Firstly it may be possible to use standard deviations from the baseline study data,  or secondly, so long as you have outcome means and sample size, you can use a trial and error approach to find the best standard deviations that give the published p value for example using ttesti in stata. 

After that, I would impute standard deviations from studies of similar sample size. It may be good practice to undertake a sensitivity analysis including and excluding studies with imputed data to see if your findings change.

I hope this is of help.  

WOuld you post which study stated it is safe to impute data for meta-analysis? 

If your data is large enough, perhaps a multiple chain equation is a good try. The good (and bad) of multiple imputation is that usually it may simulate many datasets with imputed data. Good because you may run many sensitivity analysis based of the data you have, bad because there is an additional source of variance. 

If you are wondering to impute the outcome data, which is the core of trial interpretation, I wounder why the original author did not provide the SDs in first place. Perhaps he/she did not have symetric data or not reliable data. In that case you may be pooling stuffing together that should not be pooled. This issue always concerns me. 

Hi Pedro

You make some valid points about caution in imputing values, which is why I would check findings against those excluding imputed data for validity.

BW

Chris

I think the Cochrane manual has some comments on this sort of thing, if you want to take a look.