Did you try to Association Rules in machine learning specially Python? 

Go through the following links also: https://ourcodingclub.github.io/2017/03/15/mixed-models.html https://www.spss.ch/upload/1126184451_Linear%20Mixed%20Effects%20Modeling%20in%20SPSS.pdf https://stats.idre.ucla.edu/other/mult-pkg/introduction-to-linear-mixed-models/ https://web.stanford.edu/class/psych252/section/Mixed_models_tutorial.html http://users.stat.umn.edu/~helwig/notes/lmer-Notes.pdf http://www.stat.cmu.edu/~hseltman/309/Book/chapter15.pdf http://www.stat.rutgers.edu/home/yhung/Stat586/Mixed%20model/appendix-mixed-models.pdf http://edwardlib.org/tutorials/linear-mixed-effects-models

Good luck.

Hi Andrea,

Point 1) The techniques are software specific, but your goal is to copy the baseline values for each row of the case. Suppose your wide form data is of the form:

y1 y2 y3 x1 x2 x3 z d1 d2 d3 d id

10 12 11 5 6 7 23 0 0 0 0 1

11 . . 2 . . 21 1 . . 1 2

12 10 . 3 3 . 15 0 1 . 1 3

where y1-y3 are values of a continous outcome, x1-x3 are values of a predictor, z is a variable measured only in baseline, d1-d3 are measurement-specific dementia indicators (0 = no, 1 = yes), d is an overall dementia indicator (0= no / 1 = yes, whether dementia was observed for subject at any point of the study) and id indicates subjects (dot indicates missing value). Your goal is to bring the data into the following long form:

id de d y x z t

1 0 0 10 5 23 1

1 0 0 12 6 23 2

1 0 0 11 7 23 3

2 1 1 11 2 21 1

3 0 1 12 3 15 1

3 1 1 10 3 15 2

where de collects the values from d1-d3, d from d, y from y1-y3, x from x1-x3, z from z and t is constructed to indicate measurement timepoint.

To get z into this form you can make three copies of the baseline measurement. For example in SPSS you can achive the same thing by using the restructure procedure and in the window titled "step 3 or 7", move variable z into the fixed variable box.

If you think that the predictor effects change over time, you might consider using interaction indicators for time indicators, but be careful not to make the model over-parameterized. In this way it is possible to consider a wide array of models for the longitudinal associations.

2) From the layout of the long format you can easily see that you can choose cases for analysis in various ways. For example, by selecting de=0, case data is selected for analysis prior to onset of dementia (analysis set will include pre-dementia data from those who later were diagnosed with dementia). By selecting subjects with d=0, only those cases who remained dementia-free over the whole course of the study will be selected.

For mixed models a useful mostly highly accessible reference is Singer and Willett. Applied Longitudinal Data Analysis (https://oxford.universitypressscholarship.com/view/10.1093/acprof:oso/9780195152968.001.0001/acprof-9780195152968). It includes many models that might be useful for you with good examples.

3) The mixed models with data constructed in this way will assess the impact of associations between timepoints of the data and it does not directly address the impact of changes. If you are specifically interested in the impact of changes, you might consider e.g. latent change score models (see, e.g. https://www.tandfonline.com/doi/abs/10.1080/10705511.2018.1562929?journalCode=hsem20)

or possibly growth curve models. The easiest program to use for these analyses is Mplus, but you could do them also in R (package lavaan). Change score models could also be considered by computing change scores (e.g. z2-z1 and z3-z2 or z2-z1 z3-z1 depending on the point of reference) and using time point indicator interactions.

Best regards,

Timo

You can read my article in research gate as PDF about this subject

good Luck