IRE
Information Retrieval Experiment
The pragmatics of information retrieval experimentation
chapter
Jean M. Tague
Butterworth & Company
Karen Sparck Jones
All rights reserved. No part of this publication may be reproduced
or transmitted in any form or by any means, including photocopying
and recording, without the written permission of the copyright holder,
application for which should be addressed to the Publishers. Such
written permission must also be obtained before any part of this
publication is stored in a retrieval system of any nature.
Decision 9: How to analyse the data? 93
estimators[OCRerr]the sample mean and the sample proportion when the sample
[OCRerr]iie is large. It can be shown that, in both cases, these estimators are normally
(listributed and that the means are equal to the corresponding population
v[OCRerr]iIues. The standard deviations, which are the standard errors, can be
tpproximated by the following formulas:
standard error of the sample mean: s/\/;;, where S is the sample standard
deviation and n is the sample size;
standard error of the sample proportion: [OCRerr]p)/n, where p is the
sample proportion and n is the sample size.
The standard error, by itself, does not mean much. It is more usefully
employed in setting up confidence intervals. A confidence interval is an
interval or range on either side of the sample estimator which contains the
population value with a given confidence. Confidence is usually expressed as
`I percentage between 0 and 100. A 95 per cent confidence interval, for
example, is an interval determined in such a manner that 95 times out of 100
it will contain the population value. With large samples, a 95 per cent
confidence interval for the population mean or population proportion will be
[OCRerr]pproximately 2 standard errors (more exactly 1.96) on either side of the
sample value. A 99 per cent confidence interval will be approximately 3
(more exactly 2.57) standard errors on either side of the estimator. For
example, suppose, in a survey of users attitudes to an online retrieval system,
it was found that 96 out of the 120 users surveyed were satisfied with the
service they received. The standard error is thus given by
96/120(1-96/ 120)[OCRerr]1/2
120 j
- 0.036
a 99 per cent confidence interval for the satisfied proportion of the
population would be
96
120i2.57(O.036) (0.707, 0.893)
`I'hese methods assume, also, a large population. Other methods must be used
to set up confidence intervals for small populations.
Tague and Farradane15 have shown that if one estimates system recall and
system precision by searching random samples of n queries against m
documents and calculates estimated recall, say, by using microaveraging to
obtain a sample estimator p of recall, the estimator will have a standard error
of approximately
( p(1 p)[OCRerr]1/2
mnyl
where p is the average system or population recall, y is the average system
generality, m is the size of the database, and n is the number of queries in the
sample. This may be approximated by
{ Zi[OCRerr]=i a[OCRerr][OCRerr][OCRerr]=1 C[OCRerr][OCRerr]112