People angry that Apple rejected Opera on the iPhone should probably read John Gruber at Daring Fireball, who investigated this and found out that it doesn't seem to have happened at all, since Opera hasn't submitted the browser to Apple yet, let alone had it rejected.
You can be angry at Apple for their ham-handed handling of the App Store as much as you like, but the "Opera rejected by Apple" story is, so far, from the Precrime files.
Fair enough. Given that he provided copious links to the original articles in his blog post and that he provided a reasonably thorough summary (for a journalist) of the work, I can't say that this really bothers me at all. However, as is clear, I'm not overly familiar with Slashdot etiquette, so I'll just take your word for it that it's a Bad Thing.
The article is full of links: to the Nature article webage, to the authors' homepages, and even to a link with a.PDF containing the article itself. Did you actually read the ZDNet article?
I think that's one of the problems with Linux, actually: that a "quick" solution involves checking out code from a Subversion repository. As a more practical note, one of the problems I've had with Ubuntu is that (over two months into using it), my DVD burner is *still* not full recognized (it burns at about -100x speed, and reads even slower). When I've tried to fix this, about a dozen times now, I invariably end up trolling through about 6 different "HOW-TOs" or suggested fixes at a given time, most of which give different and sometimes contradictory advice to solve the same problem! Some of this is because one of them is out of date, or the author was just plain wrong, or I just didn't understand enough about low-level workings of the kernel to realize that they were actually talking about slightly different problems. It's not that the community is unhelpful, or unresponsive, because they're not - but they are disorganized, and many times the left user doesn't know what the right user is doing.
This all reminds me of an article I read when I started using Linux: http://linux.oneandoneis2.org/LNW.htm. The author makes a great point about how Linux is not Windows, and I think until Linux decides that it *wants* to be Windows or that it wants to be something better FOR THE AVERAGE USER, it will never have widespread adoption. (And by average user, I don't mean the kind of person for whom checking out code from a repo is common practice, I mean the kind of person who's never heard of a "forum" and gets frustrated if something takes more than two clicks which are highlighted with shiny red arrows. Linux people don't understand that kind of user very well. Neither do Windows or OS X people, sometimes, but at least they're trying.)
You're probably right - as I mentioned earlier, it's entirely possible that he just got lucky, and he does display a rather uncertain grasp of many of these concepts. I never said that he was good at this, I just said that the results of the correlation and the chi-square are effectively interchangeable in this case - if you test the significance of the correlation, you'll get exactly the same p-value as you'd get from doing a chi-square. Everyone around here piled on because *obviously* the only way you could do this was a chi-square (or Fisher's exact), and I just don't think that is either correct or fair.
But maybe my wording was poorly chosen, and I implied that I thought he was actually entirely correct throughout the article. If so, I regret that, because I don't.
P.S. A correlation coefficient itself doesn't take sample size into account, but if the author *had* done a significance test as he could have, *that* would have taken the sample size into account when calculating the standard error. I honestly think that the author simply doesn't know the difference between practical and statistical significance - he effectively argued that 0.046 wasn't significant because it wasn't practically significant, but he used the language of statistical significance.
Causality isn't implied in the experimental design, it can only come from the experimental design. In any case, the null hypothesis in Fisher's test doesn't test anything about causality. Fisher's test (like its approximate cousin chi-squared, and the phi correlation) will tell you if there is an association between the row (seeded with a yawn) and column (yawned) variables. It doesn't tell you if seeding caused yawning. You can reverse the variables, and you will get exactly the same response from the test. This is just the same as correlation. If you have data from an experimental design, you can calculate a correlation on it to measure the strength of the relationship and still claim causality (because of the design, not the test).
As to you being half-way through your Ph.D, argument from authority is fallacious. If it weren't, though, I'd be happy to compare credentials with you. I guarantee you that mine are just as good as yours, if not better.
The correlation test is justified, just as justified as the chi-square test would be. It doesn't say anything about causality, and neither did I. The article's conclusion that there was no significant relationship between the two variables is correct (and I am not talking about the design of the experiment, nor do I have any wish to), and the method that TFA employed works exactly the same as a chi-squared would. As to causality, I've explained about this elsewhere in this thread, and so did another person (quite well).
Where did all this causality stuff come from? I didn't raise it, I didn't imply it, I had nothing to say about it. All I was trying to do was point out the relationship between the test that was being howled for (chi-square) and the correlation that was actually computed in TFA. Nobody has actually addressed that, except for the first reply to my original post.
As for what you wrote (which must have been as I was writing the other reply), you're right - I said exactly the same thing just below your reply.
Causality comes from the design of the experiment. If you randomly assign to treatment conditions (yawn / no yawn) and then measure a response variable, you are doing a randomized experiment and you can make inferences about causality. When you have an experimental design (not an observational study), you can do whatever test you like and it won't change the cause-and-effect relationship. There is nothing about Fisher's exact test that will "determine causality" if the data is collected from an observational study.
I shouldn't have to tell you this; would you like some references on basic research methods and statistics?
I never even came close to implying that correlation implies causality. Can you tell me what I said that makes you think that? I'd be happy to explain further.
Fisher's exact test is indeed a good test for this data, but I'm trying to point out that tests on contingency tables give the same results as calculating a Phi correlation coefficient (because they're related!). Can you demonstrate that I'm wrong?
1). I wasn't talking about the design of the "experiment". I'm just trying to point out that the statistical test employed reduces to *exactly the same thing* as the test that everyone else is championing.
2). What exactly do you think a chi-squared test is doing? It's measuring the relationship between the two variables. The null hypothesis in a chi-squared is that the row and column variables are independent. The correlation equivalent (phi) is doing the same thing with dichotomous variables, except it puts a number on the strength of the relationship which can be interpreted in the same fashion as the Pearson correlation coefficient.
I completely agree that the author of the article has a shaky grasp on these concepts, but the fact is that there wasn't actually anything wrong with the test employed - it works just as well as the chi-squared for this data set. True, the numbers regarding what would be required for significance are completely off-base, but that still doesn't mean that s/he originally employed the wrong test.
As to significance, it's common usage to refer to a large correlation coefficient as indicating a "significant" relationship between the two variables. I agree that it's confusing, but I wouldn't call it entirely incorrect unless you want to get really picky.
Well, a chi-squared test would have worked too, but so would Phi correlation (a correlation between two dichotomous / binary variables), which can be computed exactly the same as... Pearson correlation, which TFA used. In fact, if you take the chi-square value you worked out to a few more decimal places: 0.10504 (from R), divide by 50 (=N, the sample size), and then take the square root, you get 0.046, which is the phi (and hence Pearson) correlation coefficient for the TFA's data. I can't tell if OmniNerd knew this or if s/he got lucky, but there's nothing wrong with the test employed.
So, TFA's conclusion was correct, and so - whether intentionally or not - was the method. It's just not nearly as common as chi-squared test for a 2x2 table.
Slashdot Mirror
People angry that Apple rejected Opera on the iPhone should probably read John Gruber at Daring Fireball, who investigated this and found out that it doesn't seem to have happened at all, since Opera hasn't submitted the browser to Apple yet, let alone had it rejected. You can be angry at Apple for their ham-handed handling of the App Store as much as you like, but the "Opera rejected by Apple" story is, so far, from the Precrime files.
Way to validate Godwin's law.
Fair enough. Given that he provided copious links to the original articles in his blog post and that he provided a reasonably thorough summary (for a journalist) of the work, I can't say that this really bothers me at all. However, as is clear, I'm not overly familiar with Slashdot etiquette, so I'll just take your word for it that it's a Bad Thing.
Ah - I see. Well, thanks for the explanation. :-)
If by "new" you mean "I usually read the articles and don't comment often", then you're right. Are you trying to tell me that I did something wrong?
The article is full of links: to the Nature article webage, to the authors' homepages, and even to a link with a .PDF containing the article itself. Did you actually read the ZDNet article?
I think that's one of the problems with Linux, actually: that a "quick" solution involves checking out code from a Subversion repository. As a more practical note, one of the problems I've had with Ubuntu is that (over two months into using it), my DVD burner is *still* not full recognized (it burns at about -100x speed, and reads even slower). When I've tried to fix this, about a dozen times now, I invariably end up trolling through about 6 different "HOW-TOs" or suggested fixes at a given time, most of which give different and sometimes contradictory advice to solve the same problem! Some of this is because one of them is out of date, or the author was just plain wrong, or I just didn't understand enough about low-level workings of the kernel to realize that they were actually talking about slightly different problems. It's not that the community is unhelpful, or unresponsive, because they're not - but they are disorganized, and many times the left user doesn't know what the right user is doing.
This all reminds me of an article I read when I started using Linux: http://linux.oneandoneis2.org/LNW.htm. The author makes a great point about how Linux is not Windows, and I think until Linux decides that it *wants* to be Windows or that it wants to be something better FOR THE AVERAGE USER, it will never have widespread adoption. (And by average user, I don't mean the kind of person for whom checking out code from a repo is common practice, I mean the kind of person who's never heard of a "forum" and gets frustrated if something takes more than two clicks which are highlighted with shiny red arrows. Linux people don't understand that kind of user very well. Neither do Windows or OS X people, sometimes, but at least they're trying.)
You're probably right - as I mentioned earlier, it's entirely possible that he just got lucky, and he does display a rather uncertain grasp of many of these concepts. I never said that he was good at this, I just said that the results of the correlation and the chi-square are effectively interchangeable in this case - if you test the significance of the correlation, you'll get exactly the same p-value as you'd get from doing a chi-square. Everyone around here piled on because *obviously* the only way you could do this was a chi-square (or Fisher's exact), and I just don't think that is either correct or fair.
But maybe my wording was poorly chosen, and I implied that I thought he was actually entirely correct throughout the article. If so, I regret that, because I don't.
P.S. A correlation coefficient itself doesn't take sample size into account, but if the author *had* done a significance test as he could have, *that* would have taken the sample size into account when calculating the standard error. I honestly think that the author simply doesn't know the difference between practical and statistical significance - he effectively argued that 0.046 wasn't significant because it wasn't practically significant, but he used the language of statistical significance.
You know, googling "Anonymous Coward" and "Nature" didn't return the results you might have expected...
Causality isn't implied in the experimental design, it can only come from the experimental design. In any case, the null hypothesis in Fisher's test doesn't test anything about causality. Fisher's test (like its approximate cousin chi-squared, and the phi correlation) will tell you if there is an association between the row (seeded with a yawn) and column (yawned) variables. It doesn't tell you if seeding caused yawning. You can reverse the variables, and you will get exactly the same response from the test. This is just the same as correlation. If you have data from an experimental design, you can calculate a correlation on it to measure the strength of the relationship and still claim causality (because of the design, not the test).
As to you being half-way through your Ph.D, argument from authority is fallacious. If it weren't, though, I'd be happy to compare credentials with you. I guarantee you that mine are just as good as yours, if not better.
The correlation test is justified, just as justified as the chi-square test would be. It doesn't say anything about causality, and neither did I. The article's conclusion that there was no significant relationship between the two variables is correct (and I am not talking about the design of the experiment, nor do I have any wish to), and the method that TFA employed works exactly the same as a chi-squared would. As to causality, I've explained about this elsewhere in this thread, and so did another person (quite well).
Where did all this causality stuff come from? I didn't raise it, I didn't imply it, I had nothing to say about it. All I was trying to do was point out the relationship between the test that was being howled for (chi-square) and the correlation that was actually computed in TFA. Nobody has actually addressed that, except for the first reply to my original post.
As for what you wrote (which must have been as I was writing the other reply), you're right - I said exactly the same thing just below your reply.
Causality comes from the design of the experiment. If you randomly assign to treatment conditions (yawn / no yawn) and then measure a response variable, you are doing a randomized experiment and you can make inferences about causality. When you have an experimental design (not an observational study), you can do whatever test you like and it won't change the cause-and-effect relationship. There is nothing about Fisher's exact test that will "determine causality" if the data is collected from an observational study.
I shouldn't have to tell you this; would you like some references on basic research methods and statistics?
I never even came close to implying that correlation implies causality. Can you tell me what I said that makes you think that? I'd be happy to explain further.
Fisher's exact test is indeed a good test for this data, but I'm trying to point out that tests on contingency tables give the same results as calculating a Phi correlation coefficient (because they're related!). Can you demonstrate that I'm wrong?
1). I wasn't talking about the design of the "experiment". I'm just trying to point out that the statistical test employed reduces to *exactly the same thing* as the test that everyone else is championing.
2). What exactly do you think a chi-squared test is doing? It's measuring the relationship between the two variables. The null hypothesis in a chi-squared is that the row and column variables are independent. The correlation equivalent (phi) is doing the same thing with dichotomous variables, except it puts a number on the strength of the relationship which can be interpreted in the same fashion as the Pearson correlation coefficient.
I completely agree that the author of the article has a shaky grasp on these concepts, but the fact is that there wasn't actually anything wrong with the test employed - it works just as well as the chi-squared for this data set. True, the numbers regarding what would be required for significance are completely off-base, but that still doesn't mean that s/he originally employed the wrong test.
As to significance, it's common usage to refer to a large correlation coefficient as indicating a "significant" relationship between the two variables. I agree that it's confusing, but I wouldn't call it entirely incorrect unless you want to get really picky.
Well, a chi-squared test would have worked too, but so would Phi correlation (a correlation between two dichotomous / binary variables), which can be computed exactly the same as ... Pearson correlation, which TFA used. In fact, if you take the chi-square value you worked out to a few more decimal places: 0.10504 (from R), divide by 50 (=N, the sample size), and then take the square root, you get 0.046, which is the phi (and hence Pearson) correlation coefficient for the TFA's data. I can't tell if OmniNerd knew this or if s/he got lucky, but there's nothing wrong with the test employed.
So, TFA's conclusion was correct, and so - whether intentionally or not - was the method. It's just not nearly as common as chi-squared test for a 2x2 table.
Actually, a relationship with God can apparently be had through your local intercom. :-)