubtful that’s the whole reason, but for a few it may be enough to tip the scales.

It will be interesting to see if this corrects itself at some point…will those not getting their license at 18 now get it at 28 instead, or will they forego it entirely? The initial survey most (about 2/3rds) still plan on pursuing one at some point, but whether or not that happens remains to be seen.

The Real Dunning-Kruger Graph

/ bs king / 5 Comments

I’m off camping this weekend, so you’re getting a short but important PSA.

If you’ve hung out on the internet for any length of time or in circles that talk about psych/cognitive biases a lot, you’ve likely heard of the Dunning-Kruger effect. Defined by Wiki as “a cognitive bias wherein persons of low ability suffer from illusory superiority, mistakenly assessing their cognitive ability as greater than it is.”, it’s often cited to explain why people who know very little about something get so confident in their ignorance.

Recently, I’ve seen a few references to it accompanied by a graph that looks like this (one example here):

While that chart is rather funny, you should keep in mind it doesn’t really reflect the graphs Dunning and Kruger actually obtained in their study. There were 4 graphs in that study (each one from a slightly different version of the study) and they looked like this:

Humor:

Logic and reasoning (first of two):

Grammar:

And one more logic and reasoning (performed under different conditions):

So based on the actual graphs, Dunning and Kruger did not find that the lowest quartile thought they did better than the highest quartile, they found that they just thought they were more average than they actually were. Additionally it appears the 3rd quartile (above average but not quite the top), is the group most likely to be clearsighted about their own performance.

Also, in terms of generalizability, it should be noted that the participants in this study were all Cornell undergrads being ranked against each other. Those bottom quartile kids for the grammar graph are almost certainly not bottom quartile in comparison to the general population, so their overconfidence likely has at least some basis.  It’s a little like if I asked readers of this blog to rank their math skills against other readers of this blog….even the bottom of the pack is probably above average. When you’re in a self selected group like that,  your ranking mistakes may be more due to a misjudging of those around you as opposed to just an overconfidence in yourself.

I don’t mean to suggest the phenomena isn’t real (follow up studies suggest it is), but it’s worth keeping in mind that the effect is more “subpar people thinking they’re middle of the pack” than “ignorant people thinking they’re experts”. For more interesting analysis, see here, and remember that graphs drawn in MS Paint rarely reflect actual published work.

 

5 Things You Should Know About Orchestras and Blind Auditions

/ bs king / 6 Comments

Unless you were going completely off the grid this week, you probably heard about the now-infamous “Google memo“.  Written by a (since fired) 28 year old software engineer at Google, the memo is a ten page long document where the author lays out his beliefs about why gender gaps in tech fields continue to exist. While the author did not succeed in getting any policies at Google changed, he did manage to kick off an avalanche of hot takes examining whether the gender/tech gap is due to nature (population level differences in interests/aptitude) or nurture (embedded social structures that make women unwelcome in certain spaces). I have no particular interest in adding another take to the pile, but I did see a few references to the “blind orchestra auditions study” that reminded me I had been wanting to write about that one for a while, to deep dive in to a few things it did or did not say.

For those of you who don’t know what I’m talking about, here’s the run down: back in the 1970s, top orchestras in the US were 5% female. By the year 2000, the were up to almost 30% female. Part of the reason for the change was the introduction of “blind auditions”, where the people who were holding tryouts couldn’t see the identity of the person trying out. This finding normally gets presented without a lot of context, but it’s good to note someone actually did decided to study this phenomena to see if the two things really were related or not. They got their hands on all of the tryout data for quite a few major orchestras (they declined to name which ones, as it was part of the agreement of getting the data) and tracked what happened to individual musicians as they tried out. This led to a data set that had overall population trends, but also could be used to track individuals. You can download the study here, but these are my highlights:

  1. Orchestras are a good place to measure changing gender proportions, because orchestra jobs don’t change. Okay, first up is an interesting “control your variables” moment. One of the things I didn’t realize about orchestras (though may be should have) is that the top ones have not changed in size or composition in years. So basically, if you suddenly are seeing more women, you know it’s because the proportion of women overall is increasing across many instruments. In the words of the authors ” An increase in the number of women from, say, 1 to 10, cannot arise because the number of harpists (a female-dominated instrument), has greatly expanded. It must be because the proportion female within many groups has increased.”
  2. Blind auditions weren’t necessarily implemented to cut down on sexism. Since this study is so often cited in the context of sexism and bias, I had not actually ever read why blind auditions were implemented in the first place. Interestingly, according to the paper written about it, the actual initial concern was nepotism. Basically, orchestras were filled with their conductors students, and other potentially better players were shut out. When they opened the auditions up further, they discovered that when people could see who was auditioning, they still showed preferential treatment based on resume. This is when they decided to blind the audition, to make sure that all preconceived notions were controlled for. The study authors chose to focus on the impact this had on women (in their words) “Because we are able to identify sex, but no other characteristics for a large sample, we focus on the impact of the screen on the employment of women.”
  3. Blinding can help women out Okay, so first up, the most often reported findings: blind auditions appear to account for about 25% of the increase in women in major orchestras. When they studied individual musicians, they found that women who tried out in blind and non-blind auditions were more successful in the blinded auditions. They also found that having a blind final round increased the chances a woman was picked by about 33%. This is what normally gets reported, and it is a correct reporting of the findings.
  4. Blinding doesn’t always help women out One of the more interesting findings of the study that I have not often seen reported: overall, women did worse in the blinded auditions. As I mentioned up front, the study authors had the data for groups and for individuals, and the findings from #3 were pulled from the individual data. When you look at the group data, we actually see the opposite effect. The study authors suggest one possible explanation for this: adopting a “blind” process dropped the quality of the female candidates. This makes a certain amount of sense. If you sense you are a borderline candidate, but also think there may be some bias against you, you would be more likely to put your time in to an audition where you knew the bias factor would be taken out. Still, that result interested me.
  5. The effects of blinding can depend on the point in the process Even after controlling for all sorts of factors, the study authors did find that bias was not equally present in all moments. For example, they found that blind auditions seemed to help women most in preliminary and final rounds, but it actually hurt them in the semi-final rounds. This would make a certain amount of sense….presumably people doing the judging may be using different criteria in each round, and some of those may be biased in different ways than others. Assuming that all parts of the process work the same way is probably a bad assumption to make.

Overall, while the study is potentially outdated (from 2001…using data from 1950s-1990s), I do think it’s an interesting frame of reference for some of our current debates. One article I read about it talked about the benefit of industries figuring out how to blind parts of their interview process because it gets them to consider all sorts of different people….including those lacking traditional educational requirements. With many industries dominated by those who went to exclusive schools, hiding identity could have some unexpected benefits for all sorts of people. However, as this study also shows, it’s probably a good idea to keep the limitations of this sort of blinding in mind. Even established bias is not a consistent force that produces identical outcomes at all time points, and any measure you institute can quickly become a target that changes behavior.  Regardless, I think blinding is a good thing. All of us have our own pitfalls, and we all might be a little better off if we see our expectations toppled occasionally.