Australia's 1990s productivity surge: a response to Keith Hancock's Challenge *.

In the penultimate issue of the Australian Bulletin of Labour, Keith Hancock challenged the view that Australia's productivity growth improved markedly in the 1990s (Hancock 2005). He found evidence for a surge in productivity growth to be 'far from compelling' and concluded that, if there has been a surge, 'it is certainly weak'. He considered that to make a surge in Australia's productivity growth 'a significant ingredient of one's understanding of recent economic history is to rely on faith, rather than evidence'.

This is quite a challenge to analyses undertaken independently by a range of government agencies (such as the Australian Treasury, the Reserve Bank of Australia and the Productivity Commission), international organisations (such as the OECD and the IMF) and academics (such as Steve Dowrick, Mark Wooden and John Quiggin (1)). Analysts have employed a variety of methods, from simple growth rate comparisons to sophisticated econometric techniques. Whilst they differ in technique and detail of application, they all have provided evidence of an acceleration in Australia's productivity growth from the late 1980s or early 1990s.

Hancock based his alternative view on his own statistical curve-fitting exercise. He used the same source of productivity estimates--the official ABS estimates--that most of the institutions and analysts have used. Hancock's challenge and its validity therefore have nothing to do with the 'data', but everything to do with how the data are interpreted and analysed.

It matters whether Australia's productivity growth improved in a 'structural' sense in the 1990s. A structural improvement in productivity growth raises the prospects for ongoing improvements in Australian living standards. Whether or not there has been an economically-significant surge matters in a policy context, given the additional policy focus that has been placed on lifting Australia's productivity growth over the past two decades.

In this response, I take issue with Hancock's methodology and its ability to test the productivity surge thesis. I also point to the evidence that there has indeed been a productivity surge.

The Hancock Challenge

Hancock analysed observations on annual growth in both labour productivity (output per hour of labour input) and multifactor productivity (output per unit of combined labour and capital input). The observations are from series that run from 1964-65 to 2003-04.

First of all, Hancock noted that annual productivity growth shows considerable volatility and that the observations depict no obvious time pattern over the period. He highlighted the need to look beneath the volatility to examine underlying trends and chose to do this by fitting a straight line to the observations. Hancock reported that the fitted line has a slope that is not statistically different from zero--a result that he took as confirmation of the impression that there has been no overall increase in the rate of productivity growth since the mid-1960s.

Hancock acknowledged that this did not provide a basis upon which to assess whether productivity growth accelerated in the 1990s. He stated that the surge thesis 'does not necessarily entail a continuous rising trend across the period. Rather, it asserts a 'speeding up' over some or all of the period since the late 1980s or the early 1990s.

To do the surge thesis 'justice', he then fitted several non-linear curves of quadratic and cubic form. Non-linear forms could allow the fitted curves to exhibit an increase in productivity growth over some part of the observation period.

Hancock found that such fitted curves do not differ with statistical significance from a straight flat line and that they provide poor explanations for the variation in the observations as measured by the goodness-of-fit ([R.sup.2]). On this basis, he concluded that the evidence for a surge in the 1990s is, at best, weak.

Before commenting specifically on Hancock's method as a test of the surge thesis, it is worth taking another look at the data.

Another Look at the Data

Hancock is quite right in highlighting the need to look beneath the volatility of annual changes for evidence of underlying trends in productivity growth.

A lot of the volatility is spurious and uninteresting from an underlying trend point of view, since it reflects the influence of droughts, recessions, recoveries and other short-term factors.

There is a variety of ways to take out short-term 'noise' from series. A simple moving average is one technique. More complex methods have also been devised to filter out cycles of various lengths--for example, to take seasonal influences out of quarterly or monthly series or business cycle effects out of annual series. Hancock's approach of fitting a straight-line, quadratic or cubic function can also be viewed as a method of filtering out a lot of variation to expose underlying change. At issue, though, is whether this approach takes out not only the short-term spurious variation, but also meaningful longer-term variation that might occur over say 5-10 years.

The ABS uses a fairly simple but well regarded approach to filter out noise in the official estimates. It uses a Henderson 11-period moving average to create a 'trend' or underlying productivity series in which spurious variation over a few years is filtered out, but meaningful variation over longer periods is preserved.

The important difference between the ABS trend series and the original series can be illustrated by presenting the annual changes in productivity from the two series. Figures 1 and 2 (for labour and multifactor productivity respectively) use dots to show the annual changes in the raw series, as used by Hancock, and bars to show the annual changes in the ABS trend series. What might be difficult for the eye to pick up in annual changes in the original series is easy to pick up in annual changes in the trend series. A deceleration in trend productivity growth from around the midor late-1970s and an acceleration from the late-1980s or early-1990s are plain to see. (2)

But is this evidence of a productivity acceleration in the 1990s just an artefact of the 'filtering' process? No. In fact, the evidence of a productivity acceleration in the 1990s can be readily found in the original raw data.

Table 1 shows the mean values of, and the standard deviations in, the annual changes in the original series. The observations are grouped according to what are now conventionally regarded to be three stages in Australia's productivity growth in the post-WWII era--rapid productivity growth in the 'golden age' of the 1960s and early 1970s, a slowdown through the late 1970s and 1980s and then a surge in the 1990s. The marked differences between these periods are evident in Table 1. The means show the high-low-high growth pattern across the three periods. The standard deviations show another feature of the most recent period--much lower volatility in annual growth. This evidence constitutes a strong prima facie case for the 1990s productivity experience to be viewed as 'structurally' different.

Incidentally, an international perspective on these three periods also suggests that there has been a structural shift in Australia's productivity performance since the early 1990s. Although Australia s productivity growth was high in historical terms in the golden age, it was below the OECD average over that period. It remained below the OECD average in the late 1970s and 1980s when, as in Australia, there was a productivity slowdown in the OECD generally. After 1990, however, Australia started to out-perform the OECD average, suggesting that Australia tapped a productivity source that other advanced nations, in general, either did not tap in the 1990s or had already tapped some time before the 1990s.

The ABS also provides another method of abstracting from spurious short-term variation in order to detect changes in underlying trends in the original series. It calculates average rates of growth from peak to peak in successive productivity growth cycles. Use of the same point in successive cycles ensures that short-term influences are standardised at the start and end points of periods used for the measurement and comparison of underlying trends. Peaks are identified as points in the original series where the actual level of productivity reaches a (local) high point above trend. The ABS estimates of annual average rates of growth over productivity cycles are published in the annual national accounts.

In the last productivity growth cycle that the ABS has identified--1993-94 to 1998-99--productivity growth was at record highs. Average annual labour productivity growth was at 3.2 per cent over this period or 1.6 percentage points higher than the average across cycles starting in 1981-82 and finisning in 1993-94. Average annual MFP growth was at 2.0 per cent in the period or about 1.2 percentage points higher than in the 1980s and early 1990s.

In summation, three ways of examining the data--annual changes in a trend series (as formed by the ABS with a conventional filtering technique); means and standard deviations for three sub-periods in the series; and annual average rates of growth over productivity cycles (as published by the ABS)--all show clear evidence of a speeding up in Australia's productivity growth in the 1990s.

On Methodology

In fairness, and lest I be accused of hypocrisy (3), lines and curves are sometimes fitted to data as a summary descriptive device. For example, when observations appear at least vaguely to form a line or a curve, a fitting exercise can capture the extent to which they do indeed form a line or curve; or, if the goodness-of-fit is generally high, the exercise can highlight observations that lie outside the regular pattern.

But it is a different matter when poorly-fitting lines and curves that are essentially arbitrary (bearing no theoretical connection to a process that could generate the observations) are used as the basis for hypothesis testing.

Hancock attempted to test the 1990s surge hypothesis in the context of fitting curves to nearly 40 years of data. Implicitly, the actual hypothesis tested was that Australia s productivity growth since the mid-1960s conforms to a quadratic or cubic form.

First, is there any theory that could underpin the notion that a country's long-term productivity growth describes a quadratic or cubic pattern over 40-year timeframes? It is very difficult to think of one.

Second, the implicit hypothesis is rejected by the data. The coefficients are insignificant and the fit is poor (and there are most likely other problems, such as serial correlation in the residuals). Given the first point, it is not surprising at all that the data reject this view of very long-term productivity growth.

Third, a rejected 'model' of four decades of productivity growth simply does not provide a basis for testing the significance of a change in productivity trends from one sub-period to another.

There are several ways in which the surge hypothesis could be validly tested. Given that an increase in productivity growth in the 1990s is there in the data, the issue is not whether the surge exists but whether observed increases in productivity growth are part of a systematic increase in productivity growth or merely some random short-term shifts. A simple start would be to test the statistical significance of any difference in the patterns of productivity growth in the 1980s and in the 1990s. A modelling approach would require the specification of a model that has some connection to a plausible theory and that is capable of capturing and testing the significance of any change in productivity trend. The stability over time of estimated parameters (or the behaviour of residuals) from the model could be investigated for structural breaks. This could lead to more elaborate specifications to allow for breaks, or even gradual change, in productivity growth and provide a basis for valid tests of the statistical significance of a surge.

Previous studies of Australia's productivity growth in the 1990s have taken a number of approaches, including specification of econometric models along these lines. All have found an economically and statistically significant acceleration in productivity growth in the 1990s. (4) The weight of evidence favours an acceleration of 1 percentage point in annual productivity growth.

Hancock's method could also be questioned on purely practical grounds. Although Table 1 gives some support for a 'U-shaped' view of Australia's productivity growth record, Figures 1 and 2 suggest that the reason that a quadratic curve, fitted to the entire period, does not differ significantly from zero probably has a lot to do with the fact that the deceleration phase at the end of the 1970s started (in trend terms) from around 1974-75 and the acceleration phase in the 1990s ended around 1996-97. As a consequence, the inclusion of observations prior to 1974-75 and after 1996-97 would tend to 'flatten out' the fitted curve, compared with the case in which those observations are excluded. That is, the shape of the fitted non-linear curve, and its ability to reflect the acceleration in productivity growth from the 1980s to the 1990s, is likely to be sensitive to the selection of data to include in the estimation. (Even though a cubic form might address this issue, at least to some extent, at the pre-1975 end, it would not address this issue at the crucial 1990s end.)

Conclusion

Keith Hancock's view, that the evidence of a surge in Australia's productivity growth in the 1990s is far from compelling, does not do justice to the rigour and breadth of analysis that has been undertaken. On the one hang, his view is based on a method that has serious flaws as a test of whether there has been a significant increase in productivity growth. On the other hand, evidence of a productivity surge in the 1990s lies in the data and in sound analytical methods used by a range of institutions and individual analysts.

The flaws in Hancock's method stem from the fact that a quadratic or cubic form, whilst employed to capture an acceleration in productivity growth at some point over the observation period, is not consistent with a sound theory of productivity growth over the very long term (40 years in this case). It is not surprising that both forms are rejected by the data. Importantly, though, that rejection says nothing about the existence or otherwise of a 1990s surge. A rejected model of 40 years of productivity growth does not provide the basis for a test of a structural change from one decade to another. Moreover, on purely practical grounds, the validity of the test used is undermined by the likely sensitivity of results to the selection of the period over which curves are fitted.

The evidence that Australia did experience a productivity surge in the 1990s is, in fact, very strong. The evidence lies in the data. Three forms of demonstration, including the method used in constructing the official estimates published by the ABS, have been outlined in this response. The evidence also comes (without exception) from a number of assessments, analyses and formal tests. These include studies that have specified econometric models that can validly test the existence and significance of a structural improvement in Australia s productivity growth in the 1990s.

References

Hancock, K. (2005),' Productivity Growth in Australia 1964-65 to 2003-04', Australian Bulletin of Labour, vol. 31, no. 1, pp 28-32.

Parham, D. (2000), 'A More Productive Australian Economy', Agenda, vol. 7, pp 3-16.

Parham, D. (2004), 'Sources of Australia's Productivity Revival', Economic Record, vol. 80, pp 239-257.

Parham, D. (2005), 'Is Australia's Productivity Surge Over?', Agenda (forthcoming).

Quiggin, J. (2001),' The Australian Productivity Miracle: A Sceptical View', Agenda, vol. 8, pp 333-348.

Endnotes

(1) A self-professed sceptic, Quiggin (2001) has come to accept that there was a productivity surge, although his is at the low end of the range of estimates of its magnitude. See Parham (2004) for an interpretation.

(2) The figures also give rise to another question that is quite separate from the issue of whether there was a surge in the 1990s. Does the apparent reduction in trend productivity growth since the late 1990s mean that the productivity surge is now over? I address this issue in Parham (2005).

(3) In Parham (2000), I plotted pair-wise observations of labour productivity and the capital-labour ratio. Movement from one observation to another could be explained in terms of a relationship between labour productivity, capital deepening and multifactor productivity. I fitted a curve to the observations to summarise a strong historical pattern. The goodness-of-fit was very high.

(4) See the appendix in Parham (2004) for a review of studies of the existence, magnitude and timing of the productivity surge.

Dean Parham, Productivity Commission, Canberra

* Jonathan Pincus and Ralph Lattimore provided helpful suggestions in framing this comment. The views expressed are the responsibility of the author alone and do not necessarily reflect the views of the Productivity Commission.

Table 1: Three stages in Australia's productivity growth Per cent

              1964-5 to 1978-9   1978-9 to 1993-4   1993-4 to 2003-4

Annual change in labour productivity

  Mean              2.86               1.82               2.78
  Std
  deviation         2.84               2.13               1.44

Annual change in multifactor productivity

  Mean              1.54               0.84               l.:
  Std
  deviation         2.64               2.19               1.00