![Sample our Economics, Finance,Business & Industry journals, sign in here to start your access, latest two full volumes FREE to you for 14 days]({"type":"image" , "src" : "/sda/5931/TOC-Subject-Sample-2021-Economics-Finance-Business-Industry.jpg"})
Abstract
This article analyses the effects of the real value of the dollar on investment in US domestic manufacturing using aggregate data for 1973–2004. Econometric estimation shows a negative effect that is much larger than has been found in any previous study. The exchange rate affects investment mainly, although not exclusively, through the channel of financial or liquidity constraints, rather than by affecting the desired stock of capital. Counterfactual simulations show that US manufacturing investment would have been 61% higher and the capital stock would have been 17% higher in 2004 if the dollar had not appreciated after 1995.
Acknowledgements
The author would like to thank Bob Chirinko, Steve Fazzari, Ray Mataloni, Bill Noellert, Arslan Razmi, Rob Scott and Christian Weller, as well as participants in the 2004 Post Keynesian Conference at the University of Missouri, Kansas City, for comments on earlier versions. Amos Golan and Maria Heracleous provided econometric advice and two anonymous referees made exceptionally helpful suggestions. The usual disclaimers apply.
Notes
1. US manufacturing employment fell by 3.3 million jobs, from 17.6 million in 1998 to 14.2 million in 2005 (US Congress, Joint Economic Committee, Citation2006, p. 14). As of 2004, the US trade deficit for goods had reached US$650.9 billion (on a Census basis), of which US $550.8 or 84.6% was accounted for by manufactured goods (US Census Bureau, Citation2005, Part B, Exhibit 14).
2. Manufactures represented 85.6% of US ‘domestic exports’ (excluding re‐exports) and 79.9% of US imports of goods in 2004 (US Census Bureau, Citation2005, Part B, Exhibit 14).
3. Unless otherwise stated, ‘exchange rate’ in this article always means the trade‐weighted, price‐adjusted, real value of the US dollar.
4. Goldberg (Citation1993) and Campa and Goldberg (Citation1995) included both the level and volatility of the real exchange rate in their models; since the estimated effects of volatility were generally small or insignificant, the discussion is restricted here to the results for the level.
5. Goldberg (Citation1993, pp. 579–580) stated that she used lags, but did not report exact lag lengths.
6. The authors converted their data to growth factors (ratios of current/lagged values) for stationarity reasons, and used lagged exchange rates to avoid endogeneity problems. They also controlled for accelerator effects and real interest rates.
7. Campa & Goldberg (Citation1999) dropped exchange rate volatility, measured other variables differently, and used a different weighting specification. Three‐stage least squares (3SLS) was used to control for endogeneity of the mark‐up rate, which interacted with the exchange rate.
8. The signs of the exchange rate effects were reversed in this article because the exchange rate was defined as the home currency price of foreign exchange. In the following discussion, the reported results are translated into the equivalent effects of changes in the value of the dollar.
9. Campa & Goldberg (Citation1999) included a measure of the total import share of domestic consumption in their equation for the mark‐up rate, but not in their investment equation.
10. Campa & Goldberg (Citation1999) also applied the Beveridge & Nelson (Citation1981) filtering procedure to estimate the ‘permanent’ and ‘transitory’ components of exchange rate changes and used only the permanent component in their regression models. The authors did not report any sensitivity tests for how this procedure affected their results.
11. The user cost of capital is designed to measure the theoretical rental cost of a unit of capital in a neoclassical production function (see Hall & Jorgenson, Citation1967). See below for the precise definition used in this article and comparisons with other definitions in the literature.
12. In this vein, Lavoie et al. (Citation2004) compare neo‐Marxian and neo‐Kaleckian models of investment using aggregate Canadian time‐series data. They focus on the dynamic adjustments of capital stock growth rates and capacity utilisation rates to long‐run ‘normal’ levels.
13. This view originated in the heterodox tradition of Kalecki (Citation1937), Steindl (Citation1952) and Minsky (Citation1986), but was also found in mainstream Keynesian investment models of the 1950s (for example, Meyer & Kuh, Citation1957); it was later revived in a neoclassical framework in models of credit rationing with asymmetric information (for example, Stiglitz & Weiss, Citation1981). See Fazzari et al. (Citation1988) for a survey.
14. Many (although not all) of these studies assume that the effects of profits (especially when measured by cash flow) reflect financial or liquidity constraints based on asymmetric information. Chirinko & Schaller (Citation1995, p. 529) find evidence in support of this view in their study of Canadian firm‐level data, in which the firms that have greater ‘difficulty in communicating private information to outsiders’ also face stronger liquidity constraints.
15. Crotty and Goldstein (Citation1992) included a foreign relative cost variable in their theoretical model, but then used the import penetration ratio as a proxy in their empirical work.
16. Some of Campa & Goldberg’s trade‐weighted exchange rate measures were interacted with the price‐cost margin (in their Citation1995 article) or profit mark‐up rate (in their 1999 article), for reasons implied by their theoretical models, but these profit variables were not included as separate regressors in their investment equations.
17. This effect is greater, the greater are the ‘Marshall‐Lerner’ elasticities of demand for exports and imports, although strictly speaking the Marshall‐Lerner condition for a devaluation to improve the trade balance is not relevant to the effect on investment per se.
18. The traditional literature on investment functions effectively imposes the restriction that all the β 4i coefficients in Equation (Equation1) are zero. The literature on the effects of exchange rates on investment can be viewed as imposing the restriction that the β 5i coefficients are all zero. If any of these restrictions is incorrect, then the restricted estimates of the investment function could be biased.
19. Chirinko & Schaller (Citation1995) and Chirinko et al. (Citation1999) emphasise that variables that affect the desired capital stock should enter the investment function in differences (or percentage changes) for this reason, whereas variables that reflect short‐run financial constraints should enter the investment function in levels. By this logic, if either the exchange rate, cash flow ratio, or net profit rate affects the desired capital stock, that variable should be expressed in difference form, whereas if it affects financial constraints it should be expressed in levels.
20. A disadvantage of this approach is that it imposes the same long‐run dynamics for all the independent variables, a restriction which is not found to be true in studies that allow different lag lengths for different variables. For example, Chirinko et al. (Citation1999) have up to six annual lags on some variables, and four lags on most. However, most of these studies gain extra degrees of freedom by using either firm‐ or industry‐level panel data.
21. Quarterly data for investment, capital stock, depreciation and profits in US manufacturing are not available for recent years anyway. The quarterly data series for industry‐level investment used by Goldberg (Citation1993) are no longer produced.
22. More detailed definitions and sources for all variables are given in the Appendix below.
23. The series were spliced by multiplying the ratio of the NAICS measures for 1999 to 1998 times the SIC‐based measure for 1998 to obtain the value for 1999, and then applying the annual percentage changes in the NAICS series to the lagged spliced series to create data for 2000–2004. Although the levels of net profits and cash flow were quite different in the two data‐sets (for the overlapping years), the ratios of these variables to the lagged capital stocks were much more similar, thus giving us greater confidence in the spliced measures created using the ratios.
24. The SIC‐based estimates are not reported here, but are available from the author on request.
25. Standard definitions of UC also make adjustments for the tax treatment of investment, but tax policy variables are omitted here because the focus is on exchange rates rather than fiscal policy. In Chirinko et al. (Citation1999, p. 57), the entire expression on the right‐hand side of our Equation (Equation4) is multiplied by the ratio (1 − xt − zt )/(1 − τt ), where x is the investment tax credit, z is the discounted value of depreciation allowances for tax purposes, and τ is the tax rate.
26. Details on the unit root tests along with descriptive statistics, correlation coefficients and Granger‐causality tests are presented in an unpublished statistical Appendix, which is available from the author on request.
27. This statement includes the percentage change in the user cost, which is the variable actually included in the estimated investment functions, although stationarity is rejected at the 5% level for the user cost measured in levels according to this test.
28. It should be noted that this procedure may result in the inclusion of some variables that appear insignificant according to more conventional t‐tests or F‐tests.
29. The Durbin‐Watson (DW) test is not valid with a lagged dependent variable and hence is not reported in Tables –. The Breusch‐Godfrey Lagrange Multiplier (LM) test (Godfrey, Citation1988) is used instead because it is valid with a lagged dependent variable and also allows us to test the null hypothesis of no serial correlation for any number of lagged residuals.
30. The first difference of the real interest rate was used partly for empirical reasons (using the level instead of the change produced an insignificant coefficient), partly for theoretical reasons (because the investment function literature implies that the cost‐of‐capital variable should enter the equation in rate of change form) and partly for consistency with some of the previous literature on exchange rate effects (especially Campa & Goldberg, Citation1999).
31. However, the Breusch‐Godfrey LM test with two lags has a p‐value of 0.091 in Equation (Equation1.3). This suggests serial correlation of the residuals, which would invalidate hypothesis tests using this equation.
32. One cannot reject the null hypothesis that the net profit rate (in levels) has a unit root using an ADF test with either an intercept only or an intercept plus a trend; only the first difference of the net profit rate is stationary according to this test. ADF tests also show that the cash flow ratio has a unit root in levels with an intercept only, although the null of a unit root can be rejected at the 5% level with an intercept and a trend. The alternative test due to Kwiatkowski et al. (Citation1992) shows that the cash flow ratio is stationary in levels with an intercept alone, while the results for the net profit rate are sensitive to whether a trend is included and the significance level used.
33. The current real dollar index was not significant in these equations and hence was omitted.
34. The current year changes in user cost, net profits and cash flow were insignificant (according to t‐tests) and redundant (according to LR tests) and hence were omitted.
35. As noted by Chirinko et al. (Citation1999, p. 61), with investment measured relative to the lagged capital stock and the user cost expressed as a percentage change, the elasticity of the long‐run desired capital stock with respect to the user cost is the sum of the coefficients on the distributed lags of ΔUCt−i/UC t−i−1. In the present model, although there is only one lag of this variable (i = 1), the long‐run coefficient can be calculated from Equation (Equation2) with m = 3 and β 30 = 0.
36. Equations similar to those shown in Tables – were also run with the real dollar index measured in first differences as a sensitivity test; the differenced real dollar index (either current or lagged) was not significant in any of these regressions. Following the logic of Chirinko & Schaller (Citation1995) and Chirinko et al. (Citation1999), the fact that the real dollar index is significant in levels rather than in differences suggests that its effects are short run rather than long run.
37. These difficulties with the profit equation are not surprising, since the only variable that was significant at the 5% level in Clarida’s model of US manufacturing profits was the exchange rate (Clarida, Citation1997, p. 182).
38. Clarida (1997) uses demand for domestic goods instead of GDP because the former excludes exports. The same measure was constructed and tested in our profit equations, and it was found that the measure yielded very similar results to the GDP growth rate but with a slightly worse fit.
39. Although the Breusch‐Pagan LM tests thus show no significant correlation of the OLS residuals, as a sensitivity test 3SLS was used to estimate the same four pairs of equations for investment and profits shown in Table . All the exogenous and predetermined variables in the two equations were used as instruments. The coefficient estimates were very close to those found using OLS, confirming that there is negligible simultaneity bias in the OLS estimates.
40. Since these two equations are simulated as a system, the 3SLS estimates discussed in the previous note are used rather than the OLS estimates shown in Tables and , but this makes little difference to the results.
41. The baseline simulation represents the dynamic forecast of the complete five‐equation model based on the actual values of the exogenous variables (including the real dollar index) and the forecast lagged values of the endogenous variables. These simulations assume that the other exogenous variables would have remained the same even if the dollar had stayed lower. Actual real investment in 2004 was US$165.8 billion.
42. These are ‘high end’ forecasts of the effects of the appreciation of the dollar, because Equation (3.4) has a relatively high estimated coefficient on the dollar and includes an indirect effect via lagged changes in the cash flow ratio. Using the investment equations with a lower exchange rate elasticity and no cash flow effect, the simulated effects are roughly half to two‐thirds as large.
43. Another difference is that the change in the real interest rate was generally insignificant using the nominal investment rate, although the percentage change in the user cost was significant. The coefficients on the lagged dependent variable were smaller (around 0.5 to 0.6) in the alternative estimates. The short‐run coefficients on the real dollar index were slightly larger in absolute value (around −0.04 to −0.06), but the long‐run coefficients (and elasticities) were smaller because of the smaller coefficient on the lagged dependent variable. The residuals were more problematic in these alternative estimates, requiring the use of outlier dummies in most specifications, most likely as a result of inflationary shocks.