Base level in effects coding & status quo

[lib]

Dear Ngene

First of all thanks for a great piece of software for the experimental designs. Your tool is much appreciated.

I have a remark regarding the base level when attributes are effects or dummy coded. From this forum e.g. the thread "Coding Ngene" I can understand that "the final attribute level in the list is the base level" when an atribute is effects or dummy coded. However, from the color example in the manual (page 111-115) it really looks like it is the first attribute level in the list (blue, level = 0) which is the base level - even though it is not explicitly written in the text. So I guess that the aim of this post is to 1) be confirmed in the fact that it really is the final attribute level which is the base level and/or 2) make an amendment to the manual.

Besides I would like to know how the levels are set in the status quo alternative when nothing is specified, e.g. ;alts = alt1, alt2, alt3 - but utility functions are only specified for U(alt1) and U(alt2). Are all levels in alt3 set equal to zero except from the effect or dummy coded attributes where the status quo levels are set equal to the base levels? Is it possible to deviate from the defaults e.g. if the levels in the status quo are not equal to the base levels in the effect coded attributes? (I will have to use this in a later design but have not yet tried it in any of my current designs).

Thanks in advance.

Best regards
Line Bjørnskov Pedersen
University of Southern Denmark

[johnr]

Hi

The internal structure is such that the base level is the last. You are correct, the figures and example suggest otherwise and are missleading. We will correct these.

Your second queston is a common misconception relating to dummy/effects codes. The base level of dummy/effects codes relate to an attribute, not an alternative and hence should really have nothing to do with the status quo alternative. They are in effect, supposed to be self contained within whatever alternative the dummy/effects code is associated with. I have seen many papers where people generate designs so that the dummy/effects base is the status quo alternative. Unfortunately, this is strictly not correct and causes many estimation issues. In estimating the model, changing between dummy and effects coding should return the same results. The ASCs and dummy/effects parameters will change (the latter up to a scale), however the log-like-lihood, probabilities, etc. should remain exactly the same. If however, one sets the base level for the dummy/effects coded variables, the two are no longer the same. You will obtain a different log-likelihood and model results including predictions. I understand that this is done, but it should not be and is a mistake.

That does not mean that you cannot have levels for the status quo (and hencea a non-zero utility for it). It simply means that to be theoretically correct and consistent between the two coding structures, the dummy/effects codes for the non-status alternatives should have a seperate base level.

John

[spqr83]

Hi
I'm trying to fix the status-quo level of two dummy attribute but i have no clear idea about the right Ngene-syntax. The model I'm running contains two categorical attributes, "view" and "recr", that I coded using effect-coding function and you can see below

Code: Select all

design
;alts= P1*, P2*, SQ
;rows=12
;eff=(mnl,d)
;block=2
;model:
U(P1) =  b0
        +b1.effect[0|0|0]*view[-3,-1,1,3] 
        +b2*co[5.5,7,8.5,10]             
        +b3*bio[-50,-25,0,10]
        +b4*land[-10,-5,0,2]
        +b5.effect[0|0|0]*recr[-3,-1,1,3]
        +b6*cost[0,25,50,75,100,125,150,175]/
U(P2) =  b1*view
        +b2*co
        +b3*bio
        +b4*land
        +b5*recr
        +b6*cost/
U(SQ) = +b1.effect[0|0|0]*view_sq[-3,-1,1,3]
        +b2*co_sq[5.5]
        +b3*bio_sq[-50]
        +b4*land_sq[-5]
        +b5.effect[0|0|0]*recr_sq[-3,-1,1,3]
        +b6*cost_sq[0]$

nevertheless, it is not correct because I cannot have four levels in the same attributes in the Status-Quo function. However, If I impute one level in the squared-brackets an error is obviously generates

Error: An attribute has the wrong number of levels for dummy or effects coding. 'view_sq'

. Now I'm in a dead-end truck and I'm wondering what syntax does allow to keep only one level in the status quo function for the two categorical attribute? Does anyone have any suggestions?
Thanks In advance.
Enrico

[Michiel Bliemer]

Enrico, a similar question was asked by JMeyerhoff a few topics down, where I suggested using constraints to restrict the status quo level to be a certain value. I hope that suggestion also works for you.

[spqr83]

Thanks a lot Michiel!