Running testNhoods with and without "model.contrasts" argument results in wildly different results

[jcshuy]

I have a dataset of about 40k cells with only two conditions (condition1 and condition2). When I run the following testNhoods with the model.contrasts = "condition1 - condition2" argument, I get a result of table(da_results$SpatialFDR < 0.1) as all FALSE:

da_results <- testNhoods(milor, 
                         design = ~ 0 + condition + batch, 
                         design.df = design, 
                         reduced.dim="PCA", 
                         fdr.weighting="graph-overlap", 
                         model.contrasts = "condition1 - condition2"
                         )

The above is the Pvalue histogram of that result.

However, when I run the same function without including the model.contrasts = "condition1 - condition2" argument, I get a result that looks more expected, with some significant values and some non significant. The Pvalue distribution looks as follows:

Also, if I switch the design argument to design = ~ 0 + condition (doesn't account for batch) or reorder it to design = ~ 0 + batch + condition, the result of table(da_results$SpatialFDR < 0.1) will either be all TRUE or all FALSE

Is the model.contrasts argument necessary when there is only two conditions to choose from? I did read through the provided link to categorical variables from the vignette, but I am still confused on the stark differences between the two results. Which one is the correct one to go with? We don't expect to see any extensive changes, but having them return as all insignificant or all significant would also be pretty unexpected.

My design matrix looks like the following:

design <- data.frame(colData(milor))[,c("orig.ident", "condition", "batch")]
design$batch <- as.factor(design$batch) 
design <- distinct(design)
rownames(design) <- design$orig.ident
# contrast <- c("condition1 - condition2")

> design
           orig.ident   condition   batch
sample1       sample1  condition1       1
sample2       sample2  condition2       1
sample3       sample3  condition1       2
sample4       sample4  condition2       2
sample5       sample5  condition1       1
sample6       sample6  condition2       1

[MikeDMorgan]

The default behaviour of testNhoods is to use the last variable in the design formula unless specific contrasts are provided. When you have a binary variable you don't need to use model contrasts, so your design formula just needs to be ~ batch + condition. To remove any doubt about the direction of effect sizes, set your condition variable to be a factor and define the order of levels yourself. Also, make sure that batch is a character/factor variable - if it is a numeric variable it will assume an additive relationship for each batch number (ignore this if there are only two batches).