The questionnaire was trying to explore the customer priorities leading to their purchase of a product (More like, if your are intending to purchase this, what shall be your priorities in a ranked manner) . There were 7 variables like product quality, price etc. to be ranked by rank 1 to rank 7. If a rank has been allotted to a variable, the same rank doesn't go to any other variable.
What are the preliminary set of tests (supported by SPSS) to be conducted by default to ensure the reliability and validity of this ranked data. I have tried to use Multidimensional technique to visualize this ranked dataset, but I wanted to know what are some other standard approaches in analyzing :-
1. Are these 7 variables different and unique?
2. How could we know which variable has great influence in the mindset of majority respondents of the survey?
3. Could we use any regression or analytics to correlate these rank values by appropriate weight to represent any algorithm which could predict the likelihood of buying that product.
4. If you have any thoughts on good approaches, something worth testing, please mention the test and possible outcomes or inferences possible.
Hello Vishnu,
Your seven features will exhibit some predictable characteristics based on the way you forced respondents to assign non-repeatable rank values from 1 to 7.
First, the correlations among the feature rankings will be negative. This is typical with any kind of forced choice/ipsative type measurement or scaling. It probably would have been better to allow independent ratings for each feature.
So, to your questions:
1. Different? Yes. Unique? No, because of the negative intercorrelations.
2. Great influence in the mindset? I think the best you can do is order the features based on the median rating assigned by respondents.
3. If you had either a measure of buying intention or actual purchase history, there's no reason that you couldn't see which features do or don't contribute to explaining differences in the outcome of interest. If done individually, there's no restriction on the inferences. If done as a set (all seven features, with interdependent rank values), I suppose you could treat the data in a multilevel fashion, as ranks are nested within persons.
4. I think the more important question is, what are the specific research aims you had in collecting these data in the first place? That should be the principal guide towards identifying suitable analyses.
Good luck with your work.
It is very likely that some parameters are related to each other.
I would advise a principal component analysis to evaluate whether parameters can be grouped (i.e., that they show similar behavior). Each principal component with an eigenvalue > 1 is a relevant subgroup of parameters, with the largest eigenvalue being the most relevant in your dataset. That answers both Q1 and Q2. Principal component analysis cannot be done in SPSS to my knowledge, but is easy to perform using R.
For the regression analysis it is favorable to have less parameters (principal components instead of the separate parameters) to reduce over fitting of the model.
Hope this helps.
David Morse As of now, i had just used the weighted approach where we multiply highest weight to the frequency of the best rank, then next lower weight to next rank and summing up to see how each variables stand. But when i tried to chi square goodness of fit, certain ranking distribution within some variables were having p value above 0.05 ( which means rank 1 has almost same proportion as rank 2 same as rank 3 - not perfectly same, but near ) would that impact my analysis. So should i rather try to do PCA to reduce the dimensionality and make inferences?
Hello again Vishnu,
PCA/factor analysis would accomplish a very different goal than would sorting or comparing the seven feature rankings via Friedman's or some other comparison.
What the Friedman test will do (I'm referring to your other query: https://www.researchgate.net/post/How_can_I_analyze_data_from_Qualtrics_ranked_questions_in_SPSS#view=61472bc24d6ad22b30123906) is to compare the rankings given to the seven features for equality. A significant result implies that at least one significant difference exists among all possible comparisons of the seven features (not just pairwise).
What a lot of folks will do in the presence of such a result is run pairwise comparisons (via Wilcoxon test). Of course, with a lot of tests, the likelihood of type I error increases, so you'd likely want to control for this in some way. R base system features some nice options for adjusting the (unadjusted) p-values via some preferred methods like Benjamini-Hochberg or Hommel's (https://stat.ethz.ch/R-manual/R-devel/library/stats/html/p.adjust.html).
Good luck with your work.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression