In my annual review post I promised to complete some papers in progress... so here is the second working paper for 2013. This is actually the revised version we are submitting after getting a revise and resubmit so the job was to revise rather than complete the paper. An earlier version of the paper is actually already online titled "From Correlation to Granger Causality". That was a version I prepared for a conference in 2011. This version is coauthored with Kerstin Enflo. Added to this version there are unit root tests with structural breaks and cointegration analysis using the Johansen procedure with and without structural breaks. Also all the test statistics have been recomputed and the discussion of Granger causality in general is reduced.
This is one of a couple of studies we are doing on energy-GDP causality. The other is a meta-analysis, the first paper of which will be coming soon. There is a huge literature on causality between energy use and economic output. There are literally hundreds of papers. Most studies use short time series, in which it is hard to come to strong conclusions about what is going on. And it would seem that a lot of the results are actually spurious as the true significance levels of the tests are much lower (i.e. bigger p value) than the researchers think. There are quite a few studies that use panel data - i.e. combine time series data for many countries - in order to get a bigger data set and higher statistical power. There is one published meta-analysis that combines the results of many studies and there is one published paper with a much longer time series. Our new paper also adopts the strategy of using a very long time series.
We use Swedish data from 1850 to 2000 on GDP, "gross output", capital, labor, energy, quality adjusted energy, energy prices, and oil prices. We show that Granger causality techniques are very sensitive to variable definition, choice of additional variables in the model, and sample periods. All of the following appear to make a finding that energy causes growth more likely: using multivariate models, defining variables to better reflect their theoretical definition, using larger samples, and including appropriate structural breaks. We find causality from energy to GDP using a multivariate model for the full sample period. But in data for just the last 50 years this finding is reversed.
It is possible that the relationship between energy and growth has changed over time. On the other hand, it seems likely to me that causality in fact runs both ways, but what is actually detected depends on the size of the shocks to the two variables in the sample observed. One thing that has changed is that the cost share of energy has declined over time. This means that the output elasticity of energy has declined and we would expect the effect of changes in energy use to, therefore, have less impact on output. These smaller effects are harder to detect especially as the sample size gets smaller and reduces statistical power. On the other hand, the effect of output on energy use shouldn't have declined.
The paradox of the energy-GDP relationship is that the more important that energy is in terms of being essential in production the less the quantity of energy used will change in response to price shocks. However, the inability to substitute for energy will cause demand for other inputs and goods and services to decline by more. So energy prices can have a bigger effect on the economy but the quantity of energy to GDP relationship is weaker. We do find that energy prices have a significant causal impact on both energy use and output both in the longer and shorter (more recent) sample.
Another finding is that allowing for structural breaks reduces the signfiicance level of the Toda-Yamamoto Granger causality tests but makes it easier to find cointegration in our VAR models. This is paradoxical as cointegration implies that there must be Granger causality in at least one direction.
This is one of a couple of studies we are doing on energy-GDP causality. The other is a meta-analysis, the first paper of which will be coming soon. There is a huge literature on causality between energy use and economic output. There are literally hundreds of papers. Most studies use short time series, in which it is hard to come to strong conclusions about what is going on. And it would seem that a lot of the results are actually spurious as the true significance levels of the tests are much lower (i.e. bigger p value) than the researchers think. There are quite a few studies that use panel data - i.e. combine time series data for many countries - in order to get a bigger data set and higher statistical power. There is one published meta-analysis that combines the results of many studies and there is one published paper with a much longer time series. Our new paper also adopts the strategy of using a very long time series.
We use Swedish data from 1850 to 2000 on GDP, "gross output", capital, labor, energy, quality adjusted energy, energy prices, and oil prices. We show that Granger causality techniques are very sensitive to variable definition, choice of additional variables in the model, and sample periods. All of the following appear to make a finding that energy causes growth more likely: using multivariate models, defining variables to better reflect their theoretical definition, using larger samples, and including appropriate structural breaks. We find causality from energy to GDP using a multivariate model for the full sample period. But in data for just the last 50 years this finding is reversed.
It is possible that the relationship between energy and growth has changed over time. On the other hand, it seems likely to me that causality in fact runs both ways, but what is actually detected depends on the size of the shocks to the two variables in the sample observed. One thing that has changed is that the cost share of energy has declined over time. This means that the output elasticity of energy has declined and we would expect the effect of changes in energy use to, therefore, have less impact on output. These smaller effects are harder to detect especially as the sample size gets smaller and reduces statistical power. On the other hand, the effect of output on energy use shouldn't have declined.
The paradox of the energy-GDP relationship is that the more important that energy is in terms of being essential in production the less the quantity of energy used will change in response to price shocks. However, the inability to substitute for energy will cause demand for other inputs and goods and services to decline by more. So energy prices can have a bigger effect on the economy but the quantity of energy to GDP relationship is weaker. We do find that energy prices have a significant causal impact on both energy use and output both in the longer and shorter (more recent) sample.
Another finding is that allowing for structural breaks reduces the signfiicance level of the Toda-Yamamoto Granger causality tests but makes it easier to find cointegration in our VAR models. This is paradoxical as cointegration implies that there must be Granger causality in at least one direction.
David - Nice paper. The Giles and Godwin paper that you reference is now published:
ReplyDeleteGiles, D.E.A. & R.T. Godwin, 2012, “Testing for Multivariate Cointegration in the Presence of Structural Breaks: p Values and Critical Values”, Applied Economics Letters, 19, 1561-1565
Hi Dave - Thanks for that, your help was much appreciated! We will include this reference in the final published version that we resubmitted after the revise and resubmit we got.
ReplyDelete