Structure of Questionnaire : measure 5 attributes, each with 4-5 factors having 3-5 item statements on Likert scale . Have 757 responses. No missing data.
Planned statistical analysis are :-
Factor analysis
develop a SEM equation for entire set, groups based on dependent variable
compare different group for many dependent variables - ANOVA, T-tests
Prime questions :
1. Can there be outliers for likert-type scales.
- please note, individual box plots show outliers, removing them is not solving problem of Normality)
- The overall data shows outliers (in box plot and also using skewness & kurtosis scores
2. Request to advise if normality needs to be checked
a. item-wise (88 statements),
b. factor-wise (as developed with EFA) - 5 constructs with 4-5 factors each or
c. overall constructs (5 constructs).
. With grouping based on dependant variable, normality is met for a few groups. .
. One very large group (with almost 50% respondents) has high skewness and kurtosis value - statistic /SE is outside the range of +or - 1.96
I am afraid transforming will affect my findings.
Requesting guidance, please
Hello Hemalatha,
So, your questionnaire was designed to measure, according to your query, 20-25 first order factors and 5 second-order factors, and each factor was supposed to have 3-5 indicator (manifest) variables associated with it. Further, you have a DV (no information given), and groups (which are to be identified based on this DV, as opposed to being established before analysis). Given this framework, you inquired about:
1. Outliers. Yes, outliers can exist no matter what the nature of the response variable scale. Consider this hypothetical, dichotomous scale survey item: We should annihilate all plant life on the planet. Agree or Disagree. If, in a sample of 500 persons, 499 choose "Disagree," while one chooses "Agree," then by virtually any indicator (such as those you're using), the lone respondent choosing "Agree" is an outlier (or, perhaps, an outright liar!). However, many variables, because of their nature, have rare or exceptional cases, but these are not unexpected or false. Consider family income per annum, world-wide. This is a positively skewed distribution, but it most assuredly has legitimate cases at the very highest of values. For such instances, transformation (e.g., log transform) will help to make the resultant distribution more symmetric and reduce the orders of magnitude separating the extreme values.
2. Normality. Yes, check it, though it's less an issue for the measurement model and more an issue for the subsequent statistical comparisons of groups you propose. I'd use factor scores (first or second order, depending on your research questions) as the variates for this.
3. Transformation. Don't fear this; you're just re-expressing the values onto a different scale (in this context, to achieve some desired distributional shape), not changing the fundamental meaning of the variable/construct.
Good luck with your work!
- Badges
- Science topic
- Similar topics
- Psychology