7.2.1 Bootstrapping Techniques

К оглавлению
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 
17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 
34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 
102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 
119 120 121 122 123 124 125 126 127 128 129 130 131 

By this is meant any approach in which a category system is developed on the

hoof, in the course of categorising the items being dealt with. Holsti’s original

account (Holsti, 1968) is definitive and should be referred to for background

information about content analysis in general social research; Neuendorf’s

work (2002) looks set to replace Holsti in due course. The following core

procedure summarises the basics in as much detail as you require.

Firstly, a decision is taken on what constitutes the content unit, or basic unit of

analysis. What’s the basic unit being categorised: how large or small should

the basic idea be? And less obviously but equally important, the context unit:

where should the basic idea be located: should we scan the phrase, the

sentence, and the paragraph in a search for it? Should we stick to the same

context unit, such as the sentence, throughout, or can we use the sentence at

times, and a whole paragraph at other times (‘This sentence is basically about

X’; ‘These next few sentences are basically padding, but I can see that the

whole paragraph is about Y’)?

You’ll be relieved to hear that when you content-analyse a repertory grid,

these questions have already been resolved for you during construct

elicitation. Each construct is your basic unit of analysis and, in Holsti’s

terms, the construct is both the content unit and the context unit. In other

words, each and every construct is regarded as expressing a single unit of

meaning. Of course!

The Core-Categorisation Procedure

Each item being categorised is compared with the others, and then:

(1) If an item is in some way like the first item, the two are placed together

under a single category created for them at that very moment.

(2) If an item is different to the first item, they’re put into separate

categories.

(3) The remaining items are compared with each of the categories and

allocated to the appropriate one if an appropriate category exists.

(4) A new category is created if required; when a new category is created,

the possibility that existing categories need to be redefined (combined

or broken up, with their items reallocated accordingly) is considered

and effected as necessary.

(5) This process continues until all the items have been classified.

(6) However, a small number are usually unclassifiable without creating

categories of just one item, which would be pointless, so all unclassifiable

items are placed in a single category labelled ‘miscellaneous’.

(7) If more than 5% of the total are classified as miscellaneous, consider

redefining one or more existing categories so that, at the end, no more

than 5% of the total are categorised as ‘miscellaneous’.

Now, with one proviso, this caters for our first need. By recognising similarities

and dissimilarities in the interviewees’ items as they specified them, we’ve

aggregated the meanings in the whole sample while classifying some

individual meanings as similar, and others as different. We’ve accommodated

differences in meaning and intention on the part of individual interviewees.

(We have, however, lost a lot of information, for none of the ratings have been

taken into account. We’ll see that Section 7.3 deals with this difficulty.)

The proviso is obvious. The categories we have devised are simply our own

opinion. But your category system is simply your own way of construing your

interviewees’ constructs: other people might not see the same kinds of

meaning in the constructs, and might disagree! Yet the whole point of this

chapter, you’ll recall, is to make statements which communicate meanings

effectively to other people. Something’s wrong if we can’t agree on the

category to which a particular construct belongs.

To guard against this problem, every content analysis needs to incorporate a

reliability check. This is a procedure which ensures that (though one has every

right to a private interpretation – remember Section 5.3?) the category system

shouldn’t be wildly idiosyncratic, if it is to make sense to other people. There’s

more on this later, in the section on reliability.

Don’t forget that there are several different forms of reliability. Hill (1995: 105^106)

reminds us that in content-analysis terms, three kinds exist. Firstly, there is stability,

the extent to which the results of a content analysis are invariant over time. Are your

category definitions robust enough that, if you were to repeat the core procedure all

over again, you would end up with the same categories, and within the same

constructs under each category? After all, you’re supposed to be recognising similarities

in meaning in a set of constructs. So long as you’re working with that particular

set, themeaning you recognise should be the same on both occasions.

Secondly, there is reproducibility, the extent to which other people make the same

sense of the constructs as you do. If meaning is socially defined, if you are to avoid

layingyourownidiosyncracies onto the data, your content analysisneedsto bereproducible.

Finally, there is sheer accuracy. How consistently are you applying your category

definitions, once you have fixed themas a standard to aimat?

In practice, it is sometimes difficult to distinguish between these three sources of

unreliability; however, you will notice as you use the procedures outlined below that

all three confront you as you make decisions about category definitions, and about

the allocation of constructs to categories. The procedures described have been

devised to reduce unreliability under all three headings.

This is all a little abstract. Let’s pin it all down. Firstly, what are the items we’re

talking about? In fact, the generic procedure can be applied to either elements

or constructs, though usually it’s the latter. Thus, a discussion of the elements

provided by all the interviewees in a group, getting the group members to

categorise them themselves by means of the procedure outlined above, is often

a good way of getting people to examine, debate, and challenge their ideas

about the topic, particularly in a training setting (see Section 9.2.1, step 2 of the

partnering procedure outlined there, for an example of this activity in a

personal change setting).

For general research purposes, though, it’s the constructs which are the items

being categorised, and the remainder of this chapter will deal with construct

content analysis only. First, let’s look at the generic procedure, and then deal

with the matter of reliability.

The Generic Content-Analysis Procedure

(1) Identify the categories.

(2) Allocate the constructs to the categories following the core procedural

steps 1 to 7 above. You’ll notice that this results in a set of categories which are

mutually exclusive, and completely exhaustive: all your constructs are

accounted for. A convenient way of doing this is to transcribe each

construct from all the grids onto its own file card, coding the card to

identify which interviewee provided the individual construct, and which of

his/her constructs it is, in order of appearance in that interviewee’s grid.

(Thus, the code 5.3 would indicate that the construct in question was the third

construct in the fifth interviewee’s grid.)

Now go through steps 1 to 7 above, placing the cards into heaps, each heap

constituting a different category. If you lay all the cards out on a large table,

you can see what you’re doing, and shuffle cards around as you identify

categories, allocate cards to them, change your mind and create new

categories, and so on.

(3) Tabulate the result. In other words, record which constructs have been

allocated to which categories. On a large sheet of paper (flip-chart paper,

especially if it has been ruled as graph paper, is ideal), create a set of rows, one

for each category. Create a column on the left, and in it, label each row with its

category name. Now create a new column and use it to record a short

definition of that category. In a third column, record the code numbers of all

the constructs that you allocated to that category.

(4) Establish the reliability of the category system (ignore this for the

moment; see the discussion below).

(5) Summarise the table; first, the meaning of the category headings. What

kinds of categories are these? What sorts of constructs have we here? Use the

column 2 information to report on the distinct meanings available in the whole

set of constructs.

(6) Summarise the table: next, find examples of each category heading. Are

there constructs under each category which stand for or exemplify that

category particularly well? Are there perhaps several such constructs, each

summarising a different aspect of the category? Highlight the code numbers of

these constructs among the list in column 3. You’ll want to remember these

and quote them in any presentation or report that you make, since they help

other people to understand the definitions you have proposed in step 4.

(7) Summarise the table; finally, find the frequency under the category

headings. In a fourth column, report the number of constructs in each

category. Which categories have more constructs and which have fewer? Is

this significant, given the topic of the grid? For reporting purposes, when you

have to list the categories, consider reordering them according to the number

of constructs allocated to them.

Table 7.1 provides an example, taken from a study I once did of fraud and

security issues in the Benefits Agency (Jankowicz, 1996). For the moment,

ignore the two rightmost columns (the ones headed ‘Prov.’ and ‘Met.’). Notice

how, in step 3, I’ve expressed the definitions of each category as bipolar

constructs (in the second column). Next, the codes which stand for each

construct are listed; I haven’t listed them all in this example since there isn’t

the space on the page! The ‘Sum’ column shows how many constructs were

categorised under each heading, and below that, the percentage of all the

constructs that this figure represents (for example, in the first row, 57

constructs, this being 19.1% of all the constructs, came under the first

category). This table is in fact a summary, and was accompanied by a set of

tables which listed the constructs themselves, as well as their codes. You might

argue that it’s the constructs themselves, and not the code numbers standing

for them, that matter, and you’d be perfectly right. However, a table like

Table 7.1 provides a useful summary, and the third column is, in fact, a

necessary discipline in checking the reliability of the content-analysis process

(see steps (4.1) to (4.7) below), so you might as well take this column seriously!

Design Issues: Differential Analysis

If you have some hunch or hypothesis which you want to check by your

analysis, you need to create additional columns in Table 7.1 at step 7, one for

each subgroup into which you have divided the sample. You will then count

the number of constructs under each category, subgroup by subgroup, thereby

carrying out a ‘differential analysis’. This is very straightforward. It simply

reports whether the constructs from the members of one subgroup are

distributed differently across the categories than the constructs from other

subgroups, working with the percentage figures in each category. (Where the

total number of constructs provided by each subgroup varies, you’ll need to

change all the figures into percentages of each subgroup’s total, so that you

can compare between subgroups.)

For example, do younger interviewees think differently to the older ones (in

terms of the percentage of their constructs they allocated to particular

categories)? Does a sales force think differently about the price discounts

Table 7.1 Content-analysis procedure, Benefits Agency example

Category Definition Construct Sum

%

Prov. Met.

Deliberateness

of action and

intent

Knowing what’s right and

ignoring it; lawbreaking;

deliberate fraud; errors

of commission versus

bypassing procedures; making

technical errors; mistakes and

errors of omission

2.1

17.1

18.2

35.1

etc.

57

19.1

47

21.1

10

13.3

Friendship and

other external

pressures

Divulging information to a 3rd

party; collusion versus acting

alone; no 3rd party involved

46.1

31.1

etc.

39

13.3

29

13.0

10

13.3

Pressure of

work

Shortcuts to gain time or ease

workflow; pressure of targets;

negligence versus reasons other

than workflow; breaches of

confidence; deliberate

wrongdoing

16.1

44.1

13.1

etc.

34

11.4

26

11.7

8

10.7

Internal versus

external

involvement

Staff member is the agent in

doing/condoning the offence

versus claimant the agent in

the offence

1.2

17.2

35.2

etc.

33

11.1

27

12.1

6

8.0

Risk, proof

and

obviousness

Definite and easy to prove; clear

feeling there’s something wrong;

rules clearly broken versus

unsure if fraud occurred; no

rules broken

1.1

4.5

34.5

etc.

32

10.7

25

11.2

7

9.3

Systems and

security

procedures

Using information improperly;

cavalier attitude to checking;

misuse of IT procedures versus

accidental outcomes of IT system;

inadequate clerical procedures

3.1

12.2

39.3

etc.

31

10.4

21

9.4

10

13.3

Who gains Employee reaps the benefit of

fraud; less personal need or

motive versus claimant reaps

the benefit; personal problems

provide a motive

10.5

25.2

47.6

etc.

27

9.1

20

8.9

7

9.3

Money versus

information

Personal cash gains;

clear overpayments versus

provision of information

4.1

11.4

etc

21

7.1

14

6.3

7

9.3

Outcomes Severe consequences or

repercussions versus fewer/

less severe consequences

33.5

44.5

etc.

8

2.7

3

1.3

5

6.7

Training

issues

Not preventable by training

versus preventable by improved

training

7.7

36.3

etc.

4

1.3

2

0.9

2

2.7

Where it

happens

Occurs in the agency office

versus occurs in claimant’s

home or similar

7.4

9.6

etc.

3

1.3

0

0

3

1.3

Miscellaneous 3.6

etc.

9

3.0

6

2.7

3

4.0

Totals 298

100.2

223

99.9

75

99.9

Source: Reproduced from the Analytical Services Division, Department of Social Security.

available to them than the sales office staff who set the discounts but never

meet real clients?

Take a third example. Perhaps, as a manager in a department of a municipal

administration, you suspect that clerical officers who have had private-sector

experience before deciding to work in local government have a systematically

different way of thinking about their jobs than clerical officers who have

always worked in local government. You feel there may be two distinct types

of clerical officer, and if this is the case, the implications (to do with their

attitude to the public as customers; to issues of supervision; and to the way

they exercise initiative) may be wide-ranging.

The following steps can now be completed.

(8) Complete any differential analysis which your investigation requires.

Create separate columns for each group of interviewees you’re interested in,

and record the constructs separately for each of them. Count the constructs

from each group in each category, and see! Does each group of respondents

think systematically differently?

In the Benefits Agency study reported in Table 7.1, the Agency wished to see

whether there were any differences in the construing of employees in busy

metropolitan offices as distinct from quieter provincial offices, and the

sampling was designed to pick this up. Table 7.1 shows the relevant

information in the columns headed ‘Prov.’ and ‘Met.’. Since there were

differing numbers of interviewees in each location, each providing grids

showing differing numbers of constructs, the entries in each category were

turned into percentages of the total number of metropolitan and provincial

constructs. Each entry in the ‘Prov.’ column shows the number of constructs in

that category mentioned by provincial employees, and, below, the corresponding

percentage of all the provincial employees’ constructs. Ditto for the

‘Met.’ column.

As you can see, though there were some differences, these weren’t dramatic.

‘Deliberateness of intent’ was mentioned especially frequently by provincial

employees (21.1% of their constructs in this category). Metropolitan employees

saw this as important (13.3% categorised under the same heading) but were

equally concerned about ‘friendship and other pressures’ (13.3%) and ‘systems

and security procedures’ (13.3%).

(9) Complete any statistical tests on this differential analysis as required. If

you’re familiar with null hypothesis testing, you’ll have noticed that this table

consists of frequency counts under mutually exclusive headings. Any

differential analysis in which you’re involved turns the content analysis into

an X by Y table where X is the number of categories (rows) and Y is the

number of distinct groups (data columns; Y equals 2 in the case of the clerical

officers example above). This is exactly the kind of situation which lends itself

to the chi-square statistic (or Fisher’s Exact Probability Test, depending on the

number of expected values in each cell). Other tests dependent on a

comparison of the order of frequencies in each column may occur to you. If

you aren’t familiar with these statistical tests and this last step doesn’t convey

any meaning to you, just ignore it.

Reliability

You may have noticed that we omitted one crucial step from the procedure. As

you remember from our earlier discussion, content analysis can’t be

idiosyncratic. It needs to be reproducible, and to make sense to other

people. And so, all content analyses should incorporate a reliability check.

This is an additional twist to the procedure, and takes place during the

tabulation stage.

Let’s work with a running example. This is an (invented) study of the factors

which a publishing company’s sales reps believe are important in achieving

sales. Imagine, if you will, that 50 sales reps have each completed a repertory

grid, taking eight recently published books as their elements.

Run through steps 1 to 3 as above, using the core-categorisation procedure

described earlier.

(1) Identify the categories.

(2) Allocate constructs to those categories.

(3) Tabulate the result.

(4) Establish the reliability of the category system.

(4.1) Involve a colleague: ask a colleague to repeat steps 1 to 3

independently, producing a table, like your own, which summarises

his/her efforts. Now, the extent to which these two tables, yours and your

collaborator’s, agree indicates how reliable your procedures have been.

(4.2) Identify the categories you both agree on, and those you disagree

on. You can assess this by drawing up a fresh table, whose rows stand for

the categories you identified, just as you did before; and the columns stand

for the categories your collaborator identified. This is a different table to

either of the content-analysis tables which you and your collaborator have

filled out. Its purpose is to compare the two separate ones. For the sake of

clarity, let’s call it the reliability table. Here’s our worked example, shown

as Table 7.2.

Jot down the definitions of the two category sets; discuss the categories,

and agree on which ones mean the same. Now rearrange the rows and

Table 7.2 Assessing reliability, step (4.2), before rearrangement

Collaborator

Interviewer

1

Sales

price

2

Nature of

purchasers

3

Current

fashion

4

Coverage

5

Trade

announcements

6

Layout and

design

7

Competition

8

Advertising

budget

1 Popularity of

topic

5.8

2 Buyer

characteristics

6.1

3 Pricing decisions

4 Design

5 Contents

6 Competitors

7 Promotion 7.4 4.1

Example of initial content-analysis categories from a study of the factors which a publishing company’s sales reps believe to be related to the

volume of sales they’re able to achieve. This example is developed in Tables 7.3 to 7.6.

The reliability table will be used to record how the interviewer, and the collaborator, have categorised all of the constructs. As an example, four

constructs have been placed into the table. So, for example, the interviewer has put construct 6.1 into the ‘buyer characteristics’ category. The

collaborator seems to disagree about its meaning, having put it under the ‘layout and design’ category.

columns of the reliability table so that categories which you and your

collaborator share are placed in the same order: yours from top to bottom

at the left of the table, and your collaborator’s from left to right at the top

of the table. In other words, tuck the shared categories into the top left

corner of the table, in the same order across and down, with the categories

that you don’t share positioned in no particular order outside this area (see

Table 7.3).

(4.3) Record your joint allocation of constructs. Working from your two

separate content-analysis tables prepared in step 3, record the position of

each of your constructs into the reliability table. How did you categorise

the construct, and how did your collaborator categorise it: which row and

column, respectively, was it put into? Write the construct number into the

appropriate cell of the table. Table 7.3 shows just four constructs which

have been allocated in this way, as an example, while Table 7.4 shows a

full data set of constructs recorded in this way.

As you can see, there are two parts to the reliability issue. Can you agree

on the category definitions; and, when you have agreed, are you both

agreed on the allocation of the constructs to the same categories? The

rearrangement of the reliability table, as shown in Table 7.3, is a useful

exercise, since it forces you both to think about your category definitions.

It also provides you with a measure of the extent to which you can allocate

constructs consistently, as follows.

(4.4) Measure the extent of agreement between you. Think about it. If you

were both in perfect agreement on the allocation of constructs to

categories, all of your constructs would lie along the diagonal of the

reliability table (the shaded cells); to the extent that you disagree,

some constructs lie off the diagonal. So your overall measure of agreement

is as follows:

. the number of constructs which lie along the diagonal

. in all the categories you were both agreed on (the categories which lie

above and to the left of the two heavy lines drawn in Table 7.3)

. as a percentage of all of the constructs in the whole table.

Work out this figure, and call it index A. Now, repeat this calculation but

express the number as a percentage of the constructs, not in the whole

reliability table, but just those which have been allocated to categories you

both agree on. Call this index B.

Table 7.5 provides you with a complete worked example. Index A is 54%:

you’ve only agreed on just over half of what the constructs mean! Index B

is, as it must be, larger, at 64%: when you confine your attention to

categories which mean the same to both of you, you have a better result.

Table 7.3 Assessing reliability, step (4.2), after rearrangement

Collaborator

Interviewer

1

Current

fashion

2

Nature of

purchasers

3

Sales price

4

Layout and

design

5

Coverage

6

Competition

7

Trade

announcements

8

Advertising

budget

1 Popularity of

topic

2 Buyer

characteristics

6.1

3 Pricing decisions

4 Design

5 Contents

6 Competitors

7 Promotion 7.4 4.1

1. Discussion of the definitions showed that the interviewer’s ‘popularity of topic’ category is the same as the collaborator’s ‘current fashion’; the

interviewer’s ‘buyer characteristics’ is the same as the collaborator’s ‘nature of purchasers’ category; ‘pricing decisions’ is the same as ‘sales price’;

‘design’ is the same as ‘layout and design’; ‘contents’ is the same as ‘coverage’; and ‘competitors’ is the same as ‘competition’. The interviewer has

one category not used by the collaborator, ‘promotion’; and the collaborator has two categories not used by the interviewer, ‘trade announcements’

and ‘advertising budget’.

2. The categories have now been reorganised so that the commonly shared ones are at the top left of the table.

The way in which both the interviewer and collaborator have categorised the constructs is now recorded by placing construct codes into their

appropriate cells; just four examples, the same ones which appeared in Table 7.2, are shown above.

3. Construct 5.8, ‘there’s a demand for a textbook like this – no demand for this topic’, was categorised under ‘popularity of topic’ in the interviewer’s

analysis, and as ‘current fashion’ in the collaborator’s analysis, so it’s placed in row 1, column 1, in this table: a construct on which both are agreed.

Construct 6.1, ‘ring-bound covers: bookshop buyers don’t like – conventional cover: bookshop buyers will accept’, was categorised under ‘buyer

characteristics’ by the interviewer but under ‘layout and design’ by the collaborator.

Construct 7.4, ‘advertised heavily in the trade press – not advertised in the trade press’ was placed in the ‘promotion’ category by the interviewer,

but in the ‘trade announcements’ category by the collaborator.

Construct 4.1, ‘big advertising budget – small advertising budget’ was categorised under ‘promotion’ by the interviewer but under ‘advertising

budget’ by the collaborator.

5.8

Table 7.4 Assessing reliability, step (4.3)

Collaborator

Interviewer

1

Current

fashion

2

Nature of

purchasers

3

Sales price

4

Layout and

design

5

Coverage

6

Competition

7

Trade

announcements

8

Advertising

budget

1 Popularity of

topic

1.4 3.2

2 Buyer

characteristics

4.6, 6.1 1.6, 6.4 6.6

3 Pricing decisions 7.1 7.3 6.2 7.2

4 Design

5 Contents 7.8 3.6, 5.2

6 Competitors 7.9

7 Promotion 7.4 1.7, 5.6,

6.7

2.2, 4.1,

3.4, 5.4

All of the constructs in the publisher’s example are shown here, identified by their code number.

2.3, 5.8,

3.5, 5.3

1.1, 2.5,

3.7, 5.7

1.2, 3.1,

4.4, 5.5,

6.3, 7.7

1.5, 4.5,

7.5

4.3, 5.9,

7.6

2.1, 1.3,

2.4, 3.3,

4.2, 5.1, 6.5

Table 7.5 Assessing reliability, step (4.4)

Collaborator

Interviewer

1

Current

fashion

2

Nature of

purchasers

3

Sales price

4

Layout and

design

5

Coverage

6

Competition

7

Trade

announcements

8

Advertising

budget

Total

1 Popularity of

topic

1 1 6

2 Buyer

characteristics

2 2 1 9

3 Pricing

decisions

1 1 1 1 10

4 Design 3

5 Contents 1 2 6

6 Competitors 1 8

7 Promotion 1 3 4 8

Total 5 5 9 6 6 11 3 5 50

1. Index A: number of constructs along the diagonal for the categories agreed on, as a percentage of all the constructs in the table:

4 + 4 + 6 + 3 + 3 + 7 = 27;

50 constructs in total;

100_27/50 = 54%

2. Index B: number of constructs along the diagonal for the categories agreed on, as a percentage of all the constructs in the categories agreed on:

4 + 4 + 6 + 3 + 3 + 7 = 27;

42 constructs in the categories agreed on (5 + 5 + 9 + 6 + 6 + 11, or, of course, 6 + 9 + 10 + 3 + 6 + 8; it’s the same!)

100_27/42 = 64%

4

4

6

3

3

7

But it’s still not good enough; a benchmark to aim at is 90% agreement, with

no categories on whose definition you can’t agree. So:

(4.5) Negotiate over the meaning of the categories. Look at which

categories in particular show disagreements, and try to arrive at a

redefinition of the categories as indicated by the particular constructs on

which you disagreed, so that you improve on the value of Indices A and B.

Argue, debate, quarrel, just so long as you don’t come to blows. Break for

lunch and come back to it if necessary!

For example, in Table 7.5, even without knowing what the constructs are,

you can hazard a guess that the interviewer and collaborator will be able

to agree on a single category, ‘promotion’, since announcements in the

trade press, and advertising, might both be regarded as forms of

promotion. This single redefinition would be sufficient to create a total

set of seven categories which accounted for all the constructs and on

which both were agreed.

Even if nothing else changed, a redrawing of Table 7.5 (see the result in

Table 7.6) shows an improvement to 68% agreement. It is likely that this

discussion will clarify the confusion which led to construct 7.2 being

categorised under ‘pricing decisions’ by the interviewer, raising the index

to 70%. Further discussion, concentrating on the areas of disagreement,

would tighten up the definitions of the other categories. The aim is to get

as many constructs onto the diagonal of the table as possible!

(4.6) Finalise a revised category system with acceptably high reliability.

The only way of knowing whether this negotiation has borne fruit is for

each of you, interviewer and collaborator, to repeat the procedure. Redo

your initial coding tables, working independently. Can you both arrive at

the same, including categorisation of the constructs to the carefully

redefined categories?

Repeat the whole analysis again. That’s right! Repeat step 2 using these

new categories. Repeat steps 3 and 4, including the casting of a new

reliability table, and the recomputation of the reliability index.

This instruction isn’t, in fact, as cruel as it may seem. The categorisation

activity is likely to be much quicker than before, since you will be clearer

on category definitions and you will be using only agreed categories. It is

still time-consuming, but there is no alternative if you care for the

reliability of your analysis.

(4.7) Report the final reliability figure. The improved figure you’re

aiming for is 90% agreement or better, and this is usually achievable.

There are more accurate measures of reliability, including ones which provide a

reliability coefficient ranging between _1.0 and +1.0, which may be an obscure

Table 7.6 Assessing reliability, step (4.5)

Collaborator

Interviewer

1

Current

fashion

2

Nature of

purchasers

3

Sales price

4

Layout and

design

5

Coverage

6

Competition

7

Promotion

Total

1 Popularity of

topic

1 1 6

2 Buyer

characteristics

2 2 1 9

3 Pricing decisions 1 1 1 1 10

4 Design 3

5 Contents 1 2 6

6 Competitors 1 8

7 Promotion 1 8

Total 5 5 9 6 6 11 8 50

1. The collaborator’s category no. 7, ‘promotion’, is the result of combining the previous two categories: 7, ‘trade announcements’ and 8, ‘advertising

budget’.

2. Index A: number of constructs along the diagonal for the categories agreed on, as a percentage of all of the constructs in the table:

4 + 4 + 6 + 3 + 3 + 7 + 7 = 34

50 constructs in total

100_34/50 = 68%

3. (As all the categories are now agreed on, index A is identical to what was earlier called index B.)

4

6

3

3

7

7

4

characteristic to anyone other than a psychologist or a statistician, who is used to

assessing reliability in this particular way. Probably the most commonly used

statistic in this context is Cohen’s Kappa (Cohen,1968).However, if having a standard

errorof the figure youhave computed matters to you, thenthe Perrault ^Leigh Index is

the appropriatemeasure to use: see Perrault & Leigh (1989).

The value of Cohen’s Kappa or the Perrault ^Leigh Index which you would seek to

achieve would be 0.80 or better.This is the standard statistical criterion for a reliable

measure, but, if you’re conscientious about the way in which you negotiate common

meaningsforcategories, ahighly respectable 0.90 istypical for repertorygrid content

analyses.

And that’s that: with the completion of step 4 of our procedure, you’d continue

with the remaining steps, 5 to 9, taking comfort that the categories devised in

steps 5 to 9 were thoroughly reliable.

All of this seems very pedantic, and for day-to-day purposes, most people

would skip the reanalysis of step 4.5. However, if you were doing all this as

part of a formal research programme, especially one leading to a dissertation

of any kind, you’d have to include this step, and report the improvement in

reliability (it is conventional to report both the ‘before’ and the ‘after’ figure, by

the way!).

Well and good; but haven’t you forgotten something? When you present the final

results at steps 5 to 7 (the content-analysistable,withits subgroup columnsfordifferential

analysis as required), whose content-analysis table do you present: yours, or

your collaborator’s? You’ve increased your reliability but, unless you’ve achieved a

perfect100%match, the two tables, yours and your collaborator’s, will differ slightly.

Which should you use? Whose definition of reality shall prevail?

In fact, you should use your own (what we’ve been calling the interviewer’s contentanalysis

table), rather than your collaborator’s.You designed the whole study and it’s

probably fair for any residual inaccuracies to be based on your way of construing the

study, rather than your collaborator’s. (Though if someone were to argue that you

should spin a coin to decide, I could see an argument for it based on Kelly’s alternative

constructivism: that one investigator’s understanding of the interviewees’constructs

is asgood as another’s, once the effort tominimise researcher idiosyncrasy hasbeen

made!)

Okay, this is a long chapter: take a break! And

then, before you continue, please do Exercise 7.1.