You might want to consider adding more lags, but not for the reason you specified (i.e. to obtain the short-run relations), but simply to eliminate potential bias in your estimates.  In general, under-specifying the lag length for the VECM increases the finite sample bias.

Gonzalo, J. (1994). Five alternative methods of estimating long-run equilibrium relationships. Journal of econometrics, 60(1), 203-233.

You might want to consider using a LR or AIC criterion.  You did not state how you conducted your lag specification tests, but most software these days automate this process. In addition, rather than testing up (i.e going from lag 1 to lag 2) maybe you should test down (i.e. going from lag 4 to lag 3).

Also see the link to the STATA manual below.

http://www.stata.com/manuals13/tsvecintro.pdf

Lag selection criteria are the not the last words for selecting the optimal lag length in VAR models. Depending on the number observations, one should start with a fairly large number of lag and pare it down with 'lag exclusion test'. 

Ok, thanks. I generally go with SBIC, since I'm told it's the most parsimoneous. There seems to be different schools in this. Some say you should choose fewest possible lags, others say more is better.

i just did a cointegration test at lag 1(1) and my variables did not cointegrate, but when I increased to lag 1(3) cointegration was established is that still in order?