SIGMA Y BETA-CONVERGENCIA

The effect of the introduction of EGMs on retail trade has been controversial with claims made of the adverse impact of EGMs on small businesses. As discussed in Phase 1, a visual scan of South Australian consumption expenditure excluding gambling does not show any sign of a significant impact following the introduction of EGMs (see Figure 2.1).

Figure 2.1

South Australia – Real Household Final Consumption Expenditure minus ‘Recreation and Culture’ (Gambling Expenditure)

3,000 4,000 5,000 6,000 7,000 8,000 9,000

Jun84 Jun86 Jun88 Jun90 Jun92 Jun94 Jun96 Jun98 Jun00 Jun02 Jun04 Quarter

$

millio

n

EGMs introduced in South Aus tralia (25th July 2004)

Source: ABS Cat. No. 5206.0 Australian National Accounts, Table 87, and calculations by the researchers.

Nor, as was also pointed out in Phase 1, is the scale of net expenditure on EGMs high enough for it to be expected to have any impact on overall trends in retail trade expenditures. Even at its highest relative level in 2003-04 (which is also the most recently available figure), gambling only represented 2.91 per cent of household disposable income which is not large enough to meaningfully impact on the overall figures.

It is possible that this visual inspection of the time trend in expenditure data is misleading, and that there is indeed a shift in the overall level of household final consumption due to the introduction of EGMs. For this reason a time series model of the level of consumption spending excluding gambling was estimated. The model structure used was an Auto Regressive Distributed Lag (ARDL), which is often used in modelling consumption

behaviour. A key advantage of the ARDL structure is that it allows specifications which model the impact of the past values which is necessary for consumption as modelling consistently shows that past levels of consumption are important in explaining current consumption.

The basic functional form was to model current (non-gambling) consumption as a function of current and past levels of income, past values of (non-gambling) consumption, and current values of gambling expenditure. The series used for consumption was ‘final consumption expenditure’ (minus gambling expenditure), and income was ‘gross disposable income’, both obtained from the ABS’ Household Income Account, 5220 (Table 37, South Australia). Data on gambling expenditure came from Australian Gambling Statistics 2005, produced by the Office of Economic and Statistical Research, Queensland Treasury.

There is no a priori reason to follow a particular lag structure for either consumption expenditure or income, and so different structures need to be tested as part of the model development process. Initially three lags each were included in the equation and this lag structure was then ‘tested down’ to identify the best system of lags, with lags being removed one at a time, and the explanatory power of this new lag structure being compared to the previous one using tests of model specification (using the Akaike Information Criterion and Schwarz Criterion61), as well as the value of the Durbin-Watson statistic to ensure that removing the lag hadn’t introduced autocorrelation. In this case it was found that the best model specification was the inclusion of consumption from the previous period in the regression, but not to only use the current value of household disposable income. As all of the variables increase over time, logs were used in the model and a deterministic trend was included to capture the natural propensity for consumption to grow at a changing rate over time.

The explanatory variables of the final model were (all in log form62):

• Current household disposable income;

• The previous period’s level of non-gambling consumption expenditure;

• A time trend; and

• The level of expenditure on gambling.

The model appears to be a good fit for the data, with no evidence of any systematic pattern in the residuals, and the Durbin-Watson statistic indicating that serial correlation is not a problem. The F-Statistic is a measure of the joint statistical significance of all the explanatory variables; in this case its value of 707 means that the probability that the explanatory variables are not jointly significant in explaining consumption behaviour is 0.000000.

61 _{The Akaike Information Criterion and the Schwarz Criterion are methods of comparing alternative specifications to adjust the}

residual (RSS) for the sample size and the number of variables. The use of these methods is designed to improve the ‘exactness’ or degree of fit of the model to explain or to account for changes in the dependent variable.

62 _{Reference to “log form” refers to use of logarithm in econometric analysis or specification of the equation in the exponential}

functional form, given the assumption of constant elasticity. One advantage of this form of specification is that they make it easier to assess impacts in percentage terms.

Table 2.1

Determinants of South Australian Household Non-Gambling Final Consumption Expenditure Dependent Variable is Log Non-EGM Consumption

Variable Coefficient Std. Error t-Statistic Prob.

Intercept -2.707** 0.834237 -3.244307 0.0101

Log Income 0.512** 0.158274 3.234146 0.0103

Log non-EGM consumption (t-1) 0.736* 0.117960 6.239430 0.0002

Log Time Trend -0.091* 0.024987 -3.647859 0.0053

Log Gambling 0.052 0.041394 1.250120 0.2428

Note: * significant at the 1% level ** significant at the 5% level

F-statistic 707.3702 Prob (F-statistic) 0.000000

Durbin-Watson statistic 2.08

As summarised in Table 2.1, the coefficient for gambling (log gambling) is not significant, meaning that expenditure on gambling does not appear to have any impact on the level of non-gambling consumption behaviour. Even if the impact was different from zero (and there is still a chance that it actually is) the coefficient takes a positive value, indicating that rather than consumption expenditure falling as gambling expenditure increases, it increases. It is unlikely that there would be a causal relationship with increasing gambling expenditure inducing increased other consumption expenditure, rather it would be more likely that they were both affected by some other variable such as income or changes in wealth.

It could be the case that, although the coefficient for gambling expenditure is not significant, the overall pattern of consumption behaviour changed with the legalisation of EGMs. This can be tested for with the Chow test, which tests whether there has been a significant change in a particular data relationship after a specific point in time. The point of time chosen in this instance was 1994-95, the introduction of electronic gaming machines to hotels and clubs in South Australia. By using the Chow test, we can assess whether there was any change in the pattern of overall consumption behaviour at the point in time at which EGMs were introduced. If the Chow test is significant then this shows that the factors influencing consumption behaviour were different after that point in time than they were before it. As the F-statistic is 0.688 (calculations not shown here) compared to a critical value of 4.74 we can reject the hypothesis that there was a structural break in the series in 1995. The CUSUM, and CUSUM squared tests were then used to test whether there was an apparent structural break of the series in any other year, but this was found unlikely to be a problem as the residuals were within the confidence intervals through the entire sample period (1984-2004).63

In summary, it is concluded that expenditure on gambling does not appear to have any significant impact on the level of non-gambling consumption behaviour.