RDP 2025-01: Are Investment Tax Breaks Effective? Australian Evidence 5. Identification Strategies

Nu Nu Win, Jonathan Hambur and Robert Breunig

February 2025

To identify any effect of the ITBs we use both a DD approach and an RDD approach. As explained below, each identify different impacts and rather than focus on one, we use both to see if there are any effects of the ITB policies. Our assessment of the impact of the policies is based on an examination of the combined evidence across these approaches.

5.1 Difference in differences

Our first approach is a DD model. We exploit the timing of the policies, as well as the exogenous qualifying turnover thresholds, to identify the impact of the policy, comparing qualifying and other firms during and outside of the policy period. The treatment and control groups are summarised in Table 3.

Table 3: Summary of DD Identification Strategies
Policy Turnover Dates
Treated Control
GFC 2009 < $2 million $2 million to $5 million Sep 2009 to Dec 2009
Small 2012 < $2 million $2 million to $5 million Sep 2012 to Dec 2013
Small 2015 < $2 million $2 million to $5 million Sep 2015 to Dec 2015
Medium 2016 $2 million to $10 million $10 million to $20 million Sep 2016 to Jun 2017
Medium 2019 $10 million to $50 million $50 million to $60 million Jun 2019 to Mar 2020
COVID 2020 $50 million to $500 million $500 million to $600 million Mar 2020 to Jun 2020
COVID 2021 $500 million to $5 billion $5 billion to $6 billion Dec 2020 to Jun 2021

Our regression takes the following form:

(3) I n v i , j , t = α + β 1 d t P + β 2 d i T + β 3 ( d t P × d i T ) + θ i + γ X i , j , t + D t T + I N D j × D t + ε i , j , t

where Invi,j,t is our measure of investment (intensive or extensive margin) for machinery & equipment for business i in industry j in period t, d t P is a dummy that equals one in the periods of operation of the relevant policy, and dummy d t T is equal to one for businesses that meet the relevant turnover test. We also include a number of controls in the model to help with identification and with precision. First, we include firm fixed effects θ i to capture time-invariant firm-specific factors influencing investment. We also include additional time-varying, firm-level controls to capture other factors that could be influencing investment, captured in Xi,j,t. Specifically these are full-time equivalent employment and deciles of the distribution of firm turnover. We also include a vector of quarterly dummies, D t T , intereacted with dummies for the treated and control groups, to control for seasonal patterns that might affect investment. And we include industry*time fixed effects (INDj × Dt) to capture heterogeneous industry-level trends and conditions, which might be particularly important during the GFC and COVID-19 episodes as different industries were differentially exposed to these shocks. As such, we are effectively comparing small and large businesses within the same industry.[8] Standard errors are clustered at the firm level for intensive margins and industry levels for extensive margins.[9]

To define a firm's eligibility we turn to the tax code. The code states that a firm is small if their revenue in a period is under a certain threshold. There are several different periods that can be used for this assessment under the tax code eligibility rules:

  • revenue in year t;
  • revenue in year t – 1; or
  • revenue in year t – 2 if the firm had a reasonable expectation when it undertook the investment that revenue in year t would be below the threshold.

This final category, which combines revenue in year t – 2 with expectations about revenue in year t, was designed for firms that may have had a windfall gain in year t – 1 and for whom revenue in year t – 2 is more ‘representative’ of their normal revenue. Given firms may not know time t sales during the year when investment decisions are made, they are allowed to rely on a reasonable expectation. We don't use firms that rely on this rule for the DD regressions, but they are used for the RDD identification in combination with information about actual turnover in time t and t – 1.

For the purposes of the DD regressions, we use revenue in year t – 1 in assessing whether the firm qualifies for the policy, as in Rodgers and Hambur (2018). As discussed below, for our RDD regressions we take a different approach, exploiting the third category noted above to get very tight identification.

One concern with identifying the effects of these ITB is that some coincided with changes in the corporate tax rate for firms based on the same threshold. Table 2 shows the changes in the tax rate over time. Of particular concern, the changes in 2015/16 and 2016/17 coincide perfectly with the 2015 and 2016 ITB. The reduction in the corporate tax rate would have simultaneously lowered the value of the ITB for companies, and potentially stimulated additional investment. As such, this may affect our estimated effects for these policies. To assess the degree to which this might be affecting our results, we provide some additional tests for outcomes across assets that were and were not eligible for the ITB (see Section 7.1).

5.2 Regression discontinuity design

Another way to exploit the exogenous eligibility threshold is to use a RDD approach. Similar to DD, this involves looking at differences in outcomes for eligible and ineligible firms. But instead of comparing outcomes more broadly between these groups, it focuses on an area ( η ) around the policy threshold cut-off, comparing outcomes for firms just above and just below the threshold to determine whether there is a discontinuity at the threshold.

The advantage of this approach is that the identification is sharper and it requires relatively weak assumptions – namely, that in absence of treatment there would be no discontinuity at the threshold.[10] However, the approach requires more data and the identified effect, while sharp, is very local and can be hard to extrapolate to an aggregate effect. Given the data requirements, we use the BAS data described above for the RDD estimates.

For identifying the relevant cut-off we adopt the approach of Rodgers and Hambur (2018). This approach focuses on firms that would qualify based on whether their turnover was above or below the threshold in year t – 2. This helps to minimise any possibility that firms could manipulate their way into treatment.

To isolate these firms, we remove firms that could qualify for the incentives based on their incomes in tax year t or t – 1 (i.e. businesses with turnover under the turnover limit in those years). Amongst the remaining firms, the only firms that could use their income in t – 2 to meet eligibility are those that had a reasonable expectation that their income in t would be below threshold. The tax rules note that the best way to assess this is revenue in t – 1 (Australian Taxation Office undated). That is, only firms with revenue a little bit above the threshold in t – 1, could say they reasonably expect to be below the threshold in t.

To meet this final requirement, we remove firms with income in t – 1 above a certain threshold. We use a ceiling of 1.25 times the threshold. This is equivalent to Rodgers and Hambur (2018) using a ceiling of $2.5 million for the $2 million threshold. While somewhat arbitrary, Rodgers and Hambur found estimates tended to be quite robust to the exact choice of threshold. The selection rules we use to choose our estimation sample are summarised in Table 4.

Table 4: Summary of RDD Identification Strategies
Turnover in t – 2 Turnover in t – 1 Turnover in t
GFC 2009, Small 2012, Small 2015
Dropped from sample   < $2 million < $2 million
Policy $2 million – η to $2 million $2 million to $2.5 million
Control $2 million to $2 million + η $2 million to $2.5 million
Medium 2016
Dropped from sample   < $10 million < $10 million
Policy $10 million – η to $10 million $10 million to $12.5 million  
Control $10 million to $10 million + η $10 million to $12.5 million  

In summary, we focus on the subset of firms who had turnover above the threshold in year t, and between the threshold and 1.25 times the threshold in year t – 1. This group would qualify for the ITB only if their turnover was below the threshold in t – 2, but otherwise would not. So we have a clean eligibility cut-off at the threshold to separate eligible and non-eligible firms based upon their revenue at time t – 2.

The RDD can be thought of as estimating a local polynomial model on either side of the threshold to estimate the conditional mean precisely on each side of the threshold.

As noted, for the regression we focus on a small region either side of the cut-off. To select this region we implement the automatic bandwidth selection procedure proposed by Calonico, Cattaneo and Titiunik (2014) with a triangular kernel. We also also use their bias adjusted standard errors.[11] The results are robust to selecting different bandwidths. For all RDD regressions we cluster standard errors at the firm level.

Footnotes

Industry is measured at the division level and takes 19 possible values. The inclusion of these industry dummies will also control for differing asset lives across industries, which determine how beneficial the various policies will be for different industries. [8]

We focus on a static DD model for ease of exposition. Using dynamic DD models that allow for differing effects over the treatment periods leads to similar conclusions. [9]

For discussions on the theory and practice of RDD, see Imbens and Lemieux (2008) and Lee and Lemieux (2010). [10]

We implement this using the rdrobust package in Stata. [11]