Hope you are doing well; I have to ask one question:
Generally, in survey instrument, we use Likert scale 1-5, 1-3 or may be some times 1-7. For interpretation of the response of the subjects we use mean score of specific statement to be asked from the subjects. I am doing so in my research studies. Since mean is the most fluctuated value as compared to median and mode that's why many researchers in Pakistan use percentage response for interpretation of the response marked by the subjects on Likert scale.
I need your comments please. Thanks
Hello Ejaz, Descriptively, actual response values (n of responses to each option, or per cent of responses to each option, if overall N is known) gives complete data for anyone who'd be interested in any other summary statistic or even in computing cumulative percentage/percentile points associated with the scale. For me, a median value makes more sense than a mean value for single Likert-type response scale results. This is not to say that you won't find many folks using means or some function of the mean (e.g., "relative importance index") with individual LIkert-type options, but doing so doesn't make a lot of sense. Good luck with your work.
Ejaz Ashraf
The measurement of attitudes has principles that the more these principles are observed, the more accurate the measure will be; these principles can be briefly divided into five parts:
Consistency or having one dimension, linear mode or equal intervals, reliability, validity, and stepwise or cumulative replication.
Various researchers have created different scales in which special attention has been paid to one or at most two of the five mentioned conditions, and less attention has been paid to other conditions. One method focuses on matching, another on finding equal units, a third on stepwise multiplication, and so on.
It does not seem to be possible to find a single method that has all these points. Therefore, it is crucial to know the purpose of each of these scales and the difference between them. However, it is not measured. It is possible to say which method is better. Each has important and necessary aspects, but at the same time, there is criticism of each.
The best method is always the method that is more suitable than others for the subject under investigation.
If you want to study the formation pattern or discover attitudinal theories and care about uniformity and one-dimensionality, the Likert method is probably more suitable.
Guttman's method is better if you are interested in studying the changes in attitude or the sequence of building an attitude and you care about step-by-step reproduction and being one-dimensional.
If the goal is to study group differences and pay attention to equal distance, you should probably choose Thurstone's method.
If the goal is to study social distances and stage reproduction is essential, use the Bogardus spectrum.
and based on David Morse 's comment
Each of these scales does a relatively good job, so a suitable scale should be chosen according to the research needs.
Ejaz Ashraf
I agree with David Moris that Mean and SD are not making any sense here. But I am wondering if Mean and SD are not useful how could the median be meaningful? I think multinomial probabilities should be presented (i.e. percentages).
Dear Dr. Morse would you please add your response to Dr. Sami Ullah's question. Thanks
Hello to all on this thread,
Responding to Sami Ullah's query about the utility of the median score:
Of all three of the usual measures of central tendency (mean, median, mode), the median is the only one about which one can make statements of probability without knowing or having to assume anything about the shape of the distribution (e.g., half the cases are at/above median; half are at/below). That can be a useful tidbit of information. As well, interquartile spread (or Tukey's hinge spread) can be a useful measure of dispersion.
If the variable in question is discrete, with a modest number of score values, I agree that the percent of cases at or at and below a given score can be a useful summary statistic as well.
Thanks Dr. Morse for your transparent response.I am sure, this has made us clear the situations to use the three measures of central tendency.
Dr. Ashraf and Dr. Morse, which test would you run if you wanted to break down the Likert overall score by race or gender? I am using a Likert Scale, 7 questions which has a continuous score and would like to show the differences between the races, and genders, and possibly race and gender.
Hello Jason,
For single variable comparisons, like gender, the usual options would be:
(a) Independent t-test (or one-way anova, with two or more levels of the IV);
(b) Mann-Whitney-Wilcoxon test (two levels only) or Kruskal-Wallis one-way anova by ranks (two or more levels);
(c) An exact or permutation test; or bootstrap/resampling test; or
(d) Equality of distribution tests, such as Kolmogorov-Smirnov test, Ridit analysis, or Norman Cliff's dominance statistic.
The choice of which one hinges on the specific hypothesis you wish to evaluate and your willingness to make assumptions about the nature of the scores and, for choices like (a) distributional assumptions.
For multiple IV models (e.g., gender, ethnicity, and potentially the interaction term), then the usual options would be:
(a) factorial anova;
(b) the equivalent (ordinary least squares) regression model;
(c) Exact/permutation tests, or bootstrap/resampling tests; or
(d) ordinal regression model, or robust regression method.
Again, the choice hinges on hypothesis & assumptions about data/distribution.
Finally, for future queries, it would be helpful to start a new thread, with your specific question as the title, to better help others looking for answers to comparable concerns.
Good luck with your work.
- Badges
- Science topic
- Similar topics
- Statistics
- Survey