Sequence wise it will be difficult to answer your question with a finite value. But if you look at Epigenomic data from ENCODE or UCLA genome browser and similar epigenetic platforms, you will notice that by and large most epigenetic marks on promoters do not span more than 2kb of sequence (looking at Active h3k4me3), However if you look at repressive marks - they are much more widespread. Going by what is necessary to dictate and correlate with gene activity - i can suggest that one probably doesn't need to look at more than 2-3kb in the proximal Upstream region to look for contiguous effects. Now, however we all know that 3C and Hi-C assays have defined TAD domains and complex 3-D architecture at several promoters all of which are in the linear length of mammalian promoters and do influence gene expression. Going by that 3D structre - if you do wish ti include enhancers - one will have to cover about 1MB domains and in some cases probably more! Therefore it is hard to predict what sequence one needs to THEORETICALLY scan in order to cover promoter effects inclusive of enhancers. 

Sequence wise it will be difficult to answer your question with a finite value. But if you look at Epigenomic data from ENCODE or UCLA genome browser and similar epigenetic platforms, you will notice that by and large most epigenetic marks on promoters do not span more than 2kb of sequence (looking at Active h3k4me3), However if you look at repressive marks - they are much more widespread. Going by what is necessary to dictate and correlate with gene activity - i can suggest that one probably doesn't need to look at more than 2-3kb in the proximal Upstream region to look for contiguous effects. Now, however we all know that 3C and Hi-C assays have defined TAD domains and complex 3-D architecture at several promoters all of which are in the linear length of mammalian promoters and do influence gene expression. Going by that 3D structre - if you do wish ti include enhancers - one will have to cover about 1MB domains and in some cases probably more! Therefore it is hard to predict what sequence one needs to THEORETICALLY scan in order to cover promoter effects inclusive of enhancers. 

Also be aware that for most genes there's no such thing as "the one promoter". Most genes will be associated with several transcripts that will have separate promoters. Using a 2kb region there indiscriminately might give you overlaps with coding regions. If you want to look at variants that have regulatory effects e.g. by changing transcription factor binding behaviour, it might be better to first identify regions of binding within the putative promoter regions and then just look at the variants within these.

I see some papers, and they seem to extract promoter sequences that were less than 1 kb region upstream of the annotated transcription start site.

The best experiment to do would be a wet one - where you do "run off transcription" assays and hybridize the nascent transcripts to an array or sequence them to figure out all the possible start sites near the promoters being queried - good luck 

Basically the answer to your question depends very much on the deepness (and aim) of the analysis you want to perform. In general, there is no one rule that you could apply to all genes. The most difficult problem (already mentioned above but I'd like to emphasize it) are the mupltiple transcription start sites (TSSs) that are typical for higher eukaryotes such as human and mouse. Even if you consider all of them as separate starts (most will overlap but some will be several kb up- or downstream) the problem doesn't become easier because of distant regulatory sequences. But fortunately for some analyses you can just ignore the existance of the enhancers and other distant regulatory sequences.

We analysed the occurrences of the known TFBSs in human and mouse and in average (!) the great majority of them lie in the 1 kb upstream. "In average" is an important remark. Nonetheless, if you take 1 kb for rough analysis it will be ok. If you take more, depending on what you're looking for, you can get too many false positives. For instance, looking for TF binding sites or motifs that could serve as such, you will get greater false positive rates in longer sequences, and the rate of falseness will grow with the distance to the TSS (the probability to find a functional TFBS is higher closer to the TSS.

So my suggestion would be to take 1 kb upstream of each TSS. You have to think how to merge the overlapping sequences. The data about the TSS you can find for instance in DB TSS (http://dbtss.hgc.jp/).