You and everyone else doing GWAS....

In any case, besides looking for ncRNAs and undocumented transcripts, there is a load of data from the ENCODE and related projects available most accessibly from the UCSC genome browser. In particular, you could look at overlap between DNAse hypersensitivity, transcription factor binding, maximally conserved elements, etc. and your SNP. If you want to get really crazy, you could look at the Hi-C paper by Dekker et al. that purports to give long-range genomic interactions between various regions of the genome.

Perhaps you could use some RNA-mapping technique to look for non-coding RNA's containing the SNP you are interested in. Perhaps 3'- / 5'- RACE could be suitable. At least from this you could get a general idea if the SNP is associated with a transcript. However, your SNP could be part of some long range cis-enhancer region like they found in this paper, which could make experimental validation more complicated.

You could also try mining non-coding RNA lists in various databases such as ensembl (ie. small RNA / lincRNA / etc) to see if any of them map to that region as well.

Well, Khader, you should have been at ASHG last week - and should have then run from talk to talk where all kinds of folks presented very similar situations. The MYC paper GWW cites is an example that comes to mind as well. In fact, I saw talks by two of those authors. I also saw a presentation of using ENCODE data to annotate the GWAS hits - so I agree with Sean's excellent idea. My notes from these talks are on my blog: Wasserman, Degner's use of ENCODE data, and Stamatoyannopoulos' talk on using ENCODE.

This is just a start, though. There is certainly a lot of ideas on this topic. Generally, the SNPs in gene deserts are SNPs regulating either expression of a distant protein-coding gene or are in/near to a non-protein-coding (RNA) gene. Ideally, you would have expression data (either from a genome-wide chip of mRNA probes) or RT-PCR data across the region (in which you can identify those novel transcription products) to compare to the GWAS data to look for a type of eQTL.

Added on 12 Jul 2011: The continued evolution of the epigenomebrowser.org site makes it a very good place to go to get the data mentioned here. There are now some very nice displays of Stamatoyannopoulos' DNaseI hypersensitivity sites, tissue-specific histone methylation and acetylation marks as well as transcription factor ChIP data.

In my opinion you have to be careful when you say "you have identified a causal SNP". From your description you have identified a region of high association with no transcript annotation, AKA gene dessert. This is still not a causal SNP as you don't know how this variation affects your phenotype and whether a variation in LD with your candidate is the true causal SNP.

As it was suggested in some of the answers already you can combine your genotypes with expression data, here is a nice example in the literature.

You can also take a look at this eQTL browser.

Cheers!

Just to add a different point - one probably already known to Khader. This type of question is being considered at WikiGenes and the article here. Many of us (those into genome variation) either need to join this effort or keep abreast of what they write.