Hi,

I run into such warnings with MAST when fitting the hurdle model to

a sc-rnaseq data. Data has been log2-transformed(log2(cpm + 1)).

The model is:

zlm.SingleCellAssay(~ FvF + batches + sex, sca)

After fitting, MAST reported "In .nextMethod(object = object, value = value) :
  Coefficients batchesb4, sexU are never estimible and will be dropped."

Are there something wrong with the model? How to understand the warning?

Here are the relation between predictors:

> table(colData(sca)$FvF, colData(sca)$sex)
       
          F   M   U
  Fed     7 199   0
  Ch10    0 231   0
  Fast   13   7  31
  HFD     0 147   0
  Refed   0   0  55
> table(colData(sca)$FvF, colData(sca)$batches)
       
         b1  b2  b3  b4  b5  b6
  Fed    75  36  81   0   0  14
  Ch10    0   0   0 231   0   0
  Fast    0   0   0   0  31  20
  HFD     0   0   0 147   0   0
  Refed   0   0   0   0  55   0

> table(colData(sca)$sex, colData(sca)$batches)
   
     b1  b2  b3  b4  b5  b6
  F   0   0   0   0   0  20
  M  75  36  81 378   0  14
  U   0   0   0   0  86   0

Thanks!

> sessionInfo()
R version 3.3.3 (2017-03-06)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Manjaro Linux

locale:
 [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
 [3] LC_TIME=en_US.UTF-8        LC_COLLATE=en_US.UTF-8    
 [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
 [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
 [9] LC_ADDRESS=C               LC_TELEPHONE=C            
[11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       

attached base packages:
 [1] parallel  stats4    grid      stats     graphics  grDevices utils    
 [8] datasets  methods   base     

other attached packages:
 [1] MAST_1.0.5                 SummarizedExperiment_1.4.0
 [3] Biobase_2.34.0             GenomicRanges_1.26.4      
 [5] GenomeInfoDb_1.10.3        IRanges_2.8.2             
 [7] S4Vectors_0.12.2           BiocGenerics_0.20.0       
 [9] data.table_1.10.4          magrittr_1.5              
[11] RColorBrewer_1.1-2         cowplot_0.7.0             
[13] ggplot2_2.2.1              colorout_1.1-2            

loaded via a namespace (and not attached):
 [1] Rcpp_0.12.10     XVector_0.14.1   zlibbioc_1.20.0  munsell_0.4.3   
 [5] lattice_0.20-34  colorspace_1.3-2 stringr_1.2.0    plyr_1.8.4      
 [9] tools_3.3.3      gtable_0.2.0     abind_1.4-5      lazyeval_0.2.0  
[13] assertthat_0.1   tibble_1.2       Matrix_1.2-8     reshape2_1.4.2  
[17] bitops_1.0-6     RCurl_1.95-4.8   stringi_1.1.3    scales_0.4.1    

Hi ccshao,

The warning means that some levels of your factors are never present together--they are completely confounded. When FvF == "Refed", it's always the case that Sex== "U", so you can't tell if a change in express is due to one, or the other, or both of those factors.  One of those variables must be dropped in order for the model to have a well-defined solution.  Same is true with batches=="b4" and FvF == "HFB".

If you don't care about those particular levels of those factors, then there's nothing to worry about.  If you do, then you may need to collect more data in order to identify the effect of interest--the current data aren't informative.

-Andrew