Hi everyone,
I am working on a project for the State Department of Education. We help them to calculate whether a program reaches an indicator. For example, for this particular study, we sent parents a survey with 25 items. The parents choose on a 6 likert scale from strongly disagree to strongly agree with these 25 statements. We then calculate the average score of the total responses of the 25 items. and see how many percent of parents reach the cut off percent (i.e. 4).
However some parents missed answer some of the 25 items. the difficulties of the items are different. In this case, do I need to remove some of the parents in the analysis (e.g, if some of them only answered 2 or 4 questions?). If so, why should I do so? How do I determine how many missing is the appropriate cut off criterion? Is it necessary to remove parents with missing items in this case?
THANK YOU VERY MUCH!
Hi Jane Li according to Lynch (2007), if missing data was more than 10% from a particular respondent, then that respondent or case would be considered problematic. There are some other views as well, Enders (2010) suggested that it is acceptable if the missing value from a particular respondent was less than 20%.
One method to deal with missing values is replacing them with the mean of the item. Hope it helps.
Reference:
Enders, C. K. (2010). Applied missing data analysis. Guilford press.
Lynch, S. M. (2007). Introduction to applied Bayesian statistics and estimation for social scientists. Springer Science & Business Media.
Jane Li Given that your data is on Likert Scale, you could use median replacement method for data imputation however, in your scale, there is no neutral data point hence mean value would be appropriate for imputation (as suggested by Long She ). Wishing you best with your research.
Jane Li Given that your data is on Likert Scale, you could use median replacement method for data imputation however, in your scale, there is no neutral data point hence mean value would be appropriate for imputation (as suggested by Long She ). Wishing you best with your research.
Dear Jane Li ,
Apart from the question you have asked, I would suggest that you also seek to validate your method of adding all your scores together to create an overall score. One approach would be to use Principal Component Analysis with a single component. Please see my study guide on scale reliability analysis for more information.
Hi Jane Li , there are a few other options that I don't think have been mentioned yet:
* Data imputation *
One can use statistical models to use correlations among questions to fill in the missing values. This goes beyond the mean or median, which are known to mess up the variances and/or distribution shape.
Because the questions have ordinal responses, something like the mode of the N-nearest neighbours might be be appropriate (e.g. Article Nearest Neighbor Imputation for Categorical Data by Weightin...
).Another alternative might be to use random forest, especially if you suspect the relationships among questions might be interdependent (interactions) (e.g. Article Comparison of random forest and parametric imputation models...
).Of course, these need some kind of cross-validation to check that the imputation is performing well (pretend some existing values are missing and see how close the imputation gets).
* Something completely different *
If the ultimate goal is to see whether the parents reach a cut off percent, this can be seen as a classification problem.
In this case, one could use one of many methods (e.g. classification and regression tree) to classify parents as below/above. Example: Article BEST: a decision tree algorithm that handles missing values
Of course, which method is best depends on your data (i.e. sample size, percent missing, pattern of missingness). Hope that helps!
Hi Jane Li ,
I would be wary about applying rules of thumb blindly. It should really depend on the specifics of your project (specific goals, sample size, number of variables, pattern of missingness..)
Example1: Say I have data on 20 subjects. Missing 20% of values would most likely be disastrous. If I have 20 000 subjects on the other hand, missing 20% is probably not a big deal (assuming missingness is random).
Example2: Say I have 10 variables in my analysis. If I'm missing 10% of the data but most of it is in the same variable, I essentially loose a variable. If the 10% missing are spread across all variables, chances are I can still use all 10 variables.
Does that make sense?
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA