A quarterly post-Second World War real GDP series for New Zealand

Abstract

There are no official quarterly real GDP estimates for New Zealand, for the period prior to 1977. We report the development of a seasonally-adjusted series for a period of more than 60 years from mid-1947, and evaluate statistical properties. The series were developed by linking quarterly observations from two recent official series to temporally disaggregated observations for an earlier time period. Annual real GDP series are disaggregated, using the information from two quarterly diffusion indexes, developed by Haywood and Campbell (1976). Three static econometric models with different disturbance term assumptions and one dynamic model are evaluated. Our preferred quarterly series is based on results generated from the Chow and Lin (1971) model. We assess movements in the preferred static model and dynamic model series against qualitative findings from New Zealand's post-Second World War economic history.

Keywords:

• quarterly real GDP series
• temporal disaggregation
• business cycles
• New Zealand

Acknowledgements

The authors thank Arthur Grimes and Glenn Otto for perceptive comments, Brian Easton for his very helpful insights on key data periods, two anonymous referees and the managing editor for very valuable suggestions. The authors are also grateful to Brian Easton for supplying his Haywood and Campbell diffusion index data.

Notes

 1. Briggs (2003) provides a good summary of what each of these estimates involves. The estimates are available on Statistics New Zealand's long-term data series webpage (see http://www.stats.govt.nz/browse_for_stats/economic_indicators_nationalaccounts/long-term-data-series.aspx).

 2. More detailed explanations and reviews of these methods can be found in Bloem et al. (2001), Ciammola et al. (2005), Di Fonzo (2003), Moauro and Savio (2005), Proietti (2006), and Santos Silva and Cardoso (2001).

 3. Bloem et al. (2001, ch. VI) have strongly advocated use of an expanded proportional Denton method rather than ARIMA, GLS or Chow–Lin (CL) methods, on the grounds of being simpler, more robust and more likely to avoid both over-adjustment and misspecification of the autocovariance structure. They argued this to be particularly so for large-scale applications such as nowcasting and forecasting National Accounts observations, although they conceded (Bloem et al., 2001, p. 108) that the CL method may have a role in filling minor gaps in observations. SNZ use temporal disaggregation in the ‘benchmarking’ sense (see Bloem et al., 2001, pp. 5–7), to ensure preliminary quarterly estimates are consistent with the annual growth rates. Benchmarking is not a primary aim of this paper, so we did not consider using this methodology. In addition, Moauro and Savio (2005, p. 220) state that the Denton (1971) approach is a special case of the CL approach.

 4. The disturbance term assumptions are: AR(1) for Chow and Lin (1971); I(1) for Fernández (1981); ARIMA (1,1,0) for Litterman (1983), and ARIMA(p, d, q) for Wei and Stram (1990).

 5. See also Ciammola et al. (2005, p. 1) and Santos Silva and Cardoso (2001, p. 269).

 6. SSC's method is suggested to be particularly appropriate where the relevant series are stationary or cointegrated, and by providing a direct solution for the initial conditions for the estimated series, is computationally simpler than the earlier dynamic solution procedures of Gregoir (2002) and Salazar et al. (1994).

 7. Temporal disaggregation and seasonal adjustment procedures could therefore be undertaken simultaneously rather than separately, as is currently done by national statistical agencies.

 8. Haywood and Campbell (1976) also seem to have been sufficiently confident of the quality of their diffusion indexes, to have reported judgementally derived business cycle turning points from them.

 9. For seasonal adjustment details, see Haywood and Campbell (1976, p. 5, Chart A).

10. The approach used for amplitude adjustment is similar to that of Shiskin (1961, pp. 43–44). Haywood and Campbell's computer programme also calculated deviation (or growth) cycles. Further details on their methodology are available in Haywood and Campbell (1976, ss 2, 3 and 4).

11. Our CL, Fernández, and Litterman results were computed using (as in CL, 1971) Generalised Least Squares (GLS) programmed in GAUSS. Results reported for the ADL(1,0) dynamic model were computed using the Maximum Likelihood (ML) option from the MATLAB toolbox TD, for which the software and associated documentation can be found in Quilis (2009), accessible from http://www.mathworks.com/matlabcentral/fileexchange/24438. One important practical issue we encountered was whether to use the feasible GLS estimation strategy originally used by Chow and Lin (1971) or an alternative estimation procedure such as maximum likelihood. In general, maximum likelihood would seem to be the superior choice. In the end we experimented with both methods and found there to be little difference. The simulation results of Ciammola et al. (2005) suggest that this should be the case in real-world situations with large autoregressive parameters. In particular, our results obtained from the GAUSS GLS program were not materially different from those produced by the MATLAB TD toolbox using the WLS and ML options – for example, correlations for the unadjusted CL growth rate series were 0.995 for TD WLS and ML, 0.999 for GAUSS GLS andTD ML, and 0.990 for GAUSS GLS and TD WLS. Similarly, the correlations for the unadjusted and adjusted SSC ADL(1,0) series estimated by the ML and WLS options were 0.995, with that between the unadjusted and adjusted series having been 0.999.

12. The Statistics New Zealand production-based real GDP series SNBA.2SAZAT reports annual growth of −2.6 percent for the year ending 31 March 1978. The series was discontinued in June 2000. An earlier series SNBA.SX9 (discontinued in March 1985) shows annual growth of −3.5 percent. Brian Easton suggests that even the −2.6 percent figure appears implausibly low, given other macroeconomic and sectoral data. (See ‘The 1977/78 Downturn in the New Zealand Economy’ http://www.eastonbh.ac.nz/?p=776.) In personal communication, Brian Easton has advised us that the exaggerated contraction is due to the adoption of a new inventory series with no overlap between surveys. The problem is even worse when one examines the quarterly data. The official real GDP index (SNBQ.SY299) records growth rates for September 1977, December 1977 and March 1978 as −2.1, −4.3, and 0.0%, respectively. Such a substantial decline in output would surely have been noticed at the time as a particularly major contraction, and would have been widely commented on. Having the underlying surveys change and a new quarterly series start during a significant contraction makes unravelling the truth of the matter impossible. So, given that the negative quarterly official statistics for the second half of 1977 are almost certainly implausibly low, we have not included those observations in our series. Instead we have used the estimates from the Chow–Lin procedure based on the annual data and the Haywood and Campbell related indicators. This still leaves a large contraction for the March 1978 year consistent with the official annual data. In generating our new quarterly GDP estimates we believe it is important tostay as close as possible to official statistics, while at the same time recommending caution to potential users, who may if they wish apply their own adjustments. For example, Brian Easton has guessed the annual March 1978 contraction to be more like 1to 1.5%.

13. The extent to which movements in the Fernández and Litterman unadjusted HC diffusion index series are imperceptibly different from those for the Chow–Lin series can be found in Hall and McDermott (2009b, Figure 3).

14. In Figure 5, the full linked GDP series for 1947q2 to 2008q3 have been illustrated, so as to provide a pictorial representation of movements over the period of more than 60years.

15. Results for the Fernández and Litterman and Santos Silva and Cardoso series are not materially different. Detailed statistics for the Fernández and Litterman series are available from Hall and McDermott (2009b, Tables 8, 9). The real GDP growth rate series generated from the Santos Silva and Cardoso regression is smoother than those generated from the static models. Visual inspection of Figure 4 shows that with the exception of the first two years, peaks and troughs for the CL and SSC growth cycles occur within similar time intervals. There are of course specific timing differences for some of the turning points, associated with the greater smoothness of the SSC series.

16. The equivalent results for the Fernández and Litterman series, similar to those from the SSC ADL(1,0) series, are in Hall and McDermott (2009b, Table 10).

17. A new ‘benchmark’ set of Classical business cycle turning points, for the period 1947q2 to 2006q2, and reference to the associated key events in New Zealand's economic history, are reported in Hall and McDermott (2009a). Utilising the series for 1947q2 to 2008q3 presented in this paper adds a further ‘benchmark’ turning point to those presented in Hall and McDermott's (2009a) Table 1 – a Classical business cycle peak at 2007q4 – while utilising the further extended series 1947q2 to 2010q2 provides the post-Global Financial Crisis business cycle trough at 2009q1.

18. The data series based on results generated from SSC's ADL(1,0) method are available on request from the corresponding author.