The new forums will be named Coin Return (based on the most recent vote)! You can check on the status and timeline of the transition to the new forums here.
The Guiding Principles and New Rules document is now in effect.
Lies, Damned Lies, and Statistics
ElJeffe
Registered User, ClubPA regular
Registered User, ClubPA regular
So here's a conversation I recently held with a friend of mine, paraphrased:
Her: "Hey, guess what the worst food is?"
Me: "The worst food?"
Her: "Yeah, it's the worst by 57%."
Me: "Umm. 57% of what?"
Her: "It's 57% worse than the next worst one."
Me: "I don't what that means, though. It's 57% higher in calories? Higher in fat? 57% more likely to give you cancer, or makes you 57% fatter...?"
Her: "What difference does it make? Just guess. It's 57% worse."
Me: "But that doesn't mean anything! You need to have some metric by which it's 57% worse, or else that number is useless! It's like saying a car is 57% better than every other car!"
Her: "God, stop being so analytical."
Me: "Okay, fine. Umm... pasta with alfredo sauce."
Her: "Nope, it's french fries!"
Me: "So fries are 57% worse than the next worse food. What's the next worse food?"
Her: "I don't know, but isn't that crazy?"
Me: *stabbing self in brain to hopefully end the pain*
Statistics are truly the worst thing. In fact, they are 83% worse than the next worse thing. So let's discuss statistics! How are they misused? And can we make it better? What will it take to teach people that reducing something to a single number almost never provides any useful information, because almost any statement containing a single number can be made true by clever use of metrics?
(Note: This is not the place to discuss food, or french fries, or whatnot. We are talking about statistics here, and if you don't do it right I am 100% likely to punch you.)
Her: "Hey, guess what the worst food is?"
Me: "The worst food?"
Her: "Yeah, it's the worst by 57%."
Me: "Umm. 57% of what?"
Her: "It's 57% worse than the next worst one."
Me: "I don't what that means, though. It's 57% higher in calories? Higher in fat? 57% more likely to give you cancer, or makes you 57% fatter...?"
Her: "What difference does it make? Just guess. It's 57% worse."
Me: "But that doesn't mean anything! You need to have some metric by which it's 57% worse, or else that number is useless! It's like saying a car is 57% better than every other car!"
Her: "God, stop being so analytical."
Me: "Okay, fine. Umm... pasta with alfredo sauce."
Her: "Nope, it's french fries!"
Me: "So fries are 57% worse than the next worse food. What's the next worse food?"
Her: "I don't know, but isn't that crazy?"
Me: *stabbing self in brain to hopefully end the pain*
Statistics are truly the worst thing. In fact, they are 83% worse than the next worse thing. So let's discuss statistics! How are they misused? And can we make it better? What will it take to teach people that reducing something to a single number almost never provides any useful information, because almost any statement containing a single number can be made true by clever use of metrics?
(Note: This is not the place to discuss food, or french fries, or whatnot. We are talking about statistics here, and if you don't do it right I am 100% likely to punch you.)
I submitted an entry to Lego Ideas, and if 10,000 people support me, it'll be turned into an actual Lego set!If you'd like to see and support my submission, follow this link.
ElJeffe on
0
Posts
Give or take 18%.
twitch.tv/Taramoor
@TaramoorPlays
Taramoor on Youtube
"How can you trade Thompson?! He leads the team in 8th-inning doubles in the month of July!"
"Watch out for Ramirez this season. He's never lost when his team has been up by 21 or more points going into the fourth quarter."
Also, the use of statistics should have made professional sports drafts worthless, yet they persist.
"We're picking this shitty kid for QB with our first pick, despite the fact that we already have a quarterback and this kid is really, really shitty, in a really obvious way. But what can we do? He went to Notre Dame!"
I like the birthday problem. (I know, I used it in an OP on DNA evidence last year.) How many people do you need to have in a room for there to be a 50% probability of any two people having the same birthday?
And 99%?
the "no true scotch man" fallacy.
― Marcus Aurelius
Path of Exile: themightypuck
And it's getting worse. In 1960 only 55% of statistics were made up on the spot (+- 20%).
Further, today only 15% of people understand statistics which is down from 25%.
Related to statistics, I really hate that people do not understand correlation != causation. Yes, as ice cream sales increase so do drowning deaths. But ice cream does not cause you to drown. Although drowning in ice cream sounds like a fun way to go.
Combinatorial probability is always interesting. Do we just want to discuss statistics? Or more advanced stuff? Serial correlation always bites people in the ass when they're investing. LTCM is a classic example.
A classic investment problem is:
Which investment is better? A 10% return on capital with 20% variance? Or 50% return on capital with 100% variance?
Or EGARCH models? Which account for the clustering of volatility. Some of the more interesting research is showing that the market "remembers" negative events in the volatility models much longer than it "remembers" positive events.
I would like to understand this, but I am caught on the financial buzzwords. I'm also caught on the use of the word 'variance' to refer to a percentage. Does "variance" in investing mean the same thing as it does in stats?
the "no true scotch man" fallacy.
Yes, my brain would hurt a lot less after watching a game if the announcers would stop throwing out so many pointless stats. I don't care what Sabbathia's ERA is in the 1st with runners in scoring position during day games in the month of June.
Depends on the thinktank responsible.
I know that a completely random survey will give representative results at, say, 1500 responses or so. That is usually when there is a small number of possible answers.
But does this hold up when there are hundreds of possible answers to a given question?
In other words, how much can we trust Neilson TV ratings / what would their sample size need to be?
Let's say you have a medical test for an exotic disease and you're testing the general population. If the likelihood of a false positive on the test is higher than the prevalence of that disease, then a positive result is more likely to be wrong than it is right.
In other words, let's say there's a new disease called Fabricatosis. 1% of the people in the world have fabricatosis; there's a blood test with a false positive rate of 5% and a false negative rate of 5%. If we test 10,000 people, we would expect:
95 true positives
500 false positives
5 false negatives
9,400 true negatives
There are a couple of ways (that I know of) to get around this.
1) Don't bother testing random people. Only test people who have symptoms or have been exposed. The prevalence of fabricatosis in the general population might be 1%, but we can presume that the prevalence is actually much higher among the subset of people who show the symptoms of fabricatosis. This is what we do with mononucleosis.
2) Develop a different test and use the two tests together. This reduces the false positive rate exponentially. This is what we do for HIV.
the "no true scotch man" fallacy.
"Neglecting Vitamin J correlated with a 300% increase in risk of toenail cancer!"
This is misleading compared to "The risk of toenail cancer in the population of those who neglect Vitamin J within their diet is 0.03% (or 3 in 50,000 <this should also be specified and be noted as per year or per death, however it's measured>) compared to 0.01% for those with sufficiently Vitamin J filled diets."
PSN:Furlion
But that way I can't scare people into buying Vitamin J!
This is exactly what I'm tlaking about - without the base rates or absolute percentages the relative increases mean nothing.
How dare she put her baby at risk.
How dare she.
She should rip out her ovaries and give them to somebody who is fit to be a mother.
the "no true scotch man" fallacy.
You can also, depending on the test, run it twice. Sometimes the false positive rate has to do with the test more than some quirk of whoever's getting it.
the "no true scotch man" fallacy.
Statistical Mechanics let me do science! They are witchcraft and no matter how much we pretend we know it, there are only like 10 dudes who really do, and they are probably witches!
Rock Band DLC | GW:OttW - arrcd | WLD - Thortar
Can I just mail you hundred dollar bills instead?
In theory, sensitivity and specificity of any given medical test should determine what is used for therapy, but in practice, there are a lot of logistical problems. For example, whether or not a test is available in a treatment setting, the cost of the test (compared to other tests), and the benefit of treatment (why bother doing the test if the treatment is going to be the same?).
But there's this new thing that I just discovered! It's called the internet. Apparently, you don't have to use mail to send him money! Just put all of your personal information right here...
In the absence of plausibly controlled experiments, a lot of statistics as applied to the social sciences rely on theory to fill in the gaps. Theory varies by ideology.
One hypothesis: disentangling the financial issues caused by unexpected negative shocks (i.e., at least someone out there is going to be overleveraged) is itself costly - paying unexpected opportunity costs is a lot easier than paying unexpected costs.
For one thing, it'd help if the texts were done a bit better. At least in my second semester college course, there was actually a homework problem assigned that involved comparing the rate of different types of paint drying. I am not sure if the statisticians were trying to troll students or something, but god man, really?
You can use information like that, but it's going to depend a lot on how you define language and what you're trying to figure out. You're not going to be able to prove specific points this way; you'd have to be going for something fairly general. Most likely, if you're trying to make decisions on information like this, you should be figuring out a better way to sample, or a better methodology of questioning that can provide more discrete information. Or you can figure out ways to mine that data, but that's again going to heavily depend on what you have and what you're looking for. I don't know too much about Nielsen ratings other than extremely basic stuff so it's hard to comment.
Presumably that's one instant where percentages above 100 are actually useful? I.e. 120% score being above the level expected of the employee/shop?
Whilst I'm here can I recommend Bad Science by Ben Goldacre and Irrationality by Stuart Sutherland. Both cover statistics (amongst other things) and their abuses.
Misused stats usually fall into two categories. The first is when a critical piece of information that explains the statistic is being withheld, like in the absolute vs relative percentage discussion above, but the number isn't exactly wrong just misleading. The other is bad data from the start. If I poll 5 Americans on their views on gay marriage, it doesn't matter how perfect my study and questions were because the data is fundamentally screwy (not a big enough sample if I'm trying to estimate for all of America).
I'd love for there to be a basic "Surviving in the Real World" class that you have to take during your Senior year at HS that covers subjects like basic statistics, basic borrowing/debt, etc.
But that's boring/unfunny stuff! I'm currently studying for a degree in Statistics and the sheer amount of hilarity I run into (from professors, other students, and, I'll admit, my own ineptitude) is surprising.
For example: in one of my stat courses we had an end of the year project to gather up data, come up with a hypothesis about that data, and run a bunch of tests. It's real purpose was to make sure we had a basic understanding of how to run the tests and knew how to get the computer to run them for us. My group - being composed of two Economics majors and myself - decided to gather up a three year stretch of housing prices from both before and after the recession and create a model to estimate that end housing price based off of a bunch of factors (gross lot size, number of beds and baths, etc). We'd then compare them to see if, say, a bathroom was worth more or less pre-recession or post-recession.
After a few days of pain and tears as we manually trudged through a list of central florida housing prices and manually transcribed 10ish variables from a few different sources into a big excel worksheet our data was finally collected. One sunday afternoon spent yelling at a computer later, we were finally had our model!...and promptly found that it suggested that adding a bathroom to your house, both pre- and post- recession, took 200,000 dollars off of your house's sale price. In fact, our model predicted that the only thing that would actually increase the price of your house was gross lot size and heated lot size. In non-housing terms, our several days of work ended up stating that a house the exact size of it's lot without any bedrooms or bathrooms was the best thing since sliced bread, at least price wise.
Needless to say, we didn't end up suggesting that anyone go out and use that model to predict housing prices.
In a different stats course, our professor randomly asks us to raise our hands if we had smoked Marijuana recently. No one actually raised their hands, which lead him to confess that he thought that *all* Americans smoked weed whenever they were depressed or stressed or anything. He was certainly an odd one.
-edit-
I have no doubt in my mind that my intro stats book was trolling students. After finding the probability that Chad guesses Mary's birthday perfectly on the first date, the very next question asked what the probability of a second date would be if Mary found out that Chad already knew her birthday and was just trying to impress her.
A house with fewer rooms has larger rooms, and it seems quite plausible that differing layouts like these are rrelated with other factors (geography, proximity to desirable features?) that may be correlated with house price.
i.e., omitted variable?
just a guess, anyway
In the end, we came to the same conclusion you did: We just didn't have enough variables, mostly because we were running off of public available data and just didn't have that many variables to work with. The story, thankfully for my grade, has a happy ending. Our model presentation was pretty funny due to the silly numbers we kept getting and, perhaps more importantly, we only took a few minutes after a long string of twenty to thirty minute presentations by people who were trying far too hard to be funny. We all ended up getting an A on the project.
The rules according to my statistics textbook:
It must be a random sample. This means that every potential thing that could be under observation has an equal chance of being picked.
It must have parameters. Basically, you have to gather the data from the relevant population. If you're judging how people will vote, and you ask 12 year old's, it won't be that useful.
Variables have to be defined properly. If you use a nebulous word in a poll like whether or not people like taxes, obviously everyone will say no. If you say, taxes on a specific group, that group will say no, the rest will be more mixed. Misleading variables is usually the "damn lies" part of statistics. You can see really dishonest stuff here.
Now, if you go with the sample size 3, your confidence interval (the percentage of confidence that you'll have that all further observations will fall into your conclusion) could be pretty weak with a high margin of error, but you could also have a distribution that is surprisingly close to a normal distribution depending on what you're testing for.
Sometimes I'll see people on these forums refer to anecdotes with a "lol sample size," but, actually, that's very rarely the problem. If they have more than 3 or more people, they actually have enough to make a inference. What you should say is "lol non random sample."
SteamID: devCharles
twitter: https://twitter.com/charlesewise
Yeah, variance is the average squared distance from the mean. It's usually calculated off of the expected return. Where the expected return is your first moment/population mean.
Volatility is simply the price variation in an instrument/stock. There's a lot of different ways to calculate it and different types. Usually, unless someone mentions otherwise, it's the yearly annualized volatility.
One of the better things my newspaper has in it's overall pretty mediocre science section is a quarterpage examining whatever the 'hyped science news of the week" was. Statistics comes up a lot, and this week it was a claim how sleeping on your left side during the latter months of pregancy reduces stillbirth. The small sample study produced a rate decrease of 0.37% to 0.19%. It apparently made headlines around the world as "Sleeping on left-side halves stillbirth!"
Edit: Most of the mentions of it imply a clear causal relation, where the researchers only find correlation.
PSN:Furlion
― Marcus Aurelius
Path of Exile: themightypuck