Appendix

The possibility of bias due to missing data on ethnicity

Data were missing from the records of 10.2% (2719/26 644) of the consultants in the study. Missing data on ethnicity accounted for almost all the missing items: for example, ethnicity was not recorded for 1747 (9.2%), 901 (7.9%), and 536 (7.8%) of the data for our three calculations (consultants receiving any award, A or Aplus award, and Aplus award). Excluding consultants who were first appointed after 1996, because they did not form part of our analyses (see Method), we compared the 17 230 consultants with known ethnicity with the 1747 with unrecorded ethnicity. Differences between them in relevant respects were fairly small. Consultants whose ethnicity was unrecorded were a little more likely than those with recorded ethnicity to be women (23.3% (407) compared with 20.4% (3517) consultants) and to have trained abroad (22.1% (386) v 17.5% (3020) consultants). They were more likely to work outside the former Thames regions (31.2% (545) v 18.7% (3220)), to work on an honorary contract (14.5% (253) v 8.5% (1460)), to work in a teaching hospital (32.0% (559) v 27.5% (4733)), to be recent graduates (48.0% (839) v 39.0% (6729) since 1992), and not to hold an award (95.3% (1665) v 80.8% (13 926)). Each of these comparisons was significant (P<0.01).

Multivariate adjustment: the example of recent B awards

We fitted multifactor models by logistic regression. Table A illustrates the effect of successive adjustments on the odds ratios for recent B awards shown in table 6. For example, the difference between women and men decreased after adjustment for length of time as a consultant (odds ratio changed from 0.72 in model 0 to 0.80 in model 1) and decreased more after adjustment for specialty group (odds ratio 0.84 in model 2) and type of contract (odds ratio 0.96 in model 3). Further adjustment to take account of type and location of hospital had little effect. For British trained consultants, the differences between white and non-white consultants decreased and lost statistical significance after adjustment for year of appointment (see table A, transition from model 0 to model 1). For consultants trained abroad, no appreciable difference existed between white and non-white consultants after multivariate adjustment (table A, transition from model 0 to model 5); but both the white and the non-white groups of consultants trained abroad were less likely than those trained in the United Kingdom to receive an award (table A, model 5).

Strength of associations and interactions in the models

Table B (top) shows the regression statistics for the final model (model 5) of table A. To further validate this model we extended it to include significant (P<0.05) interaction terms; two interactions, those between the type of contract and the year of appointment and between the type of contract and the type of hospital, were found to be significant (P<0.05).

Table B (bottom) shows the regression statistics including the interaction terms; the results displayed as model 5 in table A correspond to this version of the model. These results suggest that differences between rates of award for consultants with different types of contract vary with two factors—the type of hospital and the year of first appointment. The first is confirmed, for example, by examining the subgroup data. Consultants with honorary contracts have similar rates of award in teaching hospitals (30.5%; 236/775) and district general hospitals (24.6%; 93/378), but consultants with whole time contracts show a larger relative difference (14.1% (249/1767) in teaching hospitals and 5.9% (414/7014) in district general hospitals).

Table A: B awards in the last five years (1998-2002) by sex, ethnicity, and place of training, showing that the significance of these factors decreases with adjustment for specialty group, year of appointment, contract type, hospital type, and location of work. Values are odds ratios (95% confidence intervals)

*Differs significantly (P<0.05) from 1.

Model 0: effects of sex, ethnicity, and place of training on odds of receiving the B award.

Model 1: as model 0, with adjustment for grouped year of first appointment.

Model 2: as model 1, with adjustment for specialty group.

Model 3: as model 2, with adjustment for type of contract.

Model 4: as model 3, with adjustment for type of hospital.

Model 5: as model 4, with adjustment for location of hospital and interactions between contract type and year of first appointment and between contract type and type of hospital.

Table B:  Logistic regression for B awards in the period 1998-2002

All factors

All factors with significant (P<0.05) interaction terms added

Wald denotes the Wald statistic (see method), which is the analogue of the t test used to establish significance in least squares regression.