@article{Horridge:333002,
      recid = {333002},
      author = {Horridge, Mark},
      title = {Aggregating CES elasticities for CGE models},
      address = {2018},
      year = {2018},
      note = {Presented at the 21st Annual Conference on Global Economic  Analysis, Cartagena, Colombia},
      abstract = {It is expensive to prepare a database for a large CGE  model -- so we should plan that the database will be used  for many (perhaps unanticipated) purposes. For this and  other reasons it is wise to construct a database with as  much regional and sectoral disaggregation as possible. But  a model with too many sectors and regions is annoyingly  slow to solve. Thus it is common practice to aggregate a  large 'master' database before use. Can we make  results  computed with aggregated data more closely resemble those  from disaggregate data? A CGE database consists mainly of  matrices of flow values, and elasticities (mostly  pertaining to CES nests). Having defined many-to-one  mappings from the (many) old sectors/regions to the (fewer)  new sectors/regions, it is easy to aggregate the flows  matrices by simply adding. Aggregated elasticities are  usually constructed as weighted averages of the  disaggregate elasticities -- using as weights the flow  values associated with each elasticity (that is, the total  cost of inputs to that nest). However, we present examples  where this simple weighted averaging yields odd results. We  focus on the case where a number of users (or nests) each  combine (using the CES) the same set of inputs (but with  different cost shares). Seeking a better method of  averaging elasticities, we propose as a criterion that  aggregated CES elasticities imply a local own-price  substitution response that is close to the average of the  own-price responses in the corresponding disaggregated  nests. We see that elasticity aggregation bias arises from  aggregating nests, rather than aggregating inputs to nests.  We develop some simple formulae, which imply that the  aggregate elasticity should be K (0<K<1) times the  disaggregate elasticity where K is smaller as the variance  of disaggregate cost shares is larger. We show how such  formulae could be applied to the GTAP model.},
      url = {http://ageconsearch.umn.edu/record/333002},
}