Use the various non-parametric analyses as appropriate for your design. For further help, see statistics.Laerd.com. This is a subscription site but well worth it.

Best,

Vic Landry

What do you want to do with the variables?

One option would be to dichotomise your variables (for each variable X, transform it into another variable Y where Y=1 when X=100 and Y=0 when X

Thank you Victor, thank you Miguel.

The objective is to determine if there is a relationship between several categorical variables (location, sex, gender, income stratum, unemployment, etc.) and variables that seek to measure "sensation/fear" of violence. I use non parametric test (Kruskall-Wallis for example). But I'm trying to find out other statistics analysis tools adequate for the survey that was done.

I thank the collaboration of all

1) Am I right to assume that the variables that measure "sensation/fear" of violence are the ones that are skewed?

2) Are you

a) looking to do an exploratory analysis where you want to check the relationship between each categorical variable and each of the variables that seek to measure "sensation/fear" of violence?

b) looking to do an analysis where you try to find how much the categorical variables, taken together, influence the variables that seek to measure "sensation/fear" of violence?

3) What is your research question?

Dear Miguel,

1) You are right: the variables that measure "sensation/fear" of violence are the ones that are skewed

2) I´m looking to do an exploratory analysis where you want to check the relationship between each categorical variable and each of the variables that seek to measure "sensation/fear" of violence (but I also really like the way you frame b & I would like to do that)). Hence non-parametric test (K-W).

3) What is the influence of locality (where you live, work or study) on the "sensation/fear" of violence

Thank you very much

1- I don't really like the idea of using K-W when the variable is constant for most of the individuals.

2- What does it mean that the "sensation/fear" of violence variables are 100 for most people? Is it that they're afraid?

a) if you can dichotomise those variables, you can use fisher test (or chi-square) to check the association between each of the categorical variables and each of the "sensation/fear" of violence variables.

b) it could be even better if you could reduce all of the "sensation/fear" of violence variables into one variable defined by

[option A] the mean of all of them

[option B] the proportion of times an individual expresses fear

[option C] something smarter,

as long as that new variable carries meaning.

If you're able to do this, then you could check how much that new variable changes between the groups of each categorical variable, indeed using K-W if it is a quantitative or ordinal variable or fisher/chi-square if it is a categorical variable.

c) again if you could do like in b, maybe you can use a regression to check how much the independent variables (what you called the categorical variables) influence the dependent variable (that one created from the variables expressing "sensation/fear" of violence, if it is possible to create one).

1) OK (you are right)

2) Example: afraid to walk in the neighborhood when dark? or fear of being in the midst of the crossfire (a confrontation between policemen and bandits)? where 0 is NO fear at all and 100 is the most

3) is that the distribution of the variables does not compromise the use of traditional regression methods (even non-parametric ones)?

One more: Thank you very much

You're welcome! Wow, then most people are really afraid most of the times in your study. I think you might be able to

a) dichotomise your variables expressing fear, so that they are 1 when the person is completely afraid (the original variable was 100) vs 0 when the person is not completely afraid (the original variable was does this new variable have enough variation? Treat it as a quantitative variable which is very interpretable: it's the number of times that your participants express being totally afraid

-> does this new variable not have enough variation? Dichotomise it so that it now is something that everyone can understand what it means, e.g., it's now 1 for people who express being totally afraid at least 3 times and it's 0 for people who express being totally afraid 2 times or less

Dear Miguel

Unfortunately, problems of public insecurity are very intense in Brazil (although there are huge variations between spaces and I try to understand how this is reflected in the perception of violence).

Following your suggestions:

1) I will dichotomize (1 = 100, 0 = others) and run a logistic regression

2) I will also create ordered intervals (for example, 0 to 19, 20 to 39, 40 to 59, 60 to 79, 80 to 89 and 100). The independent variable would be ordered Could I use ordered regression with such a distribution of the independent variable & given that the independent variables are categorical?

3) I will create - as you suggest in b) - an average variable of "fear", dichotomize it (and I think that I can create another ordered dependent variable) and pay attention to variability.

Thank you very much

an average variable of "fear", dichotomize it (and I think that I can create another ordered dependent variable) and pay attention to variability.

A naive suggestion: have you tried taking the logarithm of the skewed variable? A linear relationship in a log-variable has a straight forward interpretation & to a 1st order the log transform might normalize the data.

As an aside, it does seem to me that quantifying "fear" is a subjective endeavor & has the potential to be non-scientific. I am not sure that any statistical results that you have from statistical tests are meaningful.

Dear Jack

Thanks. Yes, I tried the Log transformation. Following a few steps recommended by Miguel, I'm starting to have results, which I believe, interesting. I'm still debating which type of regression is most appropriate (ordered logit, quasi-poisson, censered). But it's interesting.

As for the sense of using statistical tools to deal with the issue of fear, some results are not with relevance (at least it seems to me). In some cases, the living space has an important (and statistically significant) effect, in others it does not. The same goes for gender, social class or age.

But I recognize that it is not a trivial matter and constructive criticism is always welcome.

Thank you very much

Think carefully when trying multiple models. To some extent every model tried increases the likelihood of getting a false positive result. If one tries enough models one will always tend to find a statistically significant one. Effectively you are lowering the bar for acceptance when trying multiple models.

Ok & Thanks. I take note. But along with the statistics results, I look at graphs (boxplot for example). I hope to reduce Type 1 & 2 errors. It´s a bit of craft´s, but it´s not useless.

Best