P-value NaN using chi2gof to validate lognormal distribution in dataset

Question

I am trying to model a dataset of mine with a lognormal distribution using Matlab. I estimated the parameters via 'lognfit' and my generated datapoints with the fitted distribution look quite good compared to the observed data.

To further verify, I wanted to use chi2gof to test the goodness-of-fit. Everything seems to work fine and the null hypothesis is not rejected but the p-value is NaN.

I am using the approach applied to the the Weibull distribution under http://www.mathworks.de/de/help/stats/chi2gof.html

So you can reproduce my result, this is an example that does not work.

>> test = lognrnd(-3.6299, 1.30985, 100, 1);
>> b = fitdist(test, 'Lognormal')

b = 

LognormalDistribution

Lognormal distribution
mu = -3.62425   [-3.86888, -3.37962]
sigma =   1.2329   [1.08249, 1.43223]

>> [h, p, st] = chi2gof(test,'CDF', b)

h = 0

p = NaN

st = 

chi2stat: 0.0183
      df: 0
   edges: [0.0015 0.2151 2.1372]
       O: [94 6]
       E: [93.6703 6.3297]

Also transforming the data to normal distributed datapoints and checking for a normal distribution results in a p-value of NaN.

I am thankful for any help!

Best regards

EDIT:

I was just able to make the transforming-to-normal approach to work:

Would still be nice to know if this approach is valid and why the original approach does not work.

>> test = lognrnd(-3.6299, 1.30985, 100, 1);
>> test = log(test);
>> b = fitdist(test, 'normal')

b = 

NormalDistribution

Normal distribution
mu = -3.68114   [-3.92861, -3.43368]
sigma =  1.24717   [1.09502, 1.44881]

>> [h, p, st] = chi2gof(test, 'CDF', b)

h = 0


p = 0.5330


st = 

chi2stat: 4.1149
df: 5
edges: [-6.1930 -5.5713 -4.9497 -4.3281 -3.7064 -3.0848 -2.4632 -1.8415 0.0234]
O: [8 7 18 16 17 20 8 6]
E: [6.4814 8.9730 14.7440 18.9931 19.1820 15.1883 9.4282 7.0101]

Answer 1

It looks like in the first case you supplied a variable "changes" to chi2gof, with only 15 observations. You wound up with only 2 bins. MATLAB approximates the number of degrees of freedom as the number of bins minus the number of estimated parameters (2) for a total of 0.

In the second case you supplied the same data as used in the fit, with 100 observations.

If you intended to use a different dataset for the fitting and testing, you could tell chi2gof that you did not estimate the parameters, since you didn't estimate them from the data being tested.

[ADDED]

With the updated question, the test data looks very skewed, and chi2gof struggles to find good equal-sized bins to use. So it still winds up with just two bins. You can specify bin edges yourself, for example:

[h, p, st] = chi2gof(test,'CDF', b, 'edges', [0 .008 .02 .04 .1 100])

Working on the log scale is a perfectly reasonable idea, though.