Natural Gas and U.S. Industrial Production: A Closer Look at Four Industries

We consider the relationship between natural gas prices and production in the U.S. resins, agricultural chemicals, cement, and aluminium industries. Overall, our analysis using various tests and regressions does not allow for generalizations about the association between natural gas prices and production across these four energy-intensive industries. Rather, the relationships between natural gas and production appear to be driven by the particular institutional details of each industry. Introduction How important are natural gas prices to production in U.S. energy-intensive industries? How beneficial have the recent increases in American natural gas supply been to these industries? If approved, will exports of U.S. liquefied natural gas (LNG) negatively impact the industries which use this fuel? These are important questions that have received much attention since the beginning of the U.S. shale gas revolution. But there is little empirical work on which to base any answers. In this paper we consider the importance of natural gas prices for output in four specific U.S. energy-intensive industries: resins, agricultural chemicals, cement, and aluminum. Our primary goal is to shed some light on the questions above using data that includes the time period after shale gas became an important component of U.S. natural gas supply. We also seek to understand and account for the results in terms of the key features specific to each of these industries. ∗The analysis and conclusions expressed here are those of the authors and not necessarily those of the U.S. Energy Information Administration. †Katherine Calais provided excellent research assistance. We have also benefitted from the comments and suggestions of Joe Benneche, Stephen Brown, John Conti, Peter Gross, Fred Joutz, Thomas Lee, Kay Smith, Russell Tarver, and members of the American Statistical Association Advisory Committee on Energy Statistics. Vipin Arora and Elizabeth Sendich | U.S. Energy Information Administration | This paper is released to encourage discussion and critical comment. The analysis and conclusions expressed here are those of the authors and not necessarily those of the U.S. Energy Information Administration. 1 Many recent studies use both macroeconomic and input-output models to evaluate the economic importance of natural gas for specific U.S. industries, but empirical work is limited.1 In one example, Kliesen (2006) finds that natural gas prices have historically been unable to predict industrial production in U.S. manufacturing industries. This paper was completed before the large increases in natural gas production due to shale gas. In additional research completed before this time, Costello et al. (2006) analyze the interaction between industrial natural gas prices, natural gas consumption, and industrial sector activity. They conclude that industrial sector firms respond to relative energy prices, including natural gas prices. Weber (2012) considers the economic impacts of the natural gas market after the shale gas boom, but on a regional level.2 Arora and Lieskovsky (forthcoming) also consider the period after 2008, although they focus on aggregate impacts of the natural gas market. Resins, agricultural chemicals, cement, and aluminum are chosen in the analysis because each is an energy-intensive industry that vary in how much they use natural gas. Both resins and agricultural chemicals are heavy users of natural gas. Aluminum is more of an intermediate natural gas user, both by volume and as a percentage of total fuel use/mix. Cement uses very little natural gas volume relative to the other three, and has a broad fuel mix. By studying these four industries we are able to get an overview of how different parts of the manufacturing sector may respond to natural gas prices. With these various considerations in mind, we begin our analysis by conducting a variety of different Granger causality tests of natural gas prices on each of the four industrial production indices. For each industrial production series we vary the tests by changing the natural gas price, altering lag lengths, and modifying the specifications to include the real oil price and total industrial production. This results in a total of 144 tests each for resins, agricultural chemicals, cement, and aluminum. In general the results are inconclusive, and with the exception of resins only a small fraction of tests are consistent with the ability of natural gas prices to predict industrial production: 25% for resins, 10.4% for agricultural chemicals, 7.6% for cement, and 12.5% for aluminum. The factors behind the inconclusiveness of these results may be due to well-known limitations of Granger causality tests, namely the omission of important variables and uncertainty over the proper lag length. However, in this case they may also be due to the assumptions of a linear relationship between natural gas prices and industrial production or that the estimated relationships between these variables have been constant over time. The latter explanation is a particularly likely culprit in light of recent events in U.S. natural gas markets. Examples of studies focused on the economy-wide and industry-specific impacts of natural gas include ACC (2011), ACC (2012), Arora (2013), CitiGPS (2012), CRA (2013), ICF (2012), IHS (2011), IHS (2012), IHS (2013), McKinsey Global Institute (2013), NERA (2012), Peterson Institute (2014), PWC (2011), PWC (2012b), PWC (2012a), PWC (2013), and Kinnaman (2010) and citations therein. In research before deregulation of the U.S. natural gas market, Leone (1982) claims the impact of natural gas price increases on the northeast regional economy are at least offset by gains due to greater revenue for natural gas producers. Vipin Arora and Elizabeth Sendich | U.S. Energy Information Administration | This paper is released to encourage discussion and critical comment. The analysis and conclusions expressed here are those of the authors and not necessarily those of the U.S. Energy Information Administration. 2 These limitations lead us to use non-parametric methods, specifically locally weighted regression (LOESS) to understand the relationship between natural gas prices and production in our chosen industries. Our primary specifications have three variables (industrial production, natural gas prices, and total industrial production), and consider three and six month lags of natural gas prices. The results from regressing each specific industrial production index on the other two variables vary between industries. Both resins and aluminum show an inverse association between natural gas prices and production, while the results for agricultural chemicals and cement indicate an association in the same direction. Overall, neither the Granger causality tests nor the non-parametric regressions allow for generalizations about the relationship between natural gas prices and production in these four energy-intensive industries. Data and descriptive statistics This section outlines the data series used in subsequent analysis, including the sources, time periods under consideration, and any modifications. It then describes the modified series in more detail, and presents correlations between industrial production and natural gas prices. Data We use nine different monthly series in our analysis. There are four industry-specific production indices and a total U.S. industrial production index. The four specific indices correspond to resins [NAICS 3252], agricultural chemicals [NAICS 3253], cement [NAICS 3273], and aluminum [NAICS 3313] sectors. The total industrial production index includes the U.S. manufacturing sector as well mining and utilities. Each of the industrial production indices are taken from the Board of Governors of the Federal Reserve over the period 1990M01-2013M04, and are seasonally unadjusted.3 Three different natural gas prices and one oil price are also utilized. The nominal price of natural gas at Henry Hub is taken from the Federal Reserve Economic Data (FRED) system at the Federal Reserve Bank of St. Louis from 1993M11-2013M04.4 We deflate this by the seasonally unadjusted Producer Price Index (PPI) for fuels [WPU05], which comes from the U.S. Bureau of Labor Statistics (BLS).5 The seasonally unadjusted nominal U.S. industrial price for natural gas is obtained from the U.S. Energy Information Administration (EIA) for the period 2001M01-2013M04.6 This price and the composite refiner acquisition costs of crude oil (also from the EIA) are both deflated in the same manner as the Henry Hub price. The final natural gas price we use is the PPI for the industrial natural See http://www.federalreserve.gov/releases/g17/download.htm. See http://research.stlouisfed.org/fred2/series/GASPRICE/. See http://data.bls.gov/timeseries/WPU05. See http://www.eia.gov/dnav/ng/ng pri sum a EPG0 PIN DMcf m.htm. Vipin Arora and Elizabeth Sendich | U.S. Energy Information Administration | This paper is released to encourage discussion and critical comment. The analysis and conclusions expressed here are those of the authors and not necessarily those of the U.S. Energy Information Administration. 3 gas price [WPU0553] from the BLS, which is available from 1990M01-2013M04.7 Each of the nine series appear to be trending over the sample period.8 This sample period is 1990M01-2013M04 for the industrial production indices, PPI natural gas price, and real refiner acquisition costs of crude oil; 1993M11-2013M104 for the real Henry Hub price; and 2001M01-2013M04 for the real industrial natural gas price. Each of the series save the real oil price also have seasonal movements. In our analysis we work with the annual percent change in each variable to remove both the trend and seasonality in the series. Descriptive statistics The annual percent changes in the specific industrial production series and each natural gas price are plotted in Figure 1. Both sets of series move together in general, but have pronounced differences. Especially notable is that before 2003, there are large changes in the real Henry Hub price which are not necessarily reflected in either the real industrial or PPI industrial prices. But the price series tend to move similarly after 2003. In terms of industrial production indices, before 2009 the series move in a similar fashion. After this time resins and agricultural chemicals show large variations, and aluminium and agricultural chemicals production have movements which are relatively synchronized. Some of these differences and similarities are reflected in the descriptive statistics in panels (a) and (b) of Table 1. Panel (a) shows that the average growth rate in resins production (%∆RS) has a mean value close to zero over the sample, and a standard deviation of 9.2%. Resins also has strong negative skewness. The kurtosis for agricultural chemicals production (%∆AC) shows that the observed values of changes in this variable have a distribution which differs from the normal (kurtosis = 0), at 12.39. Changes in this series have positive skew, indicating that values fall below the mean more often than above. This mean value is slightly negative, at -0.4%, and the standard deviation is just over 8%. The annual percent changes in cement production (%∆CM ) are also close to zero, and the series is as volatile as resins. Changes in cement are also highly negatively skewed, implying the values in the sample tend to fall above the mean, with fewer below. And the kurtosis value of 7.50 shows that the observed values of changes in organic chemicals production have a distribution which also differs from the normal. Changes in aluminum production (%∆AL) have the lowest standard deviation of the three series, at 1% over the sample period. The mean is a slightly positive 0.6%, with positive skewness of 0.166 and kurtosis similar to organic chemicals at 4.83. Panel (b) of Table 1 shows that changes in the natural gas price series are much more volatile than the corresponding changes in the industrial production indices. Consistent with Figure 1, changes in real See http://data.bls.gov/timeseries/WPU0553?include graphs=false&output type=column&years option=all years. In each case unit root tests indicate I(1) series: the Augmented Dickey-Fuller (ADF) test cannot reject the null of a unit root, and the KPSS test rejects the null of stationarity. Vipin Arora and Elizabeth Sendich | U.S. Energy Information Administration | This paper is released to encourage discussion and critical comment. The analysis and conclusions expressed here are those of the authors and not necessarily those of the U.S. Energy Information Administration. 4 Henry Hub prices (%∆HH) are the most volatile, with a standard deviation of over 31%, the PPI for industrial natural gas (%∆PP ) has a standard deviation of 21%, while changes in the real industrial natural gas price (%∆IN ) have a standard deviation just below 19%. The mean values of %∆HH and %∆IN are negative over their respective samples (-3.6% and -6.2%), while the mean value of %∆PP is positive (1.9%). All three natural gas price series are positively skewed, and their kurtosis values indicate that the observed values have a narrower distribution than the normal. In each of the price statistics, changes in the real Henry Hub price stand out from the other two series. Panel (c) of Table 1 begins to explore the relationship between changes in each industrial production index and each natural gas price through non-parametric correlations. We use non-parametric correlations because of the large changes which have occurred in U.S. natural gas markets over the past 25 years. Each row shows the Spearman correlation coefficient for changes in the respective industrial production index and changes in the three natural gas prices at lags of three and six months.9 These particular lags account for the fact that changes in natural gas prices take time to have an impact on industrial production. The results indicate that there is a negative relationship between growth in resins or aluminum production and growth in any of the natural gas prices. These are all statistically significant and hold at either lag of the natural gas price (with the exception of resins and aluminum production and three lags of the real Henry Hub price). Each of these two industries clearly have an association with past changes in natural gas prices. The same is not true for either cement or agricultural chemicals production. The correlations with cement are positive, some of which are statistically significant, while agricultural chemicals production does not have a statistically significant correlation with any of the lagged natural gas prices. Quantitative analysis In this section we further explore the relationship between changes in natural gas prices and changes in U.S. resins, agricultural chemicals, cement, and aluminum production. The section begins with standard Granger causality tests of each industrial production index on each natural gas price. Concerns about the stability of each price series, as well the assumption of linearity inherent in the tests leads us to move to non-parametric methods. The results of non-parametric regressions of the The Spearman correlation coefficient is calculated using the same formula as the standard Pearson correlation coefficient. The difference is that it is applied to ranked data and not each industrial production/natural gas price pair. To construct the correlation each of the observed values for changes in a specific industrial production index are arranged in ascending order, as are the observed values for a specific natural gas price. These ordered values are paired, and then used in calculation of the correlation coefficient using the standard Pearson formula. Vipin Arora and Elizabeth Sendich | U.S. Energy Information Administration | This paper is released to encourage discussion and critical comment. The analysis and conclusions expressed here are those of the authors and not necessarily those of the U.S. Energy Information Administration. 5 industrial production indices on the natural gas prices concludes the section. Granger causality tests Granger causality tests are a standard method used to explore the ability of past values in one series to predict another. Although we do not interpret these tests as inferring causation from one series to another, they do provide important information on the association between changes in natural gas prices and changes in industrial production indices. We use several specifications of these tests to account for other possible factors and estimation issues as well [see for example the discussions in Brown and Yucel (2008) and Joutz and Villar (2006)]. Technical details and specifications We begin by conducting an analysis similar to Kliesen (2006) and attempt to quantify the importance of changes in real natural gas prices for the respective industrial production indices. The methods we use are based on the literature on oil price shocks and economic activity [see Hamilton (1996) and Hooker (1996) for notable examples], specifically the use of bivariate Granger causality tests as in Cunado and Perez de Gracia (2003). We also extend the tests to three and four variables to control for the impact of total industrial production and the real oil price. Our initial bivariate specification tests for Granger causality from past annual percent changes in each natural gas price (∆xx, where xx = HH, IN , or PP ) to the annual percent change of each industrial production index (∆yy, where yy = OC, AC, or AL) are specified as: %∆yy,t = α0 + Σ k j=1α1j%∆yy,t−j + Σ k i=1α2i%∆xx,t−i + t (1) where t is a normally distributed error term with an expected value of zero. We test the hypothesis that α21 = α22 = ... = α2k = 0. Rejecting this hypothesis indicates that changes in the respective natural gas price Granger cause changes in the particular industrial production index. The lag length (k) is set at k = 3, 6, and 12 months. For the single-equation bivariate case specified above, each index is regressed against the three natural gas prices at three different lags, for a total of 9 specifications. The sample is also varied, from the beginning of the price data to 2013M04, and from the beginning of the price data through 2007M12. This gives a total of 18 Granger causality tests for the single-equation bivariate case of each industrial production index. These same tests are also repeated with the equation as part of a vector auto-regressive (VAR) system of equations, for a total of 36 Granger causality tests of each industrial production index in the two variable specification. We then extend the two-variable case to consider the annual percent changes in real oil prices Vipin Arora and Elizabeth Sendich | U.S. Energy Information Administration | This paper is released to encourage discussion and critical comment. The analysis and conclusions expressed here are those of the authors and not necessarily those of the U.S. Energy Information Administration. 6