I agree with Mrs Dorea. Furthermore, it’s a good thing to choose a region with a large number of observation to reduce standard error and decrease confidence interval width of other region coefficients.

Maybe some more senior researchers will give you better insight. But as far as I know, there is no obvious choice, you should just choose it in a way that makes interpretation more intuitive. For instance, if you chose the region with the lowest possible value for what is your outcome, then you know that the model figures represent "how greater than that". it is also easier to interpret the significance of the differences, if all values are compared to the lowest (or the greatest) value. If for instance if you have 3 regions, A, B and C (with increasing means), and you are comparing, say, milk yield. If A is not significantly different from B, and B not significantly different from C, but A and C are significantly different from each other, it may be easier to interpret when the reference is either A or C. Of course in the end, the significance of "region" as a variable will not depend on the p-values printed for each region, but on the pooled p-value. So actual significance will NOT depend on which you use as reference, but interpretation may be much easier when you chose one of the extremes.

So' agora reparei que e's do Rio... mas enfim devia escrever em ingles por causa do forum. Boa sorte na sua pesquisa!

As above. I typically choose a referent class at one end of the continuum or the other so you can easily interpret the coefficients. I might not do that if I were regressing treatment group membership and had a control group with multiple other groups. In that case I would use the treatment group as the referent class.

I agree with Mrs Dorea. Furthermore, it’s a good thing to choose a region with a large number of observation to reduce standard error and decrease confidence interval width of other region coefficients.

In my opinion, the best choice is first to think if there is one category to which you would like to compare all the other categories (e.g., an experimental condition). If not then you might indeed choose a reference category on one of the ends, though with categorical (not ordinal) data might be complicated. You might, however, also choose to use the one in the middle (for example, income), so you can see whether low incomes differ significantly from middle high incomes, and whether high incomes differ from middle high incomes. If you can confirm that both sides of the continuum differ from the one in the middle, you can also conclude that these differ from each other.

I agree with the previous comments; selection must be based on theoretical/conceptual grounds. I must assume "no obvious choice" means no guidance from the literature? You must have a reason for including this variable in the first place. What contrasts are you interested in? What makes sense in the context of your problem? The comment above on sample size is very important. Too few cases in a category may result in unstable estimates.

Although no precise and concrete reasons to select the reference, it is usually better to start thinking which region is one with normal characteristic and then comparing with the rest will be easily interpretable. This might include like area, population, educational/literacy, market size, economic size near to the national average, highest or lowest. But this is all dependent on your thinking and notion, not any stylistic fact. So may be other views can very well help to select a level which you can rely.

It all has to do which type of contrast you choose. Choosing one group as the reference is called treatment or simple contrast. Alternatives are Anova contrasts, which sum up to 0, or Helmert contrasts, where each factor level is compared to the mean of the subsequent levels. many others exist, and the choice mainly depends on your research question. If there is no natural choice for a reference group (e.g. a control group) then you may just want to set different contrasts, e.g. Anova type.

IMHO, a categorical variable, whatever the number of levels it contains, cannot be used as any other numerical variable for the simple reason that it is not numerical: it simply cannot be input as an independent variable. The information it conveys, however, can be taken into account by running one multivariate analysis for each one its levels.

Another analysis you can run, however, especially if you are interested in non-linear relationships, is using a neural network system, something completely different from multivariate analysis.

I faced this problem some years ago and tried to find the solution, just to discover what we can see here: there is no difinitive solution!

I would like to thank all answers! All contributions reinforce, to me, the need to look carefully to all steps in the analysis. If you choose a reference level of the variable without think about it, you can be missing an important information from your data.

Thank you all!

There are a number of options available to you and as the analysts it is your decision to be made, but often in consultation with your client. Choose either the middle (or median possibly) category of your explanatory variable, so you highlight differences from the median reference category, Otherwise choose the highest or lowest category, depending on whether you prefer interpreting ORs above one or below one. I have found most people find it easier to interpret ORs above one.

I agree with previous comment.

I will start from choosing according

1. the meaning (the one most representative)

2. then the number of observations (the larger the better)

3. then value/ relationship with the outcome (e.g. the highest or the lowest)

4. then natural order of the category (e.g. highest or lowest for ordinal category)

5. finally, I will also evaluate to see which one can give better result (easier to understand or significance of the result)

Normally we have Strongly Disagree, Disagree, Neutral, Agree and Strongly agree and we coded them 1 to 5, we calculate their means. if it is 3 then people are neutral on certain question and if it is 1 then people vote for strongly disagree along with some standard deviation

You could:

1- choose the category with the largest number of subjects as the reference category

2- choose the category with the lowest prevalence of the outcome as the reference category

@Zahid Iqbal. In such sitation, usually we can use ORs where we compare the difference between the initial (1or 5) which way you begin/refer. So as to compare the relative difference between the worst to best or from Best to worst. Usually the reference is the initial category but when there is no such categorization as indicated by the regional dummies, then very careful decision is required.

Select one by one category arbitrarily as reference and study the results. The category that gives most contrasting results in terms of the outcome/ dependent variable should be the one of interest.

If you have a variable that you consider explanatory "a priori", then its explantory value (or interest) would not be established on statistical grounds, but because of your theoretical model. So, it should be your model o theory which provides a clue as regards the choice of the level to be used as reference.

If there is no meaningful choice for a reference, and especially if you are not interested in comparing each level to the reference only, I would choose an ANOVA framework to test the overall effect of the categorical variable. Pairwise posthoc tests can then be used for comparisons within each possible pair of categories. Some help for one-way and two-way ANOVAs can be found there:

http://cran.r-project.org/doc/contrib/Faraway-PRA.pdf

If you have at least 2 categorical variables in an unbalanced design and want to study their interaction, you should be aware what sum of squares you are using :

http://goanna.cs.rmit.edu.au/~fscholer/anova.php

For logistic regression can use dummies and switch the reference to find significance.

When a regression model has a categorical variable it is necessary to take into account the effect of its categories. If you use only one variable with 5 categories then the regression coefficient for this variable will indicate the effect per each unit of increase in the categorical variable. So, if the regression coefficient is 0.5, the change from v=1 to v=2 will be 0.5, equal to the change from v=4 to v=5. What happens is that many times this is not the case, and it could be a lower increase from v=1 to v=2 and a high increase from v=4 to v=5. In this situation you need to use dummy variables. These variables indicate the effect on variable y of being of each one of the categories. The number of dummy variables for each qualitative variable is one less than the number of categories included in each, with the reference/base group as the omitted category.

I suggest running a step-wise i.e. forward or backward step-wise regression, is seems to test each variable one by one and gives an output that highlights the relative importance of the variables. In other words there is no arbitrary reference point it simply has to be tested statistically.