RDP 2024-01: Do Monetary Policy and Economic Conditions Impact Innovation? Evidence from Australian Administrative Data 3. Data and Methodology

Omer Majeed, Jonathan Hambur and Robert Breunig

February 2024

3.1 Data

We focus on four different measures of innovative activity.

The first three are narrow measures of innovative activity that have been considered in the earlier literature.

  • The (log) flow of new patents filed in Australia by Australian residents from IP Australia's IPLORD database.[3] This is a fairly narrow measure of innovative activity that is more likely to capture the creation of ‘new-to-world’ innovation and which has previously been linked to economic growth (Atun, Harvey and Wild 2007).
  • The (log) flow of new trademarks filed in Australia by Australian residents from IP Australia's IPLORD database. This is a slightly broader measure of innovation that will likely capture ‘new-to-firm’ innovation (Mendonça, Pereira and Godinho 2004; Malmberg 2005).
  • Aggregate (log) R&D spending data from the Australian Bureau of Statistics (ABS) national accounts. R&D has been linked to both higher novelty of innovation and adoption of innovation (D'Este, Amara and Olmos-Peñuela 2016; Majeed and Breunig 2023). As such, this is a slightly broader measure of innovative activity that will capture spending on the creation of new-to-world innovation, as well as potentially some spending to adopt innovations.

All three variables are observed quarterly from March 1994 through December 2019.

Our fourth measure of innovative activity is a survey measure of innovation collected in the ABS Business Characteristics Survey (BCS). This is a broad measure of innovation based on the Oslo Manual (OECD and Eurostat 2018), the OECD benchmark for innovation measurement. It captures around 8,000 firms each year from 2005/06 to 2019/20. Each year firms are asked whether they introduced new or significantly improved: goods or services; operational processes; organisational/managerial processes; or marketing methods. This measure includes adoption of existing technologies or processes, which is an important mechanism in models such as that of Moran and Queralto (2018). It is also particularly important in Australia, which tends to be an importer of innovation: around 5 in 100 firms in the survey report introducing a new-to-world innovation, whereas 50 in 100 firms report some form of innovative activity.

The BCS is a census of firms with more than 300 employees and a stratified random sample for firms with less than 300 employees. Stratification by industry and business size is implemented to produce data that are representative of Australian businesses. The ABS do not provide the sample weights so we use unweighted data. As such, when reporting results across all firms, rather than by size, large firms will be somewhat overweighted.

Firms with fewer than 300 employees are included on a rotating five-year basis. This could create some biases in our results, particularly if small and large firms respond differently to shocks, as the share of large firms in our sample will be larger as we consider longer time horizons. Taking any given year as a base period, only two-fifths of small firms will still be in the sample in four years' time. Splitting the sample into small and large firms helps to limit any potential bias.[4]

Another potential bias could come from attrition. If monetary policy shocks cause firms to exit, and if the remaining firms are more likely to be innovative, it may appear that the share of firms innovating has increased, but this result would just reflect a positive survivorship bias. To avoid this issue, our main regressions focus on firms with at least five observations. These are firms who do not exit during their period in the sample. Our firm-level results thus measure the intensive margin effect of monetary policy on firm innovation only. That said, the results are very similar if we do not impose this restriction. We also directly test whether monetary policy affects the probability of firm exit, and whether the effect differs by innovator status. We do not detect any difference between innovators and non-innovators in terms of exit post-shock.

We exclude microbusinesses from our study by removing all firms with one full-time equivalent (FTE) employee or less. Excluding micro firms is standard practice in the literature.

The BCS data are available at the firm level and are linked to the ABS Business Longitudinal Analysis Data Environment (BLADE), which contains demographic and tax data from administrative sources. This allows us to model innovation at the firm level, accounting for firm-level covariates and potential heterogeneous effects across firm types. See Tables A1 and A2 for descriptive statistics for the firm-level sample.

3.2 Methodology

We do not explicitly develop our own theoretical model for the analysis but the channels of monetary policy that we consider are those of the models developed in other papers such as Moran and Queralto (2018) and Ma (2023). Our econometric methodology is motivated by testing the implications of such models.

3.2.1 Monetary policy shocks

An inherent difficulty in examining the effect of monetary policy on innovation is that the official cash rate will be endogenous. That is, innovation activity and monetary policy are co-determined by other factors. For example, the central bank is likely to raise rates if it expects economic activity and inflation to increase. But improvements in economic activity might also spur further innovative activity. As such, it might appear that higher interest rates lead to more innovation when both are moving with economic conditions.

To get around this endogeneity issue we use a monetary policy shock measure developed in Beckers (2020).[5] This is a Romer and Romer (2004)-style shock, in that it measures shocks as divergences of the observed policy rate from what would be expected based on an estimated policy reaction function. This approach is widely used in the literature (Ramey 2016).

Specifically, Beckers (2020) estimates an augmented Taylor rule that includes a forecast for economic conditions and a number of indicators of financial conditions (e.g. bond spreads, option-implied volatility). The shocks are then constructed as the deviation of the actual policy rates from that implied by the rule. As such, this approach removes the anticipatory component of monetary policy by purging the changes in the policy rate of the central bank's systematic response to its own forecasts. We use the continuous measure that also orthogonalises the shock with respect to market expectations for the policy rate, though our results are near identical using the measure without this step.

We adopt this measure as our preferred shock measure because previous research found that it is able to overcome the so-called price puzzle in Australian data: that contractionary monetary policy is often estimated to raise prices. This is not the case for simpler Romer and Romer-style shocks (Bishop and Tulip 2017) or measures based on high-frequency changes in bond yields over a 90-minute window around announcements (Hambur and Haque 2023). We nonetheless consider these other measures for robustness in Tables B1–B4.

3.2.2 Estimation

We employ a local projection regression (Jordà 2005) to trace out the effect of a monetary policy shock at time t on our measures of innovative activity over a number of different time horizons h. This is a common approach in the literature (Ma 2023; Jordà et al 2020; Durante, Ferrando and Vermeulen 2022). Our regression takes the following form:

(1) In n i,t+h = β h shoc k t + α h In n i,t−1 + ∑ j=1 1 x h,j Control s t−j + v i,t+h

where h ≥ 0, denotes the time horizon, t denotes time and i indexes the firm (for firm-level regressions). We also estimate Equation (1) using aggregate variables which we can represent identically to the above, simply dropping the i subscripts. Inni,t+h is our measure of innovation. This is either at the aggregate level for (log) R&D, (log) patents or (log) trademarks, or at the firm level for our 0/1 indicator of whether a firm innovates. shockt is our continuous measure of monetary policy shocks described above. Typically, local projections control for macroeconomic variables to improve efficiency (Jordà 2005). We control for standard aggregate measures: gross domestic product (GDP), the consumer price index (CPI) and the trade-weighted index (TWI) exchange rate. For our firm-level regressions we also control for: (log) employment, turnover growth, (log) capital expenditure and age (a dummy if a firm is more than four years old). We include these with one lag.[6]

For firm-level regressions we cluster the standard error at the period level, reflecting the fact that the key variable of interest varies across time but not across firms. As a result, we have a small number of clusters, which can bias our standard errors downwards. To address this, we assess significance based on a T ‐distribution with t – n degrees of freedom, where t is the sample length and n is the number of coefficients on variables that do not vary across firms, as discussed in Cameron and Miller (2015). We do not allow for serial correlation as this is captured in the control variables.

Footnotes

See <https://data.gov.au/dataset/ds-dga-d61374dd-1a62-4132-913d-35d90cfdac81/details>. [3]

As a robustness test, we also estimate the models including firms in the regression for horizons 0, 1 and 2 only if we could observe their innovation at horizon 3. This ensures that the share of large and small firms remains balanced. The results by firm size are similar to our baseline, though there is slightly more evidence of an immediate effect. [4]

For a chart of this shock measure see Figure A1. [5]

The results are robust to including more or fewer lags of the RHS variables, as well as to including the contemporaneous controls, which imposes the implicit assumption that monetary policy cannot affect current conditions (Ramey 2016). Exclusion of the control for age does not affect the results. [6]