Results

Response rates and reasons for exclusions from the four cohorts are presented in table 1. Mothers and fathers who did not participate caused exclusion of 18% and 27% of the adolescents from the SES analyses and 38% and 43% from the education analyses, respectively. Among those adolescents excluded because of lack of parental data, significantly fewer were living with the parent (tables 2 and 3). In the SES cohorts, there were also significantly higher proportions of girls among those excluded because of non-participation of the parent (table 2). Judged by the data provided by the adolescents, the cohorts contained higher proportions of parents classified as having middle/higher non-manual occupations or college/university education, compared with those excluded because of lack of parental data (tables 2and 3). Those excluded from cohort 1 (fathers' occupation) also had a higher proportion of non-responses than expected by chance (table 2).

Within the cohorts, the adolescents not responding or giving incomplete answers did not differ from their peers with complete answers with respect to physical activity, smoking, gender, or living with the parent in question (results not shown). Among the non-responding 13 year olds approximately one third of their parents also did not respond, and there were lower proportions of adolescents with parents in the middle/higher non-manual category than expected by chance (see ).

Descriptive statistics (tables 4 and 5) show the classification into SES groups based on occupation and educational level according to the reports given by the adolescents and parents. Table 4 shows a high resemblance in the proportion categorised into the different SES groups based on responses to the occupational question given by adolescents and their parents. The main difference was observed in the proportions of unclassified and no response answers, which were higher among the adolescents than among the fathers. However, the proportion of unclassified answers relating to fathers' SES decreased from age 13 to age 15 for the adolescents. Table 5 shows the results from the questions on parental education. All parents replied to this question, while 13%–14% of the 15 year old adolescents did not. This non-response rate is of about the same magnitude as the sum of unclassifiable and no response answers to the occupational question in table 4. The distribution of parents' educational level based on the adolescents' answers appeared to be systematically lower than the distribution based on the parents' own reports.

Adolescents' reports at ages 13 and 15 on their parents' SES group showed high interrater stability, with κ statistics above 0.7 (results not shown). Similarly, there were only small or no changes depending on whether adolescent reports at age 13 or 15 were used for comparison with parents' reports on SES groups (table 6). This finding was persistent across gender of the parent and different combinations of SES categories.

More than 50% of those not responding or giving unclassifiable answers at age 13 were in the classified group at age 15. Still, some 40% of those who did not respond at age 13 also did not respond at age 15 (see).

Discussion

The results of this study support those of Westet al 9 by providing evidence that adolescents as young as age 13 or 15 are able to produce valid answers to questions about their parents' SES based on an open ended question about parent occupation. The strength of agreement between adolescent and parent reports on SES groups was good according to the rule of thumb for evaluating κ values.15 The strength of agreement was fairly similar whether pertaining to the SES of the fathers or SES of the mothers, indicating that for validity reasons there is no reason to prefer one to the other. In contrast with the review by Looker,8 in which she found increasing accuracy of proxy reports with increasing age of the child, we did not find significant differences in strengths of agreement when using data obtained from the adolescents at age 15 compared with data obtained at age 13. However, the lower value of the κ statistic found for education does support Looker's8 notion that children's proxy reports of parental characteristics remote in time (for example, education) may be less valid than their reports on present characteristics (for example, occupation). As the κ statistic is known to be sensitive to both prevalence in each category and the number of categories,15 it was surprising to find that the strength of agreement did not vary much between the three different ways of grouping the SES categories.

There were especially two types of attrition in this study that could potentially cause a threat to the generalisability of the results—parents who did not participate and non-response or unclassifiable answers from the adolescents. Based on the adolescents' responses the groups of parents who did not participate seemed to contain lower proportions of parents in middle/high non-manual occupations and parents with college/university education than the groups of parents who participated. This type of response bias towards high SES/education is not unusual in health surveys,13 16 17 and could bias the results of analytical studies based on such data. However, we do not think children of parents not participating in health surveys would provide less valid answers on measures of SES than those whose parents do participate in health surveys. Difficulties in reporting could occur in low SES families when parents have multiple jobs or frequently change jobs, but they may in general have an easier time giving a single label description of the job in question. High SES families on the other hand could have difficulties in giving a short and comprehensive description because of the complexity of the jobs.

In agreement with results from other studies,7 8 the proportion of non-responses or unclassifiable answers among the adolescents whose parents did participate was substantial (11%–16%). The χ² analyses of adolescent non-responders and those providing unclassifiable answers indicated that a higher proportion of the parents of these children were also non-responders, which supports Looker's8 theory that some families may have more problems reporting on this measure. As outlined above both low and high SES families could have difficulties providing a simple answer to a question on type of work. However, based on the data presented here we cannot conclude that one of these explanations is more probable than the other.

Among the parents who did not participate in the study there were lower proportions of the adolescents who reported that they were living with the parent in question. This could partly explain the lack of response from the parent as well as the tendency towards higher proportions of non-response/unclassifiable answers by the adolescents. However, when the parent did participate whether the adolescent was living with the parent in question did not seem to differ between responders, non-responders, or those giving answers that were not classifiable. These apparently contradicting results may be explained by the type of contact the adolescents have with the parent they are not living with. Frequent contact or good relations may ensure that the parent participates and that the adolescent knows the work of the parent. Thus, the adolescent's ability to provide a valid answer may not be a simple question of living with the parent or not, as suggested by the studies reviewed by Looker.8

The values of the κ statistic for education were much lower than those for SES (based on occupation), but within the interval of fair values according to Altman.15 There are at least two obvious and plausible explanations for these low agreements between adolescents and parents on this question. Firstly, low agreement could be caused by lack of relevance and closeness in time of parental education in the daily life of an adolescent. The data were, however, collected half a year before the adolescents had to make a major choice about their own future plans, as mandatory school in Norway ends at age 16. The issue of work versus education, and for those opting for education the issue of vocational versus general studies, could therefore be expected to be relevant to the adolescents.18The second explanation is methodological, and pertains to the different answer categories provided in the adolescent and parent questionnaires. The options in the adolescent questionnaire reflected former versions of the school system. The answer categories thus differentiated between the amount of years of mandatory school (7 or 9 years) and also provided three different options at the voluntary upper secondary level (vocational, trade and general), while there was only one option for higher education. In addition to being confusing in itself, the number of options for lower level education could have caused a systematic reporting bias relating to parents with college degrees in business or engineering, for instance. As the adolescents themselves were about to make a choice about their upper secondary education, they may have been less aware of the possibilities beyond this level.

Although the κ statistic is considered better than percentage agreement, its widespread use, or misuse, has been criticised by Maclure and Willett.14 Firstly, their criticism relates to the extension of the use of the κ statistic beyond dichotomous data for which it was originally intended, and thus to its application to nominal as well as ordinal polychotomous data.14 Secondly, they point out that the κ statistic was not intended for assessing validity, because in validity studies one of the instruments is expected to provide a truer value than the other, and thus other measures of agreement may be more informative.14 However, the measures of agreement suggested to assess validity are better for analysing continuous or true ordinal polychotomous data than quasi-ordinal data like the ones presented in our study. Furthermore, Nelson et al 19 point out that comparison of index and proxy respondent data often has been considered as an issue of reliability rather than validity. Westet al 9 argue that on some issues like parental occupational status, the children's reports may be more truthful than their parents', thus questioning the parental answers as the gold standard by which to measure validity.

Despite his demonstration of the κ statistic's sensitivity to both the prevalence in each category and the number of categories, Altman15 argues strongly for the use of κ statistics, as other statistics often used (for example, Pearson's correlations coefficients and the χ² test) measure association rather than agreement. However, the requirement for a perfect agreement between the answers given by the two responders does make the κ statistic a conservative measure of agreement. The weighted κ has been constructed to take into account the off diagonal distance between mismatches.15 Because of the characteristic of the data in this study, assigning weights to distances between the SES categories would have been arbitrary, and thus of little use. The high and similar values of the κ statistic across the three models may be taken as an indication that grave mismatches between parents and adolescents' reports were not common and thus the influence of the number of categories less important. These results are encouraging for future analyses on differences in self reported health and health related behaviour of these adolescents.

Our study shows that an open ended question about parental occupation in a questionnaire answered by adolescents at the age of 13 and 15 provides valid data on parents' SES among the 85%–90% who gave classifiable answers. Researchers may be able to increase the proportion of classifiable answers regarding parental occupation by including more specific questions to cue adolescents as to what information to provide. More specific questions could consequently also increase the precision of the answers, thus making assignment of job codes less dependent on personal judgement of researchers. This approach would be in line with the intended use of the SES system, which requires data from a range of questions to assign a person to one of 33 categories.11 In addition, such an approach might be less sensitive to economic changes than developing additional questions on material wealth as suggested by Currie et al,7 even though social status related to occupation clearly is not indifferent to economic and cultural changes. An alternative to including more questions on the survey about parent occupation is to have the researcher (or the person distributing the questionnaire) emphasise the importance of answering the question carefully and encourage students to ask for help. This would be a similar, but less costly approach to the “mini-interview” performed by West et al.9 However, the ethical aspects in relation to the anonymity of the adolescent would have to be considered carefully.

Because of the methodological problems related to the question on education in our study, further research should investigate the possibility of adolescents' ability to provide valid proxy reports of their parents' education. This may be an especially salient issue when considering health related behaviour in adolescents as a mediating factor between social status and inequalities in mortality and morbidity later in life,20 as research on health related behaviour among adults has shown a relation between those outcomes and level of education.21 22