Download the Full Version of textbook for Fast Typing at textbookfull.

com

Analysis of Binary Data, Second Edition David

Roxbee Cox

https://textbookfull.com/product/analysis-of-binary-data-
second-edition-david-roxbee-cox/

OR CLICK BUTTON

DOWNLOAD NOW

Download More textbook Instantly Today - Get Yours Now at textbookfull.com

 Recommended digital products (PDF, EPUB, MOBI) that
you can download immediately if you are interested.

Analysis of Survival Data First Edition Cox

https://textbookfull.com/product/analysis-of-survival-data-first-
edition-cox/

textboxfull.com

Visual Six Sigma Making Data Analysis Lean Ian Cox

https://textbookfull.com/product/visual-six-sigma-making-data-
analysis-lean-ian-cox/

textboxfull.com

Biota Grow 2C gather 2C cook Loucas

https://textbookfull.com/product/biota-grow-2c-gather-2c-cook-loucas/

textboxfull.com

Trading Binary Options Second Edition Abe Cofnas

https://textbookfull.com/product/trading-binary-options-second-
edition-abe-cofnas/

textboxfull.com
Complex analysis Second Edition. Edition David Orme Tall

https://textbookfull.com/product/complex-analysis-second-edition-
edition-david-orme-tall/

textboxfull.com

Uncertainty analysis of experimental data with R 1st

Edition Benjamin David Shaw

https://textbookfull.com/product/uncertainty-analysis-of-experimental-
data-with-r-1st-edition-benjamin-david-shaw/

textboxfull.com

ggplot2 Elegant Graphics For Data Analysis Second Edition

Hadley Wickham

https://textbookfull.com/product/ggplot2-elegant-graphics-for-data-
analysis-second-edition-hadley-wickham/

textboxfull.com

Quantitative Risk Analysis of Air Pollution Health Effects

Louis Anthony Cox Jr.

https://textbookfull.com/product/quantitative-risk-analysis-of-air-
pollution-health-effects-louis-anthony-cox-jr/

textboxfull.com

Clinical trial data analysis with R and SAS Second Edition

Chen

https://textbookfull.com/product/clinical-trial-data-analysis-with-r-
and-sas-second-edition-chen/

textboxfull.com
 M O N O G R A P H S ON
STATISTICS A N D A P P L I E D P R O B A B I L I T Y

General Editors
D.R. Cox, D.V. Hinkley, D. Rubin and B.W. Silverman
1 Stochastic Population Models in Ecology and Epidemiology
M.S. Bartlett (1960)
2 Queues D.R. Cox and W.L. Smith (1961)
3 Monte Carlo Methods J.M. Hammersley and D.C. Handscomb (1964)
4 The Statistical Analysis of Series of Events D.R. Cox and
P.A.W. Lewis (1966)
5 Population Genetics W.J. Ewens (1969)
6 Probability, Statistics and Time M.S. Bartlett (1975)
7 Statistical Inference S.D. Silvey (1975)
8 The Analysis of Contingency Tables B.S. Everitt (1977)
9 Multivariate Analysis in Behavioural Research A.E. Maxwell (1977)
10 Stochastic Abundance Models S. Engen (1978)
11 Some Basic Theory for Statistical Inference E.J.G. Pitman (1978)
12 Point Processes D.R. Cox and V. Isham (1980)
13 Identification of Outliers D.M. Hawkins (1980)
14 Optimal Design S.D. Silvey (1980)
15 Finite Mixture Distributions B.S. Everitt and D.J. Hand (1981)
16 Classification A.D. Gordon (1981)
17 Distribution-free Statistical Methods 7.5. Maritz (1981)
18 Residuals and Influence in Regression R.D. Cook and S. Weisberg (1982)
19 Applications of Queueing Theory G.F. Newell (1982)
20 Risk Theory, 3rd edition R.E. Beard, T. Pentikainen and
E. Pesonen (1984)
21 Analysis of Survival Data D.R. Cox and D. Oakes (1984)
22 An Introduction to Latent Variable Models B.S. Everitt (1984)
23 Bandit Problems D.A. Berry and B. Fristedt (1985)
24 Stochastic Modelling and Control M.H.A. Davis and R. Vinter (1985)
25 The Statistical Analysis of Compositional Data 7. Aitchison (1986)
26 Density Estimation for Statistical and Data Analysis
B.W. Silverman (1986)
27 Regression Analysis with Applications G.B. Wetherill (1986)
28 Sequential Methods in Statistics, 3rd edition G.B. Wetherill (1986)
29 Tensor Methods in Statistics P. McCullagh (1987)
 30 Transformation and Weighting in Regression R J. Carroll and
D. Ruppert (1988)
31 Asymptotic Techniques for use in Statistics O.E. Barndoff-Nielson
and D.R. Cox (1989)
32 Analysis of Binary Data, 2nd edition D.R. Cox and E.J. Snell (1989)
33 Analysis of Infectious Disease Data N.G. Becker (1989)
34 Design and Analysis of Cross-Over Trials B. Jones and
M.G. Kenward (1989)
35 Empirical Bayes Method, 2nd edition J.S. Maritz and T. tw in (1989)
36 Symmetric Multivariate and Related Distributions K.-T. Fang,
S. Kotz and K. Ng (1989)
37 Generalized Linear Models, 2nd edition P. McCullagh and
J.A. Nelder (1989)
38 Cyclic Designs y.v4. John (1987)
39 Analog Estimation Methods in Econometrics C.F. Manski (1988)
40 Subset Selection in Regression A.J. Miller (1990)
41 Analysis of Repeated Measures M. Crowder and D.J. Hand (1990)
42 Statistical Reasoning with Imprecise Probabilities P. Walley (1990)
43 Generalized Additive Models T.J. Hastie and R.J. Tibshirani (1990)
44 Inspection Errors for Attributes in Quality Control
N.L. Johnson, S. Kotz and X Ww (1991)
45 The Analysis of Contingency Tables, 2nd edition B.S. Everitt (1992)
(Full details concerning this series are available from the Publishers.)
 Analysis of
Binary Data
SECOND EDITION

D.R. COX
Nuffield College
Oxford

and

E.J. SNELL
Department of Mathematics
Imperial College London

CHAPMAN & HALL/CRC

Boca Raton London New York Washington, D.C.
Published in 1989 by
Chapman & Hall/CRC
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 1970 by D.R. Cox, 1989 by D.R. Cox and EJ. Snell

Chapman & Hall/CRC is an imprint of Taylor & Francis Group
No claim to original U.S. Government works
Printed in the United States of America on acid-free paper
10 9 8 7 6 5 4
International Standard Book Number-10: 0-412-30620-4 (Hardcover)
International Standard Book Number-13; 978-0-412-30620-4 (Hardcover)
This book contains information obtained from authentic and highly regarded sources. Reprinted material is
quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts
have been made to publish reliable data and information, but the author and the publisher cannot assume
responsibility for the validity of all materials or for the consequences of their use.
No part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic,
mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and
recording, or in any information storage or retrieval system, without written permission from the publishers.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only
for identification and explanation without intent to infringe.

Library of Congress Cataloging-in-Publication Data

Cox, D.R. (David Roxbee)

Analysis of binary data/D.R.Cox and E.J, Snell - 2nd ed.
p. cm. - (Monographs on statistics and applied probability)
Includes bibliographical references and index.
ISBN 0-412-30620-4
1. Analysis of varience. 2. Probabilities. 3. Distribution. (Probability theory) I. Snell, E.J. II.Title.
QA279.C68 1989
519.5’352—dcl9

T&F informa Visit thè Taylor & Francis Web site at

http://www.taylorandfrancis.com
Taylor & Francis Group and thè CRC Press Web site at
is the Academic D ivision o f T&F Informa pic. http://www.crcpress.com
 Contents

Preface to first edition IX

Preface to second edition xi

1 Binary response variables 1

1.1 Introduction 1
1.2 Examples 2
1.3 Dependency relations for probabilities 13
1.4 Some statistical considerations 18
1.5 Some numerical comparisons 20
1.6 Some generalizations of the logistic model 23
Bibliographic notes 24

2 Special logistic analyses 26

2.1 Introduction 26
2.2 Simple regression 33
2.3 2 x 2 contingency table 43
2.4 Matched pairs 52
2.5 Several 2 x 2 contingency tables 56
2.6 Multiple regression 66
2.7 Residuals and diagnostics 69
2.8 Factorial arrangements 84
2.9 Cross-over designs 92
2.10 Rasch model 94
2.11 Binary time series 96
Bibliographic notes 102

3 Some complications 106

3.1 Introduction 106
3.2 Anomalous dispersion 106
vi CONTENTS

3.3 Empirical Bayes methods 115

3.4 Errors of measurement 119
Bibliographic notes 124

4 Some related approaches 126

4.1 Introduction 126
4.2 Prospective studies 127
4.3 Retrospective studies 128
4.4 Relation between discriminant analysis
and logistic regression 132
4.5 Some different theoretical formulations 140
Bibliographic notes 148

5 More complex responses 149

5.1 Introduction 149
5.2 Paired preferences 149
5.3 Nominal data 153
5.4 Ordinal data 158
5.5 Multivariate binary data 161
5.6 Mixed binary and continuous responses 163
Bibliographic notes 166

Appendix 1 Theoretical background 167

A 1.1 Introduction 167
A 1.2 Ordinary least squares theory 167
A 1.3 Extensions of least squares 170
A 1.4 Likelihood functions 171
A 1.5 Techniques for inference 179
A 1.6 Some warnings 183
A 1.7 Some further developments 183
A 1.8 Conditioning 184
A 1.9 Bayesian approach 186

Appendix 2 Choice of explanatory variables in multiple

regression 188
A2.1 Introduction 188
A2.2 Types of explanatory variable 189
A2.3 Formation of explanatory variables 190
A2.4 Small numbers of explanatory variables 192
A2.5 Large numbers of explanatory variables 193
CONTENTS vii

Appendix 3 Review of computational aspects 195

A3.1 Introduction 195
A3.2 Specific programs 195
A3.3 Expert systems 196

Appendix 4 Further results and exercises 198

References 215

Author index 228

Subject index 231

 Preface to first edition

This monograph concerns the analysis of binary (or quantal) data, i.e.
data in which an observation takes one of two possible forms, e.g.
success or failure. The central problem is to study how the probability
of success depends on explanatory variables and groupings of the
material.
Many particular methods, especially significance tests, have been
proposed for such problems and one of the main themes of the
monograph is that these methods are unified by considering models
in which the logistic transform of the probability of success is a linear
combination of unknown parameters. These linear logistic models
play here much the same role as do normal-theory linear models in
the analysis of continuously distributed data.
Some knowledge of the theory of statistics is assumed. I have
written primarily for statisticians, but I hope also that scientists and
technologists interested in applying statistical methods will, by
concentrating on the examples, find something useful here.
I am very grateful to Dr Agnes M. Herzberg and Dr P.A.W. Lewis
for extremely helpful comments. I acknowledge also the help of Mrs
Jane Gentleman who programmed some of the calculations.
D.R. Cox
London
April 1969
 Preface to second edition

We have added new material partly to amplify matters dealt with only
very cryptically in the first edition and partly to describe some of the
more recent developments, for example on regression diagnostics. In
addition the contents of the first edition have been rearranged; for
example, the method of least squares with empirically estimated
weights is now of much less importance than it used to be because
computational developments have, for many problems, brought
maximum likelihood fitting within the painless grasp of most users of
statistical analysis. By giving some prominence to examples we have
aimed to make the book accessible to a range of readers. One of the
Appendices summarizes the theoretical background.
When the first edition was written it was feasible to give a relatively
complete annotated bibliography of work on the analysis of binary
data. The number of papers on this topic is now so large and is
increasing so rapidly that no such bibliography has been attempted.
Instead, in the Bibliographic Notes at the end of the each chapter we
have aimed to give just a few key references for further reading and for
details omitted from the main text.
We are grateful to Professor N. Wermuth, Mainz, for thoughtful
comments on a portion of the manuscript and to Professor J.K.
Lindsey, Liège, for advice over ordinal data and to Professor P.
McCullagh.
The first author is grateful to Science and Engineering Research
Council for a Senior Research Fellowship held at Department of
Mathematics, Imperial College, London.
D. R. Cox
E. J. Snell
London
July 1988
 CH APTER 1

Binary response variables

1.1 Introduction
Suppose that on each individual we have an observation that takes
one of two possible forms. The following are examples:
1. an electronic component may be defective, or may be
nondefective;
2. a test animal may die from a specified dose of a poison, or may
survive;
3. a subject may give the correct reply in an experimental situation,
or may give a wrong reply;
4. a test specimen may fracture when struck with a standardized
blow, or may not;
and so on. If for the ith individual we can represent this observation,
or response, by a random variable, Y;, we may without loss of
generality code the two possible values of 7; by 1 and 0 and write
E{Yi) = prob(y; = 1) = 9i, prob(7i = 0) = 1 - 0,, ( 1. 1)
say. It is often convenient to call 7^ = 1 a ‘success’ and 7^ = 0 a ‘failure’.
It is reasonable to call such observations binary; an older term is
quantal.
We assume that such binary observations are available on n
individuals, usually assumed to be independent. The problem is to
develop good methods of analysis for assessing any dependence of 0,
on explanatory variables representing, for example, groupings of the
individuals or quantitative explanatory variables.
We have followed the usual terminology and have distinguished
between (a) response variables and (b) explanatory variables, the
variables of the second type being used to explain or predict variation
in variables of the first type.
Sometimes a binary response variable arises by condensing a more
complex response. Thus a component may be classed as defective
2 BI NARY R E S P O N S E V AR I ABL E S

when a quantitative test observation falls outside specification limits

or, more generally, when a set of test observations falls in an
unacceptable region. When this is done we need to consider whether
there is likely to be serious loss in treating the problem in terms of a
binary response.
In addition to a binary response variable there may be further
response variables for each individual. Thus in a psychological
experiment, as well as the rightness or wrongness of a reply, the time
taken to make the reply may be available. Joint analysis of the
response variables is then likely to be informative.

1.2 Examples
It is convenient to begin with a few simple specific examples
illustrating the problems to be considered.

Example 1.1 The 2 x 2 contingency table

Suppose that there are two groups of individuals, 0 and 1, of sizes «q
and « 1 and that on each individual a binary response is obtained. The
groups may, for example, correspond to two treatments. Suppose
further that we provisionally base the analysis on the assumption that
all individuals respond independently with probability of success
depending only on the group, and equal, say, to (/>o and (/>i in the two
groups. Note that we use 0 for a probability referring to a group,
reserving 9 for the probability of success for an individual.
In this situation we need consider only the random numbers of
successes R q and in the two groups; in fact {Rq^R^) form sufficient
statistics for the unknown parameters ((j)o,(t>i)- It is conventional to
set out the numbers of successes and failures in a 2 x 2 contingency
table as in Table 1.1. Nearly always it is helpful to calculate the
proportions of successes in the two groups.
In Table 1.1 the columns refer to a dichotomy of the response
variable, i.e. to a random variable in the mathematical model; the
rows refer to a factor classifying the individuals into two groups, the
classification being considered as non-random for the purpose of the
analysis. It is possible to have contingency tables in which both rows
and columns correspond to random variables.
As a specific example, the data of Table 1.2 refer to a retrospective
survey of physicians (Cornfield, 1956). Data were obtained on a group
of lung cancer patients and a comparable control group. The numbers
1.2 EXAMPLES

Table 1.1 A 2 x 2 contingency table

Failures Successes Total Propn successes

Group 0 Ro ^o/^o
Group 1 Ri/n,

Total n Q n ^ — R q — Ri R q -{-Ri no + Wj

Table 1.2 Numbers of smokers in two groups of physicians

Smokers Non-smokers Total Propn non-smokers

Controls 32 11 43 0.256
Lung cancer
patients 60 3 63 0.048

Total 92 14 106

of individuals in the two groups are approximately equal and are in

no way representative of population frequencies. Hence it is reason­
able to make an analysis conditionally on the observed total numbers
in the two groups.
An essential initial step is to calculate the sample proportions of
successes in the two groups Ro/riQ and Riln^, and these are shown in
Table 1.2. Further analysis of the table is concerned with the precision
of these proportions.
When two groups are to be compared using binary observations, it
will often be sensible to make an initial analysis from a 2 x 2
contingency table. However, the assumptions required to justify
condensation of the data into such a form are not to be taken lightly.
Thus in Section 2.4 we shall deal with the methods to be followed
when pairs of individuals in the two groups are correlated. The most
frequent inadequacy of an analysis by a single 2 x 2 contingency table
is, however, the presence of further factors influencing the response,
i.e. nonconstancy of the probability of success within groups. To
ignore such further factors can be very misleading.
For many purposes it is not necessary to give the numbers of
failures in addition to the numbers of successes and the numbers of
trials per group. In most of the more complex examples we shall
therefore omit the first column of the 2 x 2 table.
4 BI NARY R E S P O N S E V ARI ABL E S

There are many extensions of the 2 x 2 table; there follow some

examples.

Example 1.2 Several 2 x 2 contingency tables

Suppose that to compare two treatments we have several sets of
observations, each of the form of Example 1.1. The different sets may
correspond to levels of a further factor or, as in the following specific
example, may correspond to different blocks of an experimental
design. In general, in the sth set of data, let «o,s be the sample
sizes in the two groups and let /?o,s and be the total numbers of
successes.
Table 1.3 (Gordon and Foss, 1966) illustrates this. On each of 18
days babies not crying at a specified time in a hospital ward served as
subjects. On each day one baby chosen at random formed the
experimental group and the remainder were controls. The binary
response was whether the baby was crying or not at the end of a
specified period. In Table 1.3, not crying is taken as a ‘success’ and
the observed numbers rQ^ and ^ are therefore the numbers of
babies in the two groups not crying; the common convention is

Table 1.3 The crying of babies

No. of
No of control No. not experimental No. not
Day, s babies, Mq.s crying, Tq., babies, ^ crying,

1 8 3 1 1
2 6 2 1 1
3 5 1 1 1
4 6 1 1 0
5 5 4 1 1
6 9 4 1 1
7 8 5 1 1
8 8 4 1 1
9 5 3 1 1
10 9 8 1 0
11 6 5 1 1
12 9 8 1 1
13 8 5 1 1
14 5 4 1 1
15 6 4 1 1
16 8 7 1 1
17 6 4 1 0
18 8 5 1 1
1.2 EXAMPLES 5

followed of denoting observed values of random variables by

lower-case letters. The special feature of this example is that ^ = 1,
so that 5 takes values 0 and 1; usually there are several individuals

in each group.
The object of the analysis is to assess the effect of the treatment on
the probability of success. The tentative basis for the analysis is that
there is in some sense a constant treatment effect throughout the
experiment, even though there may be some systematic variation
from day to day. The experiment has the form of a randomized block
design, in fact a matched pair design, but the binary nature of the
response and the varying numbers of individuals in the groups
complicate the analysis.
For the reasons indicated after Example 1.1 it would not in general
be a sound method of analysis to pool the data over days, thus
forming a single 2 x 2 contingency table with entries ZR q.s»
One simple, if approximate, method of analysis that is not distorted
by systematic differences between groups is to calculate for the sth
group the difference in the proportions of successes, i.e.

0 ,s
( 1.2)
«0,/
This is an unbiased estimate for the sth set of the difference between
the probabilities of success. When (1.2) is averaged over all the sets, an
unbiased estimate of the mean difference between groups results. A
difficulty of this analysis is that the quantities (1.2) have in general
different precisions for different s.

Later, further examples will be given whose analysis requires the

combination of data from several 2 x 2 contingency tables.

Example 1.3 A 2 x 2^ system

Suppose that there are p two-level factors thought to affect the
probability of success. Let a binary response be observed on each
individual and suppose that there is at least one individual corre­
sponding to each of the 2^ cells, i.e. possible factor combinations;
usually there are an appreciable number of individuals in each cell.
It would be possible to think of such data as arranged in a (p + 1)-
dimensional table, with two levels in each dimension. Alternatively
and more usefully, we can think of a 2 x 2^ table in which the two
columns correspond to success and failure and the 2^ rows are the 2^
standard treatments of the factorial system. One of the problems of
 BI N AR Y R E S P O N S E V AR I A BL E S

Table 1.4 2 x 2 ^ system. Study of cancer knowledge

No. No. Propn No. No. Propn

successes trials successes successes trials successes

1 84 477 0.176 d 2 12 0.167

a 75 231 0.325 ad 7 13 0.538
b 13 63 0.206 bd 4 7 0.571
ab 35 94 0.372 abd 8 12 0.667
c 67 150 0.447 cd 3 11 0.273
ac 201 378 0.532 acd 27 45 0.600
be 16 32 0.500 bed 1 4 0.250
abc 102 169 0.604 abed 23 31 0.742

analysis is to examine what order of interactions, in some suitable

sense, is needed to represent the data; that is, we think first of 2^
probabilities of success, one for each cell, and then try to represent
these usefully in terms of a smaller number of parameters.
Table 1.4 is a specific example of a 2 x 2"^ system based on an
observational study by Lombard and Doering (1947) (see Dyke and
Patterson (1952) for a detailed discussion of the analysis of these
data). In this study, the response concerned individuals’ knowledge
of cancer, as measured in a test, a ‘good’ score being a success and
a ‘bad’ score a failure. There were four factors expected to account
for variation in the probability of success, the individuals being
classified into 2"^ cells depending on presence or absence of exposure
to A, newspapers; B, radio; C, solid reading; D, lectures. In Table 1.4
the standard notation for factor combinations in factorial experi­
ments is used; thus ac denotes the cell in which A and C are at their
upper levels and B and D at their lower levels.
Some general conclusions can be drawn from inspection of the cell
proportions of successes. This is an example where both the response
variable and the factors are reduced to two levels from a more
complex form.

Example 1.4 Serial order

Suppose that a series of independent binary responses is observed and
that it is suspected that the probability of success changes systemati­
cally with serial order. One application is to some types of data in
experimental psychology, where a subject makes a series of responses,
each either correct or incorrect, and where the probability of correct
response is suspected of changing systematically. In this context, the
1.2 EXAMPLES 7

treating of successive responses as independent may, however, be

seriously misleading. Another application is in human genetics where
each child in a family is classified as having or not having a particular
genetic defect. Here, except for possible complications from multiple
births, each family leads to a sequence of binary responses. In this
application, data from many families are required. We then have a
number of sequences, usually not all of the same length, and it is
required to examine the data for systematic changes, with serial order,
in the probability of a genetic defect. It is not usually reasonable to
suppose that in the absence of such effects the probability of a defect is
the same for all families.

Example 1.5 Stimulus binary response relation

The following situation is of wide occurrence. There is a stimulus
under the experimenter’s control; each individual is assigned a level of
the stimulus and a binary response then observed. One important
field of application is bioassay, where, for example, different levels of
stimulus may represent different doses of a poison, and the binary
response is death or survival. Similar situations arise in many other
fields.
In such applications it is often possible to choose a measure x of
stimulus level such that the probability of success is zero for large
negative x, unity for large positive x and is a strictly increasing
function of x. In fact it has the mathematical properties of a
continuous cumulative distribution function; see Fig. 1.1(a). If the x
scale is suitably chosen, the distribution function will be symmetric;
for example, in the particular application mentioned above it is often
helpful to take x as log dose.
Table 1.5 gives some illustrative data; at each of a number of
dose levels a group of individuals is tested and the number dying
recorded. In Fig. 1.1(b) the proportions dying are plotted against log

Table 1.5 Simple form ofhioassay

Concn Log2 concn No. of deaths No. of indivs Propn deaths

Co 0 2 30 0.067
2co 1 8 30 0.267
4co 2 '15 30 0.500
8t'o 3 23 30 0.767
16co 4 27 30 0.900
 BI N AR Y R E S P O N S E V AR I A BL E S

(b) Log^ (concenrrahon/CQ)

Figure 1.1 Stimulus-response curves, (a) Idealized theoretical curve, (h)

empirical curve from Table 1.5.

concentration and, except for random fluctuations, give a curve

similar to the idealized theoretical one of Fig. 1.1(a). The object of
analysing such data is to summarize and assess the properties of
the stimulus response curve. Sometimes, as in many bioassay
problems, the aspect of primary importance is the level of x at which
1.2 EXAMPLES

the probability of success is or sometimes the level Xp at which

some other specified value, p, of the probability of success is achieved
In other applications, for example in experimental psychology, the
steepness of the response curve is the aspect of primary interest. We
shall refer tc x variously as a stimulus, or as an explanatory variable.

Example 1.6 Grouped data with binary response

Suppose that there are /c = p -h 1 groups of individuals, each indivi­
dual having a binary response. If individuals’ responses are inde­
pendent, and the probability of success is constant within each group,
we can, just as in the discussion of the 2 x 2 table, condense the data,
giving simply the number of trials n^ and the number of successes
for the 5th group (5 = 0 ,..., g). In contingency table form we then have
Table 1.6.
Now suppose that the groups are meaningfully ordered and that it
is reasonable to expect that any change in the probability of success is
monotonie with group order. If, further, scores Xq, . . . , x^ can be
allocated to the groups such that a smooth relation between the
probability of success and the value of x is reasonable, the situation is

Table 1.6 A contingency table for grouped data

Propn
Failures Successes Total successes

Group 0 «0 -7 Ro Ro no Rq/^O
Group g ^9 R jn ,

Table 1.7 Nasal carrier rate and tonsil size

Propn of
Non-carriers Carriers Total carriers x

Tonsils present.
not enlarged 497 19 516 0.0368 - 1
Tonsils enlarged
+ 560 29 589 0.0492 0
Tonsils enlarged
269 24 293 0.0819 1
10 BI N AR Y R E S P O N S E V AR I A BL E S

then formally the same as that of Example 1.5. A distinction is that in

the previous example the main interest is in the location and shape of
a response curve, whereas in the present example there is usually
substantial interest in the null hypothesis that the probability of
success is constant.
Table 1.7 gives a specific example quoted by Armitage (1955). The
data refer to children aged 0-15 and the binary response concerns
whether a child is or is not a carrier for Streptococcus pyogenes; the
children are grouped into three sets depending on tonsil size. No
objective x variable was available and the tentative scores —1,0, 1
were therefore assigned to the three groups. A difficulty in the
interpretation of such data is that other possibly relevant explanatory
variables - age, sex, etc. - should be considered.

Example 1.7 Multiple regression with binary response

The previous three examples have involved a dependence between a
probability of success and a single regressor or explanatory variable,
x; they are thus comparable with normal-theory regression problems
with a single regressor variable. Suppose now that there are for the ith
individual, p, explanatory variables Xji,...,Xjp, regarded as non-
random, and a binary response. It is necessary to assess the relation
between the probability of success and the variables x,i,...,x^p.
This type of problem occurs especially in medical contexts. The
binary response may represent success or failure of a particular
treatment; death or survival over a specified time following treatment;
or death from a particular cause, as contrasted with death from some
other cause. The variables x,i,...,X jp represents quantitative, or
qualitative, properties of an individual thought to influence the
response; possible x variables are age, time since diagnosis, initial
severity of symptoms, sex (scored as a zero-one variable), aspects of
medical history of the individual, etc.
One particular example is a study of the factors affecting the
probability of having coronary heart disease (see Walker and Duncan
(1967) who give a number of earlier references). Another example
concerns perinatal mortality (Feldstein, 1966).

Example 1.8 Regression on two explanatory variables

Table 1.8 summarizes a two-factor 5 x 4 industrial investigation in
which the number, r, of ingots not ready for rolling out of m tested is
1.2 EXAMPLES 11

Table 1.8 Number, r, of ingots not ready rolling out of m tested

(First figure in each cell is r, second m)
Heating time, Xi
Soaking
time, X2 7 14 27 51 Total

1.0 0,10 0,31 1,56 3,13 4,110

1.7 0,17 0,43 4,44 0,1 4,105
2.2 0,7 2,33 0,21 0,1 2,62
2.8 0,12 0,31 1,22 0,0 1,65
4.0 0,9 0,19 1,16 0,1 1,45

Total 0,55 2,157 7,159 3,16 12,387

shown for combinations of heating time, and soaking time, X2 . This

is a rather simpler situation than most of those sketched in
Example 1.7 because there are only two explanatory variables. Note
also that the data are grouped into sets all with the same values of the
explanatory variables.

In some situations, especially with observational data, we need

to analyse changes in the probability of success, there being a
considerable number of potential explanatory variables. Two broad
approaches are possible, illustrated by Examples 1.3 and 1.7.
In the first approach the values of the explanatory variables are
coarsely grouped, in the extreme case each explanatory variable
taking only two values, as in Example 1.3. With p explanatory
variables there will thus be at least 2^ cells, for each of which the
proportion of successes can be found. An advantage of this approach
is that quite complicated ‘interactions’ can be detected. Possible
disadvantages stem from the necessity of coarse grouping and from
the fact that if p is at all large many of the cells will either be empty or
contain very few observations.
In the second approach, a regression-like model is taken, express­
ing a smooth and simple dependence of the probability of success on
the values of the explanatory variables. No grouping is necessary, but
a disadvantage is that relatively complicated interactions may be
difficult to detect. In practice both methods are useful, separately and
in combination. With both, problems concerning alternative choices
12 B I NAR Y R E S P O N S E VAR I A BL E S

of explanatory variables, familiar from normal-theory regression

methods, arise in essentially the same form.

Example 1.9 A binary time series

In the previous examples it is a reasonable provisional assumption to
suppose that the responses of different individuals are independent. In
the analysis of a binary time series we are directly concerned with the
lack of independence of different observations.
A specific example concerns daily rainfall. It is often reasonable first
to classify days as wet (success) or dry (failure); there results a sequence
of Is and Os, a binary time series. The amounts of rainfall on wet days
can be analysed separately.
More generally, if we consider a response 1 as the occurrence of an
event and a response 0 as non-occurrence, a binary time series is a
series of events in discrete time. If, further, the proportion of Is is
low, the series approximates to a series of point events in continuous
time; in particular, a completely random binary series, in which all
responses have independently the same probability, 6, of giving a
success, tends, as 0->O, to a Poisson process. Cox and Lewis (1966)
have summarized statistical techniques for the analysis of point
events in continuous time.

In this section a number of relatively simple problems have been

described which can be generalized in various ways:
1. we may have situations of more complex structure. Thus, instead
of the single response curve of Example 1.5, we might have several
response curves and be interested in comparing their shapes;
2. we may have multivariate binary responses and consider prob­
lems analogous to those of multivariate normal theory;
3. we may have responses taking not just two possible values, but
some small number greater than two. Can the techniques to be
developed for analysing binary responses be extended?
Some of the more complex problems will be considered later. Others
are indicated as further results or exercises.
As in other fields of statistical analysis, problems of two broad types
arise. We require techniques for efficient analysis and assessment of
uncertainty in the context of an assumed probabilistic model. Also we
need techniques for tabular and graphical display and condensation
of data, sometimes with the objective of finding a suitable model for
1.3 D E P E N D E N C Y R E L A T I O N S F O R P R O B A B I L I T I E S 13

more detailed analysis. This aspect is relatively more important with

extensive data. This book deals mainly but by no means entirely with
the first type of technique.

1.3 Dependency relations for probabilities

The primary theme of this book is the study of the dependence of
0 = prob(T = 1), the probability of ‘success’ of a binary response
variable, on a vector x of explanatory variables, including
groupings of the data. Therefore a key issue is how such dependencies
are to be specified mathematically.
There are broadly three approaches. The first is to use fairly simple
empirical functions that express 6 in terms of x and some unknown
parameters in a reasonably flexible way, such that the parameters
have a clear interpretation and preferably such that the resulting
statistical analysis is straightforward. The second approach is to
relate problems for binary responses to those for some underlying
latent (i.e. not directly observable) continuous response variable.
Unless the latent variable is in principle observable via some other
type of investigation this approach is best regarded as a device for
discovering suitable functions for use in the first approach. The third
route is to look for smooth relations, usually not specified parametri­
cally, uncovered essentially by merging information from individuals
whose explanatory variables are close together in some sense. We
shall not discuss this approach in detail.
At first sight the simplest empirical relation is to suppose that the
value of 6 for the ith individual, is linearly related to the explanatory
variables, i.e. for the ith individual

Oi (X X lP (X ^ X , 5 ^ 5 , (1.3)

where x, is a row vector of explanatory variables for the iih individual,

is a column vector of unknown regression coefficients and a is an
unknown intercept. Note that in general theoretical discussions it is
sensible to omit the separate a and to incorporate it into the vector P
by making one component explanatory variable a vector of Is. We
shall not do this in this initial discussion.
Equation (1.3) is exactly the standard sort of linear model used for
quantitative response variables; this makes it tempting to treat the
binary responses as if they were quantitative, i.e. to score them 0 or
14 BI N AR Y R E S P O N S E V AR I A B L E S

1, and to apply ordinary regression and analysis of variance methods;

we comment further on this below.
The most serious restriction on the usefulness of (1.3) arises from
the condition
O ^ 0 ,^ l. (1.4)
This may cause some difficulties in fitting, especially if ordinary least
squares is used, but there is the more major matter that the
parameters in the model (1.3) inevitably have a limited interpretation
and range of validity. Consider for instance the analysis of the
stimulus-response curve of Example 1.5 and suppose that we take a
model (1.3) in which the probability of success is a linear function of
the variable x. Even if the data were satisfactorily linear over the
observed range of stimuli, it is certain that the linear relation will fail
outside a restricted range of stimuli; further it is usually rather
unlikely that the underlying relationship has an ‘inclined plane’ form,
i.e. a range with zero probability of success, a linear portion of positive
slope and a range with unit probability of success; the general shape
illustrated in Fig. 1.1(a) is much more common. The use of a linear
model is likely to mean that even small-scale extrapolation is
hazardous and that a different experimenter, using a different range of
stimuli, would be likely to find an apparently different stimulus-
response relationship. Similar remarks apply to the other examples of
Section 1.2.
Of course, all postulated models have at best limited and approxi­
mate validity. The point here is that the use of a model, the nature
of whose limitations can be foreseen, is not wise, except for very
restricted purposes.
We now turn to models in which the constraint (1.4) is automati­
cally satisfied.
We shall use the notion of a distribution of a latent response
variable to motivate some alternatives. Suppose then that for a given
vector X of explanatory variables the latent variable, U, has a
continuous cumulative distribution function F(m; x) and that the
binary response T= 1 is recorded if and only if U > 0. That is
6 = prob(y = l;x) = 1 —F(0;x). (1.5)
Note that since U is not directly observed there is no loss of generality
in taking the critical point to be zero and that also without loss of
generality we may take the standard deviation of U or some other
1.3 D E P E N D E N C Y R E L A T I O N S F O R P R O B A B I L I T I E S 15

measure of dispersion, if constant, to be unity. If U is normal with

mean a + xjS it follows that
0 = O(a + xiS), ( 1.6)
where <I)(.) is the standard normal integral

1/2
m = (2n) -
exp{ —^v^)dv. (1.7)
i'
J-c
In particular if x consists of a single (scalar) component, then
0 = 0)(a -f xp) = 0{(x - /i)/(j}, (1.8)
say. Equations (1.6) and (1.8) define the probit model which has a very
extensive literature, especially in connection with bioassay (Finney,
1952). Note that the relation is linearized by the inverse normal
transformation
(D-i(0) = a-hx,jS,
i.e. the model can be written in a form analogous to (1.3), namely
O"^(0^) = a + xjS,
where x^ and P are in general vectors.
We have in this derivation regarded the critical level of U as fixed
and the distribution of U as changing with x. It is clear that one could
have used a complementary formulation in which the distribution of
U is fixed and the critical level varies with x. This second version is
rather more natural in bioassay when dose, or log dose, is the
explanatory variable. Here V can be taken as the dose that would just
produce a response, sometimes called the tolerance, and 7 = 1 if and
only if the dose actually employed exceeds that tolerance. Note that in
this formulation, if the ‘dose’ is a H- xjS, then
p ro b (7 = l;x) = p ro b (F ^ a -h xj?), (1.9)
relating the probability that Y = 1 directly to the distribution function
of V. In particular if x is scalar and jS > 0, then the curve of
p ro b (y = l;x) versus x is quite directly related to the distribution
function of V.
We shall, however, for the most part use the first formulation
because U thereby is more directly related to the observed binary
response. An econometric illustration (Amemiya, 1985, p. 269) con­
cerns the choice between two methods of travelling to work. Here U
16 BI N AR Y R E S P O N S E V AR I A BL E S

might be the difference between the utilities of the two methods,

assumed to have normal-theory linear regression on explanatory
variables. The sign of U determines the preferred method.
The normal form is only one possibility for the distribution of JJ.
Another is the logistic distribution with location a -h xjS and unit
scale. This has cumulative distribution function
exp (m—a —xjS)/{ 1 -h exp (w —a —xjS)}, ( 1. 10)
so that
F{0;x) = 1/(1 -h exp(a + x^)},
from which it follows that
6 = prob(y = l;x) = exp(a + x/?)/(l H-exp(a + xjS)},
( 1. 11)
1 —0 = p ro b(7= O ;x)= 1/{1 + exp(a-f xjS)}.
Note that this relation is linearized by the transformation
log (0/(1 -6 )} = a + x/?. ( 1. 12)
For scalar x and j?> 0 (1.11) defines via (1.9) a probability density
function on differentiation with respect to x, namely
,0L+XP
(1 +

In a reparametrized form this is

exp [(x - /2 )/t]
(1.13)
t{1 + exp[(x-/z)/T ]}

This is called the logistic probability density function; it is a

symmetrical distribution of mean and standard deviation ut/^ 3 ,
having a longer tail than a matching normal distribution. Statistical
analysis of a random sample of continuous observations from (1.13) is
not particularly simple, because (1.13) is not within the exponential
family of distributions. From a manipulative point of view, the
distribution has, however, the property that both density and
cumulative distribution functions are expressed explicitly in terms of
elementary functions. This is sometimes useful, for example in dealing
with censored data.
In particular, one might have a mixture of binary data, observing
success or failure at a set of preassigned x values, with a second set of
data in which a sample of tolerances is measured directly. A combined
Exploring the Variety of Random
Documents with Different Content
The Project Gutenberg eBook of Kaarilan tytöt
This ebook is for the use of anyone anywhere in the United States
and most other parts of the world at no cost and with almost no
restrictions whatsoever. You may copy it, give it away or re-use it
under the terms of the Project Gutenberg License included with this
ebook or online at www.gutenberg.org. If you are not located in the
United States, you will have to check the laws of the country where
you are located before using this eBook.

Title: Kaarilan tytöt

Kertomus nuorisolle

Author: Annikki Virvatuli

Release date: January 3, 2024 [eBook #72604]

Language: Finnish

Original publication: Hämeenlinna: Arvi A. Karisto Oy, 1920

Credits: Anna Siren and Tapio Riikonen

*** START OF THE PROJECT GUTENBERG EBOOK KAARILAN

TYTÖT ***
KAARILAN TYTÖT

Kertomus nuorisolle

Kirj.

ANNIKKI VIRVATULI

Hämeenlinnassa, Arvi A. Karisto Oy, 1920.

SISÄLLYS:

I. Pitkin Päijännettä ja Jämsänjokea.

II. Kotona.
III. Äidin-äidin kirstu.
IV. Pyryn ystävätär.
V. Aurinkokylpyä ja lehmiä.
VI. Ilta Satukannaksella.
VII. Pollen syntymäpäivä.
VIII. Ruusunpunainen puku.
IX. "Synninlukko."
X. Kreets käy maailmalla.
XI. Tehdas.
XII. Untuvasta, kotoalähdöstä ja Pyryn palkasta.
I.

Pitkin Päijännettä ja Jämsänjokea.

"Huu — uu — iih!" Juna pysähtyi Vesijärvelle. Toisiaan työnnellen

kiiruhti matkustajain joukko tavaroittensa kimppuun. Oli toukokuun
viimeisiä päiviä. Päijänteen kauniit seudut vetivät taaskin puoleensa
kesävieraitten tulvan sekä kotiin palaavien opettajain ja koululasten
joukot.

"Puh — huh, shuh — huh", puhkuivat satamassa laivat. Ne

liikahtelivat levottomasti ja syöksyttivät vahvoja savuja, aivan kuin
kiirehtiäkseen matkustajia. Ja kiirettä matkustajat pitivätkin: he
hätyyttivät asemamiehiä ja hoputtivat laivamiehiä, jotka ketterästi
pujahtelivat joukossa ja kärräsivät tavaroita laivaan.

Pitkä, hintelä ylioppilas keikaroi tyhjäntoimittajana joukossa ja

kiikaroi neitosia. Hän seurasi mielenkiinnoin nuorta tyttöä, joka
tunkeusi vaunusta vaikeasti laahaten suuria kantamuksia. "Sininen
kävelypuku ja valkoinen hattu, hoikka ja hauska!"

Ylioppilas harppasi avuksi.

 "Suvaitsetteko, neiti?" Hän kumarsi ja otti tytön raskaat tavarat.

"Olen hyvin kiitollinen! — 'Heinola'-laivalle!" Ja tyttö kääntyi

takanaan kurkkivan vanhan mummon puoleen. "Menkää tuon herran
perässä. Hän vie tavaranne perille. Tuolla noin on 'Heinola'. Hyvää
matkaa!" Hän puristi iloisesti vanhuksen kättä. Ja mummo lyllersi
aika vauhtia auttajansa perässä.

Ylioppilas palasi pian takaisin ja tapasi neitosensa toimittamassa

tavaralippuaan "Lahti"-laivan miehelle. Sitten hän näki hänen ottavan
parkuvan pienokaisen eräältä vaimolta, että tämä sai etsiä
tavaroitaan.

"Paljon kiitoksia!" huusi hän ylioppilaalle, ja silmissä välkkyi

veitikka.

Ylioppilas oli hieman äkeissään. Mutta tytön hymy oli niin

houkutteleva, että hän kysyi, voisiko auttaa. "Kyllä, ehkä tahdotte
pidellä tätä lasta!"

Nuorukainen kääntyi poispäin, ja tyttö nauroi.

Satamassa keinuivat kaikki Päijänteen kyntäjät: heinolalainen

paikassaan, Jyväskylän-puoleiset peräkkäin, komeasti kohoileva
"Suomi", äsken maalattu "Jämsä", jolla on taipumus tutkia
Päijänteen pohjia, korkea, keikkuva "Toivo", pieni, solakka "Aallotar",
ja "Lahti" kuin hymyilevä vanha tuttu. Jo irroittautuivat laivat
laituristaan ja alkoivat hiljalleen solua Vesijärven valkoista vettä
kanavaa kohti.

"Lahti"-laivan yläkannella nojautui kaidetta vasten äskeinen

vallaton tyttö ja ahmi aukenevaa selkää. Viima pörrötti ruskeita
kiharoita hatunreunan alla ja hulmutti sinistä hametta. Kostea tuuli
tuli pitkin järveä ja saattoi tytön silmät loistamaan. Tätä tuulta ja vettä
hän oli kaivannut pitkän talven kuivassa, järvettömässä paikassaan.

Viimein hän istuutui ja kaivoi käsilaukustaan kirjeen.

"Kotiopettajatar Riitta Kaarila" oli kuorella. Riitta Kaarila luki ja
hymyili. Se oli äidiltä, hellältä, vilkkaalta äidiltä, joka iloissaan Riitta-
tyttönsä kotiintulosta unohtui kirjoittamaan liikoja. "Isä on maalannut
venettä, samanlaiseksi kuin ennenkin. Mutta tämä on yllätys sinulle,
jotta älä ole tietävinäsi!" Niin äidin tapaista. Ja isän tapaista. Riitta oli
talvella kirjoittanut, että maalauttaisivat veneen; se oli jo viime
kesänä maalin tarpeessa; mutta isä oli vastannut, että "mitäs
joutavia". Oli aikonut yllättää, mutta äidin oli mahdoton pitää
salaisuutta. Riitan silmissä kimalteli onnellinen nauru hänen
pistäessään kirjettä laukkuunsa. Häntä vastapäätä istuva vanha
herra nauroi myöskin. Teki hyvää katsella iloisia kasvoja.

Hitaasti keinui "Lahti" kanavaan. Sukelsi sulusta toiseen, vuoroin

laskien, jotta rantalaituri jäi korkealle yli yläkannen, vuoroin kohoten,
niin että yläkannelta ei ulottunut kurkoittamaankaan laiturille.
Iltapäivän aurinko sai rannan uhkeat koivut loistoon,
hiirenkorvalehdet heloittivat vaaleanvihreinä ja lähettivät toukokuisen
tuoksun yli kanavan.

Riitta heitti makeisen laiturilla seisovalle repaleiselle pojalle, joka

jäi aukomaan sen käärepaperia, iloissaan unohtaen kiittää. Laiva
soljui hiljalleen Päijänteeseen, ottaen tasaisen vauhdin. — Päijänne!
Hämeen ylpeys ja Sisä-Suomen komea kulkutie! Toisinaan avautuu
selkä taivaanrantaan hukkuvaksi mereksi, jonka äärillä ui sininen
auer ja yllä välkähtelee kalalokin valkoinen rinta, tai kun laiva rantoja
sivuaa, rikkoo se milloin salmen kirkkaan pinnan, jossa koivut
kuvaistaan katselevat, milloin solahtaa ohi kalliosaaren ja sille
pystytettyjen valkeiden merimerkkien.

Riitta Kaarila seisoi yhäti kannella juoden Päijänteen uljasta

kauneutta sielunsa täydeltä. Hän oli yksin ja uneksi avoimin silmin.
Yhtä yksin seisoskeli vanha herrakin ja muuan keskenkasvuinen
koulutyttö, joka arkana oli vetäytynyt syrjään sinne tänne liikkuvasta
meluavasta joukosta.

Joku veti esiin eväänsä, ja heti tunsivat kaikki nälkää. Ryhmityttiin

pöytien ympärille ja koetettiin laivan täristessä pysyttää pystyssä
munakuppia tai maitolasia.

Riittakin oli etsinyt eväskorinsa ja alkoi tyhjentää sen sisällystä,

kun hän samalla huomasi syrjässä seisovan koulutytön.

"Ettekö tahtoisi olla niin ystävällinen ja syödä kanssani!" pyysi hän.

"Minun nimeni on Riitta Kaarila."

Tyttönen oli Hilja Berg, ja hän istui mielihyvin uuden tuttavansa

viereen. Kuinka hyvältä munat ja voileivät maistuivatkaan!

Hilja-tyttönen tuli Hämeenlinnasta, eikä hän arvannut matkaa näin

pitkäksi. Ei hän ollut osannut ottaa evästä mukaan. Siitäpä tulikin
nälkä ja kauhean ikäväkin. Hän oli ensi kertaa matkalla. Oltiin kesää
nyt täällä Päijänteen rannoilla. Kotiväki oli jo tullut. Hän vain oli
jäänyt koulun loppuun asti kaupunkiin ja niin joutunut yksin
tulemaan. Hyvänen aika, kuinka hänestä tuntui äkkiä matka
hauskalta ja mukavalta! Riitta Kaarilaa hän ihaili. Hän vasta osasi
kysellä juuri sellaista, mitä ihan ilokseen kertoi.
 "Nyt me tilaamme teetä", sanoi Riitta, kun he olivat lopettaneet, "ei
missään tee maistu niin hyvältä kuin laivalla kuivain voileipäin
lisäksi". Ja kun tee sitten tuli, oli se oikein ihanaa, niinkuin Hilja
vakuutti.

Riittipä heillä rupattelua. Ujous oli kadonnut, ja vilkas tyttönen

juoksi vuoroin puolelta toiselle; milloin oli sorsaparia, milloin
kummallista pilvenlonkaa näytettävä toverille.

Mutta Riitta huomasi pikku ystävänsä käyvän viluiseksi ohuessa

puvussa.
"Sinun pitää hakea päällystakkisi", sanoi hän.

"Ei minulla ole", tunnusti toinen, "äiti käski lähettää tavarat

edeltäpäin, ja minä pakkasin päällystakkinikin, kun oli niin lämmin
ilma".

"Sepä oli pahasti", tuumi Riitta, "matkoilla aina tarvitsee

päällysvaatetta lämpimilläkin ilmoilla". Hän mietti. "Jopa keksin!" Ja
hän juoksi hakemaan sadevaippansa. Se oli pehmeä ja vahva, ja
Hiljalla oli siinä varsin lämmin, alkoi vain raukaista niin kovin. Yö
olikin jo puolessa kohta. Useat matkustajat olivat vetäytyneet
makuusijoilleen.

"Mihin sinulla on piletti?" kysyi Riitta.

"Tänne yläkannelle."

"Niinpä saat tulla naisten salonkiin nukkumaan."

"Kuuluuko se pilettiin?"

"Kuuluu. Tulehan."
 Naisten salongissa virui leveä rouva yhdellä sohvalla. Muut paikat
hän oli sälyttänyt täyteen matkatavaroitaan, siten varustaen itselleen
turvaavan linnoituksen muita matkustajia vastaan. Hilja vetäysi
arkana pois.

"Mitä vielä!" Riitta työnsi hänet reippaasti sisään. "Hyvä rouva!

Luvallanne autan teitä tavaroittenne järjestelyssä!"

Tuokiossa olivat tavarat sohvan alla rouvan mulkoilusta

huolimatta, ja
Hilja-tyttönen pehmeään vaippaan käärittynä sohvalla.

"En minä peitettä, ota sinä, ettet palele", vastusti hän.

"Nukkuessa pitää aina olla peite, muuten tulee kylmä! Hyvää yötä!
Kyllä minä ajoissa herätän!"

Tyttönen sopersi jotakin puoli-unessa. Kun on koko päivän ollut

ihmistungoksessa, nälissään ja viluissaan, ja sitten saa ruokaa ja
lämmintä ja oikaista itsensä pitkäkseen, niin ei ole ihme, jos nukkuu
heti, varsinkin kun on lapsi vielä.

"Hyvää yötä teille ja tavaroillenne!" Riitta niijasi rouvalle

ilkamoiden. "Se oli ilkeästi kyllä", tuumi hän, "mutta ilkeä oli se
rouvakin".

Kansi oli tyhjä. Vain vanha herra oli siellä kääriytyneenä vahvaan
turkkiin. Riittaakin alkoi puistattaa yön kylmyys.

"Voin tarjota teille hyttini, jos suvaitsette." Vanha herra seisoi

hänen vieressään.
 "Kiitos", sanoi Riitta, "en voi nukkua Päijänteellä. Tahdon katsella
sitä."

"Sen jo arvasinkin", murahti vanha herra. "Käärikää sitten tuo

huopa ympärillenne."

Riitta otti tottelevaisesti huovan hänen vierestään, kääriytyen

siihen.

Äkkiä häntä alkoi naurattaa.

"Mikä on?" kysyi vanha herra.

"Niin vain, kun me olemme kuin kaksi karhua näin köytettyinä!"

Riitta nauroi heleästi. Vanha herra nauroi myös. Ja sitten he

alkoivat jutella. Ja vanha herra jutteli niin hauskasti, että Riitta unohti
Päijänteen.

Laiva vihelsi, ja Riitta meni herättämään pikku ystäväänsä. Isä ja

äiti olivat Hiljaa vastassa. Ja Hiljan isä juoksi laivaan kädestä
kiittämään Riittaa ja pyytämään häntä kesällä heille tulemaan. Ja
kaikki he heiluttivat nenäliinojaan niin kauan kun laiva oli näkyvissä.

"Ihmiset ovat sentään hyviä", sanoi Riitta.

Vanha herra hymyili. "Niin, tiedättehän: kaikki, mitä te tahdotte,

että…"

"Minulle se oli lapsena kompastus, se lause."

"Kuinka niin?"
 "No, kun minä en osannut sanoa profeetat, vaan aina tuli
porohetat!" nauroi Riitta.

Vanha herra hymähti. Hän oli merkillinen, hänelle puhui kuin

isälleen.

"Ei olekaan sanoista väliä, kunhan elämässään toimittaa sitä

enemmän."

"Mutta kun ei tule toimitetuksi mitään suurta."

"Pientä sitten. Esimerkiksi auttelee vanhoja mummoja ja pikku

lapsia, opettaa nuoria herroja, holhoo koulutyttöjä ja nuhtelee
itsekkäitä rouvia!" — Vanhan herran suupielissä piili veitikka.

"Oh", sanoi Riitta hämillään.

"Niinpä niin, nuori ystäväni, joka pienessä on uskollinen, sille

uskotaan suurtakin. Mutta ehkä nyt autatte minuakin. Tässä tulee
minun rantani!"

Riitta auttoi häntä laiturille. Rantatiellä surisi auto.

"Jos joskus haluatte tavata, olette tervetullut", sanoi vanha herra

hyvästiksi ja ojensi nimikorttinsa. Riitta niijasi hämmentyneenä.
Hänelle hän oli uskaltanut puhua tyhmyyksiä! Mitä hän mahtoikaan
tuumia, siitäkin, kun hän vertasi häntä karhuun! Kuin vastaukseksi
vanha herra heilutti autosta kättään.

"Hauskaa kesää, karhutoveri! Kiitos ilosta!"

Riemuiten liehutti Riitta nenäliinaansa. Ei hän ollutkaan puhunut

tyhmyyksiä.
 Tehin aukeella selällä muutettiin laivaa. Pian alettiin kulkea
Jämsänjokea. Rehevät rannat puistoisine taloineen, kaikki uhkeata
ja kukkeata. Tuossa on tumma vuori, Virmapyhä, se kohoaa ihan
eteen. Eipäs! — joki polvistuu, väistää vuoren, ja jo näkyy kirkko
kellotorneineen. Sitten laiva sätkyttää Jämsän komean kirkonkylän,
Seppolan, satamaan. Joku viluinen odottelija harmaalla rantasillalla.
Koko kylä nukkuu toukokuisen aamuyön valoisassa hämärässä.
Rannan puissa vain linnut livertelevät. Nukkuvat vehmaat,
heleänvihreät ruispellot vanhan kirkon ympärillä, nukkuvat
rakennuksetkin: punakattoinen yhteiskoulu niittyjen keskellä, sen
takana kansanopisto, lännempänä kauppatalot ja matala
kansallispirtti pilareineen, nukkuvat poikki joen kaartuvat korkeat
sillat, kivinen ja komea Valtasilta ja puinen, runollinen kirkkosilta —
palanen isänmaata, sirpale Suomea.

"Hyvää huomenta, isänmaa!" Riitta huiskuttaa rannoille. Tuntuu

taas niin suloiselta saapua kotipaikkaan. Laiva sivuuttaa Paunu-
puiston ikivanhat koivut; — entisaikain kalmisto, nyt
kansanjuhlainpito-puisto. Nukkuvat siellä vanhat jäämit ikuista
untaan. Mitä miettinevät haudoissaan, täytyneekö heidän kääntää
kylkeään ja murahtaa: "Sanoja, sanoja, tekisivät työtä!" Oliko Riitta
tehnyt työtä isänmaalle?

Vihellys. Tuolla ne tupruttavat, tutut pitkät savutorvet, kuin

tulivuoret. Kohoavat tehtaiden tiiliseinät, Jämsänkosken
teollisuuskeskus, Sydän-Suomen Tampere. Kuuluu rytinä ja ryske.
Työ käy yötä päivää. Pitkin rantaa mustia proomuja ja pikku höyryjä;
niitä täytetään tehtaitten tuotteilla maailman markkinoille vietäviksi.

Sirpale Suomea siinäkin, palanen sen kohoavaa teollisuutta.

 Kosken varrella olevia tehtaita katselee suojaavasti hymyten
Linnasten vuori, jonka rinteillä ammoisina aikoina on taisteltu,
katselee lastensa taistelua nytkin, taistelua töin ja toimin, pyrkimystä
sivistyskansain rinnalle.
II.

Kotona.

Eläköön! Siellä ne ovat rannalla, isä, Polle, Ter, Kreets ja kultainen

Untuva. Äiti on jäänyt kotiin. Tietysti kahvin tähden.

"Ei saa mennä vielä!" huutaa laivapoika.

Mutta Riitta on jo lautoja odottamatta hypännyt laiturille.

"No, opettajatar, kauniisti!" toruu isä.

"Minä en ole mikään opettajatar täällä, olen vain Pyry!" — Se oli

Riitan kotinimi.

"Sinä olet niin kovin hyvä, kun olet vain Pyry", sanoi Untuva ja
taputti ison siskon käsivartta. Hän olisi kovin mielellään kiivennyt
syliin, mutta ei ilennyt laivapoikain nähden, kahdeksan vanha
ihminen.

Isä ja Polle ottivat kuljettaakseen matkakoria ja hakkasivat sillä

vuoronperään jokaista. Kreets möykytti Pyryä selkään eväskorilla,
kun ei muuten sopinut taputtaa. Pyry itse huimi käsilaukullaan
jokaista ja jutteli sekaisin matkastaan. Ainoastaan Ter käveli
moitteettomasti, luultavasti siitä syystä, ettei hän ollut saanut mitään
kannettavakseen.

"Voi minun pikku tyttöäni, kuinka sinä raukka olet laihtunut!" —

Äidin mielestä lapset aina kotiin tullessaan olivat laihtuneita, vaikka
he olisivat olleet ympyriäisiä kuin kerät.

Sitten äiti taputteli Pyryä ja auttoi häntä riisuuntumaan siten, että

rypisti hatun naulaan ja tunki takin hattuhyllylle. Ja Pyry kysyi, että
ripustetaanko hänet naulaan vai nostetaanko hattuhyllylle.

Ja kaikki nauroivat.

Ja sitten he joivat kahvia auringon noustessa. Ja Untuva kiipesi

Pyryn syliin ja sai kuumaa kahvia päälleen, mutta ei huomannut
itkeä.
Missukka kehräsi ja pyyhkieli Pyryn jalkoja. Mutta kesken juttelun
torkkui Pyry ja nukkui auttamattomasti vaatteet yllä sohvalle.
III.

Äidin-äidin kirstu.

Satoi. Tuli ihan tulvimalla. Untuva oli kastellut itsensä läpimäräksi

kattokourun alla ja istui nyt vilttiin käärittynä keinutuolissa, sillä aikaa
kun Ter etsi hänelle kuivia vaatteita. Kreets oli sohvannurkassa jalat
alla ja luki "Välskärin kertomuksia". Pyry parsi sukkia. Se oli kaikkien
mieliksi hänen lempityötänsä. Polle liimaili kasveja.

"Kuulkaapas, tytöt!" sanoi äiti, "nyt, kun sataa, saatte mennä

vinnille leikkelemään matonkuteita. Poutapäivinä siitä ei kuitenkaan
tule mitään."

Kreets ei ollut kuulevinaan. Hän oli juuri Ester Larssonin kanssa

kiikkumassa ikkunaluukussa. Eivätkä muutkaan näyttäneet
halukkailta. Mutta kun Pyry, "näyttääkseen hyvää esimerkkiä", kääri
sukkansa kokoon ja valmistautui lähtemään, täytyi toistenkin
seurata.

"Yksi jää tänne auttamaan päivällisen laitossa", julisti äiti.

 Jokaiselle tuli tavaton into leikkaamaan matonkuteita. Oli nyt niin,
että he tahtoivat aina olla yhdessä. Siitäpä riitti kahnausta, kun ei
kukaan olisi yksin ryhtynyt mihinkään. Jos yksi määrättiin auttamaan
Vanhaa-Maria pyykillä, olivat kaikki mukana, jotta siitä tuli oikea
sekamelska pesutuvassa. Jos yhden piti mankeliin, menivät he
kaikki, ja Vanha-Mari väitti tulevansa kuuroksi. Puhumattakaan
jostakin huvimatkasta tai iltamasta, jollaisissa he olivat harvoin. Sillä
kun vain yksi kerrallaan saattoi saada pääsylipun, tunsi tämä itsensä
niin onnettomaksi kuvitellessaan kotiin jääneitten ikävää, että
mieluummin hänkin pysyi kotona. Kävelyllä he olivat kaikin, ja heitä
oli tien täydeltä. Ikävintä oli pyöräily. Pyöriä oli vain kaksi, joten
pyöräiltiin vain asioissa.

"Eihän ole saksiakaan kaikille", yritti äiti puolustautua.

"Onhan meillä kolmet, yksi saa ommella, ja minä haen Hietalasta

keritsimet!"

Kreets oli nuolena ulkona. Mutta Ter tunsi tunnonvaivoja. Vaikka

vasta viidentoista, oli hän järkevin joukossa.

"Ehkä me kaikki ensin laitamme päivällistä ja menemme sitten

vasta", ehdotti hän.

Se olisi ollut vihoviimeistä. Niin he saivat kaikki kavuta vinnille.

Siellä he pian istuivat piirissä lattialla suuri kudevasu keskellä ja

toinen vieressä keriä varten. Sade ropisi hauskasti kattoon, ja seinän
toiselta puolen kuului pääskynpoikasten piipitys pesästä. Keritsimistä
vain oli hetkisen kiistelyä.

"Untuva saa ne", julisti Pyry.

Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

textbookfull.com