Economic data are often revised. Sometimes the revisions change the story; sometimes they do not. Either way, they argue for humility when using data in policy debates.
- Recent revisions to US PCE inflation, UK productivity and Canadian population estimates are reminders that statistical measurement is imperfect. Revisions are both a fact of life and, when they improve measurement, desirable.
- Australian data will also be revised over time, though not necessarily in the same ways as these latest revisions abroad. The ABS tends to favour quality over speed, which reduces the need for early revisions as more data become available. But revisions can still matter. Population estimates implied in the initial releases of labour market and GDP data are especially likely to be revised in the near term.
- Estimates of current trends can change as future data are published. Shifting seasonal patterns are another source of revision, especially for monthly CPI components with short histories. Series that are derived from multiple other imperfectly measured series are more prone to revision than quantities that can be measured directly. All these considerations point to a need for humility when using recent data to make strong policy pronouncements.
On September 30, the Fed’s preferred measure of inflation, PCE, was revised down. New methods and data sources were applied to three components of services inflation (financial advice, software and legal services), large enough to result in downward revisions to the overall read on core PCE inflation. These changes had been pre-announced in June and were therefore not a surprise. The CPI measure of inflation was not affected, nor should the revisions materially change your view of US inflationary pressures.
Some revisions are more consequential for interpretation and policy. A couple of weeks before the US PCE revision, it was revealed that the UK Office for National Statistics was revising down its estimates of average hours worked per worker. The result was that measured productivity growth was nowhere near as weak as had been widely believed. As my friend and counterpart, TS Lombard’s Chief Economist Freya Beamish commented on social media, “Oh well. Not like any major policy decisions were taken on the basis of those numbers.” Quite.
Also in September, Statistics Canada revised up its estimates of Canada’s population, taking the recent growth rates from outright contraction to still-slow growth a little below ½%yr. The number of temporary residents in Canada is still falling, but the revised data do a better job of tracking them, especially those seeking an extension on their permit. (An alert for customers who have seen the cross-country population graph I often use at client events: the Canadian line now looks different but still shows a sharp slowing.)
There is no fault to be attributed here. Revisions to key statistics are a part of economic life. Sometimes the underlying data arrive with a lag, perhaps because of reporting deadlines. Sometimes the compiling agency improves its methods or gains access to a better data source. The US and UK examples above fall into this category. In the Canadian case, better data and a more accurate definition of scope were involved. All these revisions should be welcomed. They provide a more accurate picture of the underlying reality.
Australia, revised
Revisions are a fact of life for Australian statistics, too. I am still a little scarred by that unexpected –0.5%qtr GDP print when the September quarter 2016 data were first released on my first day as RBA Assistant Governor; the figure has since been revised to 0%qtr and was followed by a big bounce-back in the December quarter.
The recent revisions mentioned above are less likely to be replicated in Australia, though. Australia’s CPI is designed such that the headline (not seasonally adjusted) measure is never revised. This is because it is used in many legal contracts. If an error is detected, the ABS makes the adjustment in subsequent months and quarters rather than revising the history. Our labour force survey has a much higher response rate than its UK equivalent – Australians are a compliant lot when the ABS comes knocking – and measures actual hours worked more reliably. And our population statistics already handle the array of temporary resident visas with the Bureau instead applying a standard 12-month-in-16-month residency test to count as a resident for the population statistics rather than using a visa-classification basis.
This last case illustrates Australia’s broader tendency to favour quality over speed in official statistics. Our CPI and national accounts data tend to be published a bit later than peer countries’ data for the same reference period, and our estimated resident population data is quite a bit lagged relative to Canada’s (the 12-month-in-16 criteria taking over a year to confirm). The trade-off is that observers of the Australian economy must wait longer but can be less worried about “flash” and “advance” releases of data that can be quickly revised to tell a completely different story.
Revisions to data still can matter in Australia, and it helps to be aware of the specific cases where it might occur. For example, precisely because the population data are very lagged, the population estimates underlying the initial releases of the labour force survey and national accounts are based on preliminary data and get revised later as a matter of routine. This is why ratios such as the unemployment, employment-to-population and participation rates provide a better real-time gauge of labour market conditions than the monthly change in employment. Similarly, GDP per capita can be revised even when GDP is not. This type of revision is also a more likely source of short-term changes in measured hours worked in the Australian productivity data than the conceptual change that shifted the UK data.
The usual reporting lags – and occasional correction of reporting errors by the reporting entity – will always be a factor in Australia just as it is in other countries. The annual national accounts, due on 23 October, incorporate additional data, revise the quarterly estimates and provide first estimates of important quantities such as the capital stock. Just as importantly, though, the annual accounts reconcile the different ways of measuring GDP. In doing so, what we thought we knew about the recent history of output growth can change.
Every purchase is also a sale, so conceptually, you should get the same answer for final output regardless of whether you count how much is produced, how much people spent on it, or how much income people got for producing it. In practice, though, the “production”, “expenditure” and “income” approaches do not always line up until the ABS goes through a more thorough reconciliation process in compiling the annual accounts.
When the future changes the past.
Future data can also change our view of the past. Two good examples are trend estimation and seasonal adjustment. Future observations help reveal whether an apparent trend is stable. As a result, real-time trend estimates can differ considerably from those calculated with the benefit of later data.
Identifying stable seasonal patterns also requires a reasonable run of data. Future observations affect whether an earlier movement is judged to be normal seasonal variation. This problem is more pronounced for newer series, where there is not yet enough history to establish what “normal” looks like.
Consider, then, the monthly CPI. This relatively new measure has only a short run of data with many individual components having too little history to use standard seasonal adjustment methods. As Westpac’s senior economist Justin Smirk has previously highlighted (PDF 468KB), some components such as footwear are being ‘seasonally adjusted’ by applying Henderson trends, similar to those used in the ‘trend’ measures of labour market and national accounts data. For some other components, including books, furniture, audio-visual and computing equipment and children’s clothes, certain price falls – usually associated with sales periods – are just arbitrarily removed from the seasonally adjusted series. The index is held at its previous month’s level but otherwise matches the non-seasonally adjusted series.
The seasonally adjusted measures of these CPI components will inevitably be revised over time, and thus so could the monthly trimmed mean estimates. This helps explain why the RBA has said it would, for now, focus on the quarterly trimmed mean based on the old seasonal factors. That measure will at least not be buffeted by shifting measurement of seasonality in a short history. If seasonal patterns are themselves also shifting, the current underlying trend might look a bit different in a few years, not just a smoother month-to-month profile.
The broader point here is to remember that economic statistics are a lens on reality. Sometimes the lens is smudged, sometimes the scope is off and sometimes what is being read as underlying trends is wrong. Measures constructed from several imperfect series – productivity from output and hours worked, or unit labour costs from productivity and labour costs, or savings rates derived from income and spending – are especially prone to revision compared with concepts that can be measured more directly, like a price. That is a strong argument for humility before making firm policy pronouncements based on simple data comparisons.
