|
Article Excerpt
Introduction
The 2001 Census data, which have become available during 2003, contain for the first time in the history of the Census, the Social Grading classification, which is in common use by marketing and market research practitioners. It is often referred to as the ABC1 classification, though in fact it consists of six grades which are labelled A, B, C1, C2, D and E. This is in addition to the Government's own new socio-economic classification system, developed for the new Census, known as NS-SEC (National Statistics Socio Economic Classification). Both systems use occupation as their foundation, but how they group occupations, together with other relevant information, is different.
The Social Grades which have been applied to the Census are 'approximated' Social Grades, derived electronically from the relatively limited data available on the Census, using an algorithm that allows them to reflect as closely as possible actual Social Grades used in market research. Social Grading as used in market research was developed for the National Readership Survey and makes use of potentially much more data available from the personal NRS interview, and other market research interviews, than is available from the Census.
The algorithm was developed by a working party of the Market Research Society's Census and Geodemographics Group. An MRS conference paper of 1998 by the present authors on behalf of the working party, entitled 'Can the Census make the grade?', contained a report on the first stages of the project. At the MRS conference the following year we presented a paper entitled 'Social classifications--a new beginning or less of the same?' in which we demonstrated the continuing value of Social Grading and pointed to the practical difficulties for most data collection methods currently in use by market researchers, if the Government's NS-SEC were to be used instead.
The case for Social Grading is that it is still widely used as a general demographic variable, that it is still of strong discriminatory power for most goods and services, and that possible alternative systems are, for general use, of less (or at best of equal) discriminatory power and, as is the case with NS-SEC, that they are less practicable for use with standard data collection methods.
The case for adding Social Grading to the Census is that it will make Census data, particularly area statistics, much more useful to marketing and market researchers than they were previously, by providing them with the industry's own 'common currency' of social classification.
In the following, we will describe the main development stages and validation of the algorithm used to approximate Social Grading, relate the decisions taken by the working party as a result of the work undertaken, and assess the outcome--that is, comment on the results as now available on the Census.
The data now available on the Census, published by ONS (Office for National Statistics) online as tables SO66, CASO66, SO67, CASO67, UV50 and UV78, are not perfect. A flaw was discovered, after publication, affecting the data of the population aged 65 and over. The cause was the lower than expected accuracy of the implemented algorithm among the non-working and retired population, due to the absence, in many cases, of previous occupation details. Our comments will include a discussion of this problem and how we plan dealing with it for the future.
We will note how the current data are disseminated and cross-referenced to the MRS website, and recommend likely uses. Finally, we will discuss how the new data may be updated year by year between now and the next Census.
Developing the algorithm
The working party was formed in 1994. There were four distinct phases to our work. The first was an exploration phase in which we checked whether it was at all feasible to approximate Social Grading from the data likely to be available with the 2001 Census. In the second and third phases we concerned ourselves with identifying the most appropriate of various possible algorithms and their validation, first for workers, for whom occupation details were universally available, and then for non-workers. The fourth phase consisted of fine-tuning the finally recommended algorithm ready for application by ONS to the Census data collected.
The first two phases were described in our MRS conference paper of 1998. In the following we give a brief overview of its contents. We then describe the third and fourth phases.
Exploration
In its standard form, Social Grading, as applied to the National Readership Survey and followed by other surveys, is a household classification. All members of a household receive the same grade, which in the first place is based on the chief income earner's current occupation, if working, or previous occupation (if applicable), if not working. Apart from occupation, there are additional criteria, which are taken into account by trained interviewers and, on the NRS at least, by office staff employed to verify the interviewers' coding. These can be status, qualifications, source of income (but not income per se), lifestyle and such like. In marginal cases, where the rules seem ambiguous or not fully applicable, a final subjective coding decision may be taken following the spirit of the rules.
Thus, if a group of survey respondents is classified as, say, C1 (supervisory or clerical, and junior managerial, administrative and professional), this means that they are all members of C1 households, based on a household's chief income earner's occupation and other relevant characteristics.
In our exploratory phase we looked for a relatively straightforward way of assessing whether a respondent's grade could be replicated by a combination of just a few variables. We made use of NRS data, not only because the NRS sets the standard for Social Grading, but also because it allowed us to simulate Census variables, including the Government's SOC coding (Standard Occupational Classification). A further advantage was that the NRS, in addition to the Social Grade of the chief income earner of a respondent's household, allocates an individual Social Grade to the respondent (his or her own grade), if he or she is working and is not the chief income earner. This was of great value to us because SOC and all other variables of the NRS, apart from Social Grade (of the chief income earner), relate to the respondent.
From the outset, we agreed that it would be very difficult and indeed not necessary to look for separate approximations for grades A and B. They were to be amalgamated as ABs.
Our first test was based on 100 full-time working NRS respondents interviewed in December 1994. It showed that in 69 out of 100 cases, we were able to predict conclusively from 11 selected variables the actual individual Social Grades of these respondents. Besides SOC, the main predictors were shown to be employment status and size of establishment. This we found very encouraging.
The algorithm for full-time workers
Our second test consisted of the examination of five possible solutions for an algorithm for full-time workers using just SOC, employment status and establishment size, the main relevant variables we knew would be available from the Census.
The first of these was the SEG (Socio-Economic Grades) solution, following an established, if crude, practice of many marketers of assigning Social Grades to particular groupings of the Government's SEG classification. The other four we called SOC solutions. For these, two basic combinations of SOC groupings together with employment status and size of establishment (with variable size bands) were used to assign Social Grades. However, we did not force a Social Grade coding. Where there was not sufficient information to assign a Social Grade conclusively by using the chosen variables, we coded, say, BC1 instead of either B or C1, and so on. These cases, together with other marginal cases of missing information, we called 'doubtfuls'.
The test was carried out on 14,837 NRS respondents in full-time employment in January-December 1994. The five solutions were coded onto the NRS data and their outcomes compared with the actual individual Social Grades of test respondents.
The results showed that, after removing the...
|
