My colleagues and I have delivered a 4-wave survey (day 1, day 30, day 60, day 90), which makes use of multiple psychometric measures. We have the complete data for each survey collected, but also have entirely missing surveys. For example, we may have the data for day 1, day 30 and day 90 only. In that case, can we impute the data for the day 60 survey from the data collected? If so, which technique would be the most appropriate? The missing data does not form a pattern.
Hi,
It depends of envisaged statistical analyses.
If you analyse the change over time, the mixed model can be a very efficient method to analyse longitudinal data and this method can deal with missing data (Article A Comparison of the General Linear Mixed Model and Repeated ...
).On larger dataset and to evaluate complex structural model on longitudinal data, Latent Growth Curve models can be used.
Dear Kellie
In fact there is no pure "Reliable" method available to impute entire missing data. There are methods that can be of helpful in dealing with such issues. Each method have its shortcomings attached to it.
The mixed model might be used, but the number of time points is quite low. Maybe even too low as you have a repeaed manova design with fixed timepoints. Or are they in reality a little more random? You might also want to have a look at Loh and Zheng (2013) REGRESSION TREES FOR LONGITUDINAL AND MULTIRESPONSE DATA in The Annals of Applied Statistics, Vol. 7, No. 1 (March 2013), pp. 495-522. This article uses regression trees to analyze longitudinal data, but can also accommodate for missing values.
Thank you Nicolas, for the suggestion and reference.
Rajesh, do you have any suggestions as to what might work in this instance?
Peter, in reality the timepoints are more random. The participants are recruited in a staggered fashion, and while they may have been given the link to the survey at the time points mentioned, that doesn't mean that this is when they filled it out. We are having to make decisions as to what timepoint a survey can be allocated to or even if it can be allocated to one. I will have a look at the reference you mention. Thanks for your help.
@Kellie, Did you register the dates on the surveys (for instance if you have collected data using an online tool)? Depending on what you want, you could see these as random, but it depends on the type of relation in time you would expect. With more points in time, the function to be estimated can be more complex. For instance with two points, one can fit a line, with three points a curve, and with many point, one could fit a line, a curve and a signoid (seasonal influence) and ..... But that would depend on the theory giving indication on the relation between the dependent variable and time.You would not need to be restricted to a 'timepoint', but to a dates (1-1-2018, 1-2-2018 etc) or timeframes (week01-2018, week02-2018,etc, January 2018, February 2018 etc or quarter01 2018, quarter02 2018). How fine grained you can make it, depends on the theory and amount of data.
- Badges
- Science topic
- Similar topics
- Psychology
- Psychometrics