The classification of education in surveys: a generalized framework for ex-post harmonization

The first approach is to convert national education variables into the corresponding years of education (following Duncan and Hodge 1963). The ISCED mappings provided by UIS (UNESCO 2021) include the required information on the cumulative duration of educational programmes across countries. Sometimes years of education are also derived from harmonized categorical education variables. The ex-post harmonized International Social Mobility File (ISMF, Ganzeboom and Treiman 2019), only accessible on request, for example provides codes for derived ‘virtual’ years of education (Ganzeboom 2019). In this derivation, vocational education is penalized by not counting fully the related number of years of education so that the resulting variable better reflects how much general education a respondent has obtained. The resulting information is easy to include in linear statistical models, and is particularly popular among economists analyzing returns to education (Mincer 1974; Flabbi et al. 2008).³

Regarding the second approach, levels of education, various coding schemes are used. For national data from different sources, harmonization can be supported by national education classifications (which not all countries have), while for international data, an international classification is needed.

At the most simple level, data harmonized using levels of education distinguish just a few broad, ordinal categories such as ‘less than primary’, ‘primary’, ‘secondary’ and ‘higher’ (or ‘tertiary’) education, which are often not explicitly related to an international standard. For example, the IPUMS Demographic and Health Surveys (DHS, Boyle et al. 2019) and IPUMS International (Minnesota Population Center 2019) harmonize education into just four broad categories, in addition to years of education. What specifically constitutes these broad categories unfortunately differs across projects, which is due to different levels of development and educational expansion across countries. Furthermore, the specific content of broad categories is not always documented, and if the source variables were translated into English, the link with the educational system of the country in question becomes obscure. The exact contents of categories is also often unclear because there is no universal understanding of terms like ‘secondary’ or ‘higher’ education. Therefore, with data using just broad education levels for measuring and/or harmonizing education data, interoperability with other data, including harmonization with other sources, is severely limited, unless they use the same broad education categories. Some projects use standard aggregations of ISCED levels to produce three broad levels (e.g. Barone and Ruggera 2018), which are also commonly used in official reports concerning education across countries (e.g. OECD 2017).

At the most complex level, there are multi-digit coding schemes reflecting not just levels but also types of education within levels. The most common ones here are ISCED 2011 (UNESCO 2012, see further details in Sect. 3) and its predecessor ISCED 1997, developed for international official education statistics, and the CASMIN education scheme, which was developed in the academic project ‘Comparative Analysis of Social Mobility in Industrial Nations’ (König et al. 1988; Brauns et al. 2003). While the CASMIN scheme has been used mostly as a coding scheme for ex-post harmonization in selected European countries (Erikson and Goldthorpe 1992; Breen 2004; Breen et al. 2009), ISCED is by now widely used for ex-ante and ex-post harmonization (e.g. Barone and Ruggera 2018). For CASMIN, mappings linking national education categories with the international ones are only available for a limited set of countries. In the case of ISCED, such mappings are available for almost all countries in the world (UNESCO 2021) and in quite some detail specifically for Europe (Eurostat 2021).

To give some examples, the European Social Survey (ESS, 2020c), the European Values Study (EVS, 2020), the Survey of Health, Ageing and Retirement in Europe (SHARE, Börsch-Supan et al. 2013), the Programme for the International Assessment of Adult Competencies (PIAAC, OECD 2016), and the European Union Labour Force Surveys (EU-LFS, Eurostat 2020), harmonize education ex-ante and use fairly detailed categorical coding schemes based on ISCED. However, the specific ISCED-based schemes still differ from each other. The multi-digit education coding scheme used in the ESS since 2010, differentiating levels and various types of education within levels in one variable, called ‘edulvlb’, has fared quite well in empirical tests and was thus recommended for use in other comparative surveys (Schröder 2014, , chap. 3). It has since then gained some acceptance beyond the ESS community and has been implemented in the EVS 2017, SHARE, and—in slightly adapted form—for the upcoming cycle of OECD’s PIAAC (Allen et al. 2017). Most of these datasets additionally have information on (actual or derived) years of education.

In between these two solutions using levels of education, there are datasets using a medium amount of differentiation, providing main education levels without differentiation of types of education. These main education levels are often defined with reference to ISCED, which is thereby being reduced to its ‘bare bones’ (Schröder 2014, p.76). The European Social Core variables project has specified ISCED main levels as the minimum requirement for official Eurostat surveys (Eurostat 2007). The Luxembourg Income Study (LIS, 2019) and the Survey Data Recycling project (SDR version1, Survey Data Harmonization Team 2017; Slomczynski and Tomescu-Dubrow 2018) use main ISCED levels, but the latter also provides additional information in separate so-called ‘harmonization control’ variables and thus in fact almost falls into the category of complex coding schemes: whether the respective education was not completed, whether it was vocational, and whether the category also includes higher levels of education. These harmonization control variables are intended to support researchers controlling for properties of the source survey items, and thus to enable researchers to ‘recycle’ rather than throw out data.

Finally turning to the third approach, scaled education variables, these are obtained by transforming categorical education variables into a linear metric using supporting variables. The ISMF project developed a linear education scale, constructed by Schröder and Ganzeboom (2014) through cause-and-effect-proportional scaling of national education variables—an approach building on prior work by Treiman (1975) and Smith and Garnier (1987)—, called International Standard Levels of Education (ISLED). The resulting scores can in principle be applied to national education variables across surveys.⁴ Given these national variables are usually not standardized within countries across surveys, this process is rather complex and introduces new errors.⁵ ISLED scores can also be derived from the ‘edulvlb’ coding scheme used in the ESS (Schröder 2014, chap. 3),⁶ as well as from 3-digit ISCED codes (Schröder 2014, chap. 5), making it more widely applicable but still relying on detailed harmonized measures of levels of education.

This summary of existing harmonization schemes reveals that both scaled education variables as well as derived years of education rely on harmonized measures of levels of education. The latter thus are not just important for data users who wish to analyze levels of education, but equally so for data users preferring other ways of analyzing educational attainment. However, as will be shown in Sect. 4, the existing international coding schemes for education are not easily applicable in ex-post harmonization, which motivates the development of the more generally applicable coding scheme GISCED.

It can be debated whether survey data and harmonized datasets should rather provide detailed or aggregate education variables. Aggregate variables contain information on many different educational qualifications in very few summary categories, which often carry a more vague meaning. Detailed variables in contrast contain many categories that are more specific. An argument in favor of aggregate data is that mis-classifications will be less common than when providing detailed variables: some ‘noise’ will disappear due to the aggregation. Producing aggregated variables is also arguably less labor intensive than producing detailed variables. A lot of public reporting (see e.g. OECD‘s ‘Education at a Glance’-series⁷.) is also limited to aggregated measures, and it is easier to communicate about broad distinctions in public debate.

Cross-national survey data involving many European countries, however, notably chose the more complex coding schemes, because European educational systems tend to be more complex than the US-American system (Shavit and Müller 1998) or education systems in most low and middle income countries. When using years of education to measure educational attainment in complex educational systems, there is a concern that important distinctions between people who have followed different educational pathways with the same number of years of education would not be measured, leading to hidden inequalities and low validity of the data (Braun and Müller 1997; Schneider 2010; Kerckhoff and Dylan 1999; Kerckhoff et al. 2002; Schröder 2014). The same criticism may be applied to projects using a low number of broad education categories, and research shows that types of education within levels can make a difference for various outcomes (Triventi 2013a; van de Werfhorst 2017; Delaruelle et al. 2020). We can only study the effects of extreme disadvantage by adequately representing the periphery of the socio-structural concepts we are interested in. Avoiding loss of information, be it by losing entire cases because of incomplete data, by aggregating categories, or transforming a categorical measure of educational qualifications into hypothetical years of education, is therefore an important guiding principle of output harmonization (whether ex-ante or ex-post).

So, while both years of education and broad education categories are commonly used and may be perfectly adequate for data analysis in many research projects, they may be deficient as the sole harmonized education variable(s) for projects aiming for a high level of data reuse, since important categorical distinctions cannot be derived from years of education or broad education levels. Then, educational effects and inequalities may be underestimated, and hypotheses concerned with unveiling such inequalities cannot be tested. In contrast, when using a detailed categorical coding scheme, it is possible to derive years of education from these categories, or simplify the detailed scheme into broad categories in whichever way needed for a specific set of countries or a specific research question. Even for deriving years of education or linear education scores, the amount of detail covered in the source variables is crucial (Schröder and Ganzeboom 2014). Such derivations are not possible the other way round. The degree of ‘interoperability’ and ‘re-usability’, core principles of FAIR data (Wilkinson et al. 2016), are thus highest when applying a detailed categorical scheme for harmonization, or (which is in terms of information content equivalent), harmonizing information into a whole set of harmonized variables. In contrast, especially when combining data from countries at different stages of development and different time points, the problem of harmonization becomes rather severe when only broad levels are available in the data, and the boundaries between the broad levels do not match across datasets. The GISCED framework therefore follows the strategy of coding at a high level of detail.

In upper secondary education (ISCED level 3), subject specialization increases substantially, especially in vocationally oriented programs. Program choice and achievements at this level predetermine to a large extent whether an individual will gain access and be able to successfully complete higher education. ISCED level 3 in most countries ends at age 18 or 19, after about 12 years of schooling. All individuals pass through ISCED levels 1 to 3, if they do not stop their education before the end of upper secondary education. ISCED level 4, post-secondary non-tertiary education, in contrast, is an ‘optional’ level. It mostly contains educational programs that allow individuals to change the specialization they had at level 3, in order to either gain access to tertiary education programs they would otherwise not be admitted to, or get vocational training to then enter the labor market. It also contains programs that may nationally be considered as ‘tertiary’, but that last less than two years and thus do not fulfill the requirement for classification in tertiary education.

ISCED level 5 is the first out of four levels summarized as ‘tertiary education’. It contains all programs leading to qualifications below the level of a Bachelor’s degree but that take at least two years of full-time education to complete. These are mostly vocational programs.⁹ Higher education ‘proper’ starts with ISCED level 6, the Bachelor’s level. Next to the prototypical Bachelor’s degrees, it is also used for other qualifications in higher education that last 3 to 4 years since the end of upper secondary education, such as polytechnic diplomas. ISCED level 7, the Master’s level, contains not just Master’s degrees and other post-graduate qualifications, but also qualifications from long (5 to 6 years) first degree programs, which were common in many countries before the onset of the Bologna reforms and in many countries still exist in certain subject areas such as medicine or law. ISCED level 8 finally is reserved for doctoral programs. Code 9 is foreseen for education that cannot be classified in any of these levels.

These levels are also referred to as ‘main’ levels. Very often, nothing else is actually used, or not even that: official data often only use the three broad reporting categories ‘low’ (up to ISCED 2), ‘medium’ (ISCED 3 and 4) and ‘high’ education (ISCED 5 and upwards). However, since depending on the country, many respondents may accumulate in one or a few ISCED levels despite heterogeneous educational experiences, and specific types of education (such as vocational vs. general) are also of theoretical interest in research, it is worthwhile to also consider the next two digits of ISCED in data coding and analysis. Especially for data producers that provide harmonized data for researchers with diverging measurement requirements, a more detailed than first digit coding is advisable.

The second digit has a different meaning depending on the level of education we are talking about. For ISCED 0, the second digit distinguishes (0) ‘no education’ (at all) from (1) ‘some pre-primary’ and (2) ‘some primary education’—also a new feature of ISCED 2011. At ISCED levels 2 to 5, it distinguishes between (4) general and (5) vocational programs, without any ordering implied. A vocational orientation applies if a program prepares for a specific occupation or class of occupations. All other programs are considered to be ‘general’. At ISCED levels 6 to 8, the terms ‘academic’ and ‘professional’ are sometimes used in a similar vein. However, since the distinction of programs preparing for a specific occupation is not of major importance here, and internationally agreed definitions for distinctions within higher education lacking, the second digit is rarely used in official data at these levels. This is why ISCED provides code (6) ‘orientation unspecified’ specifically for these levels.

The third digit of ISCED 2011 for attainment (ISCED-A) is only defined for levels 2, 3 and 4, and identifies whether the educational program (at least partially) completes the ISCED level (code 2). This is usually not the case for programs of very short duration of e.g. just one year (which get code 1 on ISCED-P and are not considered for classification in ISCED-A). If the program in question completes the level, ISCED distinguishes whether it provides access to a higher level of education (code 4) or not (code 3). These codes are thus ordered and can be regarded as sub-levels. At ISCED levels 3 and 4, access refers to either of levels 5, 6 or 7. So if a vocational program at level 3 only gives access to a program at level 5 but not 6 or 7, it will be classified as providing access, even though it will likely be at a lower standard than a general program providing access to levels 6 and 7.

For the second and third digit of ISCED, code 0 is used when these digits are not specified at the respective ISCED level. For example, at ISCED 1, there is no distinction of vocational and general education and sub-levels don’t apply either, and at ISCED 0, 1, and in tertiary education, all programs are assumed to complete the respective level and to give access to a higher level.

Therefore, another digit is added after the ISCED main level, which indicates the (included) upper bound of a category, while the main level indicates the (included) lower bound.¹¹ If there is no range, the code for the applicable level would simply be repeated. Code 9 on the first (and/or second) digit of GISCED should thus only be used if the applicable main ISCED levels are absolutely unknown. Table 2 shows how the first two digits of GISCED would then look like, in relationship to ISCED main levels.¹²

This leads to a classification where categories are not entirely mutually exclusive, because the ‘spanning’ categories include the non-spanning categories, which is usually undesirable (Hancock 2013). However, this is exactly needed when survey data are not collected or harmonized with the ISCED main levels as a coding target in mind. Education categories that can be coded into a non-spanning category should always be coded there and spanning categories only be used if otherwise missing data would be produced.

For harmonization projects for which the supplementary dimensions of ISCED—i.e. all information beyond main level—is not relevant, this scheme that turns the first ISCED digit into two digits may be entirely sufficient. It will be referred to as ‘GISCED2’. Codes 02 (low), 34 (medium), and 58 (high) would represent the three broad ISCED levels commonly used in statistical reporting. GISCED2 can also be used to identify specific binary distinctions often used in the analysis of educational transitions, such as less than upper secondary (02) vs. upper secondary or more (38), or less than Bachelor’s level (05) vs. Bachelor’s level or more (68). For more demanding harmonization projects, ISCED second digit codes can be appended to these codes, too, e.g. when an education category clearly identifies vocational education, but mixes vocational education at two ISCED levels.¹³

GISCED2 can be approximately transformed into ISCED main levels by dropping the second digit. With this solution, there will be some systematic underestimation of education when simplifying data using the lower bound only, i.e. ignoring the upper bound. However, this is often preferable to the alternative of producing missing data. Sometimes a level between the lower and upper bound may be a more adequate simplification.

Sometimes we do not know whether a qualification is vocational or general because of incomplete information, or a response category contains both. For ISCED 0, the distinctions foreseen by ISCED of whether any pre-primary or primary education was attended are hardly ever available in survey data (although for developing countries, sometimes ‘some primary education’ is separately measured). In a similar vein, we may not know whether a qualification or program gives access to a higher level of education or not—or a response category mixes these different categories.

There are several possible solutions to this problem. Firstly, if we happen to know that one of the sub-categories is clearly dominant in the respective education category, i.e. much more common than the other, the respective sub-category could be coded, disregarding the measurement error for those respondents for whom another sub-category would be more adequate. Secondly, a specific code could be used to signify that we do not have this piece of information. This would be advisable if no specific sub-category dominates the education category found in the questionnaire. Following the solution for ISCED main levels, we will use code 9 across all digits and levels to identify unknown further specifications.¹⁴

Tables 3 and 4 show the codes available for the third and fourth digit of GISCED, corresponding to the second and third digit of ISCED. The only difference compared to the official ISCED coding (UNESCO 2012, Tables 2 and 3, p. 21–22) is then the changed label of code 9. Table 5 shows some exemplary GISCED codes and their labels to illustrate how the different digits come together, and how the lower and upper bounds (first and second digit) can be used to flexibly classify education categories that span ISCED levels.

There are (at least) five situations in which source education categories can be more differentiated than ISCED (see also Schneider and Kogan 2008): the existence of 1. external stratification or ‘tracking’ in secondary and 2. higher education, 3. different types of vocational education, 4. short higher level programs not considered for level completion, and 5. specific categories for incomplete education (dropout). In such situations, harmonizing data using ISCED leads to information loss. The GISCED framework therefore includes separate ‘extension’ variables, in order to allow this information to be retained in the harmonized data, without changing the GISCED code itself.¹⁵ They carry substantive information that goes beyond information harmonized in ISCED. These extensions will be useful for labor market and social stratification research, but are less relevant when education is just a background variable. Data users can use these variables to custom-build harmonized education variables that include the respective kind of information. More specifically, the extension variable ‘edustrat’ (see Sects. 4.3.1 and 4.3.2, and Table 6) allows researchers to complement the GISCED scheme with information about stratification in secondary and higher education. The extension variable ‘vettype’ (see Sect. 4.3.3 and Table 7) complements GISCED by adding information about different types of vocational education and training.

Like ISCED, GISCED normally assumes source variables representing levels of education. With some pragmatic inference, variables including information on completed years of education can, however, be recoded into ISCED levels and thus GISCED as well. In this case, the correspondences provided in Table 9 should be used, which use the cumulative years of education by which ISCED levels are generally defined.²² In such a case, GISCED does not add anything to ISCED main levels apart from the distinction of ‘some primary education’, which forms the second digit of ISCED-A at ISCED level 0 (and is thus included for ISCED-A in Table 9). Then, it may well be better to use derived years of education as the harmonized target variable anyway in order not to throw out any information. For recoding ISCED levels into years of education, following the Luxembourg Income Study (2019), the years shown in bold here should be used.²³

Next, information for links between variables more directly relating to ISCED are presented together. The harmonized variable ‘edulvlb’ is used by the ESS since round 5 (2010). This coding scheme was implemented in the ESS while ISCED 2011 was still in development, but the most important features were already known. Therefore, the relationship between ISCED 2011 and the ESS education scheme is very close, even though the exact codes used at the second and third digits differ. However, the ESS education scheme adds information that is not covered in ISCED, along the lines of the differentiations mentioned here for the extension variable ‘edustrat’ (see Sect. 4.3). The mapping of country-specific education categories to ESS education codes is documented in Appendix A1 of the Data Documentation report of each survey round (European Social Survey 2018a, b, c, 2020a, b). These may thus also help for the harmonized coding of the extension variable ‘edustrat’ in other datasets. Table 10 shows how ‘edulvlb’, ISCED-A and GISCED correspond, using ‘edustrat’ as supporting variable. The close relationship between edulvlb and ISCED-A is very apparent, resulting in GISCED codes that do not actually need the new digit for categories spanning ISCED levels (see Sect. 4.1). New information is located almost exclusively in the extension variable ‘edustrat’.

From ‘edulvlb’, and thus also GISCED with the extension variable ‘edustrat’, it is also possible to derive the so-called European Survey-Version of ISCED (ES-ISCED). This variable has a much lower number of categories than ‘edulvlb’ and is thus suitable for statistical analysis. It has been shown to produce more valid and comparable results than ISCED main levels (Schneider 2010; Schröder 2014), and is thus an attractive alternative to ISCED main levels.

The final comparative education coding scheme to link with GISCED is the CASMIN education scheme, which has its roots in social stratification research. It is not possible to code CASMIN into ISCED directly because some CASMIN categories span across ISCED levels, which GISCED, however, has a solution for (see Sect. 4.1). The mapping of CASMIN to GISCED is more difficult than of the ESS education scheme because in contrast to ISCED, CASMIN is constructed on a relative education scale in order to capture social selectivity effects of education (Brauns et al. 2003; König et al. 1988; Brauns and Steinmann 1999). It thereby identifies both class-specific barriers in educational systems and labor market signals of educational qualifications. This means that the mapping of qualifications to CASMIN is somewhat a ‘moving target’ since selectivity changes over time as educational systems expand. The duration of education as well as vocational vs. general education play a considerable role in both CASMIN and GISCED (as well as ISCED) though, allowing the construction of a rough correspondence.

CASMIN category 1b (‘general elementary education’) is defined as the (non-selective) ‘social minimum’ of education in any given country. To achieve a correspondence, I assume ISCED level 2 to correspond to this ‘social minimum’, even though this may differ across countries and time. Also, CASMIN differentiates general elementary education and intermediate general education, which, historically, maps to primary education (ISCED 1) and lower secondary education (ISCED 2) respectively, but today rather corresponds to differentiations within ISCED 2. This distinction is achieved by combining the extension variable on educational stratification with ISCED. Similarly, when general elementary or intermediate general education is combined with vocational training (categories 1c - Basic vocational qualification or general elementary education and vocational qualification and 2a - Intermediate vocational qualification or intermediate general qualification and vocational qualification), this can only be distinguished using ‘edustrat’, using the same values as with the qualifications without subsequent VET. Additionally, these two CASMIN categories span across ISCED levels. We therefore map both with vocational education at ISCED 2 to 3, resulting in code 2359²⁴. Finally, while CASMIN 3a corresponds to short tertiary education in most countries, which can be mapped to ISCED level 5, for Germany, it refers to the lower tier of higher education, classified as ISCED level 6. As a consequence, CASMIN 3a will be mapped to ISCED 5 and 6 lower tier, and CASMIN 3b to ISCED 6 – 8 upper tier. Table 11 shows how CASMIN coded data can be coded into GISCED, including ‘edustrat’. It is apparent that here, the harmonization using GISCED is much less straightforward than for the ESS education coding scheme, but nevertheless, it can be done (unlike with just ISCED).