Good morning,
for my research project, I am using school meal data selection. I would like to investigate the children's food selection patterns using multiple time series using the K-means method. Given the remit of the study, in my data, I have missing data due to data collection during school and bank holidays, weekend generating breaks in food selection values. When you investigate a phenomenon on a daily scale, how do you manage these kinds of missing values? Do you change the temporal scale for example month rather than day, keeping breaks in graphics, or perform an imputation?
Rather imputation, for certain data, you can perform interpolation (linear by default, if you know the "nature" of your variables, you can use other functions).
I don't know how much data you have or more, on which scale (several years or only a couple of months?) you work, but for the weekends, you could just set your temporal scale to "week", where it corresponds to the mean of the five (or less due to bank holidays) days during the week.
In case of design of experiments, we have specific equations to calculate missing values, and if not, you can take the average of other replicates!!. Regards.
Thanks everyone for your answers I work with categorical variables I will calculate the percentage of selection for each category and I would like to use the menu cycle as a temporal unit (1 menu cycle is equal to 3 weeks, but sometimes the menu cycle can be 2 weeks when the menu change, thus, I think that I will keep only complete cycle to avoid biases of selection due to shorter period)
I don't agree with interpolation. The K-means method does not exploit the time ordering so that ignoring that there are missing data has no effect on the result. Even if you used methods that exploit past observations (e.g. selection of the previous day), I don't think that taking care of bank and school holidays will be important. It is only if you take into account the selection of the same day of the previous week that the presence of missing data will be annoying. In that case, if you have long series (let us say longer than 100), the treatment of missing data in Tramo-Seats will perhaps be enough.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series