Hi all,
It is a research to test for participants preference on 5 given categorical options. Participants are divided into 3 different stakeholder groups, each participant is given 5 options and rank the options with from 1 - 5 (with 1 being 1st choice)
I have attached the data set below in SPSS format.
I want to test
1) if there is an association between stakeholder groups and the 5 categorical options.
2) For individual stakeholder group, Is there a preference on certain options
3) Is there an aggregrate preference on certain options.
For the 1), I tried to test for the significant difference using crosstab chi-square test, tho I struggle to perform crosstab chi square test as it gives out 5 individual chi-square test result instead of 1. (they test the association between each option and the stakeholder group)
And I am not sure what test will be situable for 2) and 3)
At this stage I prefer using SPSS for analysis, would be grateful for any suggestions, thank you!
Hello W. Chan,
For the first question, create a new variable for each item that you wish to analyze. Here's a crude syntax set that would create one consistent with your example:
compute Q11 = 0 .
if (optionA = 1) Q11 = 1 .
if (optionB = 1) Q11 = 2 .
if (optionC = 1) Q11 = 3 .
if (optionD = 1) Q11 = 4 .
if (optionE = 1) Q11 = 5.
execute.
crosstabs
/tables = Category by Q11
/format = avalue tables
/statistics = chisq
/cells = count
/count round cell.
This will test whether groups rank options alike as their preferred (#1) rating. Given the number of missing rankings in your sample, this is likely preferable to trying to account for all 5 ranking values.
For the second and third question, if everyone had ranked every option, the Friedman two-way anova by ranks would be suitable (for question 2, you'd first select only the group you wanted to analyze: Data / Select cases / Select if condition...; whereas you'd use all cases for question 3). Since there are so many missing values, you'll likely have to try something different. One option would be to assign rank values of k to all unranked alternatives, where k = sum of unused ranks / number of unranked options. Hence, case #1, which only assigned a rank value (1) to option D would have a new profile of 3.5, 3.5, 3.5, 1, 3.5 (for A-E, respectively). Case #2 would have a new profile of 4.5, 3, 2, 1, 4.5 (and so on).
Good luck with your work.
David Morse thank you david for your suggestion. As I set 1 as the largest value, I have transformed the ranked data in reverse order. Inside SPSS I see there is a friedman test but it does not specify whether it is friedman one-way anova or two-way. To perform your suggestions, do I go from nonparametric test > Legacy > K-related sample: friedman ?
As to find out which specific option that the group prefers I may apply a post-hoc test afterwards, is Wilcoxon signed-ranks tests a situable test for that?
Thank you very much.
W. Chan, people may be reluctant to open files that are not plain text (e.g., CSV files). I risked opening your file, and here is what I found:
ID Category OptionA OptionB OptionC OptionD OptionE Questions
1 1 0 0 0 1 0 Q11
2 1 0 3 2 1 0 Q11
3 1 0 0 1 0 0 Q11
4 1 1 2 3 0 0 Q11
5 1 2 4 3 1 5 Q11
6 1 1 0 0 0 0 Q11
7 1 2 4 3 1 5 Q11
8 1 0 3 2 1 0 Q11
9 1 5 3 4 1 2 Q11
10 2 1 3 2 4 5 Q11
11 2 1 3 2 0 0 Q11
12 2 2 1 0 0 0 Q11
13 2 1 2 0 0 0 Q11
14 2 0 1 0 0 0 Q11
15 2 1 3 2 0 0 Q11
16 2 2 1 5 4 3 Q11
17 2 1 3 2 4 5 Q11
18 2 0 1 0 0 0 Q11
19 2 2 3 1 4 5 Q11
20 2 1 2 3 0 0 Q11
21 2 1 0 3 2 0 Q11
22 2 0 1 0 3 2 Q11
23 2 3 1 2 5 4 Q11
24 2 1 2 3 4 5 Q11
25 3 4 3 2 1 5 Q11
26 3 0 3 1 2 0 Q11
27 3 0 0 1 3 2 Q11
28 3 0 3 1 2 0 Q11
The code David Morse posted tells me he was expecting your OptionA to OptionE variables to be indicator variables (i.e., to have values of 1 and 0 for Yes and No respectively). But the listing above shows that the values can range from 0 to 5. What cannot be seen in this listing is that you instructed SPSS to treat 0 as missing. I gather from all of this that if there were no missing data, each respondent would have the numbers 1-5 across options A to E, as respondent 5 does. And in that case, Friedman's test would be one option.
So, I think your question perhaps boils down to, "How can one use Friedman's test when there are missing data?" I just did a Google search on , and it appears that there are methods for dealing with this. Whether any of them have been implemented in SPSS is unclear (and doubtful).
HTH.
EDITED (22-Mar-2020): I meant to say "implemented in SPSS", but left out "in SPSS" the first time!
Hello again, W. Chan,
0. Yes, as Bruce Weaver mentioned, I presumed that a value of "0" was a missing value for your data set.
1. Yes, you can access Friedman test via legacy dialogs in the Nonparametric tests menu option. However, going to the "K related samples" option (instead of legacy dialogs), and opting for customized tests and choosing Friedman in the dialog box will yield a richer analysis, with all pairwise tests run automatically.
2. Yes, if you need to run follow-ups for legacy version of Friedman, Wilcoxon will work.
3. You really didn't need to reverse the scale, but if that makes the results easier for you to interpret, no harm done.
Good luck with your work.
Regarding the relationship between the Friedman test and the Wilcoxon signed ranks test, see this nice blog post by Thom Baguley.
- https://seriousstats.wordpress.com/2012/02/14/friedman/
These days I'd model these sort of data as an multiple ordinal logistic regression predicting the 1-5 rating by option. This would be a bit tricky with only 28 participants and missing data.
A simpler option might be to use multiple imputation and regression (maybe with a transformation for the rating).
- Badges
- Science method