So you have repeat measurements for the same patients across 2 years? Do you have complete data for everyone across the period, or could they have had encounters elsewhere (e.g. at other facilities)?
Hi guys, i have multiple biological parameters in multiple patients during a period of two years and when to check for correlation or relation between 2 or more biological parameter variation in time between groups of patients. What should be the more efficient test i could use ? Time series ?
Thanks for help and time.
Guyom
So you have repeat measurements for the same patients across 2 years? Do you have complete data for everyone across the period, or could they have had encounters elsewhere (e.g. at other facilities)?
Stop cowardice, ban guns!
Well of most of them (n=180) i have repeat measurements for at least 10-20 biological parameters... pretty complete data. My question is to try to explain the variation of a specific parameter in relation with 2-3 others. during a period of time. I tried different test like PCAs, several discriminate analysis, two Way ANOVA... but i think it is good for punctual relation between parameters. I want now adding time... so sounds like a 3D matrice, with X=biological parameters, Y=patients and Z=time... but not sure of the exact test.
thanks
Are variables collected at the same time point, all measures from same visit (per subject)?
Can you provide an example of your data? You can make of the values.
Stop cowardice, ban guns!
well data are collected from same visit (per subject) or, if not, we added the biological values obtained at the closer data from the visit (per subject). here is an example of the data for two patients.
Statut clinique size weight BMI Glycemia Saturation % CRP (mg/L) VEMS % etc...
P002-V01
P002-V02
P002-V03
P002-V04
P002-V05
P002-V06
P003-V01
P003-V02
P003-V03
P003-V04
P003-V05
P003-V06
P003-V07
where V01, V02 are visit number per subject. For each visit and each patient, we have Saturation (O2), FEV1 (%), CRP etc... we have numeric or character, nominal versus continue etc...
Which could be the more interesting test : Time Serie, Multivariate approach, Univariate Split-Plot Approach and Mixed Model Approach ?
thanks
Last edited by Gbouvet; 09-05-2017 at 01:57 PM.
just to add : Patients have not the same number of visit (depending of the exacerbations, some patients have more than other) and Visits have roughly the same intervals (meaning 3-4 months).
Your p002... is confusing. How many values do you have per visit, and on average how many visits does a person have?
Stop cowardice, ban guns!
Well P002 is Patient ID... I have per visit one value for each biological parameters and at least 15 parameters (so 15 values) and an average of 7-8 visites per patients, sometime less, sometime more. (range goes form 1 to 17).
Are all variables continuous? I guess what is the real purpose? What will you do with the results?
Do you have an a priori plan of which variables should be co-linear or are you going to run models with 10-20 dependent variables? Is there any relationships among 2+ variables. Is multi-collinear a threat? If you just look at X1 vs. X2; X1 vs X3;...;X9 vs. X10, for say 10 variables that is ~ 45 models and your familywise error rate is going to be pretty high. So is this a free for all or is there background knowledge to guide the process.
I am assuming values are correlated with each other (covariance structure, say P0001's first visit value is correlated with second visit data, etc.?
Stop cowardice, ban guns!
my a priori plan is to try to understand the variation of one biological parameter with other standard ones in a genetic disease context. Yes i can have some variables linked with other like inflammatory markers, glucid or lipid markers, BMI age so i can reduce a bit the number of the variables... at least to 7-8 including my target. So yes there is a background.
Yes, Values can be correlated to each other i mean intra-patient. As an example, if there is a increase of inflammation due to an infection, pulmonary capacity and BMI can decrease.
You can potentially run a bunch of multilevel models (for repeated measures clustered in patients). You will just need to realize if you do a bunch of pairwise comparisons there is a risk for type I errors.
Stop cowardice, ban guns!
Ok i used JMP for statistical analysis... i am not sure what is your multilevel models are ?
You have repeated measures that are clustered within in patients. You have to control for this dependence and its covariance structure. It is like controlling for students in different classrooms. Students in the same classroom will tend to be similar in the outcome, like values will be similar within patients. Multilevel models control for student and classroom level variables and the nested relationship. If you just ran partial correlation or multiple linear regression, you would lose this important piece of information.
Another approach besides multi-level models is generalized linear regression. These are not typically novice analytics, that take a little bit more thought.
Stop cowardice, ban guns!
Ok got it thanks!
Tweet |