As far as I remember non-parametric tests tend to be more conservative. They throw away lots of information, in the case of Man-Whitney only the rank is kept. So many people prefer to test to check whether gaussianity is fulfilled. In a t-test gaussianity is reported not be very critical, whereas it is indeed for other tests, like F.test. It is a good practise to perform the parametric test together with the non-parametric.

Generally speaking, there is no "best choice", for instance CHI-sq points on frequencies in certain classes and may be sensitive to values at the tail of the distribution, whereas as the Kolmogorov test does not care much about the tails.

In general non parametric test are less rigorous than parametric tests. They can be applied to small samples, or to dichotomic and/or ranked variables.

If one or more of these assumptions are violated:

1. non continuous variable

2. mean and variance independent

3. not normal distribution

4. dis-homogeneous variances

5. small samples

it is possible to apply non-parametric tests. They are called "non-parametric" because they do not involve the estimation of statistical parameters of a distribution (mean, standard deviation, variance, etc..).

Non-parametric (or distribution-free) tests do not require special conditions to be applied. It is not necessary that the sample contains more than a certain number of elements. It is not necessary that the magnitudes compared to follow a certain distribution

They do not have as its purpose the estimation of population parameters on the basis of the characteristics of the sample. When the type of distribution can be likened to one of the standard distributions, parametric methods are more efficient estimators. However, even non-parametric tests lead to the same conclusions. They are particularly suitable for comparing very small samples and / or not following any known distribution.

In fact, non-parametric test are not bound instead of specifying a parametric distribution model and are valid regardless of the size N of the sample.

one of the most important things in the statistical analysis is the change in the variable.

Parametric tests have several observations as

1 - . The scale of the variable has to be almost interval scale ( http://www.graphpad.com/support/faqid/1089/ )

2 - . N> 5 (for testing the Gaussian distribution with the test ... QQ

3 - . the data set has no extreme values ​​( with Z transform (mean SD divided data ) ) to be enclosed Zvalues ​​-3 to 3 ..

4 - . Dataset have a Gaussian distribution (test QQ )

Parametric tests are based on a Gaussian distribution ... distribution this distribution is called "normal" ...... in fact , Laplace and Gauss 's basically called the normalized distribution ... and is based on a Z transformation

However, normalization can be based on another approach, such as percentage or logatitmos ...

normalization is used only to describe two or more variables on the same scale ( as the proportional scale in%).

is the statistical analysis of the transformation necestita not alter the native distribution of the variable.

if a scatter plot of a data set (A ) (A vs A) ised get a perfect 45 degree line ... the interpretation of this data is that A has the same distribution as the data A )

whether to transform the data as plotted A * 1000 and A * 1000 vs A dataset both have the same distribution ...

if we made a Z transformation (data-mean/ Sd) .... and ZA vs. graph gives the same line is 45 degrees ... Z transform anomalies also called standardized ... when Z have values ​​- When Z (mean data ) transformation is performed zero means that the data operating in the middle ....

If you can generate a data set of about 3000 random numbers to transform and re- obterner Ztransfromation curve 45 interplreta enjoyments things that both data sets have the same distribution ...

if you graph A logA or loge vs The graph obtained is a curve that curves more at the ends ... in other words the logarithmic transformation does not respect the natural distribution of the data

You can use non-parametric statistical perfectly with your data ... in fact, I prefer to use this type of statistics if my data do not meet the requirements for parametric tests.

All parametric test has a similar non-parametric test .

Many people use the parametric test ( for example , in the latter paper that highest status ) for several reasons

a) just to make a comparison with previous work

b ) Just apply some statistical tests , but the conclusions are not based on the tests performed . ( the conclusions are the same with or without statistical test )

c ) the algorithm in which the test is based is not known and therefore the requirements thereof are not known

these points are not a scientific question ....

my recommendation is ...

check all hypothesis testing ( gaussinidad , no extrem data etc .....

if your data does not meet some requirement , I would apply a nonparametric alternative test

Thank you all for your answers. They provided some important context and perspective. Regards, Clint

Hi Clint, I agree with Horst’s point about non-parametric tests being more conservative because they do not make assumptions. If you are reasonably certain that the assumptions will be met (i.e. a specific transformation will normalise your data either because you can test it or you know the system you are studying) then a parametric test should give you a deeper insight. However, we tend to act without thinking and ‘follow the crowd’; it’s always good to question how appropriate the analysis we are performing is. A couple of things to remember, first any test is not absolute (even though the break points for significance make it seem so) and it is up to you to interpret it; second if you do transform then be careful with any back-transformation as it may give a minor distortion.

Suppose that we intend to find anomaly. In parametric method, we have mathematical insight about threshold as we have to know the distribution. In Non-parametric method, we need to calculate I.Q.R i.e., inter quartile range and then take 1.5 times I.Q.R, why 1.5 times? So adhoc value, not based on rigorous mathematics.