Which statistical test is used to measure the impact of categorical variable (yes/no- as independent variable) and ordinal variable (i.e likert scale- as dependent variable)?
Dear Madan,
for both you can take the chi-square-test if you perform a test with independent variables and the sign-test, if you test dependent variables (for example comparison of a liked scale pre and post intervention). If there are any more questions don´t hesitate to ask.
Best regards
Tanja
1) The prototypical test for this situation is the Cochran-Armitage test. The data would be arranged in a contingency table. With the original C-A test, the nominal variable could have only two levels, but this has has been expanded in extensions of the test.
2) Using traditional nonparametric tests (Wilcoxon-Mann-Whitney, Kruskal-Wallis) for Likert data is controversial in some circles, though I find that these tests behave well in this situation in most cases. There are some plots comparing results for simulated data from Mann-Whitney and Kruskal-Wallis to those for ordinal regression here, starting at about the middle of the page: http://rcompanion.org/handbook/E_01.html .
3) Ordinal regression is really an ideal tool for dealing with Likert item data.
Dear Madan, dear Salvatore
there are differences in Likert Scales. If you have a Likert Scale with a focus on description like 'excellent', 'good', 'indifferent' and so on, it it better to choose the sign test. Other Likert Scales have numbers and therefore, merric tests can be performed.
In each case, I would recommend to start with very simple tests and after this, Salvatore is right, more complex tests like regression analyses are appropriate.
Best regards
Tanja
Hi, Tanja.
I think I'm going to disagree with you. Likert item data should probably be treated as ordinal data. Attaching numbers to the responses is just superficial. That is, if the responses are "strongly disagree", "disagree", "neutral", etc., renaming them "1", "2", "3", etc. doesn't change anything.
I was curious why both you and Ette Etuk chose to specify the sign test. There is no difference in the assumptions about data between the sign test and traditional nonparametric tests like Mann-Whitney or Kruskal-Wallis. Both handle ordinal data. The latter treats the data with ranks. The former classifies data on either side of the median. Either deal appropriately with ordinal data. But they have different hypotheses they are testing.
A caveat is that the Cochran-Armitage test does assume that the ordinal categories are equally spaced, or the spacing is defined in the test.
Dear All,
Thank you so much for your suggestions. Are these tests available in SPSS? If yes, then how do we use it?
Dear all,
yes, we always perform our calculations in SPSS. You see, Salvatore and I would conduct the analyses in onother way - this might be, because Likert scales are not all the same .
I think tests like the Cochran-Armitage test are very special and are not obligatory to use. Madan, I think you can calculate both calculations and compare the results - probably no large differences will occur.
I did not get good hold on the data yet. If the dependent variable is categorical and binary (yes/no; present/absent; normal/ abnormal) and the independent variables are also categorical (can be more than two categories) we can apply logistic regression, which is there in spss, but make sure that how to run the data and interpret the output.
thanking you
ravi
Ravi, the question states that the independent variable is binary and that the dependent variable is ordinal.
Dear Ravi,
it is very difficult to perform a ordinal regression! (yes/no; present/absent; normal/ abnormal) look like beeing 3 different parameters?
Tables of association with options for Chi Test and Cramer's V results are very easy to execute in SPSS and are standard tests used for testing association between categorical variables. As for causation (regression), you will have to resort to models like multinomial logit etc.
Dear Ravi,
You can go with the test of proportion. The test of two proportions is used to determine if a difference exists between the binomial proportions of two independent groups on a dichotomous dependent variable.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics