I have sequentially measured the time-series Electronic Health Record patient dataset. This is 7 years of research and each patient's blood pressure, Hba1c, LDL, Hb, Baseline Age,...were measured (as the independent variable for modeling) every 4 months and at the same time, each patient was classified as healthy/not healthy (output/dependent variable). Before modeling, we assigned the output of each patient only as 0 /1 (0= if the patient is assigned 0 at least once, 1= otherwise). So we have time-series (HBA, fpg, ...) and not time-series measures (race, baseline_age) for independent variables , but only one output (0/1) for each patient. I want to model this data, there are some methods used for such kind of dataset, such as; using baseline values or using the mean value of each independent variable. In addition to these two methods, I want to use time-series analysis, but I am not sure what am I going to use? Looks like a survival analysis, but it is not, since the research didn't end when we see 0 value. You can see the visualization of the data structure below. Thanks for all your responses in advance.
What research question led you to undertake data collection? David Booth
Dear Melih Ağraz ,
Before suggesting the suitable method(s) for analysis, I have some confusion I am requesting you to clear based on the information you provided and a snapshot of your variables data. You mentioned that, I quote "So we have time-series (HBA, fpg, ...) and not time-series measures (race, baseline_age) for independent variables , but only one output (0/1) for each patient". Indeed, the independent variable age is fixed at baseline. Now the question I have is, if data on the dependent variable/output at every 4 months interval, age increases over time, and since age at the start of the process can indicate or serve as baseline age, why then keep it (age) constant at baseline because the health status can also be influenced by change in age over time? I might be wrong but please kindly clarify - thanks.
Dear David, that is true, maybe that part can be changed, we can increase the Age by the time-sequence.
Dear David, actually I wasn't into the data collection part, for that reason, my answer can be I don't know, but for now, we are trying to model the effects of independent variables on the dependent variable via conventional Machine Learning Algorithms/Time Series Modelling
Melih Ağraz ,
Thanks for the response - am glad you understood me. Time series modelling approaches cannot be suitable given the nature of your dataset. The fact that your data consists of repeated observations on the same subjects/cases at different occasions (points in time - intervals) makes it longitudinal (cross-sectional time series or panel) - the key advantages of such (panel/longitudinal) data (as opposed to cross-sectional data) being that it allows one to take into account (1) unobserved heterogeneity, (2) dynamic relationships, and (3) causal effects.
In respect of your dataset, you have a dependent or response yit variable is binary (0 or 1) in nature for each subject i at occasion/interval t, sample (n) are the subjects or individuals in the sample, with Ti (number of occasions at which each subject i was observed), and a vector (Xt) of IVs or covariates for subject i at occasion t. These features make your dataset panel, and the fact that your response variable is binary makes the panel binary logistic regression an appropriate estimation technique to apply - ie you therefore need to use a use the panel-data version of logistic regression.
Elvis Munyaradzi Ganyaupfu, thank you for your detailed answer. By the way Would your answer be the same if the length of the time-sequence is different for each patient, and the output is not dependent on the time-sequence? I mean the length of the time is 20 for patient 1 and the output is 3, the length of the time is10 for patient 2, and the output 12. The number of output is different since the patient can sometimes feel sick and go to the hospital.
Melih Ağraz ,
Regarding your question whether the answer remains the same if the length of the time-sequence is different for each patient, and the output is not dependent on the time-sequence, the brief response is that the answer will NOT change - the pivotal reason being that the manner in which your data was collected does not change it from being panel data since such data was collected for each individual (Yi) over successive time period intervals (Ti). Differences in the length of time sequence for individual patients and output being not dependent on time sequence do not change the nature (panel), Yi being not dependent on time sequence is a separate issue that has to be addressed separately using suitable approaches based on relevant preliminary checks.
However, in light of that hypothetical background you gave, it is crucial for every piece of analysis of this nature to always align your analysis approach to the primary research question and primary research hypothesis formulated in your study. So far I am just sharing ideas but without knowing what your major research question and main research hypothesis are - if you may share that information, it could help. Special attention needs to be given in checking whether the dataset is either balanced or unbalanced prior to proceeding with the analysis. A dataset is balanced if all subjects are observed at the same occasions ( T1 = = ..... Tn), otherwise it is unbalanced.
Melih Ağraz ,
You mentioned consideration to use a Machine Learning algorithm for your analysis. I am quite much interested in your study for learning purposes (Health Econometrics) , and out of interest, which programming language do you intend using (- eg. R, Python, etc) and possible package(s) you will use - eg. TensorFlow, Keras, etc? What is your primary research question, as well as your research hypothesis if already formulated prior to design of your research and data collection? Thank you
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling