Dea Ali Alamydai

I think this is a big misunderstanding. 

First, for nominal data - if we are going to talk about the Chi-squared test - we are not entitled speak about normal distribution, because we are talking about the distribution of numbers rather than the distribution of features (continuous variables). So if we talk about categories we can only talk about different variants of a chi-squared procedure like likelihood ratio or generalized techniques, but it has nothing to do with distribution.

Secondly, ordinal variable even have no right to normal distribution. These are nominal variables and not continuous variables.

Third, the most common misconception resulting from the central boundary theorem is the assumption that the distribution is normal if the population is large. NO! This statement is a Cumulative distribution function and not a distribution. By the way - pay attention to how very arbitrary is the concept of a large sample.

And the last - parametric tests are relatively resistant to deviations from normal distribution, but some distributions of our results simply will not be properly evaluated with parametric tests. But this is for another topic.

Dear researcher,

Nonparametric statistics or distribution-free tests are those that do not rely on parameter estimates or precise assumptions about the distributions of variables. you can use SPSS Nonparametric statistics to compare 2 independent groups, 2 paired samples, k independent groups, and k related samples. These are some testes: (Mann-Whitney Test, Wilcoxon’s Matched-Pairs Signed-Ranks Test, Kruskal-Wallis One-Way ANOVA and Friedman’s Rank Test for k Correlated).

However, this link have more details:

http://www.uvm.edu/~dhowell/fundamentals7/SPSSManual/SPSSLongerManual/SPSSChapter12.pdf

Can you provide more information about the nature of the qualitative data you are collecting as your dependent variable? This information may enable me to answer your question with more detail, although in general, I concur with Al-Barashdi.

A little more information about the qualitative data would be helpful, if you are coding and categorizing the data into nominal and/or ordinal variables then yes, any nad/or all of the nonparametric tests mentioned above may prove useful. 

If you are coding each participant Yes/No for the presence or absence of some trait or activity, then the simplest non-parametric statistic would be chi-square.

I do not see in how far your inquiry follows a qualitative method. If you just want to investigate into the interdependence between province (independent variable) and practice (dependent variable) you have two nominal scales. All measures suitable for nominal X nominal will do. Start with analyzing a crosstab with the independent variable across the top (columns) and the dependent along the side (rows). You can understand how the interdependence which might not be linear works. If you want to test your hypothesis, chi-square is a good parameter. You can also use chi-square derived correlation measures. If your data are really qualitative, you have to give more details.

When you say "all" the responses, is your plan to reduce a large number of responses into a smaller number of constructs (where for example methods for the GIFI models or latent class might be appropriate)? Also, are the farmers randomly allocated to those three conditions and is there any contact between them that could create problems for the independence assumption?

Thanks all for your answers. Here is some details of my data. I have three groups (group A, B, and C) of farmers in each province (3 provinces in total: P1, P2, and P3). Group A (adopt a good practice in aquaculture), group B doesn't adopt that good practice and locates nearby group A, group C doesn't adopt that good practice and locates far away (5km minimum) from group A and B. During the interview, I also ask whether each group know what other groups are doing and they said "no" So I assume my variable is independent. I have two sorts of data

1. Yes/No responses with explanation why they chose Yes/why they chose No. I am thinking about using Chi Square to analyse it.

2. 5-point likert responses. I intend to use Kruskal-Wallis but I am not sure if this analysis fits to my data.

I think you are on the right track, Duc. Those statistical tests appear to be the most relevant to the data you describe to me.

I recommend suggestions made by Charles.

For couples of variables in traditional statistics, obviously the methods recommended above are the best. But if you want to preserve all variables and see the picture with all variables, I recommend network analysis. I use a free tool called Kumu.io that I think applies for your need.

Hi researcher,

Using SPSS will help you only when you have yes/no answers. As David said earlier, you might then consider Chi-square.

Cheers

Looks like all the colleagues here reached to a consensus about the methods to use. So, simply follow the recommendations. I feel that Jose's Kumu.io could validate your chosen approach. That is, running Chi-Sq first then the instrument above and compare.

I agree with Jose.  I see a lot of references to Chi Squared for multiple yes/no questions. I am not sure that I understand. If you have only two binary variables then that may be sensible (and it is reasonable to have a table of all the pairwise associations, perhaps with kappa or the log-odds ratio), but depending on how many yes/no questions, would something like IRT (item response theory) be reasonable. Or by Chi Square are the commentators meaning using a multivariate test that has a test statistics that is distributed Chi squared? Network analysis can mean lots of things, so could work too. Either way, specify your research question can help.

Dear Duc

Is not statistical true using non parametric for sampling size more than 40 because your data are normally distributed by central limits theorey so why you choose to lost some degree of confidence about your results?!

Any way if you want to used kruskal wallis is used to compared two or more independent groups.

Good luck

Dea Ali Alamydai

I think this is a big misunderstanding. 

First, for nominal data - if we are going to talk about the Chi-squared test - we are not entitled speak about normal distribution, because we are talking about the distribution of numbers rather than the distribution of features (continuous variables). So if we talk about categories we can only talk about different variants of a chi-squared procedure like likelihood ratio or generalized techniques, but it has nothing to do with distribution.

Secondly, ordinal variable even have no right to normal distribution. These are nominal variables and not continuous variables.

Third, the most common misconception resulting from the central boundary theorem is the assumption that the distribution is normal if the population is large. NO! This statement is a Cumulative distribution function and not a distribution. By the way - pay attention to how very arbitrary is the concept of a large sample.

And the last - parametric tests are relatively resistant to deviations from normal distribution, but some distributions of our results simply will not be properly evaluated with parametric tests. But this is for another topic.

Just to add to Tomaaz's excellent answer, deviations from normality can have a large effect on power (and this is true for fairly big and small samples). This was one of the main takes from Tukey's 1960 chapter (and the Princeton robust group) and is why Rand Wilcox has several clever titles about how using non-robust methods means you are losing power (https://www.researchgate.net/publication/257568054_ANOVA_A_Paradigm_for_Low_Power_and_Misleading_Measures_of_Effect_Size, see his paper for the Tukey reference), but Fisher talk about this long ago

(outliers affect the variance more than the mean because of squaring,

Group 1:     2,3,4,5,6

Group 2:    5,6,7,8,9

gives t(8) = 3.00, if Group 2:   5,6,7,8,99 gives t(8) =1.13.). The central limit theorem will mean things will look more normal, but you need to look at the tails and differentiating for example 2.5% and 8% from a tail is tough.

Article ANOVA: A Paradigm for Low Power and Misleading Measures of E...

I agree with the contributions of Ali and Toms

What are the characteristics of the qualitative data?

Hi Duc,

Parametric and non-parametric tests are valid for quantitative data. Given that your data is qualitative in nature, it is narrative and thus not quantified. However, if you data can be categorized into ordinal scales such as a Likert scale (strongly agree to strongly disagree, for example) this could be quantified and makes more sense to apply normality testing.