Re: 4-point scale still valid?

["Karsten M. Self" <kmself@ix.netcom.com>]

Wups, second try.  Posted this too soon...

Thinking about this.

Roger Dingledine wrote:
> 
> our list of options is:
> Rank the following in terms of importance (low, moderately low,
> moderately high, high):
> 
> Here are Karsten's original arguments for using a 4-point scale:
 
<snip>
 
> I asked a political science friend her opinion on this, and she wrote:

Poli-sci can make a lot of use of survey techniques.  She may know more
than I.

> ---
> 
> 1. I think it's perfectly fair, and more to the point, useful,
> for a respondent to feel neutral about something. If they

I'm checking on some survey references, and I'm finding that the Likert
scale often is given as a five-response: strong agreement, mild
agreement, mild disagreement, strong disagreement, no preference.

I think this makes sense for what we're looking for.  I'd translate
these as:

 SA:	I would require this feature in an OS I purchase/aquire.
 MA:	I would purchase/aquire the OS offering this feature over other
OSs, all else equal
 MD:	I would not purchase/aquire the OS offering this feature over other
OSs, all else equal.
 SD:	I will not purchase/aquire an OS offering this feature.
 NP:	This feature does not affect my purchase decision at all.


This is a better alternative to one I've though of, which is an
independent weighting of each question's importance, for example:

   Q:  Feature X is important to me
     [response]   SA  MA  MD   SA  NP
     [importance] essential  preferred   unimportant

...the 5 scale response eliminates the need for the second question.

Again, more than 5 responses starts begging the question of how finely
people are aware of their own preferences, and whether or not a '3' on a
scale of 1-7 is the same for different people.  


> 2. Even if you accept his argument, it's not relevant to the scale.
> He objects to having -2, -1, 0, 1, 2 because he doesn't like the

No, that's not it.  Actually, I'm not quite sure how you'd score
results, but it wouldn't be a simple numerical summation of values. 
Quantitative response is inappropriate as these are rank ordinal, and
not quantitative measures.  Using numeric values to codify responses is
a seperate issue, but you could as easily use -2 - +2, 0 - 4, 1 - 5, 102
- 107, or A - E.  I'd prefer an alpha scheme as this prevents anyone
from doing something grossly inaccurate such as summing responses.

The results we're looking for will be expressed in terms of "X% of
respondants require feature A, X% wouldn't buy an OS with feature A". 
We should be able to come up with relative measures of significance (and
division) on specific points.

In terms of eventual specification of SEUL, we'll want to emphasize
strongly favored features, weakly emphasize moderately favored features,
and attempt to avoid introducing unfavored features.

> ---
> 
> Thoughts? Arguments? I'm beginning to think 4 is a bit small too.

How about 5 then?

> --Roger

-- 
Karsten M. Self (kmself@ix.netcom.com)

    What part of "gestalt" don't you understand?
    Welchen Teil von "gestalt" verstehen Sie nicht?

web:       http://www.netcom.com/~kmself
SAS/Linux: http://www.netcom.com/~kmself/SAS/SAS4Linux.html    

 10:51am  up 30 days,  8:20,  6 users,  load average: 1.32, 1.23, 1.09