Yes, you can, as long as you get the data harmonized with same pipeline.

TCGA has a standard pipeline to process data, so it is safe to merge them, just pay attention to the data/portal version to make sure they are same.

For the batch effect, TCGA did nothing about this, so you would need to do it by yourself, maybe by integrating clinical information, or use packages such as ruv to remove unknown variation.

Personally I would say it's not very necessary to consider batch effect here, as you already have a large enough cohort which enables you to detect relative strong signals.

TCGA pancancer RNA-seq expression matrix (n = 11768 samples) can be downloaded from UCSC Xena: https://gdc.xenahubs.net/download/GDC-PANCAN.htseq_fpkm-uq.tsv.gz

Recount also re-analyzed all TCGA cancer types ( n = 11284 samples) using the same pipeline, gene and exon level read count matrix are available from

https://jhubiostatistics.shinyapps.io/recount/

UCSC Cancer Browser based Pancan normalized gene expression datasets include all harmonized (Integrated) normalized dataset except STAD. To reduce potential batch effects UCSC team only put RNAseq data from University of North Carolina from the Illumina HiSeq in the pancan normalized dataset. Because STAD only has data from the British Columbia Cancer Agency, UCSC team did not add it to the pancan dataset.

I would recommend looking for tools to measure batch effect based on source of variability and remove batch effects in gene expression data if the source of variability is higher compared to other factors.

Hope it helps,

Mamatjan

yes, you can do it with harmonized data. If you are familiar with R, check the package TCGAbiolinks

My advice would be to look for harmonized data (I guess it means all data from raw to count matrix processed in the same manner) and then adjust for batches, covariates - e.g. study IDs at the modelling level. Going into details, you may have to address statistical issues like exchangability and ignorability.