0
$\begingroup$

This question already has an answer here:

I have an array dataset of about 650.000 points. I want to test if the dataset follow a normal distribution or any other distribution. The first thing I did was to split the data in groups, find the frequency of each one and then create a simple chart. Here is the result:

LogY axis: Frequency Graph

Normal axis: enter image description here

Then, I tried to apply a statistic package in Python for Hypothesis null.

The first one: Link 1 with result (26576.286833062259, 0.0)

The second one: Link 2 with result (1.0, 0.0)

Both of them have p-value equal to zero which means that the data doesn't follow a normal distribution. Do you think that that this is valid or am I doing something wrong?

I tried it also with all the different distributions from this link and here is the results:

anglit (1.0, 0.0)
arcsine (1.0, 0.0)
cauchy (0.97256179317853242, 0.0)
cosine (1.0, 0.0)
expon (0.99998329829920973, 0.0)
gilbrat (0.99251782771674257, 0.0)
gumbel_r (0.99998329843868239, 0.0)
gumbel_l (1.0, 0.0)
halfcauchy (0.94812454120633505, 0.0)
halflogistic (0.99996659715630376, 0.0)
halfnorm (1.0, 0.0)
hypsecant (0.9999893673670458, 0.0)
kstwobign (1.0, 0.0)
laplace (0.99999164914960492, 0.0)
logistic (0.99998329857815205, 0.0)
maxwell (1.0, 0.0)
norm (1.0, 0.0)
rayleigh (1.0, 0.0)
semicircular (1.0, 0.0)
uniform (1.0, 0.0)
wald (0.99981187008374384, 0.0)
$\endgroup$

marked as duplicate by Alexis, Andy, Scortchi, Nick Stauner, whuber Aug 11 '14 at 18:05

This question has been asked before and already has an answer. If those answers do not fully address your question, please ask a new question.

  • 1
    $\begingroup$ This link may help. $\endgroup$ – Zhubarb Aug 11 '14 at 10:55
  • 1
    $\begingroup$ You have more than half a million points, so you'll reject more or less anything with prejudice at 'conventional levels'. $\endgroup$ – conjugateprior Aug 11 '14 at 13:31
  • 2
    $\begingroup$ If you want to know whether you've run into a software (or a numerical) bug, take subsamples of rather small size and compute the test statistics. p's should shrink to zero as you increase sample size until it's basically zero. $\endgroup$ – conjugateprior Aug 11 '14 at 13:34
  • 2
    $\begingroup$ Then I think that the test is doing what it's meant to. The question that @Zhubarb's comment brings up indirectly is 'what do you want this thing to be Normal for?' If it's to run a linear regression model and interpret coefficients it often won't matter at all. The CLT should have kicked in for most relevant quantities already. $\endgroup$ – conjugateprior Aug 11 '14 at 13:50
  • 1
    $\begingroup$ If you have that many data points, you can use many more bins to think about your distribution. $\endgroup$ – Nick Cox Aug 11 '14 at 14:05