Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Does there exist measurement error when I got high Cronbach's alpha?

From   "Will Hauser" <>
To   <>, "xueliansharon" <>
Subject   Re: st: Does there exist measurement error when I got high Cronbach's alpha?
Date   Sun, 17 Jul 2011 01:04:10 +0000

Hi Sharon,
Your .6 alpha is correct, the .9 is artificially inflated by your inclusion of missing values. Respondents who didn't answer one question also tended not to answer others. Unless they are coded as missing (.) Stata assumes they are valid responses.

Additionally, the .6 score is only borderline acceptable. This is my opinion here but I would say that the index is only acceptable at .6 if you have strong theoretical reasons to expect unidimensionality and it sounds like you do not. 

Have you used the -item suffix on the alpha command? It would be nice to know what items are driving the index... I suspect some are possibly even detracting from it.

Lastly, remember that Stata will reverse code variables automatically if they fit with the index better that way. So be sure to check that as well.

Have you tried including the variables in your model as is? Do you encounter colinearity?

Will Hauser

------Original Message------
From: xueliansharon <>
To: <>
Date: Saturday, July 16, 2011 5:46:50 AM GMT-0700
Subject: st: Does there exist measurement error when I got high Cronbach's alpha?

Dear all:

I got quite high Cronbach's alphas (0.9) for five-factor personality traits
(extraversion, agreeableness, conscientiousness, emotional stability and
intellect). With such high values, can I argue that there may still exist
measurement errors in the measures for five-factor personality traits?

Another question is about the computation of Cronbach's alpha. I got
different values of alpha when using different computation procedure: the
key difference happened when I recoded the responses of five factors
personality to "missing values" when the original responses were "-3" (i.e.
no questions answered) or "-2" (i.e. information incomplete). For example,
for the extraversion measure, the range of the score should be 5 to 50
points, when I recoded the response "-3" or "-2" to "missing value", the
sample size was reduced by around 680, since the number of observations who
didn't answer the questions about extraversion or didn't provide complete
information for each item were 680, and the alpha coefficient fell from 0.9
to 0.6. So is it correct to do such recoding when computing alpha

Your response is greatly appreciated.

Thanks & Regards,

View this message in context:
Sent from the Statalist mailing list archive at
*   For searches and help try:

*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index