Professional Documents
Culture Documents
KR20 & Coefficient Alpha: Their Equivalence For Binary Scored Items
KR20 & Coefficient Alpha: Their Equivalence For Binary Scored Items
KR20 & Coefficient Alpha
Their equivalence for binary‐scored items
June, 2007
f
http://www.pbarrett.net/techpapers/kr20.pdf 2 of 7
Internal Consistency Reliability for Dichotomous Items – KR‐20 & Alpha
There is apparent confusion within some individuals concerning the calculation of internal consistency
reliability for dichotomous response, items whose cumulative sum form the scale score for a test. I’ve
heard some individuals tell others that they MUST use a Kuder‐Richardson KR‐20 coefficient instead of a
Cronbach alpha.
This is completely incorrect – in that the KR‐20 is mathematically equivalent to the formula for
coefficient alpha! What seems to be confusing these individuals is that the formula for the variance of a
binary item variable looks quite different to that of a continuous‐valued variable.
The KR‐20 formula is:
k
k pi qi
KR20 1 i 1 2 where:
k 1 test
k number of items in the test
pi the proportion of respondents answering an item i in the keyed direction [1]
qi the proportion of respondents answering an item i in the non-keyed direction [0]
test
2
the test-score variance
The Cronbach alpha formula is:
k
k i2
1 i 1 2 where:
k 1 test
k number of items in the test
i2 the variance of item i
test
2
the test-score variance
Technical Whitepaper #7: KR20 & coefficient alpha June, 2007
f
http://www.pbarrett.net/techpapers/kr20.pdf 3 of 7
The Equivalency
k k
p q
i 1
i i
i 1
i
2
Note – we are using population variance formulae here – hence the use of 2
A Worked Example
6 items and 10 respondents
KR-20 and alpha test spreadsheet
1 2 3 4 5 6 7
item1 item2 item3 item4 item5 item6 Test Score
1 0 0 0 0 0 0 0
2 0 0 0 1 0 0 1
3 1 0 1 0 1 0 3
4 1 1 1 0 0 1 4
5 1 1 0 1 0 1 4
6 0 0 1 0 0 0 1
7 0 0 0 0 1 0 1
8 0 0 0 1 0 0 1
9 0 0 0 0 1 0 1
10 0 1 0 0 0 1 2
k
For every item, pi = 0.3, qi = 0.7 = 0.21, so pq
i 1
i i 10(0.3 0.7) 1.26
The Individual item variances – calculated using the population variance rather than sample variance
formula …
n
(x
j 1
ij x )2 note the division in the equation by n
and not (n‐1) – remember, it is the
i2 where population variance which is computed
n in the KR‐20 and alpha formulae
n number of respondents
xi score in item i for person j
The individual item means are all equal to 0.3, with variances equal to 0.21, and
k
i 1
i
2
6 0.21 1.26
The total test score variance is: 1.76
Technical Whitepaper #7: KR20 & coefficient alpha June, 2007
f
http://www.pbarrett.net/techpapers/kr20.pdf 4 of 7
Thus, using the formula for either the KR‐20 or alpha …
k
k i2
6 1.26
or KR20 1 i 1 2 1 0.340909
k 1 test 6 1 1.76
So far so good … until you use STATISTICA or SPSS … where the item and test score variances are not the
same as computed above, and both report coefficient alpha.
STATISTICA v.8 : Items descriptives
Means and Standard Deviations (KR-20 & Alpha data.sta)
variable Mean Variance Std.Dev.
item1 0.300000 0.233333333 0.483046
item2 0.300000 0.233333333 0.483046
item3 0.300000 0.233333333 0.483046
item4 0.300000 0.233333333 0.483046
item5 0.300000 0.233333333 0.483046
item6 0.300000 0.233333333 0.483046
But, as noted on page 2, using the variance calculated for a binary variable, every items variance was
pi = 0.3, qi = 0.7 = 0.21, with the sum as 1.26 – rather than this new sum of (6*2.333333 = 1.4)
Then, look at the test score variance …
Technical Whitepaper #7: KR20 & coefficient alpha June, 2007
f
http://www.pbarrett.net/techpapers/kr20.pdf 5 of 7
This is 1.955555556 – whereas above we computed it as 1.76.
Putting these values into the formula ‐
k
k i2
6 1.4
or KR20 1 i 1 2 1 0.340909
k 1 test 6 1 1.955555556
So, the same alpha is achieved. Why, because the item variances and test score variances are both
computed using the sample rather than population formulae divisor of n‐1 rather than n. So, the relative
ratio between these two quantities is exactly the same. The sum of item variances using either formula:
k k
pi qi i2 1.26
i 1 i 1
k
n k 10 k 2
pi qi n 1 (0.3 0.7) 9 0.233333 si 1.4
i 1 i 1 i 1
The calculation of the item variances using the usual sum of squared deviations formula is:
n
(x
j 1
ij x )2
i2 where
n
n number of respondents
xi score in item i for person j
n
(x
j 1
ij x) 2
2.1
i
2
0.21 (population variance)
n 10
n
(x
j 1
ij x )2
2.1
si2 0.23333333 (sample variance)
n 1 9
Technical Whitepaper #7: KR20 & coefficient alpha June, 2007
f
http://www.pbarrett.net/techpapers/kr20.pdf 6 of 7
The calculation of the text score variance is :
n
(x
j 1
j x )2
test
2
where
n
n number of respondents
x j test score for person j
n
(x
j 1
j x) 2
17.6
test
2
1.76 (population variance)
n 10
n
(x
j 1
j x )2
17.6
si2 1.95555556 (sample variance)
n 1 9
The ratios between the respective sample or population variances are equal:
k 2 k 2
i si
i n1 i n1 ... 1.26 1.4
... 0.715909
1.76
1.9555556
j s j
2 2
j 1 j 1
In SPSS 15 – the same sample statistics are computed as in STATISTICA :
Reliability Statistics
Cronbach's
Alpha Based
on
Cronbach's Standardized
Alpha Items N of Items
.340909 .340909 6
Technical Whitepaper #7: KR20 & coefficient alpha June, 2007
f
http://www.pbarrett.net/techpapers/kr20.pdf 7 of 7
Item Statistics
Maximum /
Mean Minimum Maximum Range Minimum Variance N of Items
Item Means .300 .300 .300 .000 1.000 .000 6
Item Variances .233 .233 .233 .000 1.000 .000 6
Scale Statistics
Interestingly, Nunnally, J.C. & Bernstein, I.H. (1994) Psychometric Theory 3rd. Edition. McGraw‐Hill. ISBN:
0‐0704‐7849‐X., p. 234 and 235 define alpha and KR‐20 in terms of population variances.
So do Crocker, L. & Algina, J. (1986) Introduction to Classical and Modern Test Theory. Harcourt Brace
Jovanovich Publishers. ISBN: 0‐03‐061634‐4., p. 138 and 139.
Anyway, I hope the above puts to rest the old chestnut that KR‐20 must be calculated for dichotomous
response items and alpha for Likert/continuous valued response items. KR‐20 was always just a
convenient way of simplifying the calculations of reliability for binary‐response items. That’s all.
Likewise that other old chestnut – you must use a point‐biserial correlation when one variable is
dichotomous and the other continuous. Take a look at a nice statement from Denis Roberts (1992)
http://www.personal.psu.edu/users/d/m/dmr/papers/formula.PDF
David Howell in his series of textbooks “Statistical Methods for Psychology” (now 6th Edition, 2006,
Wasdworth Publishing, ISBN: 0495012874) has always stated and demonstrated this. Yet, still I hear
some people talk about a “point‐biserial” correlation. Why? The point‐biserial is a Pearson correlation.
End of story
Technical Whitepaper #7: KR20 & coefficient alpha June, 2007