Almost all students took GCSEs, only 649/13698 (4.7%) not taking any GCSEs, and those individuals were excluded from further analysis. The simplest overall measure of GCSE performance is the total number of points attained (where A = 5, B = 4, C = 3, D = 2, E = 1, else = 0). It should be noted that although for a number of years GCSEs have awarded a grade of A*, which is higher than A, those A* results are unfortunately not distinguished from A in the YCS database. Inevitably any simple points total is influenced by the total number of GCSE examinations for which a student chooses to enter or is allowed to enter, and so we also provide the mean grade attained at GCSE. Overall there is only a very small absolute difference in the number of GCSEs taken by white and non-white candidates, which just reaches significance (see Table 1), but there is a highly significant difference in the number of points and the mean grade attained, p < .001 in each case. Figure 1 shows the distribution of points for W and NW students, and although it is clear that the NW students have a lower mean, which can be seen particularly in the left-hand tail, at the top end of the distribution the differences are far less marked, such that the number of A grades attained is more similar in W and NW students, and the proportion of students gaining 6 or more A grades B the pool of those with a realistic chance of subsequently applying for medical school B is not significantly different in the W and NW candidates, 10.7% and 9.8% respectively, the effect size being B.054. Table 1 also shows the proportions gaining A grades in the main subjects taken, and for most subjects there is no significant difference between W and NW candidates.

Although almost all students take GCSE at age 16, there is an element of choice in the particular subjects chosen for study. There is also good evidence that not all subjects are equally difficult, some such as Chemistry, Physics, Biology and Latin being harder subjects in which to gain top grades than Art, Drama, and Sociology 35. We therefore calculated a measure of the average difficulty of all the GCSE subjects that had been taken by a student, based on Coe = s estimates of the Rasch difficulty for an A* grade. The choice of A* for this calculation has little impact overall, since the Rasch difficulties for different grades are, to a large extent, parallel 35, and A* has the advantage for present purposes that it is the grade which most medical students would be expected to achieve. Table 1 shows that there is no significant difference between W and NW students, meaning that NW students take GCSEs of similar difficulty to W students.

W and NW students differ in a number of demographic factors, as can be seen in Table 2, the NW students coming from a somewhat lower socio-economic group, and their parents having somewhat less education, although interestingly there are no differences in the proportions attending private schools.

Achievement at GCSE depends on a number of variables, and as Table 3 shows, those variables are themselves inter-related, both in correlational and causal terms. We therefore used LISREL to fit a path model, shown in Figure 2. The causal ordering was chosen on the basis that the number of GCSEs taken and the difficulties of the GCSEs taken were both immediately prior to the total points attained, and that type of schooling was prior to taking GCSEs. The sex of the student, being fixed throughout the student = s life, was prior to all educational variables. The remaining variables were related to the student = s families as much as to the student themselves, and hence were earlier in causal terms, with parental socio-economic group being likely to be secondary to parental education, and parental ethnicity prior to parental education. The logic is that the ethnicity of parents and offspring is generally the same, and fixed throughout the lifespan of parents and offspring, whereas education of both parents and offspring is less fixed, so that in any causal model, the ethnicity of individuals and their parents is to the left of measures of education. The number of individuals of mixed ethnicity (who are here classified as NW) is relatively small, and hence little affects the above argument, and it is still the case that an individual's ethnicity precedes their education causally.

The model was fitted by firstly using a saturated model in which all variables to the left of a variable could cause all variables to the right of the variable, through the BETA matrix in LSIREL. Paths which were non-significant at the 0.05 level were dropped sequentially from the model, with the least significant first, until all paths remaining were significant.

The final fitted model is shown in Figure 2. The overall goodness of fit was excellent, with χ² = 8.92, 8 df, p = .349, a goodness of fit index of 1.000, and an adjusted goodness of fit of 0.999. Figure 2 shows a number of readily interpretable effects. Higher parental education results in a higher parental socio-economic group, and both factors result in a child being more likely to take part in private schooling. Private schools are more likely to put children in for more difficult GCSEs, and the children also gain more points at GCSE. Total GCSE points is particularly dependent on number of GCSEs taken, and the difficulty of the GCSEs taken, and there are also direct influences of private schooling, as well as effects of sex, parental social class, and parental education which are not explained by schooling. Finally it should be noticed that ethnicity has multiple effects, as a result of NW students having less parental education and coming from a lower socio-economic group, but after taking those into account, ethnic minorities are somewhat more likely to attend private schools.

To summarise the results so far, NW students do achieve slightly less well at GCSE, but that is largely, but not entirely, mediated by a number of background variables such as parental education and type of schooling. In particular, the numbers of NW students attaining good grades at GCSE (6 or more As) is similar to that in W students, both in a simple analysis, and also after taking background factors into account. From the point of view of understanding the potential pool of medical school applicants B those performing particularly well at GCSE, in other words B W and NW students do not show significant differences in GCSE achievement.

Ethnic differences in aspirations to take A levels were explored further using multiple regression and structural equation modelling, looking at taking of A-levels in relation to the eight variables shown in Figure 2 that relate to GCSE attainment and background characteristics. Forward entry logistic regression with taking of one or more A levels as the dependent variable, and the eight background variables as predictors, found significant effects of all the measures except private schooling. Predictors of taking A levels, in order of entry into the regression (but with significance levels calculated after taking all other variables into account) were higher numbers of GCSE points (p < .001), taking fewer GCSEs (p < .001), being non-white (p < .001), coming from a higher socio-economic group (p < .001), being male (p < .001), having greater parental education (p < .001), and taking more difficult GCSEs (p = .027). NW students were 1.979 times more likely to take A levels than white students (effect size = .377), all other background factors being taken into account. Very similar results were obtained for taking at least one science A level, and for taking Chemistry A level, NW students being 2.370 times more likely to take a science A level (effect size = .477), and 2.735 times more likely to take Chemistry A level (effect size = .556). Figure 4 shows a path model for taking A levels, and not only can the effects of ethnicity be seen, but it is also clear that NW students are more likely to take a science A level, even after taking into account the likelihood of taking A levels, and are more likely to take Chemistry A level, even after taking into account the fact that they are taking science A levels.

Although overall NW students have poorer achievement at GCSE, a previous analysis here had suggested that NW students are equally likely as W students to achieve very high GCSE grades (6 or more A grades or better). The logistic regressions described in the previous section were therefore repeated only for students with 6 or more A grades at GCSE. Overall there was now no effect of ethnicity on taking one or more A-levels, although that largely reflects the fact, seen at the right hand end of Figure 3, that 97% of such a high achieving group go on to take A levels. However, a similar analysis for the taking of science at A level and the taking of Chemistry showed that even among GCSE high-achievers, NW students were 2.85 times more likely to take a science A level (p < .001; effect size = .579), and 2.45 times more likely to take Chemistry A level (p < .001; effect size = .495).

As with GCSEs, not all A level subjects are equally difficult. We therefore used the data of the CEM Centre at the University of Durham 36 to calculate a difficulty score for the subjects being taken by each student at A level. Multiple regression of A level difficulty on the six background variables and two measures of GCSE attainment shown in Figure 2 found that more difficult A levels were taken by students with more GCSE points (p < .001), but who had taken more difficult (P < .001) but somewhat fewer GCSEs (P < .001), were female (p < .001), and were non-white (p < .001). However, since Chemistry is one of the most difficult of A levels, the latter effect is perhaps not unexpected. Repeating the analysis but including a variable indicating whether or not students were studying chemistry at A level meant that the effect of ethnicity was no longer significant (p = .207).

In summary, despite having somewhat poorer overall attainment at GCSE than W students, NW students are more likely to go on to take A levels, and in particular science A levels and Chemistry A level, which tend to be more difficult. The same was also true of science A levels and Chemistry A level amongst the high achieving students with 6 or more A grades at GCSE.

For present purposes we are only interested in A levels, and not AS levels, since most medical students have at least three A levels, and they are the main criterion for medical student selection. As is also the case in many medical schools, we have also not included General Studies as a > proper = A level subject. Figure 5 shows a scatter plot of total points at A level (calculated as was conventionally done by UCAS as A = 10, B = 8, C = 6, D = 4, E = 2, otherwise 0), in relation to total points at GCSE, with a lowess line fitted separately for W and NW students. The overall correlation is 0.628 (N = 2456), and was 0.633 (N = 2131) in the W students, and 0.606 (N = 326) in the NW students. It is also clear, in Figure 5, that the fitted lowess regression lines are very similar. To a first approximation therefore, GCSE achievement predicts A level outcome equally well in W and NW students, although there is a trend towards NW students with high GCSE grades underachieving to some extent at A level. Regression of A level points on GCSE points showed that the linear component of GCSE points accounted for 39.5% of the variance in A level points. Inclusion of ethnicity did not significantly improve the fit of the model (R² change < 0.1%; p = .470), although inclusion of an ethnicity x GCSE points interaction was significant (F(1,2472) = 12.7, p < .001) accounting for a further 0.3% of the variance. Fitting lines separately to W and NW groups, showed that the unstandardised regression coefficient was somewhat higher in W students (b = .660, SE = .017; beta = .633, R² = 40.1%) than in non-White students (b = .515, SE = .038, beta = .606, R² = 36.7%). GCSE grades therefore predict A levels slightly better in W than NW students, although the effect is small, and the broad picture is of similarity. Since NW students taking A levels have lower GCSE achievement than W students, it can be predicted that NW students will perform less well at A level than will W students.

Of the 13,049 individuals in sweep 1 of YCS, only 6692 (51.3%) responded on sweep 3 when A level results were asked for, and of these only 2731 had actually taken one or more A levels. (Note that these figures are slightly different from those cited at the beginning of the section on A level achievement as those figures are raw counts of number of respondents, whereas the figures quoted here, and subsequently, refer to weighted responses). Of the 3478 individuals at sweep 1 who said they did not intend to take A levels, only 74 (2.1%) actually did, with no difference in proportion between W and NW (2.1% and 2.4% respectively). However of the 3214 individuals at sweep 1 who said they intended to take A levels, 2657 (82.7%) actually did, and the proportion was higher in W students (2314/2780 = 83.2%) than in NW students (342/433 = 79.0%), a significant difference (chi-squared = 4.73, 1df, p = .030; odds ratio = 0.760, effect size = B.152), suggesting that more NW than W students had decided their aspirations were inappropriate or impractical, although the data cannot decide between those possibilities. Those not actually taking A levels who had said they would, had significantly lower GCSE point scores (mean = 29.1, SD = 10.24, N = 557) than those who did take A levels as intended (mean = 48.56, SD = 8.68, N = 2657, t = 22.76, p < .001). Similarly, although fewer NW students actually took A levels than said they would, the NW students who actually took A levels had significantly lower GCSE points than the W students (W: mean = 38.62, SD 8.63, N = 2376; NW: mean = 36.32, SD = 10.39, N = 355; t = 4.56, p < .001).

Given the lower achievement at GCSE it is not surprising that NW students also had lower total point scores at A level (see Table 4). Numbers taking chemistry and other individual subjects were too small to make useful comparisons. The effect just described is a simple effect, without taking background factors into account. As before it is useful to summarise the data on the different variables determining A level performance by means of a path analysis, and Figure 6 shows a LISREL model for the determination of A level points in terms of nine background variables. The most important determinant of A level points is the number of A levels taken, for fairly trivial reasons. GCSE points also determine A level points, both directly, and also indirectly via the number of A levels taken, better students at GCSE taking more A levels. Better students at GCSE also tend to take slightly more difficult A level subjects, and as a result do slightly less well than if they had taken easier subjects. Students taking more difficult subjects at GCSE also tend to take more difficult subjects at A level. Schooling, socio-economic group, and ethnicity do not have direct effects upon A level performance, although they have a range of indirect effects, whereas parental education level has direct effects, on both GCSE points and on A level points. Of particular interest is that there is no direct effect of ethnicity upon A-level performance, all effects being mediated indirectly via background factors and measures of performance at GCSE.

The YCS is a useful database for looking at the entire population of individuals taking GCSEs and then A levels. However, the numbers involved are too small to be useful for looking at applicants and entrants for particular university subjects. The UCAS database is extremely large, the three years from 2003–2005 having a total of 1,484,650 applicants, 476,467 in 2003, 486,028 in 2004 and 522,155 in 2005, of whom 52,557 applied for medicine, and 23,443 were accepted at medical school. However the database is weak in terms of the richness of the background variables, having only A levels, and not GCSEs, for instance, and having only a limited number of demographic variables. And of course, of necessity, it can only look at those individuals who have applied to university. Nevertheless, the UCAS data complements and supports the YCS data.

Although most of UCAS = published analyses are in terms of its > tariff score = 37, the tariff score has several disadvantages when thinking primarily about medical school applicants. The tariff combines different qualifications which are on different metrics, but that has the disadvantage that the origin of any particular tariff is far from clear. For instance, a candidate gaining B grades at three A levels will receive 3 @ 100 = 300 points, as also will a candidate gaining two As and a D (120+120+60), as also will a candidate taking six AS levels and gaining Bs in all of them (6 @ 50). The tariff also considers all such candidates equally, irrespective of the level or number of qualifications or the subjects in which they were taken. In particular, the tariff score includes General Studies A level on an equal footing with all other A levels, whereas many medical schools do not regard it as a full A level. For the present study we have therefore recalculated attainment scores de novo, and have considered only A levels, scoring them in the same way as for the YCS, with 10 = A, 8 = B, etc.. For most medical schools it is A levels that matter, and not combinations of A and AS levels. UCAS was unable to provide data on Scottish Highers, and therefore our analysis is restricted to those taking A levels, being typically candidates outside of Scotland.

For convenience and ease of interpretation we have confined our analyses to applicants who were aged under 21 at the time of application (making them broadly equivalent to the YCS data), and who were resident in the UK, giving a total sample of 976,007. Some applicants had not taken A levels, for various reasons, in particular being resident in Scotland or Wales, and taking Highers or Baccalaureate, and we therefore considered only the 531,333 candidates who had results for at least three A level subjects (excluding General Studies). A very small number of candidates, 2642, had five or more A levels, and for practical reasons we excluded them also, leaving 528,691 applicants, 473,781 with three A levels and 54,910 with four A levels. For simplicity, for candidates with four A levels we have considered only the grades from their three best results, meaning that for all candidates the maximum number of points is 30, which is equivalent to AAA. Although this might seem to devalue the additional qualification obtained by those with four A levels, it is worth noting that while 10.5% of the applicants with three A levels had attained the maximum score of 30 points, that score was attained by 48.9% of the applicants with four A levels, showing that those taking four A levels tended to be amongst the very highest achieving of all applicants.

In order to compare UCAS applicants with all those taking A levels, the data from the YCS A levels were expressed in the same way as for the UCAS applicants, considering only students gaining three or four A levels at grade E or above, and for those taking four A levels, using the three best results. This score correlates highly with the total points score used in the earlier analyses (r = .957). Figure 7 shows the possible scores for three A levels, from 6 (EEE) through to 30 (AAA). The YCS data were compared specifically with the 2003 UCAS applicants, since that cohort was the closest in time of the UCAS cohorts to the YCS cohort. Although the distributions are broadly similar, in particular being > censored= in a similar way at the top end, the mode in both cases being at AAA due to no higher grades being available, it is also clear that the YCS group (red bars) has somewhat more individuals at lower grades, and fewer individuals at the highest grades. UCAS applicants therefore have somewhat higher grades than the population as a whole, which would be expected as only the better A-level candidates will apply to UCAS, but otherwise have a broadly similar A level distribution to that found in YCS. .

Figure 8 shows the distribution of A level grades in the UCAS dataset in relation to ethnicity. The broad distributions of each group are similar to those in Figure 7, with a mode at 30 points due to censoring. A simple comparison of W and NW students finds that W students have a higher achievement than NW students (Table 5). Of interest is that the overall effect size comparing W with NW is B 0.140, which is similar to the effect size found in the YCS data of B 0.141 (see Table 4). An achievement of grades ABB or above (i.e. 26 points or equivalent, which can be regarded as a high A-level achievement), is reached by 35.7% of W applicants overall (i.e. for any subject), compared with 32.4% of NW applicants (Chi-squared = 308.8, 1df, p < .001, odds ratio = 1.160, effect size = B.082).

Although the UCAS database has 517,300 university applicants with three or more A levels, only 97,551 (18.9%) of these have taken A level chemistry, and hence can be considered as being in the > real pool = of possible entrants to medical school. Amongst those taking chemistry, 25.2% are non-white, compared with 11.8% of those not taking chemistry (odds ratio = 2.518, effect size = .510), again, a very similar figure to the odds ratio found in the YCS data of 2.735. The mean total A level points of NW students taking chemistry is significantly less than that for W students (NW: mean = 22.92, SD = 6.35, N = 24626; W: mean = 23.83, SD = 6.03, N = 72925; t = 20.26, p < .001; effect size = B.151), as can be seen in Figure 9, where it is very clear that students taking A level chemistry are a high-achieving sub-group of students compared with the population in general (Figure 8).

For the 23,247 applicants to medical school, of whom 8,174 (35.2%) were NW, the average number of A level points was significantly higher in the W applicants than the NW applicants, with an effect size of B.282 (see Table 5). Figure 10 shows that the W applicants have a higher proportions of 30 points and 28 points, whereas NW applicants have a higher proportions of all other grades. In view of the non-normal distribution, a non-parametric test was used to confirm the difference in distributions (Mann-Whitney U = 54151179, z = 15.99, p < .001). Table 5 also shows that NW applicants were significantly less likely to achieve 26 points or more, which is equivalent to ABB, (odds ratio = 0.634, effect size = B.252), or achieve grades of AAA (odds ratio = .724, effect size = B.178).

Of the 14,318 entrants to medical school, the average number of A level points was 28.77 (SD 1.97) in the 9,945 W entrants, and 28.54 (SD 2.38) in the 4,373 NW entrants, a highly significant difference (t = 5.89, 14316 df, p < .001), with an effect size of B.114. Figure 11 shows the distribution, with W candidates being more likely to have 30 and 28 points, and NW candidates to have 28 points or less. In view of the skewed distribution, a non-parametric test was used to confirm the difference in distributions (Mann-Whitney, U = 21057283, Z = 3.453, p < .001). Table 5 also shows that NW acceptances were significantly less likely than W acceptances to have achieved 26 points or more, equivalent to a grade of ABB or more (odds ratio < .001; odds ratio = .721, effect size = -.186), or to achieve a grade of AAA (odds ratio = .919, effect size = B.047).