Seasonal adjustment of the Vacancy Survey data

Labour Market Trends, Dec 2004 by Treasure, Helen

Key points ONS is introducing seasonally adjusted series of results from the Vacancy Survey.

Vacancy Survey data have been collected since April 2001, and the total series and most component series are already showing strong and stable seasonal patterns. After three and a half years there are now enough data for seasonal adjustment.

Seasonally adjusted series are being published for both the total monthly data and the three-month rolling averages. The latter are more reliable as the sampling variations are smaller.

Seasonally adjusted data are also being published on a rolling quarterly basis by size of enterprise and by broad industry group.

Introduction

Results of ONS's survey of job vacancies were adopted as National Statistics in June 2003 (see pp349-62, Labour Market Trends, July 2003). The Vacancy Survey provides comprehensive estimates of job vacancies across the economy from April 2001. The survey, based on a sample of businesses, asks employers how many vacancies they have in total for which they are actively seeking recruits from outside their organisation. Total estimates are available on a monthly basis, and as three-month rolling averages, which have smaller sampling errors and are therefore more reliable. In addition, data are available by industry and by size of enterprise on a rolling quarterly basis.

The Vacancy Survey data show strong seasonal patterns, with vacancies peaking around September to October and dropping around January each year (see Figure 1). ONS has carried out a methodological review of the data and has concluded that most of the component series are suitable for seasonal advent, using the X-12 ARIMA program.

Total vacancies

The total vacancies data are presented on a monthly and a rolling quarterly basis. The rolling quarterly data are more reliable because the sampling errors are smaller. Approximately one quarter of the businesses in the survey (around 1,500) are large enterprises and are included in the survey every month. The remaining 4,500 or so are sampled randomly and are included in the survey for five or nine quarters depending on the size of the business. While selected, they are included in the survey every three months.

Although the series are short (spanning just over three and a half years), both the monthly and quarterly series show a strong seasonal pattern and the quality of seasonal adjustment is good.

Before seasonal adjustment, the rolling quarterly data are a direct three-month average of the monthly data. The monthly data and the rolling quarterly data are seasonally adjusted separately to give the best possible seasonal adjustment for both series. Therefore, the two series do not correspond exactly after seasonal adjustment. The review considered a number of options for reconciling the two series, but all introduced additional complications and compromised the quality of the seasonal adjustment.

Comparisons using the monthly series

As a result of the three-month rotating nature of much of the sample used in the survey, month-tomonth comparisons of the monthly total estimates of vacancies are much more affected by sampling variations than are comparisons with data three months or 12 months ago. (There is a greater overlap between the respective samples.) For this reason, short-term comparisons using the monthly seasonally adjusted series are best made in terms of the change over the latest three months.

Data by size and industry

Data by size of enterprise are available for five size-bands, based on numbers employed, on a rolling quarterly basis. (The size bands are 1-9, 10-49, 50-249, 250-2,499, 2,500+). All five series show a clear seasonal pattern and are suitable for seasonal adjustment.

The methodological review also considered seasonal adjustment of the 19 industry groups for which unadjusted data are published. The results varied, with a few series displaying a strong seasonal pattern, but with many showing limited seasonality or no evidence (yet) of any seasonal pattern. Instead, the data are aggregated into eight broad industrial groups before seasonal adjustment. These aggregated series tend to show a stronger, more stable seasonal pattern with less irregular variation, therefore improving the quality of the adjustment and reducing the size of revisions caused by new observations. However, two of the broader aggregated series (energy and water; and other services) still display no seasonal pattern and are therefore not seasonally adjusted.

Seasonal adjustment models and settings

The series are seasonally adjusted using X- 12 ARIMA. A number of tests were carried out to determine the most suitable models and settings. As the series are short, they cannot be extended with forecasts. The X- 12 procedure is therefore used without the ARIMA modelling functionality (see Box 1). Also, there are not yet enough data to determine whether the timing of Easter affects the results.

The series are modelled using an additive model; that is, the time series are conceptualised as the sum of three components: the trend, the seasonal variation and irregular variation. The three components are estimated using an iterative procedure, and the seasonal component is subtracted from the time series (see Box 1). The choice of model and other seasonal adjustment settings will be reviewed when more data are available.