There is a recent study that tries to answer that question: https://www.biorxiv.org/content/early/2017/07/28/169789

We show that 1x WGS is an efficient alternative to imputed GWAS chip designs for empowering next-generation association studies in complex traits. We demonstrate that population-scale 1x WGS allows the interrogation of a large number of low-frequency and rare variants missed by classical GWAS array imputation, resulting in potential higher association power.

What the SNP array / chip manufacturers need to understand is that they need to be more generous with the genomic locations that are genotyped on their platforms, and to not just give us a random spread of positions scattered across the genome. If they want to get serious, they'd start to genotype, for example, all ClinVar pathogenic variants and tens of thousands of other variants for which there is evidence of association with disease. I'm sure that they have already had discussions about where they are going with their platforms, and I know that, for example, Affymetrix is not by any means ditching their arrays - I heard this straight from the mouth of the CEO of the company during a meeting, which actually surprised me (they may have since reconsidered).

I believe there is a future for these types of arrays if the manufacturer's 'shape up' and start to take their roles more seriously. We are now at that point for many diseases, i.e., where we have enough evidence accumulated such that a simple array could be used to diagnose various diseases, or at best infer your risk of developing a particular disease. For other complex diseases for which evidence is scant, obviously whole genome sequencing is a favourable option.

Keep in mind that the SNP array technology is also ameanable for copy number detection, which has uses, and in those situations it is actually better to have an even distribution of probes scattered across the genome.

Kevin