Hi, Grant,

Personally, I think that the way a person responds to an item in a questionnaire is not only reflective of that person's opinion regarding that one item, but also of his or her personality. Therefore, by correcting for response tendencies, we can loose an important part of variability in the data.

I agree with Mykolas -- people have "response tendencies" because that is how they interpreted the questions you asked and the options you gave for answering them.

That said, I would examine the standard in your field -- if there is a broad belief that it is necessary to correct for response tendencies, then that is what will be expected of you. Alternatively, if there is inconsistent practice among other authors, then I would just ignore it.

I would tend to disagree with Mykolas and David but, I think that the key is exactly what do you want to characterize. ¿Do you want to understand the individual? I would go with Schwartz proposed correction. His point is well taken: to understand personal values is to understand priorities. No doubt about it in my mind.

However do you want to characterize groups? well consider other options but

What about comparing  the Schwartz correction versus the no correction at all: and then check what your data are telling you considering the focus of your analysis.

Good luck and let us know what have you concluded. Regards

Thanks Mykolas, David and Octavio for your suggestions and input.

At the minute the plan is to try and group certain types of individuals based upon these values (using cluster analysis) and then look at the differences between these groups in relation to a number of variables that influence pro-environmental behaviour (Moral norms, community concern, perceived behavioural control and more). I am just not sure whether I should use the mean scores from the questionnaire for each of the values (after centring) for the cluster analysis, or implement the correction suggested by Schwartz to get an individuals value priorities and then (after centring) use this for the analysis.

It appears that if you do not correct the scores you get a cluster that are 'HIGH' on all values and a cluster that are 'LOW' on all values (and two other clusters). Obviously, when you implement the correction for value priorities it becomes impossible for someone to be low or high on ALL values as it is relative to themselves. I can see the argument for both ways to be honest but I am thinking that if in the future these groups were to be used to tailor interventions to increase pro-environmental behaviour it makes sense that the groups are based on what the individuals prioritise (i.e. with a correction) rather than if they happen to be 'low' or 'high' compared to the sample (i.e. without a correction)...

Again, any thoughts/opinions are welcome, they are really appreciated.

I'm a bit confused about the actual procedures you want too use. It sounds like the approach is to calculate each person's mean score for each of the summary measures (Egoistic, etc.), and then do a within person subtraction between that mean and the individual items. But if I answer the items with uniformly high scores, then subtracting the mean will just convert my scores to uniformly low.

Or do you plan to calculate within-person Z-scores? I'd worry about that too, on the grounds that those with low variance could contribute the highest Z-scores in the final distribution. More basically, you would be equating at 1-point difference with a low variance to a difference of several points with high variance. Is that what you want?

Also, if you are looking at cluster memberships, that complicates the "increment to explained variance" model suggested by Octavio, I guess you could produce the two different clusterings and compare the explained variance for each.

Hi Grant: I tend to agree with your latest comment on this issue. I would say that the correction proposed by Schwartz is the simplest with no further assumptions plus the advantage of easiness of interpretation.

Thanks again for the comments. 

David - I can see your point about the uniformly high scores simple converting to uniformly low scores. That is certainly something I will consider. I think the correction is useful when trying to get a sense of the person, for example: if we take a simple example say 2 values (mean score of egoistic, biospheric) and 2 people (A, B) . If person A scores Ego = 2, and Bio = 2, whereas person B scores Ego = 2 and Bio = 6. If we just compare the mean Ego score then you could conclude these individuals are similar (Scoring 2 each), however if you use the correction Person A would score 0.5 (EgoScore(2)/BioScore(2) + EgoScore(2)) whereas Person B would score 0.25 (EgoScore(2)/BioScore (6) + EgoScore (2)). Either way, these scores would then need centring. 

I think as this is quite exploratory, and individuals have differing opinions both on here and in the literature, it would make sense and be acceptable to cluster both ways and look at how each way predicts the different variables that influence pro-environmental behaviour and see if one makes more sense. I could then use this study as an exploratory study to inform future work.

As I understand it, the example you gave doesn't really match the clustering that you reported earlier, where= there were two clusters that corresponded to uniformly high and uniformly low scores. I assumed that meant those people responded with high or low scores across all four of the dimensions. In other words, I assumed that people were being clustered according to their scores on 4 summary variables, where each set of summary variables might or might not be corrected for response pattern.

Is that what you have in mind?

One thing you might consider is something a bit more unconventional but theoretically reasonable in my opinion - that is to treat the values as being represented as vectors arranged round a circle structure, and to take the correlation as being fixed by the theoretical structure as opposed to the data. This is certainly justifiable on theoretical grounds.  The degree of correction would then depend o  the angle between points on the circle structure, so values at 90 degrees would be assumed as being independent, etc...

I would say that Tony might be right (might) only if the values structure has been shown to have circumplex properties such as Schwartz theory of values. If and only if.

Hi David,

Yes - this is what I have in mind. I guess the point of interest is does it matter that the participants score uniformly high on all values or uniformly low on all values or simply that they rate the values equally. For example, if we believe it matters whether they score high or low on all values then we shouldn't use the correction and we should have two clusters (All High and All Low). However, if it doesn't matter whether they rate them high or low, but instead all that matters is that they rate the values equally then we could use the correction and then we would possibly get one cluster of these individuals. I think by running the analysis both ways it will be possible to explore this - for example if there is a significant difference between ALL HIGH and ALL LOW in terms of Moral Norms (I feel guilty if I do not recycle) then it is possible to conclude that it is not as simple as  values 'cancelling out' each other. A person with relatively low biospheric values and low egoistic values does not have the same moral norms a person with relatively high biospheric values and high egoistic values. However, if this is not the case and there is no significant difference between ALL HIGH and ALL LOW in terms of moral norms it may be the case that it doesn't matter what level (in terms of relatively high or low) of biospheric and egoistic values you hold, but simply that if you hold the same level of bio and ego values they will cancel each other out. 

Again, thanks of all your comments and Tony and Octavio for theirs, I find it really interesting and beneficial having these debates.

It seems to me, the first step would be to do the clustering without any correction to see if you can replicate the earlier results. If you do get an all-high and an-all low cluster, then you could test for differences between them.

That would also provide a basis for a comparison to the clustering that would result after doing a within-respondent standardization. FYI since clustering can be strongly affected by relatively small changes in the data structure, my guess is that the two sets of clusterings would look very different.