Page 1 of 1

Dummy coding of categorical variables in MNL_SP example.

Posted: 06 Jul 2023, 16:21
by Bob123
Dear Prof. Hess,

This might be a simple question, but might also be useful also to others starting out in Apollo.

In the Apollo MNL_SP model example, there is a categorical attribute for service (service_rail), which has 4 levels: 1 for no-frills, 2 for wifi, 3 for food, 0 if not used. Also in this example, "b_no_frills" is kept at its starting value in Apollo_fixed. In the utility functions, these attributes are listed as: ( service_rail == 1 ), ( service_rail == 2 ) and ( service_rail == 2 ).

I have a couple of questions relating to this question and dummy coding of categorical variables in Apollo.

Is my understanding of coding categorical variables (with more than two categories) in Apollo correct:

1) The ability to write out utility functions in this way in Apollo for categorical attributes (e.g., service_rail == 2), prevents the need for creating multiple dummy variables in the data (css file). It is my understanding that in most other analyses (outside of Apollo), we would usually create k-1 dummy (binary) variables (where k is the number of levels, here 4) and select one as the reference category by listing it as 0,0,0,0 to which these would be compared/relative to? In the MNL_SP csv file for example, there are no such dummy variables which makes me think that this is the case. This was the only case I could find in the Apollo examples where a categorical variable with more than 2 (i.e., binary) levels was used in any of the csv files so wanted to be clear my understanding was correct. I think is due to ease of interpretation of binary categorical variables versus those with 2+ levels. As a follow-on the interpretation then

2) Also, listing b_no_frills in Apollo_fixed performs the same function (effectively) as selecting a reference category by using a 0,0,0,0 dummy if coded that way. Including this level (service_rail ==1) in the utility functions is therefore optional as no beta will be calculated, but having a 0.000 in the output simplifies interpretation?

Thanks in advance,

Robin

Re: Dummy coding of categorical variables in MNL_SP example.

Posted: 07 Jul 2023, 11:26
by stephanehess
Hi

yes, there is no need to create separate attributes in what some people refer to as recoding of the attributes. You can just do it directly in the utility functions.

Btw, this attribute has 3 levels, not 4. The "not used" level is really a separate thing, which is that the attribute doesn't exist at all in the RP data, but in effect, this is the same as no-frills then

Stephane