Investment–uncertainty relationship: differences between intangible and physical capital

ABSTRACT This paper disentangles the effects of uncertainty in explaining the heterogeneity of firms’ investments. In particular, following Bloom [2007. “Uncertainty and the Dynamics of R&D.” American Economic Review 97 (2): 250–255], we test the role of uncertainty and liquidity constraints extending the model to include R&D, non-R&D intangibles, as well as physical capital. The analysis is performed on a large data set of Italian firms, covering both manufacturing and services sectors, as well as large and small firms. We show that non-convex adjustment costs affect different capital inputs in different ways, depending on their degree of firm-specificity. The results confirm the Bloom model: flow adjustment costs explain investment in R&D and, to a lesser extent, in non-R&D intangibles. However, it struggles to explain tangible investment plans because of the ambiguous effect of the stock adjustment costs.


Introduction
Investigating the role that uncertainty plays in investment is quite important at times, like the present one, of economic turbulence and reduced demand. In doing so, it is particularly important to pay attention to investment in intangibles, which are considered one of the main drivers of growth (Aghion and Howitt 2008). See also the findings outlined by the OECD (2013).
In the last 30 years, the debate about the investment-uncertainty relationship has flourished, mainly in regard to tangibles, while little remains known about intangibles. Theoretical studies have struggled to come to any unequivocal conclusions about the sign and relevance of the investment-uncertainty relationship in the long-run, and the short-run effects appear to differ when comparing tangibles and intangibles.
Given the complex effects of uncertainty on investment, empirical evidence is of essential value. However, results are usually based on simulations, whereas very little evidence is based on micro data. This is mainly due to the difficulties in designing the degree of substitutability among inputs, the extent of irreversibility, of returns to scale and of market power. Moreover, the difficulty in finding convincing empirical counterparts to the concept of uncertainty is of even greater importance. Some results regarding tangibles are consistent with the negative predicted effects of real options, see Leahy and Whited (1996), Guiso and Parigi (1999), Bloom, Bond, and Van Reenen (2007), Chirinko and Schaller (2009), Bontempi, Golinelli, and Parigi (2010) and Bianco et al. (2013). See also the reviews in Carruth, Dickerson, and Henley (2000) and Greasley and Madsen (2006).
Very few papers study the relationship between R&D expenses and uncertainty. Goel and Ram (2001) use nine OECD countries over the period [1981][1982][1983][1984][1985][1986][1987][1988][1989][1990][1991][1992], and find that uncertainty has a significantly negative effect on R&D, while it is not significant for non-R&D and aggregate investment (see also the extension in Drakos 2006). Czarnitzki and Toole (2007, 2011, 2013 use different samples of 'innovative' firms drawn from the German innovation survey of the manufacturing sector between 1995 and 2001. They find that product market uncertainty reduces R&D investment, and that this effect is smaller in markets where strategic competition is stiffer, and in the case of large firms; moreover, patent protection partially mitigates the influence of uncertainty. Stein and Stone (2012) analyse Compustat data for 3965 US public companies over the period 2001-2011, and find that uncertainty depresses capital investment and advertising, but encourages R&D spending. This perhaps surprising result for R&D is consistent with the theoretical literature emphasizing that long investment lags create valuable real put options which offset the effects of call options that are lost when projects are started. (Stein and Stone 2012, 1) This paper analyses the peculiarities of investment in intangibles compared to investment in tangibles, with a specific focus on the effects of irreversibility, uncertainty and liquidity constraints. In doing so, we report for the first time, to the best of our knowledge, empirical evidence regarding the theoretical model in Bloom (2007) for R&D, and we extend it to non-R&D intangibles, as well as to physical capital.
Thus, a further novel aspect of our analysis is to compare the effects of uncertainty on different types of investments, since different capital inputs are known to be characterized by different degrees of irreversibility, which in turn is the main driver of the role of uncertainty in investment decisions. Drakos (2011) analyses buildings, machinery and equipment, and motors and vehicles for plant-level manufacturing data drawn from the Annual Industrial Survey for Greece (6119 plants recorded over the 1993-2005 period); uncertainty has a negative effect on the extensive margin, decreasing the likelihood of triggering the investment in new capital. The analysis is taken one step further by Drakos and Goulas (2010), who analyse asset-specific and industry-specific irreversibility (measured in terms of the ratio of the average percentage of used capital to the sum of new and used capital expenditures). They show that uncertainty exerts a negative effect on total investment, as well as on investment in buildings, machinery and vehicles, while it exerts a positive effect on fixtures and fittings, in line with its lower degree of irreversibility. Driver, Temple, and Urga (2006) examine UK manufacturing industries and suggest that equipment expenditures are more 'sunk' than structures, and that structures are less specific in use and characterized by a lower degree of expandability (waiting to invest involves penalties); uncertainty has a negative effect on machinery investment, and a positive one on buildings.
Our analysis is based on an unbalanced panel of large and small-/medium-sized enterprises (SMEs) in Italy, operating in both manufacturing and services sectors, observed over the 2003-2012 period. This data set is interesting for many reasons. Firstly, we dispose of a panel of survey data covering firms' planned and completed investments of different types, expected future sales and demand uncertainty, and qualitative information about productive and financial conditions. This enables us to account for unobservable individual differences between firms, macroeconomic shocks and the evolution of the investment-uncertainty relationship. Secondly, Italy is a country characterized by limited economic growth over the last 10 years, compared to its international competitors and, in particular, to the nation's past growth. Economic policy uncertainty may have exacerbated the doubts that firms face when deciding the amount and type of investment to be planned. Thirdly, the European accounting system allows for reporting as investments (i.e. to accumulate in assets) any spending on intangibles not directly related to current company operations, but representing the company's pledged resources designed to ensure future economic benefits (such as revenues or reduced future costs) (see Stolowy and Jeny-Cazavan 2001;Bontempi and Mairesse forthcoming).
This paper is organized as follows. Section 2 introduces the theoretical model of investment under the real option theory, and relates it to Bloom (2007). Section 3 lists alternative specifications of the empirical model for disaggregate investment assets. Section 4 reports the descriptive statistics of different types of investment and offers a preliminary examination of the data. Section 5 presents the estimates of investment models for R&D and for three other categories of capital goods, and assesses their robustness. Section 6 concludes. The appendices, respectively, describe the data set, define the variables in question, illustrate the additional features of R&D investment and describe the dynamics of the Italian investment cycle.

Modelling the uncertainty-investment relationship: the theoretical underpinnings
When capital is characterized by an asymmetric adjustment-cost function, the initial cost of investment will not be recoverable once investment has been undertaken. Since investment is irreversible, the corresponding investment plan has a real option value reflecting the value that a firm places on its ability to choose the timing of its investment. Under highly uncertain demand conditions, the value of the option to postpone investment increases: it is convenient to wait for new information (Dixit and Pindyck 1994), so that the decision to invest is delayed and current investment is curbed. The user cost of capital includes an irreversibility premium, and is positively affected by uncertainty about future demand conditions; positive investments are justified by higher threshold rates of return, and sunk costs create a zone of inactionrepresenting the option value of waiting which is enlarged by greater uncertainty.
In order to compare the role of irreversibility and uncertainty in determining the investment behaviour of various types of capital input, we need a general model in which different effects are nested. In particular, since we want to compare tangible and intangible capital inputs, we base our analysis on the model proposed by Bloom (2007) for R&D. A generalization of the Bloom (2007) model can be written for period t + 1, as where r is some measure of the investment ratio, y indicates the demand conditions (in logarithm), σ is both individual and time variant uncertainty about future business conditions and X is a set of controls. Uncertainty and irreversibility result in a rich short-term investment dynamics that may differ depending on the type of investment, due to the different nature of adjustment costs for tangibles and intangibles. Bloom (2007) suggests that while adjustment costs for physical capital arise directly from changing the stock (they are stock adjustment costs), adjustment costs for the stock of knowledge arise from changing the rate of change of accumulation (they are flow adjustment costs). This distinction leads Bloom (2007) to suggest investment over sales as a measure of r in the R&D case. While the investment ratio for tangibles is usually measured as investment over capital stock (see, e.g., Bloom, Bond, and Van Reenen 2007), a definition of the investment ratio such as the one proposed by Bloom (2007) is also suitable for other non-R&D intangibles, as well as for tangibles such as machinery or buildings. The distinction between adjustments in the level or growth rate of the stock also plays a crucial role in shaping the response of different investment types to uncertainty, as encapsulated by the three parameters β 3 , β 4 and β 5 .
In both the tangible and intangible cases, increases in uncertainty have a direct negative effect on investment, since higher uncertainty leads firms to postpone their expenses, producing a negative 'delay effect', represented by the β 3 parameter. According to Bernanke (1983), 'In an environment in which the underlying stochastic structure is itself subject to random change, events whose long-run implications are uncertain can create an investment cycle by temporarily increasing the returns to waiting for information'.
For both tangible and intangible capital inputs, uncertainty generates a 'caution effect' in investment behaviour, represented by the negative β 5 parameter: in the short-run, firms' responses to demand shocks are lower at higher levels of uncertainty. A positive demand shock must be large enough to ensure that capital is moved up towards its investment threshold; even so, the investment response is lower than in the case of reversibility, since it is reduced by the zone of inaction.
What distinguishes tangibles from intangibles is the additional delay effect represented by the β 4 parameter. Unlike in the case of tangibles, Bloom (2007) suggests that the region of inaction creates a dynamic link between current and past R&D (or, we may presume, other intangibles) rates: if optimal R&D is higher than lagged R&D, and firms want to raise their spending (upward adjustment), then higher uncertainty will lead firms to postpone their investments, producing an 'additional delay effect' regardless of whether the knowledge stock is decreasing or increasing. While for tangibles, lagged investment interaction with uncertainty might play no role in determining current investment under uncertainty, 1 in the case of intangibles, the impact of the additional delay effect depends on the relationship between desired and lagged investment. This effect is encapsulated by a positive estimate of β 4 which takes into account the fact that uncertainty raises the response to lagged R&D expenditure, and the persistency of R&D investment. 2 In Section 3, we estimate different specifications of model (1) for both tangible and intangible rates. Together with expectations regarding the signs of the β 3 , β 4 and β 5 parameters, we can also make predictions about their absolute values, which are presumed to be higher in the intangible case because of the causal link between adjustment costs and the degree of specificity of investment in intangibles. Intangible spending involves adjustment costs (Hall, Griliches, and Hausman 1986;Lach and Schankerman 1988) and sunk costs (Bloom and Van Reenen 2002) that are greater than those associated with tangibles. In fact, the output from investment in innovation displays a high degree of uncertainty, which is important at the beginning of a research programme/project, when costs related to the need to employ skilled workers and to implement knowledge infrastructure are high. Moreover, many intangibles are highly firm specific. In the words of Williamson (1988), intangibles are not 'redeployable' assets, that is, assets whose value when employed alternatively is almost as high as that of their current use.
Compared to physical capital, the lemon premium is particularly high for intangibles like R&D, because investors find it more difficult to distinguish good projects from bad ones. The result is that the market for R&D projects may entirely disappear if the asymmetric information problem is too large, and this 'missing markets' problem leads to the irreversibility of R&D investment. If a company suffers from an idiosyncratic negative shock (firm-specific uncertainty), it will be unlikely to sell its research operations to another firm and get fairly good value for such, and so irreversibility is severe.
Furthermore, the equality between the marginal revenue product and the marginal cost of capital (the conditions for profits maximization) is affected by the peculiarities of intangibles. Some innovations can be cost-reducing; however, expensive marketing and advertising operations designed to promote the result of the innovation, are needed. Moreover, product enhancement and diversification mean that said product is not substitutable, and this will affect the elasticity of demand; hence, it is difficult to predict the effect on demand, and such effect may vary at different price levels.
Finally, we should make one final important point regarding model (1): apart from the explicit Bloom (2007) determinants, any valid assessment of the impact of uncertainty on investment also requires financial variables as controls (in X it+1 ). The reason for this is that a negative effect of uncertainty on investment might also proxy for credit constraints and/or agency costs: inherently riskier firms may find it more difficult to finance their spending and hence they may plan a lower level of investment. Furthermore, 'to the extent that external financeboth through the debt and equity marketsis subject to agency and/or moral hazard problems, an increase in uncertainty will raise the user cost of capital, inducing a decline in investment spending' (Gilchrist, Sim, and Zakrajšek 2014, 1). The literature considers R&D investment (or, at a broader level, investment in intangible assets) as being more affected by financial constraints than investment in physical capital is (Himmelberg and Petersen 1994;Mulkay, Hall, and Mairesse 2001;Cincera 2002;Czarnitzki and Hottenrott 2011). Moreover, the possibility of encountering financial constraints raises irreversibility (Holt 2007). Hall and Lerner (2010) describe some of the unique characteristics of R&D investment that could explain why external funding of R&D might be more expensive than internal funding. Moreover, the value of any asset is contingent on the state of the world, and an asset may have a very different value to equity holders than to debt holders (Brynjolfsson and Hitt 2005). Equity holders care about the expected returns of a firm and its assets across all future states of the world: intangibles are valuable for them as long as the firm is a going concern. Debt holders, on the contrary, care about what the firm would be worth in a bad global environment where it failed to generate sufficient cash flows to cover its required debt payment: from the debt holder's point of view, intangibles may even have zero value in such situations.
The financial variables we consider are cash flow and financial leverage. If we interpret model (1) as an adaptation of the first-order condition of a discrete time value dynamic maximizing problem (the Euler equation approach derived in Bond and Meghir 1994), the effect of cash flow is expected to be negative, as it depends on the magnitude of adjustment costs. A positive effect, on the contrary, indicates the agency costs of free cash flow (the overinvestment problem of Jensen 1986) or an asymmetric information problem (the underinvestment problem of Myers and Majluf 1984). The coefficient for financial debt should be not significant under the Modigliani-Miller debt irrelevance theorem, and significantly positive if there is no separability between investment and borrowing decisions as happens in imperfect capital markets where firms rely on debt, rather than on equity, for external financing (Whited 1992). A negative value for debt is implied by bankruptcy costs and creditors'/shareholders' agency costs (Myers 1977). High levels of debts could induce firms to opt for excessively risky investment projects (Jensen and Meckling 1976), thus affecting the investment-uncertainty relationship; a negative investment debt relationship could be beneficial to shareholders of low-growth firms because banks can prevent their client firms from investing in unprofitable projects.

From the theoretical model to three alternative empirical specifications
The empirical implementation of model (1) can be accomplished through alternative specifications involving alternative measurements of the variables of interest. To organize the presentation, we start from model (2) for the generic capital input j: (2) The different empirical specifications we propose (in Table 1) are encapsulated by the proxy variables labelled R j (for investment intensity, i.e. the ratio between investment in capital j, I j and sales, Y ), G (for sales growth) and U (for uncertainty).
Besides the core relationship between investment intensity R, sales' growth G and uncertainty U, model (2) lists a number of control variables X: CF is the cash flow; D is the financial debt; m i and t t are the individual and time effects, respectively, accounting for heterogeneity between and within firms, 3 and E is the possibly heteroscedastic random error affecting investment intensity. Following the theoretical discussion in Section 2, we can make explicit predictions about the signs of the coefficients of model (2). The parameter b 1 . 0 is consistent with the positive residuals' serial correlation usually found in static models, and also gives a role to the 'initial conditions' of actual investment when the new investment is set. The parameter b 2 . 0 shows that the investment activity is positively affected by sales growth. The parameters linked to uncertainty are as follows: b 3 , 0 (uncertainty negatively affects investment intensity because it increases the value of waiting and seeing); b 4 . 0 (which indicates the additional delay effect, i.e. the fact that in the presence of greater uncertainty, firms postpone investments and keep them closer to the initial conditions); b 5 , 0 (which represents the caution effect, i.e. the fact that under greater uncertainty, firms respond less to demand conditions). Finally, as far as the effects of financial control variables on investment intensity are concerned, we expect that: g 1 . 0 indicates agency costs and asymmetric information problems, while g 2 . 0 would suggest that companies prefer not to reveal information about their investment plans. These hypotheses seem appropriate in a bank-based system such as that of Italy (Bontempi 2002).
The columns in Table 1 present the different specifications of model (2) according to alternative measures of the proxy variables R, G and U. We propose three empirical equations, respectively, labelled as the 'expectations model', the 'plans-scaled model' and the 'accounting-data model'.
The expectations model uses t I j it+1 / t Y it+1 as the dependent variable R (i.e. the investment plans in t for period t + 1 over expected sales surveyed in t for t + 1), and t g it+1 and u( t g it+1 )as the explanatory G and U (i.e. real sales' growth expected in t for t + 1, and uncertainty about firms' future demand in t for t + 1). In turn, u( t g it+1 ) is defined as the difference between the maximum and the minimum expected real sales' growth in t for t + 1 (the so-called subjective min-max range).
The plans-scaled model uses t I j it+1 /Y it as the dependent variable R, where plans are again in the numerator but Y it , the level of actual sales known in t when the plan was made, is in the denominator. As in the expectations model case, explanatory G and U are measured by t g it+1 and u( t g it+1 ). The accounting-data model uses I j it+1 /Y it+1 as the dependent variable R, so that both numerator and denominator are measured by actual accounting data. Explanatory demand conditions G are measured by actual sales' growth (g it+1 ), and uncertainty U is measured either by the subjective min-max range above, u( t g it+1 ), or by the disagreement, se( t g it+1 ). 4 In the accounting-data model, uncertainty measures are the only non-accounting information. Given that data on plans and expectations are usually not available, the accounting-data model (with some disagreement measure for uncertainty) is the most frequently used in the empirical literature, despite the fact that it disregards all information about individual expectations.
The interpretation of the stochastic error E is model specific: in both the expectations model and the plans-scaled model, it represents the shock at time t when plans and expectations were formed, while in the accounting-data model, it represents the shock from t to t + 1, that is, after plans and expectations were formed. Hence, the accounting model error is a mix of many different macroand micro-economic shocks.
The set of control variables, is the same for all three specifications, and this is also true for the measure of initial conditions, given by the ratio between the actual investment in the j capital input of firm i at time t, I j it , over the actual sales' level,Y it . As a control variable in all three specifications, I j it /Y it might clean the residuals from autocorrelation due to, for example, adjustment costs and other rigidities. However, the interpretation of I j it /Y it is model specific: in expectations and plans-scaled models, it is the actual investment intensity known at the time the plans are made (i.e. the initial conditions), while in the accounting-data model, it is the genuine lagged-dependent variable.
Although the three specifications are all reasonable empirical implementations of Bloom's (2007) theoretical model, they exploit different mixes of signal and noise. This is particularly relevant when the hypothesis of investment plan optimality is rejected, meaning that plans and realizations differ by more than the unpredictable noise that occurred after plans were made. 5 In a similar way, the measurement of the subjective uncertainty at firm-level and of the expected sales' growth disagreement within clusters can embody different amounts of information. 6 In general, alternative mixtures of information lead to different model estimates.
Among the alternatives above, the investment planned for t + 1 on the basis of the information in t, is the target variable that best encapsulates the effect of uncertainty on investment spending, due to three principal, interrelated reasons. Firstly, in the light of the Euler approach, we do not have to estimate the model by evaluating the expectation with their realized values and, in this way, we prevent the error term from embodying a miscellanea of shocks. Secondly, subjective uncertainty and expected sales growth provide information about firms' decisions, whereas when companies actually realize their plans in t + 1, such information isonce againmixed with the shocks that occur from t to t + 1. Thirdly, plans are useful as they embody expected market conditions and discount the complexity of a particular project such as, for example, construction time (Kydland and Prescott 1982) in the case of investment in buildings, or high adjustment costs in the case of R&D.

Data and preliminary analysis
Our data set of company observations combines three sources: the Survey on Investment in Manufacturing (SIM), the Company Accounts Data Service (CADS) and the National Accounts (NA) data. The last two offer additional information to that provided by the SIM. Details about the three data sources are set out in Appendix 1.
The main part of the data set is drawn from the SIM, as it publishes both firms' expected demand and the uncertainty of such, as well as the effective and planned investments subdivided into four categories: (1) non-residential buildings (buildings, j = f ); (2) machinery/equipment/vehicles (machinery, j = m); (3) software/licenses, permissions, concessions/intellectual property rights and trademarks/start-up expenses and other capitalized non-financial intangibles designed to improve productivity (non-R&D intangibles, j = nr); (4) research and development expenses (R&D, j = r). The four investment categories sum to the aggregate, labelled as j = a.
The CADS database includes highly disaggregated balance-sheet data (profit-and-loss and cash flow), as well as detailed information about the characteristics of Italian companies operating in a wide range of industrial and service sectors. The CADS covers about 72% of the SIM observations, and this is very important as this source supplements SIM information on financial variables.
Finally, NA data from the Italian National Institute of Statistics (ISTAT) provide additional information about time-and industry-specific deflators. 7 The figures shown in the various columns of Table 2 allow for a preliminary assessment of the main features of the univariate distributions of the variables listed along the rows. In particular, the different rows refer to the following variables (detailed sources and definitions are in Appendix 2). 8 The ratios t I j it+1 /Y it and I j it+1 /Y it+1 are, respectively, planned investments and completed investments over actual sales in real terms; these two ratios, respectively, measure the dependent variable of the plans-scaled model and of the accounting-data model (see Table 1). The ratio t I j it+1 / t Y it+1 (i.e. planned investments over planned sales in real terms) is also considered in order to assess how close the patterns of planned investments are, regardless of whether actual or planned real sales are contained in the denominator; this ratio measures the dependent variable in the expectations model (see Table 1). In Table 2, these variables are considered at both aggregate level, j = a, and by component, j = m, f, nr and r. Finally, t g it+1 is the expected growth rate of real sales between t and t + 1, and u( t g it+1 ) it is the subjective uncertainty of firms regarding their demand in t + 1 as perceived in t. Table 2 are skewed to the right and mesokurtic (i.e. with means always larger than medians). This fact suggests long and fat right tails: a relatively small fraction of companies experience few, albeit large, investment episodes. 9 Apart from this common feature, all the other figures in Table 2 suggest that disaggregated capital goods show deep heterogeneity in their distributions across firms and over time (see also Bontempi et al. 2004).

All distributions in
Among the categories of actual investments made, the highest coefficients of variation (corresponding to the lowest variability) are those of buildings and intangibles, while the use of investment plans reverses this outcome, as the highest coefficients of variation are those of tangibles. This last feature is robust to the use of expected sales as the scale variable for plans. In general, the fact that tangibles prevail over intangibles in the measures of dispersion, or vice versa, depends on the use of plans or of completions in the numerator of the investment ratio, and not on the use of actual or expected sales in the denominator.
The percentage of zero completed investments (column labelled as 'Null obs.' in Table 2) is particularly high for R&D, buildings and non-R&D intangibles. The same percentages computed for plans are ordered differently: first buildings, then non-R&D intangibles and finally R&D. Again, the latter score is the same regardless of whether actual or expected sales are measured in the denominator of the investment ratios. 10 The average composition of investment spending in the ninth column of Table 2 suggests that machinery ( j = m) is the most relevant investment category, followed by buildings and R&D ( j = f and r, respectively). The use of planned rather than actual data increases the relevance of R&D over total investment because of the smaller number of zeros in R&D plans than in actual R&D spending. As far as the sources of variability (between, over time and idiosyncratic and I j it+1 are, respectively, planned and realized investments in each category j; Y it and t Y it+1 are, respectively, actual sales in t and expected sales in t for t + 1; t g it+1 is the expected rate of growth of sales in t for t + 1; u( t g it+1 )is uncertainty about future sales as perceived in t for t + 1. c Total standard deviation, SD, is decomposed in three % shares by using the appropriate sums of squared deviations: the time- . SDbe exploits the cross-sectional dimension of panel data and measures the permanent differences among individuals. The sum of the last two components (SDtd and SDre) is the within variability that exploits the time dimension of panel data and measures the transitory departures from individual averages due to both business cycle and the evolution of individual-specific characteristics. This decomposition is obtained from a STATA procedure (available upon request) written by the author and inspired by Sevestre (2002). within) are concerned, the largest share of the overall variability of tangibles is explained by the idiosyncratic-within component, while intangibles (non-R&D and, mainly, R&D investments) vary more between (i.e. across firms) rather than over time within firms. This fact is robust to the measureplans or completionsof investments.
As far as regards expected sales growth and uncertainty, that is, the two main explanatory variables of the models summarized in Table 1, results suggest as follows. Uncertainty does not seem particularly affected by time variability, as between variability explains about 50% of total variability. Although expected sales growth displays the greatest time variability, its share remains below 4%. The coefficients of variation of subjective uncertainty and of expected growth are very different, as they are, respectively, the lowest (i.e. the highest dispersion) and one of the highest (i.e. the lowest dispersion) of those reported in Table 2.
The evidence emerging from our micro data can be cross-validated by comparing it with the benchmark evidence for the USA, whose stylized facts have been introduced in Bloom (2007) as emblematic of his theoretical model for R&D. Table 3 reports the estimates of the second-order autocorrelation coefficients for the growth rates of sales and employees, and for the ratios of investment over sales, in Italy (using our micro data) and in the USA (using Compustat data). 11 To highlight the effect of zero R&D observations, we present coefficient estimates for Italy with and without zero R&D expenses.
The second-order autocorrelation of sales' growth is very close to zero in both Italy and the USA, suggesting that they are generated by short-memory MA(1) processes (roughly the same result is given with regard to employment growth). The ranking of the estimates for both machinery and R&D is qualitatively the same in Italy and in the USA, with R&D being largely more persistent than machinery. In addition, the drop of null R&D expenses makes Italy even closer to the USA.
Overall, our results in Table 3 are so close to those in Bloom (2007) that we can describe the Italian case by quoting his words for the US case: 'empirical evidence is that R&D rates change slowly over time, and are more persistent than sales growth, employment growth, or investment rates'. 12

Estimation results
As far as the econometric approach is concerned, even though expectations and plans-scaled models are not truly dynamic panel models, 13 shocks 1 it to investment plans are likely to be related to explanatory sales growth and uncertainty, as they are bound to interact in t with sales expectations and their range (i.e. uncertainty) formed for t + 1 (see notes a and b in Table 1). In addition, the endogeneity of the regressors can be further exacerbated by random measurement errors. Therefore, the identification process and consistent parameter estimation require instrumental variables for all three specifications.
The efficient GMM approaches of Arellano and Bond (1991), labelled as GMM-dif, and of Blundell and Bond (1998), labelled as GMM-sys, are appropriate estimators within our context. In the following, we support the choice of either GMM-dif or GMM-sys estimator by formally testing for the assumption of valid additional moment conditions used in the levels (the so-called Difference Hansen statistic, see, e.g. Bond, 2002). Furthermore, the use of the two-steps GMM estimators, besides being the most efficient, also delivers consistent parameter standard errors in the presence of heteroscedastic and autocorrelated residuals. 14 Regarding the choice of instruments, we always use lags from t−2 backwards of the explanatory variables, of disagreement, and of firms' indicators of belonging to a group and of being credit rationed.
The first three columns of Table 4 report the estimates of, respectively, the expectations, plansscaled and accounting-data models (see Table 1) for R&D investments (i.e. j = r). In particular, the first two models are estimated with GMM-dif because the Difference Hansen test always rejects the over-identification conditions used in the levels; the opposite happens with the accountingdata model, where the GMM-sys estimator is preferred because both difference and level orthogonality conditions are not rejected by the Hansen test.
Regarding the over-identifying restrictions tests reported in Table 4, the p-values of the Hansen test never reject the null of valid restrictions. 15 Models' residuals do not show any autocorrelation of the second-order or higher, suggesting that the shocks 1 are independently distributed over time: as quite often happens, the initial conditions regressor I j it /Y it prevents residuals from autocorrelation due to adjustment costs and other rigidities. However, the very short average time dimension (about 2.3-2.4 years per firm) suggests cautious assessments of model dynamics. The explanatory ability of all the models is good, as the R 2 indicators are about 70% in models using plans and expectations, and just a little lower (about 50%) using noisier accounting data.
Estimates markedly differ for some key parameters, depending on whether model (2) variables are measured using plans/expectations or accounting data. In particular, the additional delay effect and the caution effect are relevant and right-signed (i.e. β 4 > 0 and β 5 < 0) in the two models explaining The summary of the three specifications is in Table 1. b GMM-dif = Arellano and Bond (1991); GMM-sys = Blundell and Bond (1998); OLS-pooled = omits individual effects; OLS-dif = data in difference. Below the estimates, robust standard errors are reported in Italic. c Difference in Hansen statistic that tests the additional moment conditions used in the level equations. When the test rejects, in that column the GMM-dif approach is used, while the non-rejection entails the use of GMM-sys; see the row 'Estimator'. *Significant at the 10% level. **Significant at the 5% level. ***Significant at the 1% level.
plans/expectations, while with accounting data, the delay effect β 4 is not significant and the caution effect β 5 is wrong signed. In our view, these failures to support Bloom's predictions are related to the not significant estimate of the uncertainty parameter β 3 and to the significantly wrong sign of the estimate of the sales growth parameter β 2 . As expected, the use of actual rather than planned R&D investment can jeopardize the uncertainty effect, because the noise of too many sources of shocks hides the signal coming from subjective plans and expectations.
Regarding the wrong sign of β 2 when sales growth is measured using actual data, noisy demand shocks affecting sales growth data from t to t + 1 can mechanically induce a negative correlation between the dependent variable (where the actual sales in t + 1 are in the denominator) and the explanatory g it+1 . If the effect of sales growth on R&D plans is not very high (as suggested by the two estimates of β 2 in the models using less noisy plans/expectations data), the negative correlation mentioned above can reverse the sign of the β 2 estimate in the accounting-data model. 16 In order to better understand this issue, we modified the accounting-data model specification by measuring its dependent variable with I r it+1 /Y it instead of I r it+1 /Y it+1 , and by changing both the lagged-dependent variable, defined as I r it /Y it−1 , and the corresponding interaction term, defined as I r it /Y it−1 × u( t g it+1 ), accordingly. The GMM-sys estimate of β 2 in this modified accounting-data model, where the mechanical correlation described above is broken, goes from −0.023 (in the third column of Table 4) to a 10% significant 0.029 (close to the estimates from the plans/expectations models) and, while not significant, the estimate of β 5 is −0.042 rather than 0.093. This latter outcome, together with the β 5 estimates in the first two columns of Table 4 which clearly support the relevance of Bloom's caution effect, underlines the need to use plans and expectations rather than accounting data to prevent potentially misleading shocks from giving signals that have little or nothing to do with the decision of investing in R&D.
Regarding the not significant estimate of the uncertainty parameter β 3 in the accounting-data model, we have to acknowledge that our subjective measure u( t g it+1 ) is not very meaningful in this context, while it is so in the context of the plans/expectations models. Thus, the failure to effectively represent the direct uncertainty effect β 3 in the third column of Table 4 precludes any appropriate assessment of either additional delay and caution effects, β 4 and β 5 . 17 In fact, although the estimate of β 4 is within the range of those estimates obtained using the plans/expectations data in the first two columns, the volatility of the β 4 estimator with accounting data prevents its statistical significance; the same occurs in the case of the β 5 estimates of the modified accounting-data model described above. Therefore, models using plans and expectations data embody the best information, and call into question the use of the accounting-data model in assessing the effect of uncertainty on investment. Similar evidence, in the case of tangibles, is reported in Guiso and Parigi (1999) and in Bontempi, Golinelli, and Parigi (2010).
The outcomes in the first two columns of Table 4 clearly support Bloom's model, even though comparisons between specific parameter estimates of the expectations model (in the first column) and the plans-scaled model (in the second column) require further explanation.
First of all, in view of the foregoing discussion, it is not surprising that estimates of the expected demand growth effect β 2 markedly differ, and that the smallest estimate (0.010) derives from the expectations model. In fact, this latter estimatebesides being inversely related to the level of expected growthis also a function of the difference between the plans-scaled model estimate (0.028) and the share of R&D on sales levels (which can be seen from the formula in note 16). In other words, the β 2 estimate of the expectations model is lower than that of the plans-scaled model by an amount related to the R&D/sales ratio (the larger this ratio, the larger the difference between the two estimates). This feature suggests that the specification based on plans-scaled model is probably better, from a statistical point of view, than that based on the expectations model: the latter model estimates a β 2 parameter which is related not only to the elasticity of plans to expected sales growth, but also to the R&D/sales ratio which can strongly vary across firms. The estimates of the caution effect β 5 are almost equal in the two models (while those of β 2 are quite different); thus, we can predict that an equal increase in uncertainty will offset the effect of a positive demand shock to a greater extent in the case of the expectations model than in that of the plans-scaled model.
The estimates of the direct uncertainty effect β 3 are quite similar in the two models, and emphasize the negative effect on R&D plans due to an increase in uncertainty. If we measured the uncertainty with disagreement, rather than the subjective min-max range, the results of the estimation would change quite substantially (available upon request): the estimate of expected sales growth β 2 would become negative (from positive), andaccordinglythe sign of the estimate of the caution effect β 5 would become positive. The instability of the economic environment over the estimation sample makes uncertainty measured as disagreement to behave quite differently from uncertainty measured from the subjective expectations of managers (see Lahiri and Sheng 2010), and this discrepancy is likely to prevent those models using disagreement from encapsulating genuine uncertainty shocks. 18 The estimates of the initial conditions' effect β 1 are similar in both the expectations and plansscaled models, and denote the relevance thereof: more than 80% of investment plans for the following year are, ceteris paribus, in line with what happened during the previous period. In particular, firms confidently forecasting zero growth for their next year's sales, will plan on spending only a slightly lower amount on R&D than the amount spent the previous year. On the contrary, the estimates of the delay effect β 4 markedly differ, being significantly lower in the expectations model. This discrepancy is puzzling, because classical GMM diagnostics do not show any significant signs of any specification problems. A tentative interpretation could be based on the large shocksand uncertaintythat characterized the Great Recession, when the dynamic relationship between plans and actual investment could have been prone to be characterized by outliers and non-linearity; these effects are particularly difficult to identify within the short period in which they occurred.
Since GMM-dif estimators can be biased towards OLS-in-differences estimates in dynamic models with a high degree of persistence, and GMM-sys are biased towards the OLS-pooled estimates when initial conditions are related to the individual effects (see Blundell and Bond 1998;Bond 2002), the last two columns of Table 4 report the OLS-pooled and the OLS-in-differences estimates of the expectations model parameters, respectively. 19 As far as the estimates of inertia effect β 1 are concerned, they suggest that the bias of the GMM-dif estimator is not too strong, as the β 1 estimate lies between the upwardly biased OLS-pooled and the downwardly biased OLS-in-differences estimates. Under OLS-pooled, the omission of the individual effects, possibly positively correlated with the explanatory uncertainty, would upwardly bias the estimate of the β 3 parameter and, accordingly, downwardly bias the estimate of the caution effect β 5 . Estimates in the fourth column of Table 4 confirm this prediction, as both β 3 and β 5 signs are significantly wrong. This evidence is further validated by inspecting the corresponding estimates with OLS-in-differences in the fifth column, where the first-difference transformation clears individual effects: albeit not significant, the estimate of the uncertainty effect becomes negative and that of the caution effect is in line with the GMM estimates.
The vertical axis in the upper part of Figure 1 measures a dynamic delay effect, equal to b 1 + b 4 u( t g it+1 ), which grows with uncertainty (measured along the horizontal axis) because the estimate of b 4 in the plans-scaled model is equal to about 0.53 (see the second column of Table 4). For example, when uncertainty is at very low levels (1% corresponds to the 5th centile of its sample distribution), the delay effect is about 0.87 (which is very close to the estimate of b 1 in the plans-scaled model); this means that the delay effect leads to planned R&D investment for the following year that is about 87% of the R&D investment made this year. Instead, when uncertainty is very high (25% corresponds to the 95th centile of its sample distribution), the delay effect is about 1; this means that the delay effect leads to planned R&D investment for the next year that is equal to the R&D investment made this year. In other words, the grey upper area of the delay effect gets larger as uncertainty grows.
The lower vertical axis measures a caution effect, equal tob 2 + b 5 u( t g it+1 ), which decreases with uncertainty because the estimate of b 5 in the plans-scaled model is equal to about −0.12. For very low levels of uncertainty, the caution effect is about 0.028 (which is very close to the estimate of b 2 in the plans-scaled model), while for very high levels of uncertainty, the caution effect is about 0; this means that due to caution effects, R&D investment plans for the following year are unrelated to the level of growth in demand expected for that year. In other words, the lower grey area of the caution effect gets larger with uncertainty.
In order to cross-validate the ability of Bloom's model for R&D to explain other investment decisions as well, Table 5 compares the estimates of the plans-scaled model with R&D data (in the first column), with those corresponding to other investment categories (in the other four columns), namely total, equipment, buildings and non-R&D intangibles. Model diagnostics suggest the plans-scaled model is also statistically congruent with non R&D investment data.  Table 4. The upper vertical axis measures a dynamic delay effect, defined as b 1 + b 4 u( t g it+1 ), which grows with uncertainty (measured along the horizontal axis) because the estimate of b 4 in the plans-scaled model is positive. The lower vertical axis measures a caution effect, defined as b 2 + b 5 u( t g it+1 ), which decreases with uncertainty because the estimate of b 5 in the plans-scaled model is negative. On the horizontal axis, the levels of uncertainty (measured as the range between the max and the min expected sales' growth) are reported for a range from the 5th to the 95th centile of its sample distribution. The median uncertainty is equal to 7%, corresponding to the vertical line. Both the upper area of the delay effect and the lower area of the caution effect in grey get larger with uncertainty. Table 5 support the caution effects model, and the findings are quite robust across investment type: the sales growth parameter estimates are always significantly positive, and their interaction with uncertainty is almost always negative (with the exception of buildings, which also score the lowest reactivityamong tangiblesto expected demand one-year ahead). On the other hand, the findings regarding the delay effect β 4 are mixed, as non-R&D intangibles are the only case (besides R&D) in which Bloom's theoretical a priori is met. This outcome is consistent with Bloom's assumptions: higher uncertainty increases the responsiveness of investment to lagged investment as predicted by flow adjustment costs for intangibles (both R&D and non-R&D). 20 Since, vice versa, in the cases of capital and labour the most reasonable assumption is that of stock adjustment costs, there is no reason to expect the delay effect to actually materialize for tangibles, where one would rather expect the delay effect at the stock level, not at the flow (investment) level. Therefore, in the context of Bloom (2007), β 4 parameter for capital goods should be negative, and its estimates in Table 5 are actually negative for machinery, buildings and the aggregate estimates.

Some results in
As regards the effects of financial variables, overall such effects are significant, regardless of whether the expectations or the plans-scaled model is used (see Table 4), and regardless of alternative investment assets (see Table 5). The prevalence of significantly positive cash flow effect estimates suggests the presence of agency costs and asymmetric information. The positive relationship with debt in the case of intangibles seems to underline the role of borrowing in reducing those company resources freely available to managers (Panousi and Papanikolaou 2012). Note also that bank debt is the main source of financing in Italy, and its use can prevent companies from revealing information (such as that regarding R&D projects) to a large number of investors. See Bontempi (2002) and Munari, Oriani, and Sobrero (2010) for comparative analyses of R&D investment in control-based systems (such as Italy). For j = m (machinery, equipment, vehicles), f (buildings), nr (non-R&D intangibles), r (R&D), a is the sum of the previous four categories. b The estimates of individual and time dummy parameters are omitted. Below the estimates, robust standard errors are reported in italic. **Significant at the 5% level. ***Significant at the 1% level.

Concluding remarks
In this paper, we study the empirical determinants of R&D investment by implementing for the first time (to the best of our knowledge) the theoretical model introduced in Bloom (2007). The Italian case is interesting as it offers a wealth of firm-level data, that is, disaggregated investment plans and their implementation, as well as a subjective firmand time-varying measure of expected demand and the uncertainty thereof. Even though Bloom's model has been specifically designed to explain the effect of uncertainty on R&D investment, here the model is also applied to four types of capital goods: two types are tangible (buildings and machinery equipment), while two are intangible (R&D and non-R&D intangibles). This comparison is rather relevant, since different types of capital can be differently affected by areas of inaction due to irreversibility, fixed adjustment costs, asymmetric information and financial constraints.
In modelling data, a number of issues concerning the measurement of the variables in question, and the other specifications, are tackled. Our main findings can be summed up as follows. Given that the discrepancy between investment plans and their corresponding implementation is not simply due to unforecastable random errors occurring after the plans are made, the decision to estimate the model by using either investment plans or completed investments (and, accordingly, by using either sales expectations or their ex post accounting outcomes) leads to different results. The same is true regarding the alteration of the definition of uncertainty, that is, the use of either the subjective range of expected sales growth or the disagreement between individual expectations within individual economic sectors.
As shown within the context of tangible goods (see Guiso and Parigi, 1999;Bontempi, Golinelli, and Parigi 2010), also in the case of R&D, investment plans, sales expectations and their subjective uncertainty (min-max range) measure the theoretical determinants of firms' behaviour to a better degree than actual ex post data do, as the latter are affected by additional noisy shocks not necessarily related with the issue in question. Therefore, the significant estimates we found for Bloom's model determinants support the existence of flow adjustment costs for intangibles, while the partial failure to identify valid empirical representations for tangibles suggest that, as assumed in the theoretical framework, they are subject to stock adjustment costs. Therefore, our estimation results for investment in physical capital support the empirical specification of Bloom, Bond, and Van Reenen (2007) and of Drakos and Goulas (2010), whereby uncertainty only comes into the equation through its interaction with demand and/or directly.
As regards R&D investment, the response of companies to demand shocks is significantly lower at higher levels of uncertainty because of a caution effect: the chance to wait and do nothing is more valuable for those firms that encounter a higher level of demand uncertainty. Following a negative demand shock, investment is expected to be lower, as each firm will invest less (the so-called intensive margin). According to our data, very high levels of uncertainty may completely displace positive demand shocks, and for this reason lead firms to totally rely on their existing investment-sales ratios. Therefore, the main relevance of this work for policy purposes is that it shows how uncertainty delays R&D investment, and this fact implies that nowadays, R&D spending may be strongly penalized in countries such as Italy, where the level of uncertainty is currently rather high.

Disclosure statement
No potential conflict of interest was reported by the author.

Funding
This paper is part of a general research project concerning the relationships between investment and uncertainty at both micro and macro levels, conducted jointly with Magda Bianco and Giuseppe Parigi (Bank of Italy), Cecilia Jona-Lasinio and Fabio Bacchini (ISTAT), Jacques Mairesse (ENSAE), and Roberto Golinelli from my own Department; see Bontempi, Golinelli, and Parigi (2010), Bianco et al. (2013), Bontempi and Mairesse (forthcoming) and Bacchini et al. (2015).

Notes
1. The physical capital model of Bloom, Bond, and Van Reenen (2007, 399) includes an error correction term with a supposed positive parameter, so that firms with capital stock below their target level will eventually adjust upwards. The target capital stock is the stock a firm would have chosen as a function of real sales and the user cost of capital. In other terms, in the long run, it is the evolution of demand that drives investment, and not uncertainty and the gap between thresholds. Regardless of real options and irreversibility, uncertainty can only depress the expected long-term investment through its effect on the growth rate of demand (Bloom 2000). 2. Actually, the coefficient of the lagged-dependent variable times uncertainty for physical capital is not obvious, as emerged in a private conversation with Nick Bloom: the gap between actual and desired investment ratios depends on lagged-uncertainty (how reluctant firms are to make adjustment vis-à-vis their recent past), while companies suffer from current uncertainty; these two effects will be related, but the sign is not clear. In other words, the response to uncertainty can be spread out over time, imparting complex and persistent dynamics; investment could respond to both current and past demand shocks, so that firms with a recent history of positive demand shocks will be closer to their investment threshold and will be more inclined to invest. 3. Individual effects measure firm-specific irreversibility that can amplify the response of investment to uncertainty for a given capital good. Time effects account for panel cross-sectional correlation coming from individuals' reactions to macroeconomic events, neighbourhood/industry effects, herd behaviour and social norms. 4. The disagreement over expected growth rates of individuals by groups, se( t g it+1 ), is estimated by the standard deviation of t g it+1 within groups of individuals. 5. Given preliminary results available upon request, the outcomes of the tests for partial optimality with our data always reject the null of investment plans optimality, regardless of the alternative specification of the Mincer-Zarnowitz regression (proposed since the seminal work of Mincer and Zarnowitz 1969). The discrepancy between plans and their actual realizations after one year is predictable at the time plans are made, and this suggests certain factors (such as bounded rationality or information asymmetries) that could negatively affect the planner's thinking. Consequently, investment plans and implementation do not bear the same information. 6. Note that the subjective uncertainty u( t g it+1 ), being firmand time-specific, condenses the individual and timevarying expectations of managers about future demand, and thus embodies those favourable features advocated in Manski (2004). Moreover, u( t g it+1 ) is close to the definition given by Bloom, Bond, and Van Reenen (2007) of 'uncertainty about demand and productivity conditions'. In line with the findings of Lahiri and Sheng (2010), subjective uncertainty u( t g it+1 ) and disagreement se( t g it+1 ) markedly differ: our preliminary resultsavailable upon requestfrom the comparison of these two uncertainty indicators call for a great care in interpreting the outcomes of disagreement measures of variability, as they are prone to be extremely poor proxies, especially in the presence of an unstable economic environment (a case that, unfortunately, is often the rule rather than the exception). 7. Overall, the share of missing data in SIM covered by CADS information on financial variables is about 47%, while the share covered by NA information on deflators is about 29%. 8. Besides the full sample analysis presented in this section, Appendix 3based on SIM surveys for 2010 and 2011also reports additional information on certain firms' characteristics and their R&D spending, and on the sources of R&D financing which are useful during the empirical modelling phase. 9. Given that investment inaction followed by periods of intensive adjustment of capital stock may indicate a central role of non-convex (irreversibility) and of fixed adjustment costs that lead to lumpy investment (see, e.g. Doms and Dunne 1998; Barnett and Sakellaris 1998;Cooper, Haltiwanger, and Power 1999;Cooper and Haltiwanger 2006), Appendix 4 analyses the features of volatility, persistence and co-movements over time of investments at both macro-and micro-economic levels. 10. Note also that the shares of observations with zero investments in tangibles (i.e. machinery and buildings) are quite close to those reported in Bloom, Bond, and Van Reenen (2007) for the UK. 11. The source of USA estimates is in Bloom (2007, note 5, p. 252). 12. For an extended analysis of the cyclical features of investments in Italy, see Appendix 4.
13. In fact, the initial conditions regressor I j it /Y it , being made of actual investments and actual sales, differs from the genuine lagged-dependent variable of expectations and plans-scaled models t−1 I j it / t−1 Y it and t−1 I j it /Y it−1 , respectively. The variable I j it /Y it represents a 'true' lagged-dependent variable in the accounting-data model only, and as such must be instrumented. 14. Note that with our rather large sample sizes, the use of one-step GMM estimators or Windmeijer's finite-sample corrections to the asymptotic variance of the GMM estimators would deliver qualitatively similar results. 15. Note also that, being in the 20-35% range, these p-values are not too high to evoke a lack of power induced by too many over-identification restrictions; on this point see e.g. Bontempi and Mammi (2015). 16. Define, in the plans-scaled model, the change in R&D spending I r over the lagged level of sales Y due to a change in the sales growth rate g as b S 2 = ∂(I r it+1 /Y it )/∂g it+1 . The corresponding b 2 parameter in expectations and accountingdata models is related to b S 2 by means of the following formula: . Therefore, given a positive estimate of b S 2 , the precondition for a positive estimate of b 2 as well is that:b S 2 . (I r it+1 /Y it+1 ). 17. Note that qualitatively similar results are obtained if we measure uncertainty with disagreement rather than with the subjective min-max range. 18. In order to investigate this point further, we estimated the expectations model using both subjective uncertainty andalternativelydisagreement in two sub-samples (before and after 2008) characterized by different degrees of stability in the economic environment. Results (available upon request) are quite clear cut: while the estimation results of the model using subjective uncertainty show a remarkable stability across the sub-samples, those deriving from the model using disagreement show profound parameter breaks in the estimates of delay and caution effects β 4 and β 5 . The cyclical instability of the latter model emphasizes the problems occurring when the economic environment is unstable. 19. What is discussed here also holds true qualitatively for the unreported OLS estimates of the plans-scaled and the accounting-data models. 20. In particular, for non-R&D intangibles we have the extreme case in which the delay effect only comes about as a result of the interaction with uncertainty, as the estimate of β 1 is not significant while that of β 4 is both significant and very large. 21. The SIM sample is stratified by firm size (number of employees), branch of activity and regional location. Thanks to the country-wide coverage of Bank of Italy's branches and their continuous interaction with the local productive and financial system, the SIM achieves high response rates, ranging from 70% to 80%. Each time a survey was run, no respondents are replaced by other firm in the same branch and size class. Estimates of the distribution of investments by branch of activity deducted from SIM are similar to those obtained from official sources, such as the NA and ISTAT's Survey of Enterprises (see Bank of Italy 2005). Updated and very detailed descriptions on SIM sample design, response behaviour, data quality checks in each year are available at the Bank of Italy web site: http://www.bancaditalia.it/statistiche/indcamp/indimpser/boll_stat. 22. High-technology industries are Aerospace, Computer, Electronics, Pharmaceutical; Medium-High technology industries are Scientific Instruments, Motor vehicles, Electric machinery, Chemicals, Other transport equipment, Non-electric machinery; Medium-Low technology industries are Rubber and plastic products, Shipbuilding, Manufacturing n. e.c., Non-ferrous metal, Non-metallic mineral products, Basic metals and Fabricated metal products, Coke and refined petroleum products; Low technology industries are Pulp and paper, Textile-clothing, Food, beverages and tobacco, Wood. 23. According to the European Commission (6 May 2003) 'Recommendation 2003/361/EC: SME Definition' the category of SMEs is made up of enterprises which employ fewer than 250 persons and which have an annual turnover not exceeding 50 million euro, and/or an annual balance sheet total not exceeding 43 million euro. 24. The acquisition includes (a) preventive and proactive maintenance and the share of the corrective maintenance, invoiced by the suppliers, that could be capitalized by law and (b) production and repair of own capital goods made by the firm and capitalized. 25. The expected inflation is available for total tangible investment and for software expenses. For buildings and machinery (software and R&D) t p j it+1 is estimated by t p j it+1 = t p it+1 − (p nt+1 − p j nt+1 ), where p nt+1 = (PI nt+1 − PI nt )/PI nt is the total investment price inflation on goods f and m (nr and r), and p j nt+1 = (PI j nt+1 − PI j nt )/PI j nt is the j-type investment price inflation rate for j = f, m, nr, r. 26. For an international comparison of accounting principles on intangibles, see Stolowy and Jeny-Cazavan (2001).
27. An intangible asset should be recognized at cost if and only if it is identifiable, it is probable that specifically attributable economic benefits will flow from the assets, and its cost can be measured reliably. 28. On average, expensed R&D and patent-fees represent a mere 7% of good r. Operative and recurrent advertising is excluded from good r. 29. Actually, individual sales' deflators are obtained by applying the SIM growth rate for year t to the previous-year NA deflator level of the sector to which the firm belongs. We use NA sector deflator levels when SIM growth rates are not available. 30. This is available in SIM data set only for two years (1993 and 2005). In the 2005 survey, the question about the minmax range was substituted for by a more complex one on the firms' subjective probability distribution. For this, the min-max range data for 2005 are obtained as in Bianco et al. (2013) See Guiso and Parigi (1999) for the use of uncertainty based on the subjective probability distribution of respondents in 1993. 31. The same picture with industries ordered according to the percentage of sales over advertising shows shipbuilding, motor vehicles, rubber and plastic, petroleum, aerospace at the higher positions. Pharmaceutical, ADV/R&D/Com, scientific instruments display low percentages like the low-technology industries (food, paper, textile and wood). 32. In the 2010 and 2011 surveys, there was a question on whether the R&D spending would have been of the same or more or less amount without receiving public funding. The main part of the companies declared the same or a higher amount; from 2010 to 2011 the percentage of firms declaring a lower amount increased by 6 percentage points. 33. The firms which more frequently used public finance did not apply for new loans from banks or other financial intermediaries because they were convinced that their application would be rejected. 34. Some further descriptive analysis (not reported but available upon request) shows that a high percentage of firms collaborating with Italian universities was asked by the creditors for early repayment of loans granted in the past. 35. Details about the techniques we used and the full set of results are available upon request. 36. It is worth noting that such micro variability of intangibles is not merely due to the presence of many cases of individual zero expenditure in R&D, as the R&D variability would have further increased (and not decreased) if we computed it by excluding all these zeroes. This outcome suggests that the presence of spikes explains the large R&D micro variability, and that these spikes do not only occur when the expenditure was zero in the previous year. Conversely, the exclusion of zero R&D observations drops invariant observations that, if left in the sample, would deflate the estimate of variance. 37. Oppositely to what happens for variability, the exclusion from the sample of the observations with zero R&D increases persistence. The presence of zeroes negatively contributes to the estimate of the micro R&D persistence, as the negative contribution to covariance estimates (due to positive spikes occurring when the expenditure in t−1 was zero, i.e. below the average) is larger than the negative contribution to variance estimates (due to a number of zero expenditures).
The SIM survey is conducted annually since 1984 by the Bank of Italy. It is based on a representative 21 sample of about 1000 Italian industrial firms with 50 or more employees in each cross section; since 2001 manufacturing companies with 20-49 employees were added; and, since 2002, non-financial private service firms with 20 or more employees were also added to the survey, leading to more than 4000 companies in each cross section. Together with characteristics of the firms (location, ownership structure, industrial sector, year of foundation and so on), detailed information on investment and employment decisions in the two years preceding the interview and plans for the following year is also collected. In particular, disaggregated information on tangible investments (in machinery and in buildings) is available since 1992; investment in non-R&D intangibles (see their definition below) is included since 1999; expenses in R&D are available since 2003; investment plans started to be surveyed since 1996.
The SIM collects the firms' expected demand over the next year (and also over the next three years in the 2005 Survey), and a measure of its uncertainty (very detailed in the 1993 and 2005 Surveys).
The 1984-2012 SIM sample (after deleting few outliers) is composed by 9961 companies (3690 with employees from 20 to 49 and 6271 with 50 or more employees) for a total of 61,169 observations. Table A1 describes the distribution of SIM sample according to the sector of economic activity, the size and the age of the companies. Size is based on the number of employees (the exact definition is in Section A.4), and age comes from the year of foundation. We define 28 sectors of economic activity according to an elaboration of the NACE rev. 1.1 2-digit classification. In particular, Agriculture/ hunting and forestry, Fishing, Construction, Financial intermediation, Public administration and defence and compulsory social security Education (sections A, B, F, J, L, M) were excluded from the analysis (sections N, O, P, Q are not present in the SIM database); the manufacturing industries (Section D) were classified according to their global technological intensity at the 4-digits level; 22 Electricity/Mining includes Sections C and E; section K was disaggregated into Adv/R&D/Com (with computer and related activities also adding telecommunications from section I) and Other bus. serv (renting and consultancy).
Coherently with the characteristics of the Italian economic system, the prevailing industries are Textile, Fabricated metal products, Non-electrical machinery and equipment, Food, Wholesales. The majority of enterprises is SMEs, 23 usually unlisted and privately held (see also Bianco et al. 2013; Table 1).
The classification by industry reveals the predominance of manufacturing over services (80% versus 17%), and a minor role for the Mining/Electricity sector (3%); many firms were founded in the 1970s and 1980s; old firms prevail in the Manufacturing, while young companies in Services and Mining/Electricity.
Focusing on investment in both physical and intangible capitals restricts the available time span to the 2003-2012 period. Information on both realized and planned disaggregated investments is available for 4876 companies (15,823 observations). Such a reduction (74%) of the initial observations does not alter the sample composition by sector of economic activity, size and age.
SIM data set has a good overlapping with the CADS database which is available since 1982 for about 129,660 companies (more than 1,115,000 observations). The merge of SIM with CADS does not alter the representativeness of the sample; this result is quite important, as CADS furnishes  supplementary information to SIM (to initialize the perpetual inventory method and to compute financial variables). CADS is provided by Centrale dei Bilancia company set up jointly by the Bank of Italy, the ABI (Italian Banking Association) and other leading Italian banksand collects highly disaggregated balance-sheet, profit-and-loss data and flow of funds, as well as detailed information on characteristics of Italian companies operating in a wide range of industrial and service sectors. CADS is highly representative of the population of Italian firms, covering over 50% of the value added by those companies included in the Italian Central Statistical Office's Census. Further details of this data set can be found in Bontempi (2002).

A.1. Effective and planned investments and the capital stock
Realized and planned investments at current prices are available in SIM, disaggregated in four j types of goods: (1) j = f non-residential buildings; (2) j = m plants, machinery and equipment; (3) j = nr non-R&D intangibles; (4) j = r R&D. In particular, they are gross fixed investment (the depreciation is included) referring to the acquisition 24 of fixed capital to the firms' asset in the reference period; fixed capital derives from a production process and can be used repeatedly in the production of goods and services for more than one year.
Item j = f includes the buildings under construction and new-built, and the expenditure for the renovation of already existed plants; grounds and the used residential buildings are excluded from the figure. The investment in plant under construction is equivalent to the sum of the received invoices during the reference period from the contractors and/or the value of the plant construction built directly by the firm.
Item j = m includes plants, tools and machinery, hardware (available separately in 2005 and 2010 only), means of transport and used tangible assets (expenditure for physical second-hand assets which refers to the purchase of goods that was before used by other companies in their production process; neither the purchase of second-hand land and residential buildings, nor the goods involved in a company's merger or acquisition are included). For the under construction item, only the corresponding value of the sum of the received invoices during the reference period from the contractors and/or the value of the capital good directly set up by the firm is included.
Item j = nrnon-R&D intangiblesconsists of software, mineral exploitation, copyright on entertainment and works of literature and art. Software also includes the software realized in house (in this case, the development should be valued at an estimated price or, if it is not possible, at its production cost) and the expenditure for database that was used in the production for more than one year. Mineral exploitation includes also the test drilling, survey flights or other survey, transportation cost. In j = nr copyright protected entertainment, literary and artistic originals (such as movie, audio record, manuscript and model) are included. Note that patents, marketing and advertising costs are excluded from investment in j = nr.
Item j = r is expenditure on R&D, market research, design and test products; both the purchased services from an external company and the one developed in house are included; any costs for software development and expenditure on education and training are excluded.
For the ith company (i = 1, 2, … , N ) at year t (t = 1, 2, … , T ), we indicate with INV j it and t INV j it+1 the level of effective investment realized in t, and of the investment planned in t for t + 1, respectively; the superscript j ( = f, m, nr, r, a) indicates the type of good, with a denoting the aggregation over the four typologies. figures at constant (2000) prices are obtained as I j it = INV j it /PI j nt , where the NA sectoral investment prices PI j nt for all the companies belonging to nth industry are used as deflator for realized investments. In the case of plans, we use t I j is the expected inflation of the j-type investment price (estimated from the SIM source). 25 A problem in comparing Italy and major OECD economies is related to the different accounting systems as stated by the Generally Accepted Accounting Principles (GAAP): in Anglo-American countries R&D is expensed as it is incurred, rather than capitalized and depreciated, which means that the lifetime of the investment for accounting purpose is much shorter than the economic life of the asset created. 26 Notwithstanding criteria employed in recognizing intangible assets are similar to those of the IAS/IFRS, 27 Italian GAAP (as well as other European country with a Continental accounting basis) present a specific list of intangibles that should be capitalized (classified as fixed assets) in the balance sheet: in our context, item j = nr is mainly composed by deferred charges (like start-up and expansion expenses) and items (like patents and intellectual property rights, concessions, licenses, trademarks and similar rights) that are capitalized as assets and thus treated as valuable investments. As far as item j = r is concerned, it includes applied research and development costs, and advertising costs functional and essential to the start-up phase which are capitalized; basic research expenses and regular license-fees paid for patents which are instead recognized as costs when incurred. 28

A.2. Expectations and uncertainty about future demand
The level of sales at constant prices is Y it = SAL it /PY it , where SAL it is the value at current prices of the ith company's sales in t taken from SIM database, and PY it is the individual sales' deflator taken from NA. 29 Since the 1996 survey, firms are first asked for their projected percentage change in nominal sales from t to t + 1, then they are asked for the annual growth in their sales' price, and finally the percentage changes of sales in real terms t g it+1 are given by the difference of nominal growth minus the expected inflation.
Given the base prediction t g it+1 , firms indicate a range around this figure, that is, provide a minimum and maximum expected real turnover. We can obtain our uncertainty measure about the future demand growth u( t g it+1 ) defined as the min-max range of sales' expected growth rate reported by SIM respondents.
These data are available for an unbalanced panel of 33,919 observations for 7550 individuals (on average more than 1990 firms per year in the 1996-2012 period with average T equal to 4). Although less informed than the knowledge of the whole subjective probability distribution of respondents about future demand, 30 the main advantage of u( t g it+1 ) measure of uncertainty is that of being available for each year of the whole time span 1996-2012.
From the individual expected growth rates t g it+1 , we can also proxy the disagreement of expectations, labelled as se st ( t g it+1 ), with the standard deviation of the expected growth rates in sector-year clusters. The years are 17 (from 1996 to 2012), and the sectors are 28 (classified according to the NACE rev. 1.1 classification, described in Section A.1).

A.3. Financial variables
Cash flow, net of dividends paid. Individual data at current prices are derived from the CADS database: CD it = cash flow (item 9.14) minus dividends (item 7.6); this figure has been compared and updated by the information in SIM (fi12, fi14 and fi16). In order to obtain data at constant prices, CD it has been deflated using PY st (the by-industry production deflator from NA, see, e.g. Bond and Meghir 1994): Financial debt. It is defined, from CADS, as the sum of short-and long-term bank debt (items: 4.26-678-679-683 + 4.2-648-649-684) plus the sum of short-and long-term financial debt versus other financial institutions and group (items: 4.27-477-483 + 274-601-603-604-606-676-677 + 4.22-600-602-605-646-647). Missing information has been filled by SIM (fi11, fi13, fi15, fi21, fi26, fi36). As in Figure A1 compares the change in investment composition between the two sub-periods 2004-2007 and 2009-2012. The pattern of R&D is in line with OECD (2013, p. 39) and shows that investment in R&D by both the business sector and the services generally grows more than investment in physical capital, despite negative changes in financing availability and aggregate demand due to the Great Recession. This similarity is reached even if the measure of our expenditures in knowledge-based capital is not primarily or only wages (which tend to be stickier than other forms of business expenditures).
The sectors of economic activities are ordered, from top to bottom, according to increasing labour productivity; this placement does not mirror the high-technology intensity of manufacturing and services. 31 The figure shows a widespread increase in R&D percentage, especially in computer, non-electric machinery, electronics, motor vehicles and pharmaceutical, while aerospace increased the percentage of machinery. Pharmaceutical enlarged the R&D investment and gained a higher position according to the labour productivity; shipbuilding reduced R&D and increased labour productivity.
In order to better understand the nature of the investment in intangibles, we exploited some qualitative information available in SIM database and defined in Appendix 2. Results are reported in Table A3.   Table A3 compares the shares of firms declaring to engage in R&D (RED) in the two periods 2008-2010 (2010 survey) and 2009-2011 (2011 survey). The percentage is significantly lower in the second period, as evidenced by the 'Diff' and 'P-val' last two columns. In the following rows, the 'Diff' and 'P-val' columns compare the percentage of firms that filed for a patent (PAT), engaged a production process (PROD) or a product (PROC) or an organizational innovation (ORG) conditionally on the declaration of doing or not doing R&D. We also show the frequency of: (1) firms engaging or not R&D according to having a large percentage of high-level trained workers (Highperblue); (2) being credit rationed (RAT1); (3) the ownership composition (being a family firm, FA, or being part of a group, group); (4) being present abroad (INT and INTRED). Firms investing in R&D produce more frequently products' innovation, followed by process and organizational innovations. This rank does not change in 2011, while the corresponding frequencies are reduced in accordance with the lower percentage of firms doing R&D. Firms spending in R&D have also a large percentage of university graduates among entrepreneurs, senior and junior managers, and production workers and apprentices; these companies belong to a group and have decided to internationalize their activity. Table A4 compares the operational and financial sources of R&D activity in two years, 2010 and 2011, and depending on the ownership structure (family or group) and whether a firm is credit rationed. figures show that the largest part of R&D is in-house produced and internally financed. Debt is the second source of financing, followed by public funding that however is comparatively lower. 32 Firms using more public financial support are those credit-constrained. 33 Since 2008, the enduring economic crisis reduced the yet infrequent use of equity and venture capital to finance innovation, independently from being or not credit-constrained. Collaboration with Italian Universities or Research Centres is higher for not credit-rationed companies. 34 time (2003-2012 for R&D) spanned by SIM has been tackled by backward extending SIM data with CADS, as the latter provides disaggregated information for different types of investment since 1982 (see Appendix 1).
As far as the cyclical volatility is concerned, macro aggregated investment is almost four times more volatile than GDP. Among the components, real R&D expenditure is the least volatile, with estimates significantly lower than those for machinery, non-R&D intangibles and, to a lesser extent, buildings. The averages of individual micro cycle indicators confirm the macro outcome for R&D and non-R&D intangibles, while for tangibles the variability is slightly lower than the macro one. Overall, macro data suggest that investments in machinery and non-R&D intangibles are noisier than those in buildings and R&D, while micro data highlight the larger volatility of R&D and non-R&D intangibles. The latter result is greatly amplified if we pool together individual micro data to estimate cyclical indicators rather than averaging individual indicators: the volatility of non-R&D intangibles is almost 70 times that of real sales growth, and the volatility of R&D is even bigger. The variances estimated by pooling micro data show a volatility of intangibles which is vastly larger than that of both intangible at the macro level and tangibles at the micro level. 36 As far as the persistence over the cycle is concerned, macro R&D spending over GDP is the most persistent: we can sayin cointegration jargonthat R&D and GDP levels do not share a common trend. Therefore, the share of R&D on GDP does not revert very often to its historical mean over time, and this fact induces large estimates of the autocorrelation coefficients. Micro data confirm the significant persistence found at macro level, although at lower levels (about less than half of macro estimates). The large number of observations available in the longitudinal dimension, together with the short temporal span, explains the lower levels of autocorrelation (persistence), and their high significance. Across categories, the aggregate investment is more persistent than its components (both tangible and intangible), and this difference tends to increase with the number of zeroes in the disaggregate series: investment ratios of buildings and R&D are less persistent than those of machinery and plants. 37 As far as co-movements of investments with cycle indicators are concerned, macro outcomes suggest that tangible investments are pro-cyclical and coincident with GDP. Micro data confirm this cyclical features of tangibles, while present a less evident picture for intangibles which partly contrasts with the clearer a-cyclicality of macro R&D and non-R&D intangibles. Overall, we can label intangibles as a-cyclical, although this view is more subject to lacking clear-cut evidence than that for tangibles.