I looks like you do not understand what you are getting with
Cronbach's alpha, since both questions seem to imply that you
have something backwards.
Basically, alpha is average correlation among the items,
scaled increasingly towards 1.0 as the N increases.
So, we expect the alpha to decrease for a shortened scale.
For (2):
If alpha *increases* when we drop one item, that is the sign
that the particular item is "bad" in regards to consistency with
the other items.
And that is because, for (1):
When the items are intended as parallel measures, then
"higher is better", since alpha is an estimate of the reliability
of that measure.
How parallel are the items intended to be? For a psychological
measure, there is ordinarily some "universe of meaning" that
is supposed to be tapped. Guilford, long ago, pointed out that
there is a sense in which we get a trade-off between reliability
and validity: The "most reliable" scale asks exactly one single
thing, over and over; that is not very valid for the wider dimension,
assuming that there is a wider dimension. Similarly, if two items
are perfectly correlated, they might be "too correlated" in the sense
that they distort the operational meaning of the scale from what
is intended.
--
Rich Ulrich
> Date: Tue, 28 Apr 2015 11:03:55 -0700
> From:
[hidden email]> Subject: Dealing with a high Cronbach alpha
> To:
[hidden email]>
> Hello,
> I'm wondering if anyone can provide some guidance here.
>
> I am creating measure with 10 items (variables). When I run internal
> consistency analysis., it produces a cronbach's alpha score of .91 in
> internal consistency analysis. When I exclude a particular item, the alpha
> score decreases to .89. My questions are:
> 1. Is a high alpha score (in this case .91) a problem (and I ask in the
> context of a psychological measure)?
> 2. Is it advisable to drop an item on the basis of 'cronbach's alpha if item
> deleted'?
>
> I welcome any input with thanks.
>
> ...
=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD
Locked
