2
$\begingroup$

The research question is: Is there a difference in hand preference when typing vs writing? I have 3 methods: the 'Hand preference' when typing vs when writing as Left, Right, No preference.

I need to find out if there is a significant difference in hand preference when typing vs writing. Also I need a way of comparing the data from different methods because they do not always match.

Should I be using McNemar's test because allsubjects do all 3 methods?

$\endgroup$

closed as off-topic by Michael Chernick, kjetil b halvorsen, Peter Flom Apr 12 '18 at 12:19

  • This question does not appear to be about statistics within the scope defined in the help center.
If this question can be reworded to fit the rules in the help center, please edit the question.

  • 3
    $\begingroup$ Please don't repost closed questions: just edit the original. Reposting risks making the Stack Exchange system block you from asking any additional questions. $\endgroup$ – whuber Apr 3 '18 at 18:54
  • $\begingroup$ Can you say more about your study & your data? This isn't really clear. Can you post your data? $\endgroup$ – gung Apr 3 '18 at 19:22
  • $\begingroup$ It sounds like you have 3 different 3x3 tables. Can you link to a text file of your data? I don't have access to SPSS. $\endgroup$ – gung Apr 3 '18 at 20:40
  • $\begingroup$ @gung McNemar's test tests for differences in binary outcomes in paired data. Cochran's Q test extends McNemar's from paired designs to blocked designs (a la paired t test is extended by repeated measures ANOVA). $\endgroup$ – Alexis Apr 4 '18 at 16:31
  • $\begingroup$ I'm aware of that, @Alexis. I don't believe I've advocated for using McNemar's test on a 3x3x3 table anywhere on this page. $\endgroup$ – gung Apr 4 '18 at 16:53
1
$\begingroup$

EDIT: Based on added information from the OP, the following attempts to provide an answer to the question.

As you've described your data, it sounds like your outcome variable is Preference (Left/Right/None), which could be treated as a nominal variable or an ordinal variable. Independent variables: Method (nominal, 3 levels); Activity (nominal, 2 levels). And Subject.

My first piece of advice is to think carefully about what you are trying to determine. For example, in your question, you say want to compare methods. But if the essential question is to compare Typing to Writing, it would simplify things if you could treat each method as a separate experiment.

So how to proceed?

The simplest approach, I think, would be to treat the Methods as separate experiments. This would entail conducting three McNemar-Bowker tests. However, this method won't allow you to statistically compare the Methods to each other directly.

For a more complex approach, you could put everything into a single model. This would probably be a mixed effects model or either multinomial regression or ordinal regression. This might be beyond what you want to try to tackle, and may or may not be reasonably easy in SPSS. The ordinal regression might be a little easier, and the results might be a little easier to interpret. The advantages of this approach is that you should be able to compare among Methods, and among Activities. The disadvantages are being able to build the correct model, and figuring out to get the software to give you all the results you want.

EDIT: Based on added information from the OP, this answer addresses the "one method" case.

Yes, since each subject indicates preference for each of typing and writing, you should use a version of McNemar's test appropriate for tables larger than 2 x 2. This is sometimes called McNemar-Bowker test, or other names.

People may have difficulty setting up the contingency table for McNemar's test. The important thing is that the data can be arranged in a table with the same categories on the x-axis as on the y-axis. I have included an example in R.

Note that this test can be conducted as exact multinomial exact test. See the link in the code for an explanation on how to do this. This might be a desirable way to go, since with 30 participants, you are likely to have low counts in a 3 x 3 table.

For a post-hoc analysis, you can break the table down in to 2 x 2 tables.

For effect size statistics, you can use Cohen's g or odds ratio.

### Adapted from http://rcompanion.org/handbook/H_05.html

Input =("
Writing         Left.typing   None.typing   Right.typing
Left.writing      2           0              1
None.writing      2           6              7
Right.writing     3           4              5
")

Matrix = as.matrix(read.table(textConnection(Input),
                     header=TRUE,
                     row.names=1))

Matrix

sum(Matrix)

   ### [1] 30

mcnemar.test(Matrix)

   ### McNemar's Chi-squared test
   ###
   ### data:  Matrix
   ### McNemar's chi-squared = 3.8182, df = 3, p-value = 0.2818
$\endgroup$
  • $\begingroup$ This is a good answer (+1 from earlier, & a good book!), but I don't think it is actually the answer for the OP's situation. The OP's situation was ambiguous. It is somewhat clarified now. It seems they have a 3x3x3 array. $\endgroup$ – gung Apr 3 '18 at 20:43
  • $\begingroup$ @helpame, the standard McNemar test is for 2 x 2 tables. For larger tables the test is sometimes called McNemar-Bowker, or just Bowker. Bhapkar and Stuart-Maxwell can be used in similar circumstances, although the hypotheses may be somewhat different Link $\endgroup$ – Sal Mangiafico Apr 10 '18 at 12:55

Not the answer you're looking for? Browse other questions tagged or ask your own question.