Deflation and output across sectors: results for the Czech Republic

94 2018, XXI, 4 Ekonomie DOI: 10.15240/tul/001/2018-4-007 Introduction The study of defl ation seems to be gaining ever more importance. Central banks in most advanced economies, including the European Central Bank and the Czech National Bank, have observed CPI infl ation running below their targets or even in outright defl ationary territory. These central banks argue that defl ation should be avoided at all costs and employ extraordinary policies such as quantitative easing, foreign exchange interventions or negative nominal interest rates to fi ght against it. As these policies have not always led to higher economic growth and higher infl ation, there have been calls for even more extraordinary measures. Among those is a proposal to abolish cash money in order to allow banks to carry over the negative interest rates more effi ciently on their depositors and stoke infl ation (see e.g. Buiter, (2009) or Rogoff (2014)). Therefore, a lot is at stake in knowing whether defl ation (especially mild defl ation) should be avoided by all means. There are substantial problems with studies that have tried to assess the relationship between economic growth and the price level. First, all of them have used aggregate data for countries such as GDP and CPI. While aggregate variables provide information about the economy as a whole, they cannot reveal the potential variation in output and prices that takes place ‘inside’ the economy. As a result, we know very little about whether fi rms and sectors within the economy face defl ation and where it comes from. We can also hardly assess whether defl ationrecession theories hold. Second, episodes of defl ation in the aggregate CPI or in the GDP defl ator have been scarce in the past decades. As a result, studies on defl ation have had to rely on historical data, often before World War I. Whatever these studies show, their conclusions can be criticized as irrelevant, because they are based on outdated observations. In this paper, we take a different approach. Instead of using aggregate data on GDP and infl ation, we focus on sector data on production and prices. Specifi cally, we use data from the Czech Statistical Offi ce on production, gross value added and prices in sectors of the Czech economy from 1993 to 2015. This dataset has three advantages: (1) it is rich in information on fi rms’ output and inputs, (2) it provides recent observations, and (3) it contains numerous episodes of sector defl ation, which would on the macroeconomic level be concealed under the aggregate CPI or defl ator numbers. We think this approach is novel: we do not know of another study that would analyze defl ation and growth using sector data. We fi nd that, contrary to common wisdom, defl ationary pressures in the Czech economy have been coming from sectors with increasing output and increasing gross value added, not from the deteriorating ones. This shows that defl ation was most often growth-driven, representing rising productivity. This text proceeds as follows. In Section 1, we present the current state of research on defl ation and show its main drawbacks. In Section 2, we present our sector data from national accounts. We fi rst show some descriptive statistics in Section 3. We then perform regression analysis in Section 4, where we use the fi xed effects model and General Method of Moments estimation to regress the growth in production and gross value added on the growth of prices plus control variables. The last section concludes the paper. 1. Current Research 1.1 Empirical Studies There are three main lines of reasoning why defl ation is thought by many to be harmful for economic growth. First, the expectation of falling prices may delay spending by consumers DEFLATION AND OUTPUT ACROSS SECTORS: RESULTS FOR THE CZECH REPUBLIC Pavel Ryska, Petr Sklenář EM_4_2018.indd 94EM_4_2018.indd 94 28.11.2018 13:12:5728.11.2018 13:12:57 95 4, XXI, 2018 Economics and businesses. Firms see their revenues fall and if they cannot adjust wages and other costs accordingly, their margins shrink and they have to lay off workers. Second, for a given nominal interest rate, defl ation raises the real interest rate, which in turn might deter investment. Third, if there is signifi cant indebtedness in the economy, defl ation may cause bankruptcies as it increases the real value of debt. Several studies have attempted to assess empirically whether the alleged link between defl ation and recession holds. Bordo and Redish (2003) restrict their attention to the United States and Canada in the period 18701913, which was rich in episodes of defl ation. They fi nd that prices did not affect output. Bordo, Lane and Redish (2004) add Germany to the sample and reach a similar conclusion. Other recent studies use long datasets that span more than a century and contain many countries. Atkeson and Kehoe (2004) conclude from their regression of output growth on growth of prices that in the whole sample, the correlation is close to zero. An exception is the subsample of the Great Depression in the 1930s, where the link was positive. Borio and Filardo (2004) as well as Borio et al. (2015) generally confirm this result. By contrast, Guerrero and Parker (2006) lean to the conclusion that defl ation is bad for growth, fi nding an economically small but statistically signifi cant link. According to them, lagged defl ation negatively affects output. Benhabib and Spiegel (2009) fi nd a non-linear relationship between defl ation and recession. In recent years, some researchers have focused on the link between expected infl ation (or defl ation) and consumption. Interestingly, they have produced starkly different results. Bachmann, Berg and Sims (2015), using recent data on the US economy, refuse the hypothesis that the expectation of defl ation leads to lower present consumption. By contrast, Ichiue and Nishiguchi (2014) support this hypothesis using survey data on modern Japan. Overall, the existing empirical work on deflation is rather limited – which is surprising given its importance for monetary policy – and inconclusive. As we explain below, there is a major problem of applicability of the majority of these studies to modern-day monetary policy. 1.2 ‘Good’ or ‘Bad’ Defl ation? Two lines of thought can be traced in current mainstream macroeconomics. The fi rst group considers defl ation decisively harmful and studies ways to avoid it. Here we can include infl uential macroeconomists Krugman (1998), Bernanke (2002) or Svensson (2003). The second group takes a more detailed look at defl ation. Bordo and Redish (2003) coined the terms ‘good’ and ‘bad’ defl ation. Defl ation is of the good type if is associated with rising output. In particular, good defl ation occurs when fi rms invest to decrease unit costs and increase output. As a result, if the economy operates with a constant money supply, consumers buy an increased output with the same amount of money, so the price level decreases. This is a synonym for an outward shift in the aggregate supply curve. This situation is thought to have been typical for the pre-World War I period. By contrast, bad defl ation is associated with falling output. In this case, defl ation results from decreasing nominal demand, which cannot be immediately passed on to lower prices of inputs. Firms see their profi tability decrease and cut production. This was most probably one of the characteristics of the Great Depression of the 1930s. Other works adhering to this distinction between good and bad defl ation include Beckworth (2007), Bordo and Filardo (2005) and to a certain extent also Borio et al. (2015). The terms good and bad defl ation do not describe causality – they only describe correlations. For example, bad defl ation does not necessarily mean that defl ation causes falling output. Decreasing prices may only be a symptom of falling demand, not the root cause itself. However, the terms may be useful in practice because policymakers try to determine the type and decide whether they should or should not counteract it. If defl ation results from decreasing nominal demand, central banks consider it undesirable and prefer to offset it by easing policy. On the contrary, if it results from cheaper production without direct link to demand, central banks may be more willing to let such defl ation run its course. Therefore, if we fi nd that most defl ation going on in the economy is of the good type, it may have a direct policy implication. For these purposes, the existing empirical research has a clear disadvantage. All the studies mentioned in Section 1.1 use only aggregate macroeconomic variables, most often the GDP and CPI infl ation. This causes a trade-off: the aggregate annual data do not EM_4_2018.indd 95EM_4_2018.indd 95 28.11.2018 13:12:5828.11.2018 13:12:58 96 2018, XXI, 4 Ekonomie show almost any years with defl ation in the last decades for most developed economies (For the Czech Republic, too, the aggregate price defl ator shows only one year with defl ation in the period 1993-2015.). When researches want to use aggregate data and learn more about defl ation, they have to use pre-World War I data which are rich in observations of defl ation. But then the results are based on old observations which may have little relevance for today because monetary regimes have changed, there are starkly different levels of debt, there is more fi nancial intermediation, etc. By contrast, using sector data, we have many episodes of modern-day defl ation and we circumvent the trade-off. We are thus better equipped to make judgments about a potential link between prices and output and to assess particular defl ation-recession theories. (Below we use the terms infl ation and defl ation also for the change of sector prices. Some may insist that infl ation and defl ation denote only aggregate price movements, but not sector price movements. However, all aggregate numbers have their sources in their components, and it is precisely these sources that we analyse below. Thus we use infl ation and defl ation in the broader sense.) 2. Data We use data from the Czech Statistical Offi ce (2017). It has complete data on 86 sectors of the Czech economy as defi ned in national accounts. The defi nition of each industry is based on NACE classifi cation and values are published in the statistics of National Accounts. The data spans from 1993 to 2015. As we use growth in variables rather than levels, one time period drops out, so we have 22 time periods. That gives a total of 1892 observations. The variables are output prices (implicit price defl ator), output, gross value added (GVA), employment and intermediate inputs. Output, GVA, employment and intermediate inputs are available both in nominal and real terms, but in this article we use them in real terms (i.e., in constant prices) as we are interested in assessing the effect of prices on real output of fi rms. An important feature of the NACE classifi cation is that it does not cover only sectors producing for fi nal consumption, but also sectors upper in the production chain – i.e., sectors producing capital goods and intermediate inputs. As a result, fi rms’ output does not mean only consumer goods bought by consumers, but also capital goods bought by other fi rms. Fig. 1: Production and GVA growth against change of prices Source: Czech Statistical Offi ce (2017), own computations EM_4_2018.indd 96EM_4_2018.indd 96 28.11.2018 13:12:5828.11.2018 13:12:58 97 4, XXI, 2018 Economics 3. Descriptive Statistics In our sample, positive price growth prevails and accounts for 80% of all observations (see Fig. 1). At the same time, approximately three fi fths of observations have positive growth of production and gross value added. As a result, roughly half of observations lie in the top-right quadrant in Fig. 1. In this section we focus on basic statistical relationships between the sign of price change (infl ation/defl ation) and the sign and growth rate of production or GVA. In particular, we focus on whether declines in production are associated with defl ation. In Tab. 1, we report the comparison of growth in production and gross value added under increasing versus decreasing prices. We also visualize the comparisons with density graphs in Fig. 2. Out of observations with price defl ation, 69.7% recorded positive growth in real production, while the corresponding percentage for observations with infl ation is only 58.8%. This is also refl ected in the averages: the average growth rate of production during episodes of price defl ation is 5.47%, which is sizably higher than the average growth with infl ation (2.85%). Therefore, this preliminary observation contradicts the notion that defl ation is linked to subpar output growth. To check this fi nding, we also ran a formal test of equality of means. As seen in Tab. 2, the t-test rejects the null hypothesis that the average production growth under infl ation is equal to that under defl ation, thereby confi rming the fi nding. In contrast, the standard deviations of growth rates are similar, which is also supported by the testing of equality of variances. We performed the same analysis also for real gross value added (see Tab. 1). Here the better performance under defl ation is even more pronounced: 77.8% of defl ationary observations report positive growth in gross value added, while for infl ationary observations the percentage is only 52.3%. The average growth of gross value added is 14.1% with defl ation, while only 0.68% with infl ation. The difference is also confi rmed by the t-test. The only difference with gross value added as opposed to production is the higher standard deviation with defl ation than with infl ation. The reason for these rather convincing fi ndings may be the prevalence of ‘good defl ation’ in our sample: the sectors that reported product price defl ation could be precisely the ones that invested most in production and therefore enabled cheaper and greater production. Anyway, this preliminary look runs against the claim that defl ation harms prospects for growth. In the Appendix, we reverse our perspective and ask the question how much growth in prices there is under the opposite scenarios of rise and fall in production and GVA. All data Infl ation Defl ation Observations 1,892 1,509 383 Production obs. with production increase 61.00% 58.80% 69.70% obs. with production decrease 39.00% 41.20% 30.30% Mean growth 3.38 2.85 5.47 Standard deviation of growth 12.89 12.71 13.39 Gross value added obs. with GVA increase 57.50% 52.30% 77.80% obs. with GVA decrease 42.50% 47.70% 22.20% Mean growth 3.40 0.68 14.10 Standard deviation of growth 34.86 25.74 57.05 Source: Czech Statistical Offi ce (2017), own computations Tab. 1: Growth of production and GVA under infl ation and defl ation EM_4_2018.indd 97EM_4_2018.indd 97 28.11.2018 13:12:5828.11.2018 13:12:58 98 2018, XXI, 4 Ekonomie This complements our result here that slower price growth or outright defl ation is accompanied by quicker output growth. 4. Regression Analysis 4.1 Time Averages As a fi rst look on correlations, we performed a simple regression of time averages. We regressed average annual output growth in the 86 sectors of the Czech economy over 19932015 on the corresponding average annual price growth, as seen in Tab. 3. In what follows below, we denote x the annual percentage change in variable X. That is, p is the growth in prices, q in real output (production) and gva in real gross value added. We tested the results in Tab. 3 for the presence of heteroscedasticity and whenever it was detected we used the White’s heteroscedasticity-corrected covariance matrix. Test statistic p-value Production t-test for equality of means 3.44*** <0.001 F-test for equality of variances 1.11 0.19 Gross value added t-test for equality of means 4.49*** <0.001 F-test for equality of variances 4.91*** <0.001 Source: Czech Statistical Offi ce (2017), own computations Note: We use the Welch t-test, i.e., a standard unpaired and two-sided t-test to determine whether the means of production growth under infl ation and defl ation are equal to each other. The null hypothesis is that the two means are equal. To test the equality of variances, we use an F-test, where the null hypothesis is that the ratio of the variances of the samples is equal to 1. Note: The sign *** denotes signifi cance at 1% level. Tab. 2: Tests of equality of means and variances Fig. 2: Density of production and GVA growth under defl ation and infl ation Source: Czech Statistical Offi ce (2017), own computations EM_4_2018.indd 98EM_4_2018.indd 98 28.11.2018 13:12:5828.11.2018 13:12:58 99 4, XXI, 2018 Economics The regression coeffi cient on p (percentage growth in prices) is -1.142, signifi cant at 1%, suggesting that slower price growth has been associated with faster output growth. As depicted in Fig. 3, there is one distinct outlier on top left. This sector is the manufacturing of computer, electronic and optical products. Interestingly, it is both the sector with the fastest growth rate of production and the only sector with average annual defl ation over 19932015. We also ran the regression without this outlier, but the negative coeffi cient and high statistical signifi cance remain (see Tab. 3). We tried leaving out up to 10 most extreme observations, but the negative and statistically signifi cant coeffi cient for p remained. We can observe the general pattern of the sectors in the Czech economy from Fig. 3: while All data Without outlier Coeffi cient p-value Coeffi cient p-value Regressing q on: Intercept 7.082*** <0.001 -6.118*** <0.001 p -1.142*** <0.001 -0.919*** <0.001 Observations 86 85 Adj. R2 0.237 0.167 Regressing gva on: Intercept 8.996*** <0.001 8.824*** <0.001 p -1.998*** <0.001 -1.819*** <0.001 Observations 86 85 Adj. R2 0.380 0.323 Source: Czech Statistical Offi ce (2017), own computations Note: The sign *** denotes signifi cance at 1% level. Tab. 3: Time averages: Regression of average annual growth in output and GVA on infl ation Fig. 3: Regressing time averages: (a) Output growth on infl ation, (b) GVA growth on infl ation (all data) Source: Czech Statistical Offi ce (2017), own computations EM_4_2018.indd 99EM_4_2018.indd 99 28.11.2018 13:12:5828.11.2018 13:12:58 100 2018, XXI, 4 Ekonomie the contracting sectors had both low and high infl ation, booming sectors tended to have lower infl ation and those that expanded most quickly – with growth rate of output above 10% – had exclusively low infl ation or outright defl ation. The results suggest that higher output growth is linked to slower price growth, i.e. low infl ation or outright defl ation. However, it may not be evident what happens to fi rms’ profi ts as infl ation slows or turns into defl ation. We therefore used growth in real gross value added instead of real output growth and regressed it on price growth. The results are very similar to those with output growth (see Tab. 3 and Fig. 3). Overall, this evidence too seems to run against the defl ation-recession theories. In our sample, faster output growth and profi t growth are associated with slower infl ation or outright defl ation. This could suggest that higher production most often results from investment into lower-cost production, to which consumers react by purchasing more. 4.2 Panel Data Two Models Using panel data naturally offers much more information than time averages because we can make use of the entire variation of annual observations. Our data are in the form of balanced panel data and we use the fi xed effects model for estimation. This model allows unobserved sector-specifi c effects to have any correlation with the explicit regressors. We estimated the models below using the so-called ‘within estimation’. Generally, we use two economic models for estimation. The fi rst is an autoregressive distributed lag (ADL) model which takes a macroeconomic approach to prices and production. Just as studies cited in Section 1.1 regress GDP growth on price growth (infl ation or defl ation), we regress sector output growth on the respective growth in output prices for the given sector. This directly addresses the question of correlation between prices and output. We also include lagged variables. For example, we estimate qit = â0 + â1qit-1 + â2 pit + â3 pit-1 + ai + uit (1) where qit is real output growth in sector i and year t, pit growth in output prices in the sector (both in percent terms), ai the sector-specifi c unobserved effect and uit the error term. The second model, in contrast, takes a more microeconomic, fi rm-level approach. To produce output, fi rms must employ workers and buy intermediate inputs. Therefore, the model we estimate is qit = â0 + b1 pit + â 2empit + â3inpit + ai + uit (2) where emp is growth in labour employed (measured in total hours worked) and inp growth in the volume of intermediate inputs in production. The regressors include two usual inputs in the neoclassical production function (for why we exclude capital K, see our discussion below) plus a special term – the output price change p. Productivity, or effi ciency of production, is usually accounted for as a residual, i.e., what remains unexplained in output produced if we take into account changes in labour, capital and intermediate inputs. Firms typically invest or take other measures in order to make production more effi cient. After they do so, they can produce more and with lower costs, so they can offer their products more cheaply and reach more customers. If this is the case, then the output price P acts as a proxy for effi ciency. The lower is the cost of production, the lower is the observed price P. Therefore, price growth p is not only a variable that we add to fi nd out about its correlation with output growth q, but it also has a concrete economic interpretation as a proxy for the effi ciency of production. Equation 1 above contains the lagged dependent variable qit-1 among regressors. As a result, the fi xed-effects estimator is generally not consistent. As Wooldridge (2002) shows, the bias falls at a rate 1/T as T grows, and for a time dimension high enough, the inconsistency would be negligible. We think that with our T = 22, we do have a T high enough to be sure that most of the bias disappears. However, as a check we also perform General Method of Moments (GMM) estimation in equations where there are lagged dependent variables. Our GMM estimation follows Arellano and Bond’s (1991) two-step procedure and we use lags of order 2 to 5 of the dependent variable as instrumental variables. Finally, in the fi xed effects models below, each regression was tested for heteroscedasticity and serial correlation. When either of these was detected, a robust variance matrix estimator was used to correct the variances and test statistics. EM_4_2018.indd 100EM_4_2018.indd 100 28.11.2018 13:12:5928.11.2018 13:12:59 101 4, XXI, 2018 Economics Results: Macroeconomic Approach In the fi rst, macroeconomic approach, we estimate the effect of infl ation on output growth and GVA growth in an autoregressive distributed lag model. Tab. 4 shows that contemporaneous infl ation pt is negatively linked with output growth, confi rming our fi ndings from Section 4.1 above. The GMM estimation brings an estimate of the coeffi cient of pt very similar to the fi xed effects estimation (-0.303 vs. -0.293) and confi rms its statistical signifi cance. By contrast, the coeffi cient of pt-1 is economically and statistically less signifi cant, both according to fi xed effects estimation and GMM. Similarly to output, infl ation is negatively linked to the growth in gross value added of fi rms (see the lower half of Tab. 4). The only difference is that the magnitude of the coeffi cient of pt is bigger for gross value added than for output as dependent variable. This is understandable since fi rms usually have operating leverage: a rise in output (revenues) by 1% causes a rise in gross profi t of more than 1%. Results: Production-Function Approach In the second, more microeconomic approach, we use typical inputs into production function as control variables for assessing the impact of infl ation on output growth. These should be more powerful control variables as they are directly linked to output. We estimate the equation with growth in labour emp (total hours worked) and growth in the volume of intermediate inputs inp as controls. We have also added growth in gross capital stock k to imitate more the traditional production function Y = f(L, K). Because data on capital stock have poorer availability than other data, we used an alternative source from the OECD (2016). It is both narrower (only 57 sectors) and shorter (1995-2009) than our main dataset, so it has much fewer observations (798). Results from this enlarged regression that includes k are not materially different from those reported in Tab. 5 and we do not report them here. As expected, this model shows a much higher fi t as Adj. R2 is as high as 0.638 in the output regression in Tab. 5. And again, infl ation negatively affects output growth (-0.476) with high statistical signifi cance. Labour and intermediate inputs also have the expected signs and high signifi cance. In the GVA regression, infl ation also has a negative coeffi cient and the coeffi cient is again greater in absolute magnitude. Interestingly, while the labour input has a positive and statistically signifi cant impact on GVA, intermediate inputs lose their signifi cance. Interpretation The empirical results bring, in our view, two important fi ndings. Fixed effects GMM Coeffi cient p-value Coeffi cient p-value Regressing qt on: qt-1 0.012 0.696 0.042 0.231 pt-0.293*** 0.003 -0.303** 0.014 pt-1 0.136* 0.052 0.110 0.121 Observations 1,806 1,720 Adj. R2 0.017 - Regressing gvat on: gvat-1 -0.086*** 0.002 -0.178 0.236 pt-1.614*** <0.001 -1.768*** <0.001 pt-1 0.614 0.173 0.393 0.617 Observations 1,806 1,720 Adj. R2 0.068 - Source: Czech Statistical Offi ce (2017), own computations Note: The sign * denotes statistical signifi cance at 10%, ** at 5% and *** at 1% level. Tab. 4: ADL: Regression of growth in output and GVA on infl ation EM_4_2018.indd 101EM_4_2018.indd 101 28.11.2018 13:12:5928.11.2018 13:12:59 102 2018, XXI, 4 Ekonomie Firstly, the negative relationship between growth in output and growth in prices suggests that observed defl ation has been predominantly of the ‘good’ type. If falling prices resulted from falling nominal demand, fi rms would not respond by increasing production – on the contrary, they would cut it, as the ‘bad defl ation’ hypothesis goes. Hence, we would observe a positive correlation between p and q. The alternative hypothesis is that as fi rms invest and improve effi ciency, they manage to cut unit costs and prices. This allows them to sell more products, i.e., reach new marginal demand through lower prices. This ‘good defl ation’ hypothesis is consistent with our fi ndings – we see repeatedly a negative relationship between p and q. It could be argued that the negative relationship between p and q may in some more infl ationary years refl ect rather the comparison of high versus low infl ation than infl ation versus defl ation. That may be true. The shift of the aggregate demand curve – for example, through central bank easing – may move the whole mass of observations to the right in Fig. 3, bringing many of them out of the defl ationary section. Symmetrically, lack of aggregate nominal demand may push them to the left, which was the case for instance in 2009 when the economy as a whole had a 1.4% defl ation (measured as output defl ator). However, rather than focusing on the position of the mass of observations, we are interested in the slope of the regression line – that is, in the shape of the mass. Its negative slope suggests that the tendency to cut prices or at least have slower price increases than the average is associated with growing fi rms, not declining ones. Overall, we are not claiming that ‘bad defl ation’ does not exist or that it cannot be harmful. We are only showing evidence that there is much more ‘good defl ation’ than ‘bad defl ation’ in our sample. Secondly, using gross value added in addition to output, we have shown that lower prices are not only associated with greater output, but also with greater gross profi t of fi rms. Output in itself is not the fi rms’ goal, but profi t is. If fi rms invest to increase production and cut prices, they also succeed in increasing profi ts, our results show. This fi nding is important because one line of reasoning (starting with Fisher, 1933) is that defl ation or too low infl ation ultimately erodes fi rms’ profi ts and leads them to bankruptcies. Our results show otherwise, again providing support for the case that defl ation in our sample stems from fi rms’ own initiatives, not from negative shocks that would squeeze profi ts. As pointed out in Section 2, the sectors used are not only those that produce consumer Fixed effects Coeffi cient p-value Regressing qt on: pt-0.476*** <0.001 empt0.280*** <0.001 inpt0.582*** <0.001 Observations 1,700 Adj. R2 0.638 Regressing gva t on: pt-1.663*** <0.001 empt0.889*** <0.001 inpt0.062 0.637 Observations 1,700 Adj. R2 0.112 Source: Czech Statistical Offi ce (2017), own computations Note: The sign *** denotes signifi cance at 1% level. Tab. 5: Production function: Regression of growth in output and GVA on infl ation EM_4_2018.indd 102EM_4_2018.indd 102 28.11.2018 13:12:5928.11.2018 13:12:59