1.7. In principle, balance of payments statistics should be compiled by summing the foreign accounts, whether actual or notional, of every economic unit (i.e. every government body, business and household) in Australia. Alternatively, each transaction, whether actual or notional, between a resident of Australia and a non-resident could, in principle, be measured. Other things being equal, if this were possible, the resulting statistics would be a true measure of balance of payments activity. In practice however, neither approach is feasible. Instead, the statistician produces estimates of the notional true values by combining data from a wide variety of sources reflecting varying valuations, coverage, frequency, detail and timeliness.
1.8. Frequently, the data available to the statistician may only approximate the concept that it is desired to measure. For example, the conceptual framework calls for the inclusion in the balance of payments of merchandise exports and imports that have changed ownership between a resident and a non-resident. In practice, international merchandise trade statistics, compiled from customs documents which measure goods crossing the customs frontier whether or not they change ownership, are used as the main source in estimating merchandise trade activity in the balance of payments. Although adjustments to a change of ownership basis are made for all identified large transactions that are not on this basis, it is not practical to identify all such cases.
1.9. Because of this variability in the appropriateness of source data for balance of payments purposes, the data available at any particular stage in the estimating cycle will inevitably contain errors. Accuracy in this context therefore refers to how closely an estimate measures the activity it purports to measure.
1.10. The errors that are present in a statistic can be thought of as being of two types. First, the difference between a fully-revised (or final) estimate and the notional true value and, second, the difference between an earlier (or preliminary) estimate and the fully-revised estimate. Because the true value is not known, the first type of error is theoretical only and cannot be directly quantified. The second type of error is measured by the difference between a given preliminary estimate (most commonly the first estimate) and the final, fully-revised estimate that is the statistician's best approximation of the notional true value. These latter errors, that are eventually removed by revision, and are therefore quantifiable, are the focus of later analysis in this paper.