Use the numerical generative process to calibrate the model

[CastielZhao]

Does the false positive rate we claim (e.g. 0.05) correspond to 5% of false positives given our no-association, no-outlier simulated data?

Calibration:

• inference of associations. (read https://doi.org/10.1093/nargab/lqab005)
• inference of outliers

[stemangiola]

Calibrate inference of associations

• Generate 100 datasets with the same total counts per subject (M size vector, where M is the number of subjects), for each dataset
• Number of subjects 30, number of categories 20
• Design matrix would have an intercept column and a factor of interest between -1 and 1
• Setup coefficient to have same intercept (for simplicity), and zero slope
• Generate the data
• Execute sccomp (visit homepage of this repository)
• • FOR INSTALLATION DO: devtools::install_github("stemangiola/sccomp")
• • library(sccomp)
• • Follow the readme
• Count how many categories were labelled as significantly changing (by default we are using the 95% credible interval. Which means that we expect 5% of calls to be false)

[CastielZhao]

"Setup coefficient to have same intercept (for simplicity), and zero slope"
Are there any other constraints on coefficient? i.e. integer ? Range ?
Also, I assume that "zero slope" means coeff=(beta0,beta0,...,beta0; beta1,beta1,...,beta1); that the first column repeats 20 times.

[stemangiola]

"Setup coefficient to have same intercept (for simplicity), and zero slope"
Are there any other constraints on coefficient? i.e. integer ? Range ?

Execute the code at the homepage of this repository and you will see what coefficients you get for a real dataset. You can get the range from those (except the intercept that should be zero for this test)

[stemangiola]

About integer or not, it is exactly the same. When you do matrix multiplication between design and coefficient is the same.

[CastielZhao]

Hi Stefano,

I have successfully created 100 data frames from my function. To detect the change, do I need to use sccomp library? Or I shall find out a way to do that ?

[stemangiola]

Hi Stefano,

I have successfully created 100 data frames from my function. To detect the change, do I need to use sccomp library? Or I shall find out a way to do that ?

Yes, run sccomp on your data set. See example dataset from github README. Start from a few and try to draw descriptive statistics.

[CastielZhao]

which function in the sccomp is used for detecting variation ?

[CastielZhao]

As I noticed the fuction: res =
counts_obj %>%
sccomp_glm(
~ type,
sample, cell_group, count,
approximate_posterior_inference = FALSE
)
When analyzing multiple data frames, do I need to merge the data frames, or specifying different data frame by "cell goup " above? Also, type=category, count=count, sample=subject in our dictionary, right?

[stemangiola]

if you analyse different studies no, you analyse them independently. I don't know what you mean by data frames. Data frame can be anything. Please be more precise.

Also, type=category, count=count, sample=subject in our dictionary, right?

yes

[CastielZhao]

if you analyse different studies no, you analyse them independently. I don't know what you mean by data frames. Data frame can be anything. Please be more precise.

Also, type=category, count=count, sample=subject in our dictionary, right?

yes

By data frames, I mean the output simulated data frames from my numeric generation process.

[stemangiola]

one data frame includes M categories and N subjects.

another data frame includes M categories and N subjects.

one subject does constitute a very small dataset that cannot be used for regression, size = 1