Has Monetary Policy Accelerated Job Polarization Part 2

3 3.1 Data Occupational, industry, and investment data Aggregated monthly employment by industry is available directly from the U.S. Bureau of Labor Statistics (BLS). Historically comparable occupational employment data from the Current Population Survey (CPS) are available from the BLS only back to 1983. For the years 1967–1982, I compiled monthly occupational employment data from tables in the BLS’s monthly Employment and Earnings publication. Although the detailed occupational categories differ across the time period, the division of occupations into routine and nonroutine occupations can be constructed in a consistent way, another appealing feature of Autor et al.’s (2003) task-based framework. Data on investment prices are from DiCecio (2009) and are updated on a quarterly basis available at Federal Reserve Bank of St. Louis’s FRED database. Data on the equipment capital stock are constructed from Bureau of Economic Analysis (BEA) data, adjusted as suggested by Gordon (1990), following Krusell, Ohanian, Ríos-Rull and Violante (2000). 3.2 Measures of monetary policy To address the question of how monetary policy affects employment by occupation and industry, as a baseline I use monetary policy shocks identified by Romer and Romer (2004) and extended to 2008 by Coibion, Gorodnichenko, Kueng and Silvia (2012). Romer and Romer (2004) identify monetary policy shocks as changes to the Federal Funds target rate not predictable by the economic information in the Federal Reserve’s “Greenbook” forecasts. Specifically, their monetary policy shock series is given by the residuals of the following regression: ∆f fm = α + βf f bm + 2 X f mi + γi ∆y i=−1 + 2 X 2 X  λi f mi − ∆y f m−1,i ∆y  i=−1 φi π̃mi + i=−1 2 X θi (π̃mi − π̃m−1,i ) + ρũm0 + εm , (1) i=−1 where m indexes FOMC meeting dates, f f bm denotes the level of the Federal Funds target f mi denotes forecasts of real output growth, π̃mi denotes rate at the time of meeting m, ∆y forecasts of inflation, ũm0 denotes forecasts of current unemployment, and εm , the residual, is the monetary policy shock. The index i is the horizon of the forecast, and horizon i = −1 7 may be a true forecast or a realized value of the variable, depending on when the actual data were available. The shock series is from 1969–2008, which determines the sample period in all the estimates that follow. An independently identified shock series—as opposed to one identified in a structural vector autoregression (SVAR)—facilitates estimation of nonlinear impulse responses (in this case, asymmetric responses to contractionary versus expansionary shocks). The use of this series is not without drawbacks, however. For example, the series ends in 2008, but Romer and Romer’s (2004) methodology is not amenable to extension to the zero lower-bound period since it estimates the shock based on the change in the Fed Funds Rate, which exhibits essentially zero variation for almost a decade after 2008. Moreover, ignoring the post-2008 period necessitates dropping the largest decline in routine-task employment in the data. Therefore, as an alternative, I use a monetary policy shock identified using a hybrid of high-frequency and SVAR methods, as in Gertler and Karadi (2015). Their proxy SVAR method uses Fed Funds Futures as external instruments to identify a monetary policy shock. While pure high-frequency identification strategies are limited by data availability, the hybrid method of Gertler and Karadi (2015) facilitates the identification of shocks over a longer period. Specifically, they use high-frequency data to estimate the linear relationship between reduced-form and structural shocks in the VAR system. They then estimate the SVAR over a longer period under the assumption that the estimated relationship between reduced-form and structural shocks is valid for the entire period.8 I use their method to estimate a monetary policy shock over a long horizon beginning in 1973.9 Because the results from this exercise are broadly similar with the baseline linear results, I present the results from the hybrid identification approach in Appendix A.1. 4 Estimation methodology In the baseline estimates, I use Jordà’s (2005) method of local projections to estimate impulse response functions (IRFs) of occupational employment to a Romer and Romer (2004) monetary policy shock. The local projection method is extremely flexible and is partic8 This may not be the case, for example, in the pre- and post- Paul Volker chairmanship periods. The results presented here, however, are robust to the exclusion of pre-Volcker years. 9 The details of the construction of this shock measure are discussed in Gertler and Karadi (2015), so I omit a detailed description here. The start date of 1973 is dictated by the availability of Gilchrist and Zakrajsek’s (2012) measure of the excess bond premium. 8 ularly amenable to externally-identified shocks. Not only does it allow for simple linear estimates,10 but Jordà’s (2005) method also allows for a variety of nonlinearities, including sign-dependence. In what follows, I first estimate linear responses both as a baseline for comparison against the asymmetric IRFs estimated in what follows and to facilitate comparison across shock identification techniques. I then exploit the flexibility of local projections to compute signdependent impulse responses for two different specifications. 4.1 Linear estimates The baseline linear estimates are obtained as follows. For a shock ϵt and controls xt , and for horizons h = 1, . . . , H I estimate yt+h − yt = βh′ xt + γh ϵt + ut,h . |{z} (2) Linear The impulse response at horizon h to a one-unit shock, which corresponds to a 100 b.p. contractionary Romer and Romer (2004) shock, is simply given by the estimated coefficient γ̂h , and the width of the error bands is determined from the standard errors of these coefficient estimates.11 As a baseline, the vector xt includes one year of lags each of the dependent variable and shock, as well as a constant and linear time trend.12 Figure 3 displays the IRFs from (2) of total employment as well as employment in routineand nonroutine-task jobs, both in absolute terms and as a share of total non-farm employment. A contractionary shock reduces employment in routine occupations by almost 2 percent, and is statistically significant for more than four years after the shock. The IRF of nonroutine employment is insignificantly different from zero at nearly all horizons. This is reflected in the response of employment shares and total employment: a contractionary shock produces a decline in the share of employment in routine occupations of about 0.4 percentage points and a decline in total employment of about 1 percent, both of which are significant for more than four years. 10 That is, simple relative to estimates from a linear VAR, the IRFs for which are highly nonlinear functions of estimated parameters. 11 To account for serial correlation, the standard errors are adjusted as in Newey and West (1987). In addition, since the error bands for IRFs implicitly test the null hypothesis of zero effect, inference based on the standard errors of these coefficients is valid despite the presence of a generated regressor (see Pagan (1984)). 12 Section A.2 discusses alternative lag structures. 9 Figure 3: Impulse responses to a 100 b.p. contractionary Romer and Romer (2004) monetary policy shock. Estimated from Equation 2. Shaded areas are 90 percent confidence intervals. The results from this section highlight that nearly all the decline in employment after a contractionary monetary policy shock comes from the response of employment in routine jobs. The next section examines how these same variables respond when the linearity assumptions of (2) are relaxed. This is motivated by the findings of Jaimovich and Siu (2014), as well as the literature that finds significantly asymmetric effects of monetary policy discussed in Section 2. 4.2 Asymmetric estimates As discussed above, Romer and Romer’s (2004) shock series combined with Jordà’s (2005) flexible method of local projections allows for convenient estimation of asymmetric impulse responses. Equation 2, in keeping with the traditional VAR literature, imposes the assumption that impulse responses are symmetric; that is the responses to positive and negative shocks are identical except for their sign. Local projections facilitate a number of ways of 10 dispensing with this assumption. First, I estimate a specification that makes the simple assumption that positive and negative shocks are qualitatively different and allows for responses to differ depending on the sign of the shock. This simple specification is identical to that used in Tenreyro and Thwaites (2016) and Wong (2016). Specifically, I estimate for each horizon h = 1, . . . , H, yt+h − yt = βh′ xt + γh+ ϵ+ | {zt} Contractionary + γh− ϵ− | {zt} + ut,h , (3) Expansionary − where ϵ+ t := max{ϵt , 0} and ϵt := min{ϵt , 0}. To facilitate comparisons with Section 4.1, I include the same controls as in (2), and estimate the IRFs to a four-year horizon. This specification allows for straightforward tests of the null hypothesis of symmetry, i.e., H0 : γh+ = γh− for h = 1, . . . , H. One potential drawback of this specification, however, is that it imposes linearity on the impulse response conditional on the sign of the shock. This linearity assumption produces some problematic results discussed briefly below and in more detail in Appendix A.2. Therefore, as an alternative, I also estimate asymmetric IRFs using a specification that is quadratic in the shock. Specifically, for h = 1, . . . , H, I estimate yt+h − yt = βh′ xt + γ1,h ϵt + γ2,h ϵ2t + ut,h . (4) This specification dispenses with the assumption in (3) that shocks have linear effects conditional on their sign, instead allowing for more flexible asymmetries. One potential drawback of dropping this assumption involves the test for asymmetry. While the baseline specification in (3) restricts shocks to differ only with respect to their sign, the quadratic specification allows for other nonlinearities as well. For example, including the square of the shock allows for the impulse response to differ nontrivially (i.e., not simply a proportional scale change) with the size of the shock. Therefore, the test of the null hypothesis H0 : γ2,h = 0 for h = 1, . . . , H is somewhat easier to reject than the null hypotheses of symmetry in (3); that is, the null hypothesis is not that there are no asymmetries, but rather that there are no nonlinearities (up to second order), a null that is easier to reject.13 The impulse responses estimated from Equations 3 and 4 are displayed in the left and right columns of Figure 4, respectively. In response to a contractionary shock of 100 b.p., routine employment falls by 4 percent, while the point estimate of the response to an expansionary 13 Indeed, this is reflected in the point-wise asymmetry tests displayed in Figure 5. 11 Figure 4: Impulse responses to a 100 b.p. contractionary (red) and expansionary (blue) Romer and Romer (2004) monetary policy shock. Shaded areas are 90 percent confidence intervals. Left column: baseline asymmetric estimates from Equation 3. Right column: alternative quadratic estimates from Equation 4. 12 shock is at most 1 percent and is statistically insignificant. In contrast, the linear estimates implied a peak decrease (increase) of about 2 percent to a contractionary (expansionary) shock. The response of employment shares is even starker. A contractionary shock produces a full percentage-point decline in the share of employment made up by routine occupations—a response that lasts for the full estimated horizon—while an expansionary shock produces a peak effect of at most less than 0.3 percentage points and is insignificant. By comparison the peak linear response was less than 0.4 percent points.14 The point-wise tests for asymmetries in the impulse responses are displayed in Figure 5. Under both specifications, the null hypothesis of symmetry can be rejected at almost all horizons for all variables except for nonroutine employment, which displays relatively little asymmetry.15 The t-statistics for the quadratic estimates in the right column of Figure 5 are somewhat larger than those of of the baseline asymmetric specification; this is consistent with the previous observation that the quadratic specification allows for other nonlinearities than just asymmetry. The overall pattern of asymmetry is nevertheless quite similar across specifications, with larger degrees of asymmetry at longer horizons.16 5 Contribution of monetary policy to job polarization Thus far, the analysis has focused on the conditional responses of occupational employment to monetary policy shocks. Impulse responses, while they reveal the causal effects of monetary policy, do not by themselves, however, paint a full picture of the economic significance of shocks. Of equal interest is how much monetary policy shocks have contributed historically to variations in occupational employment. In a structural VAR, it is fairly straightforward to calculate the historical contribution of a shock to the time series of a variable of interest because the VAR is a structural econometric 14 It is worth noting, however, that the results from (3) imply that nominally “expansionary” shocks— i.e., shocks that produce declines in the Fed Funds Rate—are actually contractionary, at least for some periods for the variables included here. As discussed in Appendix A.2, this pattern is mitigated with the inclusion of more lags of the shock and exacerbated with the inclusion of fewer lags. As is evident from these specifications, as well as the quadratic specification, this anomaly is not a robust feature of the data. 15 Perhaps an unsurprising fact, given that neither impulse response is significantly different from zero at virtually any horizon. 16 An alternative approach is to estimate all horizons of the impulse response jointly, allowing for correlation of the errors across horizons, and to perform a joint significance test of the relevant coefficients. In such a test, the null hypothesis of symmetry is very easy to reject. The point-wise tests displayed here are biased against finding asymmetries and are consistent with the point-wise confidence intervals displayed in the IRFs throughout the paper. 13 Figure 5: Point-wise t-test against the null of symmetry for sign-dependent IRFs in Figure 4. Shaded areas indicate a 90% confidence region. Lines outside the shaded region indicate the null can be rejected at 10% significance. Left column: baseline asymmetric estimates from Equation 3. Right column: alternative quadratic estimates from Equation 4. 14 model. These very structural assumptions, however, are what lead to the inflexibility of VAR-based impulse responses relative to those estimated using local projections. On the other hand, because IRFs estimated by local projections impose few structural assumptions, the estimates provide little information on how a shock propagates through the economy. One approach is to use the one-period-ahead local projection (i.e., (2) or (3) estimated at h = 1)—which is equivalent to an autoregressive distributed lag (ADL) model—as the structural model; however, this effectively ignores the information contained in the other H − 1 projections. The approach I take here, based on forecast error variances (FEVs), exploits the full set of estimates from (2), (3), and (4), at the cost of not being able to discuss particular historical episodes. For notational simplicity, consider a more general series of local projections that nests the linear and both asymmetric specifications considered above: et + θh′ ϵet + ut,h , yt+h − yt = βeh′ x (5) et includes only lags of the dependent variable, and ϵet includes current and lagged where x − values of the monetary policy shock, potentially distinguishing between ϵ+ t and ϵt , or ϵt and ϵ2t , depending on the specification. The forecast error from (5) at time t, horizon h is given by et , ϵet ] . F Et,h ≡ yt+h − yt − E [yt+h − yt | x = ut,h .   2 The mean squared forecast error (MSFE) at horizon h is then given by M SF Eh = E F Et,h =  2  E ut,h , which can be estimated using regression residuals. To assess the importance of monetary policy shocks, I compare this MSFE to that of a specification in which monetary policy shocks are not included in the regression at all. The degree to which the specifications that do include non-zero monetary policy shocks improve on the forecast errors of those that do not include them can be interpreted as the contribution of monetary policy shocks to changes in the dependent variable at a given horizon. The most straightforward comparison is to consider the local projections that do not 15 include monetary policy shocks at all.17 That is et + et,h . yt+h − yt = βeh′ x (6) As above, the forecast errors are the estimated regression residuals and the MSFE at horizon   h is given by E e2t,h . Since the forecast errors are mean zero by construction, a direct comparison of the MSFEs from the two specifications at a given horizon gives the share of the forecast error variance explained by monetary policy shocks. Table 1 displays the share of the FEV of occupational employment variables at a given horizon that is due to the monetary policy shock, for both the linear and asymmetric case and for the alternative forecast discussed above. Even in the linear case, monetary policy shocks account for relatively large shares of the FEV for total employment and routine employment; they account for a negligible portion of the FEV for nonroutine employment, however. In the asymmetric cases, the role of monetary policy shocks is even larger, accounting for 40 percent of the variance in the share of employment made up by routine jobs over a two- to three year horizon. That monetary policy shocks can explain such a large share of the short-run fluctuations in routine employment is surprising. Together with the fact that the effects on routine employment of contractionary shocks displayed in Figure 4 are so persistent, this observation raises the question of what causes monetary policy shocks to have such large and long-lasting effects on routine employment. Two different mechanisms are discussed in the next section. 6 Two possible mechanisms This section discusses two possible channels through which monetary policy may have a disproportionate effect on routine employment. This discussion is by no means meant to be exhaustive, but it does highlight that the driving force of short- to medium-run changes in routine employment is very different from the long-run technological forces emphasized in the job polarization literature. The first possibility I consider is that monetary policy shocks might lead indirectly to 17 I also considered a slightly different comparison that takes into account the existence of monetary policy fh , but assumes that monetary policy shocks are always at their shocks in estimating the coefficient vector β mean value. This comparison is straightforward in the linear case, since the shock series is mean zero, but the asymmetric estimates introduce some small complications because there are terms with non-zero conditional means that must be accounted for. The details and results are described in Appendix A.3. The implied reductions in the FEVs are virtually identical to those from (6). 16 changes in routine employment by affecting the price of capital goods, which are a substitute in production for routine employment. If, for example, a contractionary monetary policy shock leads to a decline in the price of capital goods, substitution of new capital for routine labor could be driving the results in the previous sections. As discussed in Section 1, the trend decline in the price of equipment capital, reflecting technological progress, is widely thought to be an important driver of the long-run decline in routine employment. This “price-of-investment” mechanism is therefore a natural candidate to examine in the context of monetary policy. Another possibility is a simple industry composition effect. Monetary policy might disproportionately affect routine employment because it has stronger effects on those industries in which routine employment is concentrated. I find no evidence that the “price-of-investment” mechanism is driving the results from the previous section. A contractionary shock leads to an increase in the cost of capital. Thus, substitution between capital and labor cannot by itself explain the strong response of routine employment to monetary policy shocks.18 I find modest evidence in support of the “industry composition” channel, however. Monetary policy shocks disproportionately affect employment in construction and durable goods manufacturing, industries in which employment is concentrated in routine occupations. I present some evidence that the response of mortgage rates might be decreasing construction of new houses and the exchange rates appreciation could lead to declining demand for durable goods exports. 6.1 Capital-labor substitution and the relative price of investment As discussed at length in Autor et al. (2003) and Autor (2015), routine-task occupations are particularly susceptible to automation. Although the process of automation is typically thought of as a long-run trend phenomenon, it can potentially impact the cyclical dynamics of routine employment as well, at least to the extent that capital and labor are substitutable in the short run. One particular channel though which short-run substitution of capital for labor might occur is via the relative price of investment. If a contractionary monetary policy shock lowers the price of capital for which routine labor is a substitute in production, then one would expect to see very persistent declines in routine employment as firms substitute toward capital. 18 This does not necessarily mean that substitutability between capital and routine-task labor plays no role; there might still be substitution toward capital relative to labor with both inputs declining. 17 Fisher (2006), for example, finds the relative price of investment goods, particularly investment equipment, to be strongly negatively correlated with output; he argues that shocks to investment-specific technology are a large driver of the business cycle. On the other hand, Beaudry, Moura and Portier (2015) find that the relative price of investment to be acyclical or even procyclical during certain time periods. Thus, it is not obvious a priori how monetary policy shocks will affect investment prices. To test this notion, I consider the effect of a Romer and Romer (2004) monetary policy shock on measures of capital, investment, and the relative price of investment goods. Following Krusell et al. (2000), I focus on equipment investment and capital.19 The focus on capital equipment—as opposed to structures—is meant to get closer to the exact types of capital for which routine labor is a substitute; for example, an assembly line worker might be substituted for by a robotic system, but a new factory is not a (direct) substitute for that worker. This focus on equipment is motivated by the effects of automation on routine employment from the job polarization literature, but the results below hold if one considers both equipment and structures. I construct series on the capital stock, investment flow, and relative price of investment for equipment goods on a quarterly basis to cover the same time period as the data in Section 3, 1969-2012. The source data are from the National Income and Product Accounts (NIPA) Table 5.3.5 and DiCecio’s (2009) quality-adjusted price measures based on Krusell et al. (2000). Following Krusell et al. (2000), I construct a series of the equipment capital stock using base-year estimates from Gordon (1990). Figure 6 displays impulse responses to monetary policy shocks of these variables estimated from quarterly versions of Equations 2, 3, and 4.20 A contractionary shock leads to large and significant increase of about 4 to 5 percent in the price of investment equipment. It is unsurprising, given this result, that investment in equipment declines dramatically, between 9 and 15 percent, depending on the specification, and the stock of equipment capital declines gradually. Linear estimates also predict this pattern in response to a contractionary shock. As with the previous estimates, expansionary shocks have little effect.21 As discussed in Section 2, the job polarization literature highlights the important role of falling investment prices in driving the long-run trends in declining routine employment. 19 Although Krusell et al. (2000) focus on the differences between skilled and unskilled labor, their basic framework extends to the routine vs. nonroutine distinction, as discussed in Section 2. 20 As in the baseline monthly calculations, the vector of control variables xt includes a one-year lag of both the dependent variable and the shock. 21 Note that the anomalous responses to expansionary shocks discussed in Section 4.2 are present here as well. As before, these patterns do not appear in the quadratic estimates. 18 Figure 6: Impulse responses of equipment capital, investment, and prices to a 100 b.p. contractionary (red) and expansionary (blue) Romer and Romer (2004) monetary policy shock. Dashed lines are the IRFs of a contractionary shock estimated linearly. Shaded areas are 90 percent confidence intervals. Left column: baseline asymmetric estimates from Equation 3. Right column: alternative quadratic estimates from Equation 4. 19 The results in this section demonstrate that this argument does not extend to the short-run fluctuations in routine employment driven by monetary policy shocks. The large declines in routine employment after contractionary monetary policy shocks are not driven by lower prices of investment and substitution toward new technologies. In fact, the price of investment increases in response to a contractionary monetary policy shock, and investment in equipment declines. Firms are not substituting away from routine labor and toward capital; indeed, both routine labor and the stock of capital equipment decline by around 4 percent at a three- to four-year horizon.22 6.2 Industry effects Another way that monetary policy might have strong effects on routine employment is via industry effects. In particular, monetary policy shocks might have larger effects on some industries than others; if those strongly impacted industries employ a large number of routinetask workers, the responses of routine employment in Section 4.2 may be driven by these industry differences. This section presents some evidence in support of this mechanism. I find that total employment in construction and durable goods manufacturing respond strongly to contractionary shocks and only modestly to expansionary ones. In 1972,23 routine jobs made up 86 and 82 percent of total employment in these industries, and these two industries accounted for 17 percent of total routine employment.24 Other industries’ employment responses are also asymmetric, but no other industries’ responses are as large in magnitude as construction and durable goods manufacturing.25 Figure 7 displays the IRFs from (2) to a 100 b.p. contractionary Romer and Romer (2004) shock of employment in a variety of industries. While employment in some in22 Because routine employment declines more rapidly than the capital stock, firms may effectively be substituting away from labor and toward existing capital in the very short run. This is a qualitatively different mechanism, however, than substitution toward new capital that is an important driver of the long-run trends. 23 The first year for which annual employment by occupation and industry are available. 24 The same figures in 2017 are 79 percent, 66 percent, and 8 percent, indicating that these industries played an important role in the overall trend decline in routine employment, as argued in Foote and Ryan (2015). 25 Ideally, this effect would be tested directly by examining data by occupation and industry. Unfortunately, because of several industry and occupation reclassifications in the Current Population Survey (CPS), it is difficult to construct consistent time series of employment by industry and occupation; moreover, the monthly data are available from the CPS for a shorter time period and smaller sample sizes within each industry-occupation cell introduces additional uncertainty. Therefore, I focus on comparing impulse responses in construction and durable goods manufacturing with other industries that differ in the share of their employment made up by routine jobs. This exercise is supportive of the notion that the direction of causation is from industry characteristics to occupational composition and not vice versa. 20 dustries is hardly affected, other industries—in particular, construction and durable goods manufacturing—decline by large amounts, about 3 to 4 percent. As a comparison, the response of employment in finance, another industry that might reasonably be thought to be interest-rate sensitive, is statistically significant, but its peak effect is about four times smaller than that of construction. The linear estimates by themselves, given the concentration of routine employment in these two industries, suggest that monetary policy’s differential impact across industries might drive its strong effects on routine employment. Asymmetric industry-level employment estimates are displayed in Figure 8. Similar to the linear case, construction and durable goods manufacturing have larger responses than other industries, but here it is evident that this result is only true for contractionary shocks. In response to an expansionary shock, the increase in construction and durable goods manufacturing employment is modest and statistically insignificant at most horizons; contractionary shocks, on the other hand, lead to peak declines of about 6 to 7 percent, twice as large as the corresponding linear estimates. Tests of asymmetry by industry are displayed in Figure 9. Here it is evident that the asymmetric effects of monetary policy shocks on employment are broad-based across industries. Contractionary shocks have significantly larger effects than expansionary shocks at two- to four-year horizons for all the industries displayed. The industry mechanism described in this section can, therefore, only explain a portion of the results on occupational employment in Section 4. It can explain the large and persistent effects of monetary policy shocks on routine employment; it cannot by itself explain the asymmetric effects. Figure 10 presents some evidence of the channels through which this industry effect might occur. There are large and asymmetric effects of monetary policy shocks on exchange rates, exports, mortgage rates, and exports. The large dollar appreciation that occurs after contractionary monetary policy shocks could be leading to lower foreign demand for U.S.produced durable goods; indeed exports decline substantially, although the standard errors are large. Increases in mortgage rates could decrease demand for housing, and indeed residential fixed investment declines almost twice as much after contractionary shocks as it rises are expansionary ones. 7 Conclusion This paper connects two phenomena previously not thought to be related: job polarization and the asymmetry of monetary policy shocks. Monetary policy shocks have large and 21 Figure 7: Impulse responses to a 100 b.p. contractionary Romer and Romer (2004) monetary policy shock. Estimated from Equation 2. Shaded areas are 90 percent confidence intervals. 22 Figure 8: Impulse responses to a 100 b.p. contractionary (red) and expansionary (blue) Romer and Romer (2004) monetary policy shock. Estimated from Equation 4. Dashed lines are point estimates from 3. Shaded areas are 90 percent confidence intervals. 23 Figure 9: Point-wise t-tests against the null of symmetry for sign-dependent IRFs in Figure 8. Shaded areas indicate 90% confidence region. Solids lines from are from (4), dashed lines from (3). Lines outside the shaded region indicate the null can by rejected at 10% significance. 24 Figure 10: Impulse responses to a 100 b.p. contractionary (red) and expansionary (blue) Romer and Romer (2004) monetary policy shock. Estimated from Equation 4. Dashed lines are point estimates from 3. Shaded areas are 90 percent confidence intervals. 25 persistent effects on employment in routine-task occupations—jobs that are typically held by workers with a high school degree and no college—and this effect is highly asymmetric: contractionary shocks lead to large declines in employment while expansionary shocks produce only modest increases. Moreover, the pattern of employment in routine jobs explains essentially all of the effects of monetary policy on total employment; the effects on nonroutine employment are small and insignificant. Monetary policy shocks can explain up to 40 percent of the forecast error variance of the share of employment in routine occupations over a two-year horizon. Because the effects of contractionary monetary policy shocks on routine employment are large and persistent, while those of expansionary shocks are essentially insignificant, monetary policy shocks have on average accelerated job polarization: contractionary shocks lead to almost permanent declines in routine employment, while expansionary shocks have almost no impact. Although much of the trend decline in routine employment can be explained by increases in technology and the consequent declines in investment prices driving substitution toward capital and away from routine labor, the large share of short-run fluctuations in routine employment due to monetary policy shocks cannot be explained by this mechanism. In fact, contractionary monetary policy shocks lead to increases in the price of investment goods. I have presented some evidence, however, that monetary policy shocks drive short run fluctuations in routine employment because the industries on which monetary policy has large effects primarily employ workers in routine-task occupations. Employment in construction and durable goods manufacturing, in particular, respond very strongly and asymmetrically to monetary policy shocks, and between 65 and 85 percent of total employment in those industries is made up by routine jobs. Although the large and persistent effects on employment are unique to these industries, asymmetric responses occur in most industries. A fuller understanding of the sources of the asymmetric effects of monetary policy on routine employment is an important goal for future research. The results presented in this paper may help explain the particularly deep decline in routine employment during the Great Recession. In particular, to the extent that monetary policy was constrained by the zero lower bound on interest rates, then the steep decline in routine employment is consistent with the results presented above. These results might also be used to help understand monetary policy’s effects on inequality as described in Coibion et al. (2012). An understanding of monetary policy’s disparate impacts on different groups— in the labor market and in capital and financial markets—is an important area of research. This paper is a first step toward a fuller understanding of monetary policy’s contributions 26 toward job polarization and other longer-term trends in the labor market. 27