within-subject intervention design - covariates of no interest

[Evelyn M]

Dear SwE Community,

(sorry for potential cross-posting, but I haven't found the exact topic yet):

I have a cross-over within-subject design and I am interested in the intervention effect (post>pre > verum>placebo).

Would you recommend to model covariates of no interest (age, sex, socioeconomic status, BMI...) or is this accounted for by the within-subject design in SwE?

Moreover, for reporting a main effect of our task (not intervention related), I would consider including those for sure - would you agree ?

Thanks for a reply,

Evelyn

[Thomas Nichols]

Dear Eveyn,

I have a cross-over within-subject design and I am interested in the intervention effect (post>pre > verum>placebo).

I'm confused by your characterisation of the intervention effect.  In the simplest cross-over design, say, there is treatment A and B that each subject will receive in turn, i.e. some get A-then-B, and others get B-then-A.  In this setting, I don't understand how any "post" "pre" effect can have any interest.

 

Would you recommend to model covariates of no interest (age, sex, socioeconomic status, BMI...) or is this accounted for by the within-subject design in SwE?

I'm pretty sure that in a balanced design (equal number of visits/scans per subject) modelling between-subject covariates won't impact your intrasubject inferences.  However, including sequence (i.e. "pre/post") effect should help reduce intrasubject error and improve sensitivity; though if you have a severe crossover imbalance (far from equal A-then-B and B-then-A subjects) it may reduce sensitivity because you're effect of interest is aliased with the sequence effect.

 

Moreover, for reporting a main effect of our task (not intervention related), I would consider including those for sure - would you agree ?

Indeed... a main effect is crosssectional and would definitely benefit from all such cross-sectional covariates.

-Tom

 

Thanks for a reply,

Evelyn

--
You received this message because you are subscribed to the Google Groups "SwE-Toolbox Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to swe-toolbox...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/swe-toolbox/66ab37cb-0d79-4a08-b044-601ff3c9dc08n%40googlegroups.com.

[Evelyn M]

Dear Tom,

thanks for the speedy reply!

On Monday, April 12, 2021 at 7:56:52 AM UTC+2 ten.p...@gmail.com wrote:

Dear Eveyn,

I have a cross-over within-subject design and I am interested in the intervention effect (post>pre > verum>placebo).

I'm confused by your characterisation of the intervention effect.  In the simplest cross-over design, say, there is treatment A and B that each subject will receive in turn, i.e. some get A-then-B, and others get B-then-A.  In this setting, I don't understand how any "post" "pre" effect can have any interest.

 

-> Each participant has four visits, meaning: verumBL, verumFU, placeboBL, placeboFU; verum and placebo are randomized. Most have four visits, but some do not, this I why I chose to use SwE, to still include the single datapoints. Does this change your response below?

 

Would you recommend to model covariates of no interest (age, sex, socioeconomic status, BMI...) or is this accounted for by the within-subject design in SwE?

I'm pretty sure that in a balanced design (equal number of visits/scans per subject) modelling between-subject covariates won't impact your intrasubject inferences.  However, including sequence (i.e. "pre/post") effect should help reduce intrasubject error and improve sensitivity; though if you have a severe crossover imbalance (far from equal A-then-B and B-then-A subjects) it may reduce sensitivity because you're effect of interest is aliased with the sequence effect.

 

Moreover, for reporting a main effect of our task (not intervention related), I would consider including those for sure - would you agree ?

Indeed... a main effect is crosssectional and would definitely benefit from all such cross-sectional covariates.

-Tom

 

Thanks for a reply,

Evelyn

--
You received this message because you are subscribed to the Google Groups "SwE-Toolbox Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to swe-toolbox...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/swe-toolbox/66ab37cb-0d79-4a08-b044-601ff3c9dc08n%40googlegroups.com.

Best,

Evelyn

[Thomas Nichols]

Dear Evelyn,

Thanks for clarifying about your design.  I see... so treatment is randomised but "baseline->followup" isn't.

With the imbalance, I think it is certainly safest to include the cross-sectional nuisance variables just for good measure, even if you're not interested in cross-sectional effects.  

I'm curious how you plan on modelling this.  The natural approach is to have four indicator variables for these four (partially randomised) conditions; but you could then also include a sequence variable that is 0 for the first two visits (in time) and 1 for the last two visits (in time).  That would model the time effect effect that the randomisation is trying to balance out and could improve sensitivity.

-Tom

To view this discussion on the web visit https://groups.google.com/d/msgid/swe-toolbox/dbb82039-c09d-4a2b-946b-2c28e6bffd13n%40googlegroups.com.

[Evelyn M]

Dear Tom,

to be more detailed and precise, I want to suggest here the full design for the 2nd level and ask for your advice:
* subject (1-61)

* timepoint (0 1) (pre / post)

* condition (0 1) (placebo / verum)

* visit number (1 2 3 4)

* order (1 2) (which intervention is first, which is second)


- Would it matter of some of these are not perfectly orthogonal (e.g. visit number and order)?
- Would the contrast then be [0 1 1 0 0] for the condition*timepoint effect?
- For the main effect of visit or order I would model [0 0 0 1 0] and  [0 0 0 0 1] accordingly?


To acknowledge the efforts of the RCT design, I would then only in sensitivity analyses run an additional model with the aforementioned (cross-sectional nuisance) covariates.

What do you think?

Best
Evelyn

[Thomas Nichols]

Dear Evelyn,

to be more detailed and precise, I want to suggest here the full design for the 2nd level and ask for your advice:
* subject (1-61)

* timepoint (0 1) (pre / post)

* condition (0 1) (placebo / verum)

* visit number (1 2 3 4)

* order (1 2) (which intervention is first, which is second)


- Would it matter of some of these are not perfectly orthogonal (e.g. visit number and order)?

There is no need for orthogonality, but it seems like some of these are linearly dependent.  Let me show you how I would consider this:

SwE model specification:  You need to set subject and visit (*neither* of these are variables, they are just information SwE needs to configure the repeated measures covariance).  You then have the following variables:

Intercept

Treatment (placebo/verum)


PrePost (baseline/followup)   (I'd avoid calling it "time point", since that could be 0/1,0/1, or 1/2/3/4)

Session (1/2)

The Session variable captures the order effect.

It's useful to count the model DF... here you have intercept plus 3 binary variables, so that's 4 parameters.

I presume you probably *also* want the Treatment*PrePost interaction?  To detect treatment-specific PrePost effects? You could also encode the interaction directly with the variables:

Treatment*PrePost (verumBL, verumFU, placeboBL, placeboFU

Session (1/2)

but note that modelling Treatment*PrePost interaction this way you're using a 'factor mean' parameterisation, and so now the intercept is integrated into this factor.

The model DF are: 4 means for Treatment*PrePost, 1 binary variable, or 5 parameters.

- Would the contrast then be [0 1 1 0 0] for the condition*timepoint effect?

With the second, 5-parameter model, you'd easily capture all the contrasts you'd want with the first 4 elements...

FU-BL:  -1 1 -1 1 0

verum-placebo: 1 1 -1 -1 0

- For the main effect of visit or order I would model [0 0 0 1 0] and  [0 0 0 0 1] accordingly?

Again, it doesn't make sense to model visit once you have all the other parameters.  If you're interested in the order effect that would be [0 0 0 0 1].

To acknowledge the efforts of the RCT design, I would then only in sensitivity analyses run an additional model with the aforementioned (cross-sectional nuisance) covariates.

Yup!

-Tom

To view this discussion on the web visit https://groups.google.com/d/msgid/swe-toolbox/32fabd86-088b-4eec-a87c-9dcc62a71289n%40googlegroups.com.
<subject.png>

__________________________________________________________
Thomas Nichols, PhD
Professor of Neuroimaging Statistics
Nuffield Department of Population Health | University of Oxford
Big Data Institute | Li Ka Shing Centre for Health Information and Discovery
Old Road Campus | Headington | Oxford | OX3 7LF | United Kingdom
T: +44 1865 743590 | E: thomas....@bdi.ox.ac.uk
W: http://nisox.org | http://www.bdi.ox.ac.uk

[Aimee Flores]

Dear Evelyn and Tom,

I have a very similar design. I did therefore the following matrix. BCI is the treatment and After is the (pre/post). I added all of them in the GUI in the section "covariate". I tried running it with the contast [1 -1 -1 1]. However, matlab gives me an error. I wanted to check with you if in principle the matrix is correct and maybe I am missing something else in the batch or if maybe is a problem with how I enter the variables.

Best wishes,

Aimee Flores