help > Modelling group (dose) by blood concentration level
Showing 1-1 of 1 posts
Jan 14, 2025 12:01 PM | sap8
Modelling group (dose) by blood concentration level
Hi Alfonso,
Apologies for cross-posting, but I wasn't sure if the post I originally commented had been archived/would be picked up. I've copied my follow-up query below along with your response to the previous thread:
Jan 10, 2025 11:01 AM | sap8
RE: How to model dose-response effect 2nd
level
Hi Alfonso,
I've got a similar project design but am confusing myself on the best way to implement the below.
My design is: 3 conditions (within subjects) - 0mg (placebo), 9mg (low dose), and 18mg (high dose). We now want to do a regression of the connectivity and the actual PK concentration levels from the bloods taken, so we have the 3 conditions (0mg, 9mg, 18mg) as well as blood drug concentration levels at each visit. We're mainly interested in the high dose condition compared to the placebo condition.
At the moment, I've setup conditions as Placebo, 9mg and 18mg, then created 3 second level covariates of the actual blood concentration level for each subject. I end up with 4 categories under Subject Effects: Placebo, PK_conc_high, PK_conc_low, and AllSubjects
Many thanks,
Natalie
Originally posted by Alfonso Nieto-Castanon:
Hi Emily, Yes, that is perfectly fine (creating a new 'Dosage' covariate which contains the actual dosage levels for each subject, and then looking at associations between connectivity measures and dosage by selecting 'AllSubjects' and 'Dosage' and entering a between-subjects contrast [0 1] -or equivalently selecting the 'main effect of Dosage' contrast in the contrast list-) Just as a general comment, there are some subtle and interesting differences between using this approach vs. your original proposal of breaking down your subjects into three groups (e.g. low- mid- high- dosages) and then testing for potential between-group differences (e.g. selecting all three groups and entering a contrast [-1 1 0; 0 -1 1]). It all comes down to a combination of sensitivity and specificity, analysis 1 ('Dosage' regression) looks for linear associations between connectivity changes and dosage levels, while analysis 2 (between-group differences) does not assume any particular form of association. Naturally analysis 1 will be most sensitive to those cases where connectivity changes are in fact linearly related to dosage differences, while analysis 2 will be most sensitive when the connectivity differences show some form of saturation or minimal-dosage effects thay may be well matched to your chosen dosage groups/thresholds. In general, if you have good reasons to believe that the effect of the drug may show threshold or saturation effects (e.g. the range of dosages is very large) and your groups are representing well the range of expected changes then analyses 2 is likely the best choice, while if unsure (or if the range of dosage levels is relatively small) then analysis 1 is likely the most sensitive/best choice. There are, of course, other alternatives, like implicitly specifying a richer class of functions that characterize the expected form of the association between dosage levels and connectivity differences (e.g. a family of quadratic or higher-order polynomials, a family of spline curves, etc.) which may be a better fit if you want to cover a wider range of possible forms for those associations (this is typically done by including a set of second-level covariates which provide a basis for the space of desired functions, e.g. defining a constant, linear, and quadratic dosage second-level covariate fully model all possible quadratic curves characterizing the association between dosage and connectivity). Last, I do not know much about your particular design, it is possible that you might have repeated measures for each subject (e.g. were the subjects scanned before and after the drug was administered?) and/or that your three subject groups were not random (e.g. was dosage level for each subject decided based on other considerations, like subject diagnostic, or was it purely-random?), which may affect the best way to analyze these data. Hope this helps Alfonso Originally posted by Emily Stern:
I also decided to try adding another 2nd level covariate in the Setup tab that was the actual dosage of the drug (so, each subject gets a covariate value consisting of their dosage). Then, in the second level analysis I highlighted the "all subjects" variable (1s for every subject in Setup tab) and this new "Dosage" variable and put the contrast "0 1". Does this sound appropriate? Thanks!