Slashdot Mirror
Why Programmers Need To Learn Statistics
David Gerard writes "Zed Shaw writes an impassioned plea to programmers: Programmers Need To Learn Statistics Or I Will Kill Them All. Quoting: 'I go insane when I hear programmers talking about statistics like they know s*** when it's clearly obvious they do not. I've been studying it for years and years and still don't think I know anything. ... I have taken a bunch of math classes, studied statistics in grad school, learned the R language, and read tons of books on the subject. Despite all of this I'm not at all confident in my understanding of such a vast topic. What I can do is apply the techniques to common problems I encounter at work. My favorite problem to attack with the statistics wolverine is performance measurement and tuning. All of this leads to a curse since none of my colleagues have any clue about what they don't understand. I'll propose a measurement technique and they'll scoff at it. I try to show them how to properly graph a run chart and they're indignant. I question their metrics and they try to back it up with lame attempts at statistical reasoning. I really can't blame them since they were probably told in college that logic and reason are superior to evidence and observation.'"
110%.
Correlation != causation. Just repeat that and you don't need to know statistics.
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
Maybe the problem is in your presentation. Even here, you tell programmers that you want to kill them for not understanding a topic that even you are unwilling to acknowledge mastery of. Then you tell us how hard the topic is to understand, even though you've spent so much time trying to learn it.
Is it any wonder that no one takes your suggestions seriously? You are practically sabotaging yourself with self-effacement.
These aren't homework problems you're tackling here. They are business problems and you need to sell yourself and your ideas if you want to get any traction. Do you have any evidence that your methods are better than the SOP thus far? Do you have any case studies that show how effective statistic analysis is in *any* of your projects?
Or are you simply taking something that seems like a data point and extrapolating it to cover a vast swath of applications?
Statisticians need to learn programming or I will kill them all.
We know as much statistics as we need to know.
Some know more, some less. Each has traded off hours vs. knowledge in many fields.
For example: Why would a programmer who's job is to automate bean counting need to know more then basic statistics? (s)he rightfully focuses his efforts on accounting.
One post calculus statistics course gives me enough grounding to know what I don't know and punt to experts when I need to.
Fucking specialists forget all the things they don't know and only look at the world through one lens.
John McAfee 'It was like that time I hired that Bangkok prostitute; to do my taxes, while I fucked my accountant'
Programmers Need To Learn Statistics Or I Will Kill Them All
Okay, two things: First, threatening programmers never work. Management's been trying that for years. Second -- don't you mean 'kill -9' them all, or maybe demalloc(), or cast them to void*, or one of a dozen other witty things you could do besides the mundane answer of threatening stabby bits on them because you have a case of intellectual snobbery?
#fuckbeta #iamslashdot #dicemustdie
The only statistics book you'll even need
He's just as arrogantly claiming that he's right and they're wrong. Now, he may very well in fact be right, but he's taking the same obstinate position the people he criticizes do.
It's important to know when your input is not desired. Even if you think it should be.
is not because they don't understand statistics. It is because you are a dick.
Statistics is HARD, for two reasons:
(a) Probability theory, on which all practical Statistics is based it both (i) counter-intuitive and (ii) difficult
(b) The very Mathematics on which it is based is obscure
And, worst of all, it is uniformly badly taught, even in good universities, and the Statistics for XXX are uniformly awful, blind leading the blind.
Lastly it is very hard to get a staight answer from a mathematical Statistician.
A manga statistics book, eh?
I just realized I was a nerd. I looked at the table of contents and closed it down, then realized I hadn't even looked at the short skirt-wearing protagonist.
Sigh...
But to answer the article's point, elementary statistics are very easy. Advanced statistics are very hard. It's kind of like how people think "knowing the difference between circles and squares" is geometry and so analytical geometry must be just more of the same, right? It's quite possible the programmers think they know statistics because they know they're vaguely supposed to do a run multiple times, and maybe average the results or something.
It's also possible the author of the article is a know-it-all douchebag who tries to solve problems with overwrought solutions.
From TFA: "Zed: Fuck! Fuck! I have eyes! You do not! See!? No?! Exactly! Because you can't fucking see because you have no fucking eyes! Arrggh!"
Just throwing that theory out there.
I disagree that CS is just "programming and troubleshooting", but I do agree that Computer Science is a complete misnomer. It's extremely misleading, and difficult to explain to people, "I'm a computer scientist, but no I'm not actually a scientist, instead I understand how to describe formal languages in terms of strict grammar rules and transform abstract syntax trees from one representation to another."
It shouldn't be called Computer Science, it should be called Computational Mathematics, because that's what it is.
(On the other hand, there is whole branch of CS that extends very deeply into statistics called Machine Learning, but at the core I'd say it is still more mathematics than science. There is also human-machine interaction which often goes under CS, but is actually more like psychology.. so it's not so cut and dry.)
...unfortunately, they are mostly lost in the irony of statements like this:
I think women are better programmers because they have less ego and are typically more interested in the gear rather than the pissing contest.
I doubt I've seen anyone more thoroughly entrenched in a pissing contest than Zed Shaw, of the website formerly known as "Zed's So Fucking Awesome".
Don't thank God, thank a doctor!
So, since so many people don't seem to want to actually read Zed's stuff -- and I honestly don't blame you -- I'll try to summarize:
Eventually, every major science adopted an empiricist view of the world. Except Computer Science of course.
He tends to bitch a lot about computer scientists. I'm just starting a CS degree, and there is a Statistics class in the curriculum. Is he working with people with good degrees, people from a technical college with a "programming" degree, people from a diploma mill, or high school students with no degree at all?
Of course, he seems to be implying it's everyone, and doing so in a typically Zed-like way.
"All you need to do is run that test [insert power-of-ten] times and then do an average." Usually the power-of-ten is 1000...
I don't know that I've ever heard that particular statement. But it's a good point:
How do you know that 1000 is the correct number of iterations to improve the power of the experiment?
Generally because it was probably closer to a million, so I'm erring on the side of taking more, rather than fewer, measurements. But without careful consideration, I could be way off.
How are you performing the samplings?
I think this is vastly less important than how you are dealing with the data, but it is also a good point. For example, his complaint is that an average isn't enough; with detailed enough logging, he could easily go back into my data and figure out min, max, standard deviation, histograms...
How do you know that 1000 is enough to get the process into a steady state after the ramp-up period?
Not a huge deal -- the "steady state" will almost certainly be faster than the "ramp-up" period. Worst case, I'm over-optimizing.
What will you do if the 1000 tests takes 10 hours?
Either ctrl+c, or try it 10 times.
How does 1000 sequential requests help you determine the performance under load?
Very good point here. It's still a useful statistic, but you still need to measure things like 1000 simultaneous requests, not just 1000 all in sequence.
On the other hand, if your performance is acceptable with them all in sequence, you could just run it through something like Event Machine, so it's all sequential on production, too.
The most troubling problem with these single number “averages” is that there’s two common averages and that without some form of range or variance error they are useless. If you take a look at the previous graphs you can see visually why this is a problem. Two averages can be the same, but hide massive differences in behavior...
So yes, always make sure you can record enough statistics so that someone else can come along and use your data to give you something meaningful.
The moral of the story is that if you give an average without standard deviations then you’re totally missing the entire point of even trying to measure something. A major goal of measurement is to develop a succinct and accurate picture of what’s going on...
It doesn't have to be statistically accurate. It just has to be close enough.
Ah, confounding. The most difficult thing to explain to a programmer, yet the most elementary part of all scientific experimentation. It’s pretty simple: If you want to measure something, then don’t measure other shit.
This is both a very good and a very bad idea. It ties into the peeve he had before -- ramp-up time. For example:
If we want to take one single line of code and test it then we can. If we want to only verify one single query on a database then what’s stopping us?
What's stopping us is that our applications don't actually work like that.
Don't thank God, thank a doctor!
Before computers stats involved using parametric tests (t-tests, anova, etc) which made assumptions like "the data comes from an underlying normal distribution". BTW, in stats terms "normal" mean "Gaussian".
Now, with cheap and fast computers, we can actually compute the confidence intervals non-parametrically through permutation tests and bootstrapping without assuming anything about underlying distributions. In most cases, this non-parametric test is the "right thing to do". Most of the time, the results are the same as using a parametric test.
However, a HUGE disaster in empirical science has been the problem of multiple comparisons. With computers it is so easy to compute correlations and significance tests between every possible slice of your data set. Many "scientists" don't have good statistical knowledge and pray at the alter of "p < 0.05". They don't know about or understand the problem of multiple comparisons. So they do 20 tests, find one that comes out p0.05 and write a paper about it. They don't get that if you do 20 tests you are very very very likely to find one that come out p < 0.05.
Anyone who has access to excel or matlab can do this little experiment.
samp=50 normally distributed random numbers.
for x=1:100
test=50 normally distributed random numbers (mean=0, var=1);
sig(x)=ttest(samp,test);
end
now look at the sig vector. OMG, 5% of the tests came out significant!!!
Now you are writing a paper all about how x is linked to y. But you are essentially throwing dice and then writing a paper about why it came up '3-3'.
1 1 2 3 5 8 13 21 34 55 89 144 233 377 610 987 1597 2584 4181 6765
"Lies, damn lies and statistics" is all you need to know about statistics.
This is right up there with "'click on the big blue e' is all you need to know about the internet."
Speaking as both a statistician and a computer scientist, I've seen the statistics-vs.-CS argument play out many times before, and the lack of knowledge on both sides is really striking, but not all that surprising -- both are hard subjects which take a lot of work to master. The lack of mutual respect is both infuriating and pathetic, and there's no excuse for it.
The correlation between ignorance of statistics and using "correlation is not causation" as an argument is close to 1.
Wow. What class did you take that says if you don't know something you should assume equal probability?
I don't know if there is an invisible elephant in my kitchen, so I guess I should assign equal probability to both outcomes. I also don't really know how Baccarat works, I guess my odds are 50/50.
Without knowing something about he or his coworkers, you by definition cannot make any statistical statements. To make any statements, you would first need to make some observations. This is how statistics is different from logic. Statistics is grounded in data.
I don't agree with Zed, but you may have just proved his point.
"I construct two sets of n=100 random samples from the normal distribution. Now, if I just take the average (mean or median) of these two sets they seem almost the same."
So its true. The n's justifies the means.
Today's vices may be tomorrow's virtues.
"It's also possible the author of the article is a know-it-all douchebag who tries to solve problems with overwrought solutions."
That was kinda what I got from this. Sure, my powers of ten runs to determine performance isn't statistically sound. Did I say it was? No. Why don't I care? Because my samples are cheap. Spiking vs non-spiking is something pretty easy to see when you glance at the data.
I mean, he said we're going to die if we don't learn statistics, but he never gave a compelling argument for it.
The best example was users, but even that was lacking. If you design a script that's as aggressive on a system as a high use user and your system supports as many 'users' as students, you're safe, if it supports less you work on qualifying the problem better then.
You know, that particular citation has made me wonder in the past, but not enough to actually research it. So, I went off looking for more information and found it.
The statistic was generated from a July 1976 survey.
The sample group for this statistic was 1,200 dentists. These dentists were hand picked by the research company, probably with good reason.
They were asked, what advice would they give gum-chewing patients
1) sugared gum
2) sugarless gum
3) no gum at all.
Sugarless gum got 85% of the vote. Not terribly surprising. I'd be fairly confident that their time had been paid for, or at very least they were told "This survey is being done for Trident Sugarless Gum." That is only speculation, so hush up.
17/20 doesn't really sound very good. It just doesn't stick in your head. 4/5 is close enough, even though it reduces your answer to 80% (ahhh, a lie). Since these are marketing folks, I'm sure they pushed all kinds of values past focus groups, until "4 in 5" was accepted as most favorable.
As the link cites, they're fairly confident that the "sugared gum" answer got at least one response. There's always someone that'll take the obvious wrong answer. If you don't believe that, look at any Slashdot poll. :)
What they don't say is how many of the 1,200 samples were dropped. I'm sure there were non-responses, and they could have easily added any number of unfavorable answers in as non-responses. Of course, they couldn't have 100% in their favor, so they had to keep some.
Serious? Seriousness is well above my pay grade.
Thing is: You can only be expert in ONE of them. Period.
Hundreds of cryptologists prove you wrong.