Exploring determinants of the extent of long distance commuting in Australia: accounting for space

ABSTRACT

Previous research exploring the impacts of long distance commuting (LDC) or, more generally, mining on host regions, struggles to explain the variability of these impacts over time and across space. This article argues that spatial effects should be accounted for explicitly in order to improve the predictive power of contemporary research. We study the extent of LDC in a region in a spatial model disaggregating Australia into 325 subregions. We find evidence that space is an important factor in explaining the extent of LDC in a region, which challenges the validity of studying LDC impacts on host regions in isolation. With regards to the determinants of the extent of LDC, we find that residential attractiveness of a region influences the extent of LDC in a region; the size of the pool of unemployed in a region does not.

KEYWORDS:

• Long distance commuting
• mining
• regions
• spatial analysis

Acknowledgements

We are grateful to two anonymous referees and the editor of Australian Geographer for their valuable advice and suggestions.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 We modelled an OLS using maximum likelihood estimators. With the presence of spatial dependence in the data, we used LM tests to locate this dependence—either in the model or in the error term. LM tests found the presence of spatial autocorrelation (LM_lag = 62.199, p < 0.001) and spatial error dependence (LM_error = 47.350, p < 0.001). While spatial dependence was accounted for by the robust LM tests, there was no remaining spatial error dependence (RLM_error = 0.223, p = 0.63); however, the presence of lag-dependence was indicated (RLM_lag = 15.07, p < 0.001). Therefore, we concluded that the spatial dependence came from within the model, meaning that either a spatial autoregressive lag model or a spatial Durbin model would be appropriate. We ran both of these models using maximum likelihood estimators and both had no residual correlation (LM = 0.60, p = 0.43) for the lag model and (LM = 1.44, p = 0.23) for the Durbin model. Both the Durbin (BP = 35.36, p = 0.06) and the lag (BP = 22.59, p = 0.03) models were likely to have heteroskedasticity. However, the Akaike Information Criterion (AIC) scores favoured the lag model (265) over the Durbin model (273), hence we chose a spatial autoregressive lag model.

2 The study has two main limitations. The first limitation relates to the empirical operationalisation of LDC (cf. Skilton 2015). ABS statistics do not identify LDC. This article has highlighted the limitations of trying to convert the data to a purpose for which they were not intended. We used SA3 regions as the unit of analysis, as opposed to the more detailed SA2 demarcation. Our decision reduced the potential interference of daily commuting in the analysis. However, it also concealed some LDC activity—especially LDC which starts and ends within rural and remote Australia. The second limitation relates to reverse causality. Future quantitative research would benefit from the inclusion of time in combination with space (i.e. spatial panel modelling) in that it would provide the ability to disentangle cause and effect. Alternatively, on-the-ground qualitative research could be conducted to identify trends and hence achieve a similar goal. The results of this study should be treated in light of these limitations.