end of the year is it's still scary times! For DJ Bem, a psychologist at Cornell University, a preliminary version has provided a professional article, in which it is about nothing less than the experimental detection of Psi-powers! The promising title: "Feeling the Future: Experimental evidence of abnormal retroactive effects on cognition and affect" ( Feeling the Future: Experimental Evidence for Anormalous Retroactive Influences on Cognition and Affect ). And in this article nine psychological experiments and their results are presented, which have in fact in itself! For it is claimed on the evidence, not the future events, the occurrence is determined, can influence the thinking and acting. So it should be possible for human beings, the place of emergence of pornographic images ( or representations of "couples who are involved in non-violent and consensual, but explicit sexual conduct" as the scientist formulates ) on the computer screen to know before has set a random this place. Or one should better remember words you will learn in the future, even if you do not even know that they will learn it and has not yet determined a random word to be learned. The research results are of course spectacular, so I want to forget sometimes that I am neither a psychologist nor a statistician, but at times to take a closer look at the article!
Of course, the abnormal effects of current decisions on future very small. If they were not, they'd notice in everyday life. The low strength of these effects, however, will force to perform statistical calculations to determine whether the observed effect is a mere accident, or genuine "psi powers" are at work. The easiest way to show it on the first experiment of the article. There are subjects to predict which appear in two fields on the screen an image. After the election of the subject chooses a random field from, on which the image is then displayed. If only chance is at work, one should expect on average 50% correct predictions. But when it comes to porn images, the hit rate was 53.1%, slightly above the expected hit rate by chance. Is this evidence of a mysterious prediction of the future? To decide this you need at the statistics. For if one performs only a few attempts, it will not be likely to deviate from the theoretical mean, even if only pure chance in the game. If you only as a coin twenty times throws, it will not appear particularly suspicious if you have 12 times the "head" and only 8 times tails thrown. If one but two million times the coin throws, and it has 1.2 million times, "head" 000 times and 800 "number" to get, then maybe it would pay the Verdaches that the coin was easily manipulated. And we must also decide in the article, what is more normal accident, and what is a real effect. And to the author of the article, a whole series of statistical tests are used, the results are almost universally that the chances for a purely random occurrence of the test results are so at around the 1% for individual experiments. For an experiment that may not be as impressive, but in nine experiments? That is quite remarkable, and you want the results like to take a closer look! But since then the problems begin.
For the article presented the actual test results almost absent, and it is hardly possible to recalculate the probability values given for the random occurrence of the results. The only thing that is possible for some trials, the review of " binomial test," carried out by the author. And since meeting a little strange ...
Let's start with Experiment 1, in which the pornographic images. In the description of the experiment is that only 40 subjects twelve guesses with porn images 're running as a perceived negative pictures and neutral-looking images. Another 60 subjects have completed each of 18 experiments with porn images not erotic, but positive images occupied. So overall, should times 18, ie, 1560 experiments have been carried out with erotic pictures 40 times 12 plus 60th Of all the experiments with erotic pictures to have been made in 53.1% correct predictions. That would be 828 or 829 hits, just the can not tell when rounding to one decimal. This result has tested the author to a binomial test against the hypothesis that the test result is only a coincidence. He is a z-value of 2.30 and a probability p for that one has 828 (or 829?) hits or more at 1560 rate experiments at 1.1%.
we Convert to. The z-value is obtained by dividing the measured value: converts (here 828 or 829) to a normal distribution with mean 0 and standard deviation 1. The conversion should be very simple in this case, because the measurements should, if only by chance in the game is to follow a binomial distribution with p = 0.5 . In n trials (n = 1560 here is ) is the expected value then n / p and the variance is the square root of np (1-p) . To get to the z-value must be deducted from the value of Erwarungswert be, and are divided by the variance. With the given numbers (p = 0.5 , n = 1560 ) to get it to a z-value of 2.43 (for 828 hits) or 2.48 (for 829 goals), but not to 2 , 30, as mentioned in the article. No idea what is wrong here.
The probability p can be determined directly from the cumulative binomial distribution. Results for 828 would be the 0.81% for 0.70%, but not 1.1% as in the article. However, we obtain the stated 1.1% if you take the z-value of 2.30, and assuming a normal distribution. Obviously, all the chances in the article are not from the binomial distribution directly but calculated using the z-value from a normal distribution was. This is maybe not very elegant, but caused only a minor error.
Whence come the differences between the above and recalculating the article, I do not (comments welcome!). But there are still a few more cases in the article, where you can verify the binomial test. And since the observed deviations are very small, but all the more remarkable.
Take the first times in Table 2 In the second column as the results for Experiment 2 are given. In total of 5400 tests were 2790 hits, ie the hit rate was 51.7% (correct). The z value of this is given as 2.44. If you add yourself to, we find 2.45. That's not very dramatic, and the value for p result of 0.7% from the normal distribution for both 2.44 and for 2.45. Perhaps only a small rounding error.
we go to Table 3 Here are 963 results for 1800 trials, ie a rate of 53.5% (right). The specified z-value is 2.95. If you the math, but we find 2.97. That's strange. The value of p of 0.2% must then also have been calculated z = 2.95 , because with z = 2.97 would be a value of p received = 0.1%.
is still more remarkable is when we go to Table 6 of the article and then recalculate the figures for the binomial test. Here there are a total of 2304 tests results in 1105, ie a rate of 48.0% and not 47.9% as in the article. As z-value is -1.94, and p with this z-value was calculated. If one is found, however the math a z of -1.96.
Conclusion:
The probabilities p have not been calculated directly from the test results, but from the z-values. The z-values do not match the given test results. The differences are very small, in the final digit shown, but not one, but two digits next to it, which argues against a simple rounding error.
Where do these deviations, I do not, of course. But we can make a little mental game. Suppose I n have conducted trials. Suppose, further, I'm not myself so much for the actual number of hits, which was achieved, but I would rather have a specific z-value that an attempt is unlikely to result by pure chance. Then I choose a z-value that I would have liked. By then I calculate the probability p . Only I have now the right number of hits Found that I would have determined experimentally. So I count back with n and the z-value of the number of hits. Only the number of results must of course be an integer, so round Would I return my result by calculation to the nearest whole number, I do that, I came with my Table 6 for z from -1.94 to 1105.44 hit, so 1105, as indicated. For Table 3 gives the fixed z = 2.95 then 962.58, then 963 hits, as indicated. And Table 2, with z = 2.44 , I get hit 2789.65, rounded to 2790, as indicated. So I would have exactly the number of hits, z-values and probabilities p that are specified in the tables work.
The tiny differences between the recalculated z-values and z values stated in the article can be understood so easily if we assume that the conclusions of the work were not determined from the experimental results, but the test results from the conclusions.
And once again I am happy to run my anonymous blog ... ;-) But maybe I am doing a mistake. In any case, I do not trust the spectacular proof of psi powers only once no tiny little bit! And whether this will change in the future, this knowledge must probably still be hiding in my subconscious.
0 comments:
Post a Comment