Usually I use 1, to make simpler and clearer identification of categories, but sometimes I prefer 0 as baseline.

I do this way when the first group, that is the 0 category, is the null hypothesis (e.g. "normals", "no symptom" etc.) because it makes easier to perform a logistic analysis, for which the values must be dicothomous.

In this case I prepare two datasets, one with the full scale (e.g. 0 = no symptoms, 1 = low, 2 = mild, etc.) and another with 0 = normals and 1 = any other result.

Doing this way, it's easier to perform, say, a Kendall's rank correlation for the full scale and a logistic for the dicothomous one. The second shows you the result regardless of the intensity of the answer, and this can help to better understand your data.

I hope this consideration could help.

Thanks for the question! :-)

Generally I use the number 1 for better indentification.  People just do not usually pick the O on scales.

For binary/ dichotomous (categorical) classification, the usual practice (for coding) is to go for 0 & 1. But absolutely no problem, if it is started with 1 as base (1 & 2). But one thing we should be very clear in voiding the temptation of finding (and trying to interpret) the average of such codes. There is no need. Particularly for nominal/ categorical variables.

I prefer 1 because it is easier to follow. It is just like I will prefer 1 over 1,000 or 2004 or  1999997. We can use any numeral to symbolize a "quality". But i guess the easiest to follow would be 1, than 0 or 34567.

Thanks, all! Pierpaolo, I liked your explanation. Here's what I'm thinking of as a standard practice that should keep things clear. Right now I'm working with a data set that has a 6-category race variable that runs (0-5), and 5-category year variable that runs 1-5. In some Stata runs, the labels don't come through and it's always hard to remember which var starts with zero v. one particularly when doing interactions. 

1) For data cleaning, codebook, etc. make categorical vars (whether conceptually binary or mult-cat, ordinal or multinomial) start with 1. Variable name should include number of cats (e.g., var1_5cat). The rationale is that for data documentation and codebooks, things should be human readable more than ideal for analyses (Note: We didn't do this for the 5-cat year variable, which adds to the confusion.)

2) Only recode to 0 as base if needed for an analysis (e.g., dummy coding by hand). For truly binary variables, use the suffix "_bin" to replace "_2cat" if it's not possible to rename the variable to the value of 1. (e.g., sex_2cat could become sex_bin, but naming it female [or better yet, sex_female] would be better, and adding _bin doesn't seem to add anything in that case). 

3) For a multi-category variable this might look like this: 

year_5cat (ranges 1-5, representing 2003-2007)

Creating binary dummies leads to varnames

year_2003

year_2004

year_2005

year_2006

year_2006

Keeping year as the suffix (or core) reminds us what original variable it's linked to 

Just to add

It is also important how these are treated/ coded in the model and I would distinguish between your  6-category race variable and the 5-category year variable. The former is nominal and the latter is ordinal. Consequently and motivated by parsimony, I would think of coding the year variable as an orthogonal polynomial so that it would be reducible down to say a linear or quadratic trend in the model - S-Plus automatically treats such ordinal variables in this way. For some discussion on this see

https://www.researchgate.net/publication/234015221_Manual_supplement_for_MLwiN_Version_2.14?ev=prf_pub

This link has a consideration of more advanced forms of coding which can be used in Stata

http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter5/statareg5.htm

while this link is the equivalent in R

http://www.ats.ucla.edu/stat/r/library/contrast_coding.htm

Data Manual supplement for MLwiN Version 2.14

Matt,

(1) interpretation of the coefficient of a variable coded 1 and 2 does not change if you code it 0 and 1. If your regression equation is y = 4 + 3x + {error} when x is coded 1 and 2, then it simply becomes y = 5 + 3x' + {error} when x' is the copy of x coded 0 and 1. Acknowledged limitations: (a) The interpretation of interactionswill change; (b) The value and the interpretation of the constant/intercept term will change; (c) In multilevel modeling world where people are obsessed with the right centering, the question of coding 1/2 or 0/1 could matter two.

Having said that the coding does not really matter to me in statistical terms (and if you think otherwise, you should revisit your linear models class), the data management and cleaning tasks take the higher priority. Hence, other rules for me are:

(2) Being brought up using Stata, not Some Alternative Software, not the Software that Prohibits Sophisticated Sureys (and having learned C/C++ ways earlier in high school and college), this is a non-question: 0 is a logical "No", 1 is a logical "Yes", and Stata parses -if female do this- as -if female!=0 do this-, which provides concise and readable syntax. When coded 1 and 2, you have to be more explicit as in -if female==1-, which is fine, but Stata convention is 0/1 coding.

(3) For nominal or ordinal variables that have more than two categories, I use 1, 2, 3, 4, etc., with an understanding that I will never use this variable as is in the RHS of a regression equation (which would force the distance between categories 1 and 2 to be interpreted as the same as the distance between 2 and 3, which is rarely meaningful), but rather specify it as categorical (CLASS in SAS, i.agegroup in Stata, as.factor in R).

(4) With interactions of categorical variables, I prefer semi-interpretable coding: agegroup5_female

My suggestion: https://www.researchgate.net/publication/278403191_Numerical_Coding_of_Nominal_Data

Article Numerical Coding of Nominal Data

Thanks, Zenon and others! I don't get out here often, but I appreciate the thoughts and detailed responses, and references :) Some of them may get into the "research project management for social scientists" book that I may or may not ever write :) 

I think Stas gets the award for most thorough response. Nice work, SK! Kelvyn a close second. 

My habit has become to use 0,1 for no/yes, and start at 1 for any other categorical (nominal or ordinal) variable. I've tried starting at 0 for those (e.g., 0, 1, 2 for a 3-category variable), but find myself continually confused because I'm counting on that highest value to be the # of categories. 

I think this thread should also get an award for the most indepth discussion on a VERY unsexy topic :) 

We always start with 1 but as above there is not great reason for this!