Look at the correlation matrix of the items, or since you have three binaries, the 2^3 table. You must have something pulling in the wrong direction, and markedly so.
BTW KR-20 is the same as alpha for binary items.
-alpha- has several options that are helpful for diagnoaing what's going on but with your scale it should be quite obvious from the correlations. If you use the GUI it lays them all out.
If you're willing to share your raw data I would be willing to take a look. It looks like a good example for my class. I teach psychometrics and having a few examples where things go wrong is useful.
-----Original Message-----
From: "Chao Yawo" <[email protected]>
To: "[email protected]" <[email protected]>
Sent: 11/16/2008 11:19 AM
Subject: st: Getting Negative Alphas
Hi,
I am trying to create a summed scale of 3 items which are coded as follows:
Respondents were asked whether they
1. Have had an STD in the last 12 months: coded No (0) and 1 (Yes)
2. Used condom during last intercourse: coded No (0) and 1 (Yes)
3 How many non-marital partners they've had sexual relationship with:
range from 0 to 20
I want the summed scale to reflect the degree of risky sexual
behavior, so i followed these steps in recoding the variables:
1. STD experience: the original coding is ok, since the higher
numerical code (1 - that an individual has had an STD) is consistent
with the meaning intended.
2. Condom use: since the question wording is inconsistent with the
coding - ie, a code of 1 implies that the individual did use condom
and hence not engaged in risky behavior, I reverse coded the variable
so that the original 0s are 1, and the original 1s now take 0.
3. Non-marital sexual experience: since the original coding indicated
the actual number of non-marital partners, I merely recoded the
variable to reflect those who did not have any extra-marital
relationships (a code of 0), and those who had any (a code of 1). Here
again, having a non-marital experience is reflective of a risky sexual
behavior.
Now, since the individual items are dichotomous, I had to use the
Kuder-Richardson coefficient of reliability (see:
http://ideas.repec.org/c/boc/bocode/s351001.html), instead of the
regular Cronbach's alphas which is intended for multi-point items.
Stata keeps returning a negative alpha of -0.72 for the 3 items. I
am wondering why this is the case since I believe I have correctly
recoded all the variables to ensure that higher numerical codes give
consistent meanings. I am therefore stumped wondering what I might be
doing wrong.
I would appreciate any pointers to correct this anomaly.
Thanks very much,
CY
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
