Chi-square for trend in SPSS?

[Roberto Oliveri]

Hi (again),

It seems that SPSS doesn't operate with common chi-square test for trend -
probably it does, though I haven't figured out yet just how.
I am having 2 x 4 table (the latter being ordered). I tried to calculate
chi-square for trend manually, and it was pretty near the "linear-by-linear
association"-value (in crosstabs output) both of them having 1 d.f. giving
pretty much the same p-value. Are those two entities identical, i.e. is
linear-by-linear association SPSS terminology for chi-square for trend?

Thanks in advance
Roberto Oliveri

[Rich Ulrich]

On Thu, 18 Jul 2002 23:00:32 +0200, "Roberto Oliveri"
<rob...@oliveri.dk> wrote:

> Hi (again),
>
> It seems that SPSS doesn't operate with common chi-square test for trend -
> probably it does, though I haven't figured out yet just how.

I can think of 4 simple tests I might expect for trend, right off
the bat, and SPSS offers three of them. Use either nonparametric
correlation, or use the Mantel test (in Crosstabs).
- Mantel scores the categories by consecutive integers.
- Spearman scores the categories by average-rank.
- Kendall effectively scores the categories by average-rank
(I believe it works out that way) and penalizes deviations by
a linear (swap) count, instead of by squared differences.

> I am having 2 x 4 table (the latter being ordered). I tried to calculate
> chi-square for trend manually, and it was pretty near the "linear-by-linear
> association"-value (in crosstabs output) both of them having 1 d.f. giving
> pretty much the same p-value. Are those two entities identical, i.e. is
> linear-by-linear association SPSS terminology for chi-square for trend?

The Mantel is 'linear by linear' for any number of categories, with
a known and portable basis of scoring -- I like it because of that,
and because I distrust rankings with small numbers of categories.

With k=2, any of the above will test 'linear' trend for two groups.
There is no magic formula for what comprises 'linear' once you
have abandoned knowledge underlying the ordinal scoring.

See Agresti: An introduction to categorical data analysis.
In Agresti's example of fetal abnormalities, I prefer using a t-test
on drink-scores or category-number, *after* dropping the
zero-drinks group <because zero is a suspicious count, too
often a heterogeneous group: an issue he does not touch>.

Hope this helps.

--
Rich Ulrich, wpi...@pitt.edu
http://www.pitt.edu/~wpilib/index.html