Escolar Documentos
Profissional Documentos
Cultura Documentos
OBJECTIVES The objectives of this study using the more conventional approach of
were two-fold. Firstly, we aimed to model logistic regression.
typologies of student examination perfor-
mance by grouping students into like RESULTS The LCA identified three distinct
categories based on measures of prior classes or typologies of student examination
academic achievement (particularly in the performance using measures of prior academic
science subjects) and interview rating at time achievement and interview rating at time of
of entry to a medical degree course, and out- course entry. Measures of prior academic
come measures of subsequent performance achievement and score on a structured admis-
across the course. Secondly, we aimed to sions interview made significant contributions
illustrate and evidence the utility of the latent to the models ability to discriminate between
class analysis (LCA) clustering technique to typologies. Strong prior academic achievement,
provide meaningful information on the especially in chemistry, and high interview score
effectiveness of a student selection process with were positively related to the likelihood of suc-
respect to the likelihood of poor examination cessful test performance. These findings were
performance. supported by the logistic regression analysis.
METHODS For this retrospective study, CONCLUSIONS The LCA clustering tech-
anonymised data on two sequential cohorts of nique provided meaningful information on the
students who graduated from a 5-year Bachelor performance of a selection process. As a com-
of Medicine, Bachelor of Surgery degree course plementary tool to existing methods used in
were analysed using LCA. In order to triangu- this area of research, LCA has the potential to
late the findings, the same data were analysed empirically inform the selection process.
Institute of Clinical Education, Peninsula Medical School, University Correspondence: Paul Lambe, C409 Portland Square, University of
of Plymouth, Plymouth, Devon, UK Plymouth, Drake Circus, Plymouth, Devon PL4 8AA, UK.
Tel: 00 44 1752 588950; Fax: 00 44 1752 586788;
E-mails: paul.lambe@pms.ac.uk, david.bristow@pms.ac.uk
308 Blackwell Publishing Ltd 2011. MEDICAL EDUCATION 2011; 45: 308316
Predicting student performance from entry attributes
Blackwell Publishing Ltd 2011. MEDICAL EDUCATION 2011; 45: 308316 309
P Lambe & D Bristow
assume a linear relationship, normal distribution of who will be invited for a structured admissions
data or homogeneity of variance) for identifying interview. A-level and AS-level qualifications are post-
subtypes (classes) of related cases from multivariate compulsory education (1618 years) advanced-level
categorical data. The classes are defined by the examinations taken by those who want to study at
criterion of conditional independence: conditional university in the UK. All applications to UK univer-
on latent class membership, the observed variables are sities are made through the Universities and Colleges
mutually independent of one another. Thus, the Admission Service (UCAS); the UCAS tariff system
variables within any one class are uncorrelated. Pos- allocates points for qualifications, equivalised
terior membership probabilities (maximum likeli- according to the level and type of qualification and
hood estimates based on patterns of scores on the related to grades (AE) that the applicant is pre-
observed variables) assign cases to homogeneous dicted to achieve by his or her institution. In these
latent classes. Cases in a particular latent class are analyses, the UCAS22 points system was used to
similar to one another because their response patterns calculate a score for each student based on his or her
are generated by the same probability distribution. actual attained qualifications at entry to Year 1 of the
programme.
Two useful parameters are produced: latent class
probabilities and conditional probabilities. Latent Selection interview
class probabilities indicate the relative size of each
class and whether the sample population is relatively A structured interview informs the offer of places on
evenly distributed among the classes. Within each the undergraduate medical programme. A students
class there is a set of conditional probabilities relating interview total score is based on his or her answers to
to each indicator. The conditional probabilities a series of questions constructed to elicit perfor-
represent the probabilities of an individual in a mance in 10 areas (communication skills, empathy
particular class being at a particular response level for and pro-social behaviour, flexibility, decision making,
a particular indicator and thereby enable characteri- teamwork, honesty, manner of reflection, insight
sation of the nature of the types defined by each of about self, insight about illness and medicine, insight
the latent classes. Analogous to factor loadings in into stress). A panel of at least three trained
factor analysis, the conditional probabilities provide interviewers, generally comprising a practising
the measurement structure that defines the latent doctor, an academic and a representative of the
classes. community, rate each applicants performance in the
interview. The average over the three interviewers of
Specification of the basic LCA model does not involve the sum of ratings for the 10 attributes was used in
the issue of causal ordering. In these analyses, our these analyses.
interest is in the underlying association between
variables and we make no distinction between what is Examination performance
dependent and what is explanatory. Used here as a
data reduction technique, the method empirically Medical students AMK is tested regularly by progress
establishes the dimensions of any underlying tests delivered four times per year.23 The test is
associations. designed as a longitudinal assessment of the growth of
a students medical knowledge across the whole
Data programme. Scores on the fourth test in each of the
5 years of the degree course were used in the analyses.
For this retrospective study, anonymised data were
collected on two sequential cohorts of students who An integrated structured clinical examination (ISCE)
graduated from a 5-year Bachelor of Medicine, is undertaken by students at the end of Years 2 and 4;
Bachelor of Surgery degree course (n = 142; 73 this is comprised of a number of timed stations that
females, 69 males). All students included in these aim to assess a students clinical competence and
analyses were direct school-leavers at time of entry. professional ability.24 Students final scores were used
in the analyses.
Prior academic achievement
In Year 5 of study, students undertake patient-based
The medical school employs academic criteria presentations (PBPs) which assess ability to clerk,
(predicted or attained A-levels and AS-levels or present and lead a patient-based discussion. Clinical
equivalent qualifications) and, since 2006, the UK and professional competence is graded on a scale
Clinical Aptitude Test (UKCAT) to select students of 14 (1 = unsatisfactory, 2 = borderline,
310 Blackwell Publishing Ltd 2011. MEDICAL EDUCATION 2011; 45: 308316
Predicting student performance from entry attributes
3 = satisfactory, 4 = excellent) in eight PBPs. probability of the outcome scoring in the bottom
Students total scores were used in the analyses. quartile or not in an end-of-year progress test, given
an individuals level on selected predictors. The
Data analysis predictors used in the binary logistic regression
model were: grade at A-level chemistry; grade at
A retrospective study using LCA was conducted using A-level biology; having a bottom-quartile interview
Latent GOLD Version 4.0 (Statistical Innovations, score or not, and having a bottom-quartile UCAS
Inc., Belmont, MA 02478, USA).25 The indicators entry tariff or not.
included in the model are the dependent variables
used to define the classes. The inactive covariates do Ethical considerations
not influence the model parameter estimates and are
included in the model to enable further description Ethical approval for this study was sought from the
of the nature of the classes.20,26 medical schools ethics committee, which considered
that, as the research did not raise any human subject
The categorical variables included as indicators in the protection issues, ethical approval was not required.
model were:
Blackwell Publishing Ltd 2011. MEDICAL EDUCATION 2011; 45: 308316 311
P Lambe & D Bristow
Table 1 Goodness-of-fit statistics of latent class analysis models containing different numbers of classes or typologies
BIC = Bayesian information criterion statistic; AIC = Aikake information criterion statistic
Table 2 Outline of conditional and latent class probabilities for the three-class model
Conditional probabilities
Indicators
Bottom-quartile score on AMK progress test 0.25 0.32 0.52 12.77 0.00
Bottom-quartile score on UCAS entry tariff 0.01 0.97 0.57 6.52 0.04
Bottom-quartile interview score 0.23 0.34 0.33 9.68 0.01
No grade A at A-level chemistry 0.26 0.86 0.99 103.97 0.00
Only one science A-level 0.02 0.00 0.67 11.23 0.00
Covariates
Top-quartile score on Year 2 ISCE 0.36 0.32 0.19
Top-quartile score on Year 4 ISCE 0.30 0.25 0.15
Top-quartile score on PBP 0.27 0.22 0.15
Latent class probability 0.52 0.42 0.06
AMK = applied medical knowledge; UCAS = Universities and Colleges Admission Service; ISCE = integrated structured clinical examination;
PBP = patient-based presentation
therefore lower aggregate A-level grades than their score on an end-of-year progress test over the 5 years
counterparts in the best performing class. They were of the degree course than the best performing
also more likely than the best performing students students and twice as likely (OR = 2.23) as the
to have achieved a bottom-quartile interview score. By intermediate performing students.
stark contrast, the poorest performing students had
the highest probability of gaining a bottom-quartile Furthermore, the inactive covariates included in the
score on an end-of-year progress test, the highest model indicate that those in the poorest performing
probability of not having a grade A at A-level typology were half as likely to achieve a top-quartile
chemistry and the highest probability of having only score on Year 2 and Year 4 tests of clinical and
one science subject at A-level. Indeed, the poorest professional competence, and Year 5 assessments of
performing students were over three times as likely professionalism, than their counterparts in the best
(odds ratio [OR] = 3.28) to achieve a bottom-quartile performing typology (Table 2).
312 Blackwell Publishing Ltd 2011. MEDICAL EDUCATION 2011; 45: 308316
Predicting student performance from entry attributes
Table 3 Logistic regression modelling the probability of gaining a bottom-quartile score on an end-of-year progress test over the 5 years of
the degree course
* Chemistry and biology grades: 1 = A, 2 = B, 3 = other grades, 4 = did not take subject at A-level; bottom-quartile interview score:
1 = Yes, 2 = No; bottom-quartile UCAS tariff score: 1 = Yes, 2 = No
Wald tests: grade at A-level chemistry, v2 = 26.33, d.f. = 1, p < 0.01; grade at A-level biology, v2 = 31.69, d.f. = 1, p < 0.01; bottom-
quartile interview score, v2 = 11.033, d.f. = 1, p < 0.01; bottom-quartile UCAS tariff score, v2 = 3.95, d.f. = 1, p < 0.05
UCAS = Universities and Colleges Admission Service; SE = standard error
Blackwell Publishing Ltd 2011. MEDICAL EDUCATION 2011; 45: 308316 313
P Lambe & D Bristow
true that the probability of gaining a bottom-quartile interview. The LCA and the binary regression analysis
score on an end-of-year progress test increases as show that a structured interview can give valuable
grades on chemistry and biology A-levels decrease. added information about the quality of those selected
Further, having a bottom-quartile interview score with respect to performance on tests of AMK across
appears to have an effect independent of our the medical degree course. This finding concurs with
measure of prior academic achievement upon the those of previous studies.15,2830
likelihood of gaining a bottom-quartile score on an
end-of-year progress test. Additionally, the inclusion in the LCA model of
inactive covariates measuring clinical competence
and professionalism indicates that students who are
DISCUSSION least likely to rank among the top performers in tests
of medical knowledge are also least likely to rank
This study illustrates a method that has the potential among the top performers in assessments of clinical
to inform the selection process and identify the competence and professionalism. This finding
profiles of students who are most likely to need concurs with Wilkinson and Framptons contention
learning support. It adds to understanding of the that a students theoretical knowledge leads to good
importance of A-level science subjects and grades, performance in tests of practical clinical knowledge
and interview scores, in the prediction of subsequent and skills, and that good students tend to perform
performance across the undergraduate medical well across a range of assessments because they have
degree course. ability in all components of clinical practice.31
The LCA shows that the difference in performance Although the study indicates that results on science
among our sample of students on tests of AMK can A-levels may be more important than previously
in part be explained by differences in measures of thought for indicating likely academic progress, this
prior academic achievement and interview score at finding needs to be assessed in the context of the
time of entry. It identified distinct subgroups which general acceptance that considering a wide range
differ in their probability for poor examination of personal qualities and skills may be just as
performance according to scores on attributes mea- important as investigating academic achievement
sured at time of entry to the course. when assessing which candidates are most likely to
become good doctors.1 Consideration of prior
Good A-level grades in more than one science academic achievement is just one element of a fair
subject, particularly in chemistry, a high interview selection procedure. As this study shows, non-cogni-
score and a high UCAS tariff score were shown to be tive attributes measured at interview also have some
positively associated with the likelihood of better predictive validity.
performance. Having only one science A-level, a low
UCAS tariff score and a low interview score were All studies have limitations; the LCA cluster model
shown to significantly increase the likelihood of poor approach involves a trade-off between the loss of
performance. The results of this study support the information entailed by the rescaling of continuous
predictive value of A-level chemistry and biology for data into categorical data and the advantages facili-
performance on tests of AMK and the contention that tated by a more multivariate approach using a non-
good science grades, particularly in chemistry and parametric method, which enables a more nuanced
biology, are associated with success in knowledge- insight into the individual and combined effects of
based examinations on the medical degree course. selection instruments on outcomes of interest.
The findings also support the view that a lower
mean examination grade at A-level is associated In general, the application of latent class models
with greater risk for poor performance. Thus, with requires comparatively large samples. However, mod-
respect to prior academic achievement, our findings erate sample size is unproblematic if the number of
concur with those of many previous studies511 and polytomous variables included in the model does not
are not unexpected. What the study adds to this cause sparseness of data. This occurs when the number
literature is evidence indicating that, among equally of variables or the number of categories of these
qualified students, those with lower interview scores variables are large and the number of cells in the
are more likely to perform less well. resultant multi-way frequency table exceeds the
sample size. In this study, five dichotomous variables
Thus, the results of the study add to our under- were included in the LCA model, producing a 32-fold
standing of the predictive ability of the structured response pattern, a number well below the sample size
314 Blackwell Publishing Ltd 2011. MEDICAL EDUCATION 2011; 45: 308316
Predicting student performance from entry attributes
of n = 142. If the inactive covariates had been included 5 Yates J, Smith J, James D, Ferguson E. Should appli-
as indicators, the resulting response pattern of 256 cants to Nottingham University Medical School study a
would have exceeded the sample size and cast doubt non-science A-level? A cohort study. BMC Med Educ
on the precision of the parameter estimates. Clearly, 2009;9:5. http://www.biomedcentral.com/1472-6920/
9/5. [Accessed 15 April 2010].
the number of categories of the continuous variables
6 James D, Chilvers C. Academic and non-academic
rescaled into categorical variables and the number of
predictors of success on the Nottingham undergradu-
categorical variables included in the latent class model ate medical course 19701995. Med Educ 2001;35:1056
for this study was constrained by sample size.26 64.
7 Montague W, Odds F. Academic selection criteria and
Unfortunately, the latent class cluster model is also subsequent performance. Med Educ 1990;24:1517.
limited in that it is unable to add to our knowledge 8 Yates J, James D. Predicting the strugglers: a case-
about the incremental validity of interview score and control study of students at Nottingham University
can only provide evidence as to whether interview Medical School. BMJ 2006;332:100913.
score contributed significantly towards the ability to 9 Yates J, James D. Risk factors for poor performance on
discriminate between the typologies of student the undergraduate medical course: cohort study at
Nottingham University. Med Educ 2007;41:6573.
examination performance.
10 McManus C, Smithers E, Partridge P, Keeling A,
Fleming P. A-levels and intelligence as predictors of
Additionally, the analysis was carried out on just two medical careers in UK doctors: 20-year prospective
adjacent cohorts at a single institution. The study study. BMJ 2003;327:13942.
group comprised only those who had successfully 11 McManus C, Powis D, Wakeford R, Ferguson E, James
completed the course and only those who were direct D, Richards P. Intellectual aptitude tests and A-levels
school-leavers at time of entry. However, the results for selecting UK school-leaver entrants to medical
reflect the study of a homogeneous group which school. BMJ 2005;331:5559.
represents the vast majority of UK medical school 12 McManus C, Elder A, Champlain A, Dacre J, Mollon J,
entrants. Chis L. Graduates of different schools show substantial
differences in performance on MRCP(UK) Part 1 and
Part 2 and PACES examinations. BMC Med 2008.
http://www.biomedcentral.com/1741-7015/6/5.
Contributors: both authors contributed to the study
[Accessed 23 April 2009.]
conception and design, data analysis and interpretation,
13 Greengross S. What patients want from their doctors.
and the drafting and revision of the paper. Both authors
In: Allen I, Brown P, Hughes P, eds. Choosing Tomorrows
approved the final manuscript for publication.
Doctors. London: Policy Studies Institute 1997;129.
Acknowledgements: none.
14 Morris J. The value and role of the interview in the
Funding: Peninsula Medical School funded this research. student admissions process: a review. Med Teach
Conflicts of interest: none. 1999;21:47381.
Ethical approval: not required. 15 Powis D, Rolfe I. Selection and performance of medical
students at Newcastle, New South Wales. Educ Health
1998;11:1523.
REFERENCES 16 Lemay J, Lockyer J, Collin T, Brownell K. Assessment of
non-cognitive traits through the admissions multiple
1 General Medical Council. Tomorrows Doctors: mini-interview. Med Educ 2007;41:5739.
Outcomes and Standards for Undergraduate Medical 17 Emery J.A report on the predictive validity of the BMAT
Education. 2009. http://www.gmc-uk.org/education/ (2005) for first year examination performance on the
undergraduate/tomorrows_doctors_2009.asp. medicine and veterinary medicine courses at the
[Accessed 25 May 2010.] University of Cambridge. University of Cambridge Lo-
2 Roberts C, Walton M, Rothie I, Crossley J, Lyon K, cal Examinations Syndicate. 2005.
Kumar K, Tillar D. Factors affecting the utility of the 18 McManus C, Woolf K, Dacre J. Even one star at A-level
multiple mini-interview in selecting candidates for could be too little too late for medical student selec-
graduate-entry medical school. Med Educ 2008;42:396 tion. BMC Med Educ 2008;8:16. http://www.biomed
404. central.com/1472-6920-8-16. [Accessed 12 May 2010].
3 Patterson F, Ferguson E. Selection for medical educa- 19 Department of Health. Medical Schools: Delivering the
tion and training. Understanding Medical Education Series. Doctors of the Future. London: DoH 2004.
Edinburgh: Association for the Study of Medical 20 Hagenaars J, McCutcheon A, eds. Applied Latent Class
Education 2007. Analysis. Cambridge: Cambridge University Press 2002.
4 Ferguson E, James D, Madeley L. Factors associated 21 Rabe-Hesketh S, Skrondal A. Classical latent variable
with success in medical school: systematic review of the models for medical research. Stat Methods Med Res
literature. BMJ 2002;331:5559. 2008;17:532.
Blackwell Publishing Ltd 2011. MEDICAL EDUCATION 2011; 45: 308316 315
P Lambe & D Bristow
22 Universities and Colleges Admission Service (UCAS). 28 Powis D, Neame R, Bristow T, Murphy L. The objective
http://www.ucas.ac.uk/students/. [Accessed 12 May structured interview for medical student selection.
2009.] Br Med J (Clin Res Ed) 1988;296:7658.
23 McHarg J, Bradley P, Chamberlain S, Ricketts C, Searle 29 Powis D, Waring T, Bristow T, OConnell D. The
J, McLachlan J. Assessment of progress tests. Med Educ structured interview as a tool for predicting premature
2005;39:2217. withdrawal from medical school. Aust N Z J Med
24 Mattick K, Dennis I, Bradley P, Bligh J. Content speci- 1992;22:6928.
ficity is it the full story? Statistical modelling of 30 Glick S. Selection for entry to medicine and specialist
a clinical skills examination. Med Educ 2008;42:58999. training. Med Teach 2000;22:4437.
25 Vermunt J, Magidson J. Technical Guide for Latent GOLD 31 Wilkinson T, Frampton C. Comprehensive undergrad-
4.0: Basic and Advanced. Belmont, MA: Statistical Inno- uate medical assessments improve prediction of clinical
vations, Inc. 2005. performance. Med Educ 2004;38:11116.
26 Formann A, Kohlmann T. Latent class analysis in
Received 11 June 2010; editorial comments to authors 9 August
medical research. Stat Methods Med Res 1996;5:179211. 2010; accepted for publication 8 October2010
27 Long J, Freese J. Regression Models for Categorical Depen-
dent Variables Using Stata. College Station, TX: Stata
Press 2003.
316 Blackwell Publishing Ltd 2011. MEDICAL EDUCATION 2011; 45: 308316