employee choice modelling: predicting employee behaviour under varied employment conditions

26
AJHR paper final draft Mar 01 1 Word count 8600 Employee choice modelling: predicting employee behaviour under varied employment conditions Abstract The vast bulk of contemporary survey research is unable to give an indication of what employees will do, as opposed to what they might feel, in the event of any subsequent change to the work environment, for either the “better” or the “worse”. From a practical point of view, this is a major impediment to developing a business case for any organisational intervention that might be considered. Recent collaboration between researchers involved in the twin fields of marketing and HR has produced a hybrid methodology, “employee choice modelling”, that allows employees’ decisions about employment issues to be simulated, modelled and quantified. This paper describes employee choice modelling and its possible use by HR policy makers.

Upload: unsw

Post on 18-Jan-2023

0 views

Category:

Documents


0 download

TRANSCRIPT

AJHR paper final draft Mar 01 1 Word count 8600

Employee choice modelling:

predicting employee behaviour under varied employment conditions

Abstract

The vast bulk of contemporary survey research is unable to give an indication of what employees will do, as

opposed to what they might feel, in the event of any subsequent change to the work environment, for either the

“better” or the “worse”. From a practical point of view, this is a major impediment to developing a business

case for any organisational intervention that might be considered. Recent collaboration between researchers

involved in the twin fields of marketing and HR has produced a hybrid methodology, “employee choice

modelling”, that allows employees’ decisions about employment issues to be simulated, modelled and

quantified. This paper describes employee choice modelling and its possible use by HR policy makers.

AJHR paper final draft Mar 01 2 Word count 8600

Introduction

It seems almost de rigeur today that any self-respecting organisation should have in place an employee

satisfaction/attitude measurement system. Yet most applied survey research cannot indicate what employees

will do, as opposed to what they might feel, in the event of any subsequent “better” or “worse” changes to

work environments. From a practical point of view, this is a major impediment to developing a business case

for organisational interventions that might be considered. However, the related discipline of market research

has developed alternative methods of measurement and analysis that can help overcome this limitation. Recent

collaboration between researchers involved in marketing and HR has produced a hybrid methodology called

“employee choice modelling” that allows employees’ decisions about employment issues to be simulated and

modelled. The method can help HR managers understand employee needs and priorities and thus enhance the

business value of survey research programs and of the HR function generally.

Purpose

The purpose of this paper is to describe employee choice modelling and its use by HR policy makers. The

paper addresses the following main topics:

a comparison of the approach with traditional methods of employee research;

the development of the employee choice modelling approach from an adaptation of consumer choice

modelling;

the validity of the approach; and

the types of problems to which the approach can be applied.

Consumer decision making and its workplace equivalent

Employee satisfaction/attitude measurement

Employee survey research often is undertaken to help managers target or monitor organisational change

initiatives. For example, an organisation may use employee survey research to identify whether staff are

satisfied with the alignment between performance and pay, whether employee empowerment indicators are at

AJHR paper final draft Mar 01 3 Word count 8600

“best practice” for their industry or whether employees feel adequately consulted on major changes in their

workplaces.

Contemporary employee survey research relies on responses to questions/statements of the Likert scale type,

eg., the "strongly agree-strongly disagree" mode (see, for example, Schmitt & Klimoski 1991). Thus it

continues to have the limitation that Green and Tull (1978, 200) noted some time ago in a critique of attitude-

based market research, that “predictions from attitude scales, preference ratings, and the like still need to be

translated into measures (sales, market share) of more direct interest to the marketer”. In other words,

traditional research methods neither give much indication of what employees will do in the event of any

“better” or “worse” changes, nor allow prediction of employee reactions to markedly differently employment

or working conditions.

Consumer satisfaction/attitude measurement

These research limitations contrast with equivalent approaches in marketing research, such as those used to

model consumer choice behaviour. Although the analysis of consumer behaviour is arguably at least as

challenging as the analysis of employee behaviour, market researchers have developed advanced theory and

methods in their discipline that are scarcely ever used by their employee research counterparts. In contrast to

their equivalents in the employee research area, market researchers can choose from a wider array of research

approaches, including powerful methods that can address all of the above limitations associated with HR

employee survey research. More importantly, as we will show, these methods have been tested in both

academic and applied settings and found to approximate accurately a wide variety of behaviour(s) in many

circumstances. Thus customer satisfaction can be quantified and related directly to profitability (Rust, Zahorik

& Keiningham 1998), and an organisation can use such research to determine which marketing options might

lead to improved customer satisfaction, and how these will ultimately affect profitability. It is hardly

surprising, therefore, that market research tends to receives more attention and funding from management than

does employee research.

AJHR paper final draft Mar 01 4 Word count 8600

Consumer choice modelling

The basic concept

Approaches used to better understand and predict consumer behaviour can be adapted and used to better

understand and predict employee behaviour. The approach that we outline in this paper relies on consumer

choice modelling, or more accurately “probabilistic discrete choice models.” A brief description of this

approach is as follows.

Consumers value products and services for the benefits they provide or the problems that they solve, which

derive from various features of products and services. All products or services that compete with one another

e.g., airlines, telephones, fast-food restaurants) can be described by combinations of relevant features. The

objectives of consumer choice modelling are to determine the key features that matter to consumers, quantify

or model how consumers evaluate the features and trade them off, and predict how changes in any of the

features will lead to changes in consumer choices. Probabilistic discrete choice models and the associated

empirical methods that have been developed to apply them to solve these types of problems provide a very

powerful and flexible body of theory and associated methods to address these types of strategic issues

(McFadden 1973; Hensher & Johnson 1981; Louviere & Woodworth 1983; Ben-Akiva & Lerman 1985;

Louviere, Hensher & Swait 2000).

In the same way, employment situations also can be described by common (perhaps also unique) features that

are evaluated and traded-off by current and prospective employees. If it is possible to model the evaluations

and feature trade-offs of consumers that underlie their brand and product choices, it follows logically that we

also can model current and prospective employees’ employment evaluations and choices, as well as many

other forms of employee behaviour likely to be of strategic interest (such as the length of contract one is

willing to accept, the hours that one agrees to work, the implied monetary equivalent of flexitime, etc). This

has the potential to produce outcomes of greater practical utility than are gained through conventional

employee satisfaction measures.

AJHR paper final draft Mar 01 5 Word count 8600

Employee satisfaction measurement systems or surveys are similar to customer satisfaction measurement

systems; and, like employee satisfaction surveys, the overwhelming majority of customer surveys try to

measure customer satisfaction directly, without regard to how customer behaviour might change if product

features or competitor actions change. Despite the many billions of dollars spent of such efforts by firms

annually, customer satisfaction is of little abstract interest in and of itself (Louviere et al. 2000) because the

primary interest of an organisation is not in customer satisfaction per se but in the behaviour that might

develop from such satisfaction. A firm’s chief interest is or should be in how its actions and those of its

competitors are likely to influence consumers’ choices in the marketplace. Similarly, if organisations take

appropriate actions to change the features of job and employment situations, this should lead to higher levels

of employee satisfaction, which in turn will lead to a variety of behavioural outcomes that are of strategic

interest and can be quantified and predicted: but it is the behaviour, rather than the satisfaction, in which

managers are interested.

The consumer choice modelling approach is based on a rigorous and well-tested theory of consumer choice

behaviour known as random utility theory (eg, McFadden 1973; Ben-Akiva & Lerman, 1985; Hensher &

Johnson, 1981). Random utility theory (RUT) postulates that consumers associate some utility (a latent

measure of preference) with each product that they consider. This “product utility” consists of an explainable

and an unexplainable component. The “explainable” component consists of that proportion of product utility

that can be explained and predicted by changes in product features and differences in consumers (eg, age,

income, etc.). The “unexplainable” component can be thought of as random error in consumer choices, but

really reflects the fact that we cannot look into consumers’ heads and know all the things that drive their

choices. Thus at some level all decision making and choice behaviour contains a random or “stochastic”

element when viewed from the manager’s or analyst’s perspective. RUT posits that consumers try to

maximise utility, which essentially means that they try to choose the things that they think suit them best, all

else equal. “Best” means the options that most closely match what they need and want, subject to constraints

like income and time. For example, a consumer needing and wanting a new car, although really preferring a

“prestige” model, will often actually choose a mainline model because the prestige model is too expensive.

That is, consumers try to choose products/services that are “best” for them, subject to what they know about

AJHR paper final draft Mar 01 6 Word count 8600

competing options and whatever constraints (e.g., income) are operating on their choices. Similarly, current

and potential employees try to choose jobs that best match their needs and wants, subject to constraints like

their level of skill and training, experience, location, etc.

Consumer choice modelling, as distinct from RUT, requires marketing researchers to acquire and use specific

information to implement the theory in applied settings (see Louviere 1988a, 1988b; Hensher et al. 1999;

Louviere et al. 2000). The key decision criteria (called here attributes) that drive consumer evaluations and

choices must be identified, and relevant and/or managerially useful ranges (or levels of each attribute) must be

specified, since the success of any choice modelling project depends essentially on one’s ability to identify as

many relevant drivers of choice as possible and establish the relevant levels as accurately as possible. This

typically is accomplished by examining prior market research results and tapping into client knowledge of the

market place using qualitative market research activities such as focus groups.

Although many marketing academics and practitioners believe that RUT-based probabilistic discrete choice

models and associated choice experiments are just another subset of techniques in the general family of what

are known as “conjoint analysis” (CA) techniques (Krantz & Tversky 1971; Green & Wind 1971; Louviere

1988b), this is not correct. For example – to name only a few of the many differences – unlike RUT, CA is not

a behavioural theory, but a mathematical/statistical technique; and unlike RUT, there is no obvious link

between CA methods and results and behaviour in real markets (for further details see Louviere 1994;

Louviere, Meyer, et al. 1999; Hensher et al. 1999; Louviere et al. 2000; Louviere & Hensher 2000). This is

not to belittle CA’s usefulness in the analysis of consumer tradeoffs, but merely to make it clear that we

believe that RUT-based stated preference methods dominate CA techniques on virtually all dimensions except

simplicity. Moreover, we note with respect to “simplicity” that many, if not most, job-related decisions are

complex and it is unclear what benefits are to be gained by using simple techniques to study complex

problems.

An example

Suppose, for example, that XYZ Networks Ltd, a provider of mobile telephone services, wishes to assess how

it could increase its market share. Let us assume that previous research has shown that consumer choices are

AJHR paper final draft Mar 01 7 Word count 8600

influenced by a number attributes of such services, such as monthly and usage fees, free minutes/month,

equipment price, bundled service features and coverage. The research objective is to determine, firstly, the

relative importance of each attribute for each of several demographic/market segments and, secondly, how

each segment will react to changes in the attributes. Each such change creates a different product

configuration or “variant”. Typically, a sample of current and/or potential customers will be surveyed to

capture reactions to each of a number of such configurations. A common task asked of each respondent is to

examine in turn each of a number of specified configurations alongside each of a number of configurations

offered by a competitor and to indicate whether, in each case, they would remain with/switch to XYZ if that

task represented the actual market situation (i.e., Configuration A1 for the XYZ product alongside

Configuration B1 for the competitor’s, then Configuration A2 for the XYZ product alongside Configuration

B2 for the competitor’s, and so on). The number of such tasks given to each respondent is determined by the

complexity of the evaluation task and the statistical requirements of the modelling effort. The features and

levels are treated as a statistical design problem, and principles from the design of factorial and similar

experiments routinely used in medicine, engineering and other fields, are applied to select a sample of

scenarios with a desirable set of statistical properties appropriate for the research objective(s). Samples of

consumers then evaluate subsets of the possible total range of configurations, with RUT then used to develop

appropriate statistical models that describe and predict the choices of the consumers. In this way, the relative

importance of each feature in the choice decision can be inferred, measured and quantified.

This differs fundamentally from typical customer and employee satisfaction measurement surveys in which

consumers are asked to self-report the importance of each driver on satisfaction. Not surprisingly, when

customers are asked to evaluate the importance of each feature separately (say, by rating the importance of

each), it is not unusual for everything (or at least many things) to appear to be important. While this may be

true in a specific instance, it does little to help managers prioritise actions and evaluate their likely behavioural

influences. In contrast, RUT-based methods for modelling consumer choice behaviour force consumers to

reveal their product feature trade-offs, and hence their priorities.

In addition, and probably more important managerially, the resulting models reveal how purchasing decisions

vary as a function of individuals’ evaluations of attributes and their personal characteristics (eg. the market

AJHR paper final draft Mar 01 8 Word count 8600

segment to which they belong, such as income, age, education, attitudes towards risk, etc.). RUT predicts a

particular relationship between the choices consumer make in such experiments and their actual choices in

real markets (Louviere & Swait 1993). This relationship allows one to calibrate survey-derived choice models

to real choices by observing the marketplace choices made by the sample studied or another sample, and their

perceptions of product features. This relationship can be estimated in virtually all applications and can

calibrate scenario-derived models to actual behaviour, even for individuals who did not evaluate the scenarios.

The value of such strategic information to a firm can be readily appreciated. Without such modelling, a firm

can assess the costs of offering a package, but not the associated benefit(s). With RUT-based choice models,

however, managers can determine if the increase/decrease in market share would be warranted by the change

in the cost and ease of service provision for each scenario. In this way, firms can analyse a range of options

and select the package(s) with the best cost-benefit ratios.

Validity

The RUT-based choice modelling approach has an impressive record in accurately predicting what people do.

A number of studies, involving choices between such things as public transport usage and choice of place of

residence, supports the conclusion that the models predict “real world” choices in a wide range of studies, as

reported by Louviere et al. (1999) and Louviere et al (2000). A number of academic and applied research

teams have consistently shown that models derived from observations of individuals’ decisions in hypothetical

situations are highly correlated with and can predict actual behaviour accurately(eg, see Louviere 1988b,

1994; Hensher, Louviere & Swait 1999). The 2000 Nobel Laureate in Economics, Professor Dan McFadden

(2000, p. 24) recently noted that:

“Numerous travel demand studies have now been published that use market research data. Some of the

early applications are Morikawa (1989), Ben-Akiva and Morikawa (1990), Morikawa, Ben-Akiva and

Yamada (1991), Hensher and Bradley (1993), Louviere, Meyer, et al 1999, Louviere (1993, 1999) [sic],

Hensher, Louviere and Swait (1999), and Brownstone and Train (1999). These studies show that carefully

collected conjoint analysis data are on the whole measuring the same preferences as revealed preference

data …”

AJHR paper final draft Mar 01 9 Word count 8600

For example, Louviere (1974) showed that a model of managers’ recommendations of hypothetical options

was highly correlated with their recommendations of actual options. Lerman and Louviere (1978) reported

that models of preferences for hypothetical residential options significantly outperformed econometric models

estimated from actual residential choices. Louviere & Meyer (1981) showed that predictions of models of

preferences for hypothetical supermarket options were highly correlated with actual supermarket choices of

independent samples of consumers. These and other early validity evidence were summarised by Louviere,

Levin, Schepanski and Norman (1983).

More recently, research in the RUT paradigm has established the theoretical link between decisions made in

surveys and lab experiments and corresponding real market behaviour (Ben-Akiva & Morikawa 1990;

Louviere & Swait 1993; Louviere, Meyer, et al. 1999; Louviere, Hensher & Swait 1999, 2000). Specifically,

this work posits that the primary difference in decisions in hypothetical and real situations is the difference in

the amount of random error in the respective decisions (ie, the amount of error variance). The hypothetical

decision situation must capture the essence of the real situation, of course, so it is crucial to spend adequate

time in advance of a decision making modelling project to understand as fully as possible the key drivers of

decisions and choices and the ways in which to best present them to the subject population of interest. Since

Morikawa’s (1989) pioneering work, there have been numerous tests of the hypothesised relationship, and few

serious rejections (eg, Adamowicz et al. 1994; Louviere, Moore & Fox 1993; Louviere & Swait 1993;

Louviere, Swait & Williams 1994; Louviere et al. 2000). This so-called “rescaling” approach has been used in

marketing, transport and environmental and resource economics to “calibrate” decision models to predict in

real markets1.

Thus, to summarise, the random utility theory paradigm provides sound, well-tested behavioural theory to

model human judgment and decision making. Models of human judgment and decision making have been

shown to predict real behaviour in real situations in a wide array of situations and markets. Although there

have been few applications in HRM, there is no reason to believe that they would not apply equally well in

many areas of strategic HR managerial interest. The next section demonstrates, in fact, that this is the case.

AJHR paper final draft Mar 01 10 Word count 8600

Application to employee research

Employee choice modelling

HR practice in recent years has seen wave after wave of change and innovation with the objectives of

improving organisational flexibility and workforce productivity. Career and employment systems have been

changed as part of corporate restructuring, employment security has been reduced, and managers at all levels

are being asked to do more with less. As a trade-off for these changes, many employees are offered higher pay

and other extrinsic rewards like performance bonuses. More sophisticated employers include intrinsic rewards

in their “trade-off packages”, such as redesigning jobs to give greater responsibility, freedom of action and

skill development, and so on.

Most large employers, faced with the need to make dramatic changes to their employment conditions, can

only guess at how their employees will react to them. Change is sometimes achieved only after protracted and

tough negotiations between employer and employee representatives, often derived from ambit positions taken

by both sides. These processes are usually lengthy and sometimes painful and may contribute to long-term

animosity and distrust between the parties. Employee choice modelling, as a supplement to this long-

established process, has the potential both to help HR managers better understand employee needs and

priorities and thus enhance management of change programs and to provide valid data on employee decisions

that might facilitate and accelerate the negotiation process. This is now illustrated with reference to a

particular project conducted for the Australian Army with which we have been involved for the past few

years.

The case study

The need for employee choice modelling research arose from the Army’s intention to make significant

changes to the career structures of its full time (“regular”) personnel. At that stage, the Army had about 27,000

regular officers and soldiers, almost all of whom enjoyed security of tenure to compulsory retiring age.

Compulsory retiring ages varied by rank, but due to the physically active nature of military service, these are

1 “Calibration” is not really the technically correct term because one actually estimates a theoretical constant

– the rescaling constant – from two or more sources of data.

AJHR paper final draft Mar 01 11 Word count 8600

generally lower than those imposed in the civilian community (eg, junior officers are not allowed to serve on

after age 45). Since lateral recruitment is minimal, promotion and career advancement are almost entirely

from within.

The Army had concluded, like most large Australian organisations, that an employment system based on long-

term employment security was a serious impedient to organisational flexibility. It wanted to change this

“permanency” to something else. Before this was done, however, some serious questions had to be addressed.

These questions included:

Were there viable and acceptable alternatives to employment security for a generally conservative

workforce that had traditionally accepted security as the trade-off between market-based remuneration?

Given the nature of military service, with its requirement for members to give “unlimited liability” (Jans

1988), was it morally acceptable to limit the employment security of members who had willingly worked

long hours, in often arduous and dangerous conditions, and accepted forced geographic mobility for

themselves and their families?

Would a shift to a more “occupational” style of employment undermine the “institutional” culture of

military service (Jans 1988)?

What kind of compensators (such as higher pay or improved career development opportunities) – if any –

would be acceptable to members as the trade-off for removal of employment security?

To answer these and other questions, the Army adopted the RUT-based research approach described above to

develop mathematical models of the decisions their personnel make about their intended commitment to full-

time service. The results of modelling the decisions made by members at all levels in response to a career

decision analysis survey were embedded in the software package called a “Career Decision Support System”

(CDSS).

Model development

A Career Decision Analysis Questionnaire (CDAQ), was developed after an extensive qualitative research

program. Focus groups were held in a number of centres with various types of personnel, which, together with

AJHR paper final draft Mar 01 12 Word count 8600

previous studies of work adjustment in the Australian military (Jans 1979, 1982, 1985, 1988, 1989b, 1989c;

Jans & Frazer-Jans 1989), were used to identify and operationalise the key features or attributes that military

personnel use in making career decisions.

Figure 1 about here

The ten attributes used in the CDAQ are shown in Figure 1. Each attribute was varied over four levels, except

for Retention Bonus (eight) and Geographic Stability (two). In the case of the “extrinsic” attributes, such as

Pay Rise, definition of levels was comparatively straightforward (nil, 4%, 8%, 12%). For the more “intrinsic”

attributes, such as Job Satisfaction, feature levels were expressed in terms of given work outcomes (“you

expect that none/one/two/all of your next three postings will provide you with work that will give you a sense

of purpose and match your professional interests”).

Career features represent the Army’s traditional “institutional” form of employment attraction. Army

personnel tend to have a strong expectation of high job satisfaction and the Army is generally able to place its

members in career streams which suit their skills and interests. Similarly, progressive and period promotion

can usually be maintained, because of the relatively early compulsory retirement age policy. Personal

development opportunities are available and, as noted, employment security is high.

Financial features represent the more “occupational” approach to employment conditions. The concept of

retention bonuses and gratuities additional to superannuation entitlements was seen by a number of policy

analysts (eg, “Serving Australia” 1996) as an appropriate response to the counter-attraction of market forces

from the civilian community. Although these were novel to the Army, they could be modelled.

Finally, family support attributes represent recognition by the Army for conditions to alleviate/compensate for

the stresses of frequent and compulsory geographic mobility on members’ families. Such conditions could, for

example, provide assistance to spouses/partners to find employment in new locations or arrangements that

would maintain dependent children’s educational stability as they moved from school to school and from state

to state (Jans, 1989c).

AJHR paper final draft Mar 01 13 Word count 8600

Six of these attributes (employment security, Retention Bonus, Gratuity on Retirement, Partner Support for

Employment, Children’s Support for Education, and Geographic Stability) expressed policies that were

essentially a radical departure from the traditional military employment contract. Thus the proposed policy

changes reflected in the modelling process were potentially quite radical for this institution.

The CDAQ was constructed by treating each of the employment features as a variable with a certain number

of levels, and then selecting a particular sample of combinations of feature levels from the entire factorial that

would have desirable statistical properties from a modelling standpoint. That is, each scenario consists of a

combination of feature levels, selected by a statistical design procedure to ensure that each feature varies

independently (ie, “orthogonally”) to every other feature. In this way the separate and independent effect of

each feature can be captured by an appropriate generalised multiple regression model, and the model then can

be used to predict responses to the many thousands of unobserved scenarios. Each respondent received 16

scenarios to evaluate. Respondents were asked to indicate how long they would be prepared to serve under the

conditions described by each scenario (see Figure 2). The questionnaire had four versions, each with a

particular set of 16 scenarios. Each respondent, of course, were given only one of the versions, with

respondents randomly assigned to a particular version2.

The final section of the questionnaire asked respondents to self-report their current career intentions and levels

of satisfaction, as well as to describe their current career situation in terms of the features varied in the

scenarios. The responses from this section of the survey were used to calibrate the model derived from

responses to the hypothetical scenarios.

2 Different applications generally require different experimental designs, and depending on the research

objectives of any given project, designs can be more or less complex and involve fewer or many more

scenarios, as well as multiple choices and comparisons.

AJHR paper final draft Mar 01 14 Word count 8600

Figure 2 about here

A stratified random sample of 2273 was drawn to cover all ranks up to and including brigadier. The CDAQ

was administered in group, face-to-face sessions through a network of specially trained Army Survey

Coordination Officers. The final sample was 1546, representing all sample members available in locations at

the time of the survey.

Each respondent ticked a box to indicate, for each scenario, the length of time they would agree to serve

(Figure 2). Although such ratings data often are analysed by means of ordinary least squares (OLS) regression

procedures (eg, Louviere 1988b), the response data probably do not satisfy the assumptions required for OLS

regression to produce consistent and unbiased estimates of the model parameters. That is, the response data

are unlikely to be equal in measurement interval (ie, “will serve for less than 12 months more” minus “will

serve for 2-3 years more” is unlikely to equal “will serve for 4-6 years more” minus “will serve for 7-10 years

more”). Nor are the errors likely to be independent and identically distributed normal random variates.

However, since such violations of OLS regression assumptions have been the object of considerable attention

in econometrics, a number of generalised multiple regression procedures have been developed to deal with

their violation. Additionally, we needed to develop models that were consistent with RUT in order to be able

to calibrate the models to predict actual decisions.

The foregoing considerations led us to rely on statistical models for ordered dependent variables. Ordered

logit and probit are the two well-known variants (eg, Zavoina and McElvey 1975; Greene 1993, 672-676).

These two models differ only in their respective assumptions about the distribution of the errors; we chose to

use the ordered probit model only because it has received more research attention. This model is derived by

assuming that the true decision is unobserved (called a “latent” variable), and the responses are regarded as

“indicators” of that true response with normally distributed random errors. So, if we code the tick boxes in

ascending order as 1, 2, 3, …, etc., and a respondent ticks “less than 12 months more,” we say the latent

indicator is y* is less than or equal to 1 (y* = the indicator). If a respondent ticks “2-3 years more,” we say

that y* is between 1 and a new parameter 1, which is the threshold value or cutpoint that divides tick box 1

AJHR paper final draft Mar 01 15 Word count 8600

from tickbox 2. In a similar way, each of J-2 (here J=5 tickboxes) ordered categories can be defined in terms

of cutpoint values between the tickboxes or categories. These cutpoint can be interpreted as the estimated

value of the true score that divides one tickbox from another. Put another way, if the tickbox categories

constitute an equal-interval scale, each of the estimated intervals should be approximately equi-distant; hence

ordered regression models not only provide a way to test this assumption, but also yield models that can

predict the response outcomes even if the assumptions are not satisfied.

The probit model predicts the probability (and hence the frequency distribution) that each of a number of

segments (eg, junior soldiers, junior officers, single soldiers, married soldiers, female officers, etc.) will agree

to serve for a particular category of years if offered a particular employment scenario. As we discuss later, this

allows us to compare predicted and actual distributions of years served to validate the models.

Findings

Segments that were modelled separately were determined by demographic characteristics, with a minimum

sample size of at least 150 respondents per segment. This minimum was derived from considering the

expected error bounds on the proportions of responses in the mutually exclusive and exhaustive response

categories under the assumption of equal choice probabilities for each category if the experimental variables

did not impact the category choices. Models were developed for 43 segments, derived from consideration of

strategically significant characteristics like as rank, employment type, marital status, family development

stage, marketability, etc.

For most segments, the most important features (ie, those that members were least willing to trade-off) were

job satisfaction and promotion chance; and the least important were personal development and employment

security. (The method of assessing relative importance or effect of each feature in the determination of

intentions and satisfaction levels is explained below.) Thus the results suggested that a standard set of

employment conditions (or at least those based around a standard core) would suit the vast majority of

personnel. This indicates that the main objectives of Army personnel policy should be, at a minimum, to

ensure high job satisfaction and to meet or exceed personnel promotion prospects expectations. (Validity is

considered below.)

AJHR paper final draft Mar 01 16 Word count 8600

Again, however, the most useful outcome comes from the modelling and the quantitative indicators of the

effects of given policy packages. For example, Figure 3 shows the results of modelling five types of package

for junior soldiers. Package #1 (the situation reflected in Figure 1) would be moderately effective in terms of

keeping a majority of junior soldiers for at least 3 years (31% would leave within three years). Package #2,

relying only on weak career features, would be a disaster because 91% would leave within three years. The

terms of Package #3 could partly offset this outflow by offering generous financial conditions and permanent

employment security. Package #4, involving attractive career features and a modest pay rise but with less

employment security (i.e., a 20-year contract), also produces an unsatisfactory outcome. The lowest predicted

wastage rate of the six packages is derived from #5, which consists of lower security (20-year contract)

balanced by very attractive career features, a modest pay rise, and strong family support (assistance to

spouses/partners to find career-related employment in new locations and educational stability for children in

both primary and secondary school years). In fact, Package #6 shows that these “balancing” features are such

strong motivators for service that, with such conditions in place, employment security could be reduced to a

12-year contract with little adverse effect on short term retention and satisfaction.

Figure 3 about here

Packages #5 and #6 suit both employees and the employer. Although they predict the best retention outcome,

they are not necessarily the cheapest to implement (family support programs are likely to be expensive).

However, because they rely on intrinsic and family development motives for service, they have the

considerable advantage of being aligned with the Army’s culture of “professional community”. And, whilst

the cost might be high, it would be at least partly offset by savings in recruitment, training and associated

organisational functioning costs.

In deciding which policy should be implemented, the Army now can assess the relative merits of each package

(policy). Package #5 has marginally higher retention, but at the comparative cost of lower flexibility in terms

of career structures. Although junior soldiers appear to be comfortable with the level of employment security

AJHR paper final draft Mar 01 17 Word count 8600

associated with Package #6 (which arguably is better for this segment), full consideration also must be given

to the effects of each package on other segments before final decisions are made (eg, on sergeants, because

after 12 years’ service many junior soldiers will reach this rank).

Validity

The external and predictive validity of the model was assessed in a number of ways, by:

comparison of predicted wastage rates with those that actually happened in the two years following the

survey;

internal consistency of the utility estimates derived for the various segments of the sample;

stability of the internal relationships in the model over time; and

a “common sense” test; that is, whether it made sense to those familiar with the military.

Predictive validity

First, the overall predicted wastage rates can be compared with those that actually happened in the two years

following the survey3. Army personnel wastage data for the two years following the survey closely correspond

to the predictions of the CDSS, when such predictions are based on the two most common “current situations”

as indicated by respondents in 1994 (viz, employment security at Level 4, promotion chance at Level 3, job

satisfaction at either Level 3 or Level 4 and all other attributes at Level 1). Indeed, the predicted wastage was

closer to the actual state of wastage than the latter was to the stated intentions in 1994 (see Table 1).

Table 1 about here

The following caveats apply to the comparisons in Table 1:

The observation period is very short and the three sets of figures are, in any case, very closely aligned.

This probably reflects the effects of members following through with decisions made, in many cases no

doubt, some time back.

3 Because of the need to be aligned with the Navy and the Air Force, neither of whom had undergone the

same degree of research and analysis as the Army, the latter has made no changes to personnel policy in this

interim period. A current project is essentially repeating the research exercise for all three Services.

AJHR paper final draft Mar 01 18 Word count 8600

The Army’s records do not differentiate between different kinds of officers and soldiers. Thus, for

example, the “soldier wastage” statistics include all ranks (privates to warrant officers). Had we been able

to compare differences in segment prediction accuracy, a different story might have emerged.

The predictive validity results are limited by virtue of the fact that the Army has not implemented policy

changes such as reduced employment security and increased family support. The effects of these features,

after all, are of the most interest; so until the Army implements them, validity comparisons are limited to

simple scenarios.

Validity comparisons were across the total Army population because individuals were not asked to

identify themselves in the survey; hence sub-populations cannot be tracked longitudinally. (A current

project does involve such tracking, but will take some time for the results to be available.)

Internal consistency

A key consideration in assessing the internal consistency of a model is whether there is a logical progression

of utility values from levels that should be low in utility to those that should be high in utility. That is, the

responses of the sample should validate the assumed attractiveness of the levels. For example, in the case of

the “pay rise” feature, we expect level 4 to have the highest utility, level 3 the next highest, level 2 the second

lowest and level 1 the lowest. Without exception, this was the case: in every segment, the utility estimates for

every feature were lowest for the lowest defined level and highest for the highest defined level.

Stability across time

Following the original survey in 1994, two limited-sample surveys (N=178 and 170 respectively) were

conducted in 1995 and 1996, with the same CDAQ questionnaire and method of application.

The hypothesis tested in each case was that there was no difference in the relative importance placed on each

attribute in each of the three surveys. The relative importance (or relative contribution) of each attribute can be

approximated by calculating its partial (percent) contribution to the total log-likelihood, an analog to the

proportion of explained variance due to each attribute in a linear regression. A simple approximation to the

contribution can be obtained by calculating the range in the utility estimates (high versus low) for each

attribute. The range provides a good approximation because it behaves like a partial variance under the

AJHR paper final draft Mar 01 19 Word count 8600

assumption of a linear relationship between each attribute and the response measure (eg, if the utility

coefficients for the influence of job satisfaction on retention for the total sample were -.47, -.20, +.01 and

+.39, the range would be .86). The latter measures can be used in this way because the experimental design

used to generate the scenarios insures that the attributes are orthogonal; hence partial contributions can be

summed. The estimated utility values can be thought of as measures of the contribution of the levels of each

feature to the overall scenario responses. For example, a high negative utility estimate would suggest that

offering that level would be a significant deterrent towards inducing the behaviour of interest, whereas a high

positive utility estimate would indicate that offering that level would be relatively attractive.

Comparisons between the three surveys (see Table 2 ) showed that job satisfaction, career advancement

opportunities and retention incentive bonus were consistently the largest drivers of retention. The remaining

attributes were approximately equally important, except for pay rise, which had a bit more effect than the

others, but had less influence than the career attributes. There were few significant differences or systematic

trends in relative effects from 1994 to 1996.

Table 2 about here

Although there seemed to be a small shift in members’ values over the two years, it was not large enough to

call the modelled predictions into question nor to invalidate the recommendations associated with the results.

In fact, because the differences are small and most trends non-systematic, it is not clear whether the model

should be tested and revised if necessary. Nonetheless, prudence probably would dictate at this stage in our

knowledge and experience that the models be tested for stability with small samples every 24 or 36 months.

The “common sense” test

The final check on the validity of the model is whether it “makes sense” to managers in the organisation: did

the order of relative influence of the features on retention (Table 2) for both the total sample and for the

various segments make intuitive sense to the user? Army managers confirmed the intuitive logic of the derived

order of relative influence on retention.

AJHR paper final draft Mar 01 20 Word count 8600

Additionally, internal comparisons between segments were intuitively logical. For example, when comparing

the influence of the ten attributes on married and single officers’ retention decisions, married officers place

relatively more weight on family support factors, and single officers relatively more on career development

factors (Table 3), but both categories place about the same weight on financial support attributes. Similar

“logical” internal differences were observed in other comparisons. For example, junior soldiers value

promotion chance than more than warrant officers, who are older and (for many) near the end of their Army

careers. In contrast, as might be expected, warrant officers place higher values on financial features and family

support.

Table 3 about here

Uses of the modelling outcomes

The modelling outcomes have a number of uses, at the strategic and operational levels of HRM. At the

strategic level, the Army has been able to identify the extent to which it could make changes to its long term

psychological contract of paternalistic personnel management, especially in terms of employment security.

The findings give Army confidence that it could reduce employment security, and show those features of

employment that, if strengthened or introduced, would offset any adverse effects of such reductions. At the

operational level of HRM, the model has been used to determine the policies needed to retain specific

employee categories, such as middle level officers with poor long term promotion prospects but whose

technical skills in fields such as IT makes them both valuable to the Service and at-risk in terms of alternative

employment opportunities.

Other possibilities include the use of the process in determining how best to attract new recruits, or what

differential sets of career incentives will attract different types of recruits (e.g., generalist officers versus

specialists such as doctors and aviators).

AJHR paper final draft Mar 01 21 Word count 8600

Discussion and Conclusions

The current approach to employee surveys, widely used to assess employees’ sensitivity to career features, is

unlikely to achieve its potential unless greater attention is paid to ways of enhancing the useability and

immediacy of such research. As shown above, an “employee choice modelling” approach to employee

research, adapted from consumer modelling research techniques, can help HR managers better understand and

model employee needs and priorities.

In one sense, the employee choice modelling approach is not novel. After all, researchers have known for

some time that the closer an “attitude” is to an actual behaviour, the stronger its predictive power for that

behaviour. For example, intention to quit is a stronger correlate of quitting than are job satisfaction and

thinking of quitting (Muchinsky 1990 p 331). This is because employees are being asked, indirectly, to weigh

up all aspects of their job or career role and make an overall decision about their future with it. Employee

choice modelling, by simulating the conditions that might cause people to quit or to stay, exploits this

phenomenon. Indeed, employee choice modelling goes further by, firstly, including employment conditions

that may not yet exist so that the trade-off potential of novel strategies can be assessed and, secondly,

presenting results in a quantified format that facilitates their consideration by business-minded managers.

Whilst the employee choice modelling approach has obvious advantages over conventional, scale-based

employee research, it does have certain limitations. Firstly, employee choice modelling is considerably more

expensive than conventional survey methods, and projects based on this approach usually take longer. Indeed,

attitude and opinion surveys have almost reached the stage of “do-it-yourself” for many HR managers, given

that modern desk-top publishing and commercial software packages have made them easy to mount and

analyse. In contrast, employee choice modelling projects tend to take much longer; and their data must be

analysed by analysts with sophisticated statistical skills. Secondly, employee choice modelling’s validity

might be limited in cases in which hypothetical employment conditions are beyond employees’ experience-

imagination boundaries. For example, employees who have never had to change locations in order to pursue

their careers may find it difficult to evaluate their reactions to the possibility of having to do so, because they

have no concept of the benefits and drawbacks of doing so. In such cases, although people might be able to

AJHR paper final draft Mar 01 22 Word count 8600

assess the economic and life-style costs of having to relocate, they are less likely to be able to assess benefits

such as meeting new career challenges and the stimulating effect of geographic and social variety. Thirdly,

validity might also be limited in situations in which external circumstances may have much more influence on

job persistence than employment conditions (such as might apply with casual workers, vacation-job

employees, backpackers and the like). For this methodology to work properly, therefore, respondents

probably need to have a certain threshold commitment to the employment situation, so that their decisions are

significantly influenced by what the job offers in comparison to their expectations and/or to alternative

employment situations.

Early indications of the validity of the method are encouraging, but further research (some of which is

currently in progress) is needed. For example, although the application of the method has been confined to the

military so far, future research will extend the method into the civilian sphere.

The findings reported in this paper suggest that employee choice modelling can provide managers with a

means to capture and quantify the potential benefits of any given HR strategy, by showing such a strategy will

affect employee behaviour. It thus enhances the capacity of managers to make better informed and more

strategic HR policy development decisions.

AJHR paper final draft Mar 01 23 Word count 8600

References

Adamowicz, W., Louviere, J.J. and & Williams, M. (1994) Combining Revealed and Stated Preference

Methods for Valuing Environmental Amenities, Journal of Environmental Economics and

Management, 26, 271-292.

Ben-Akiva, M. and Morikawa, T. (1990) Estimation of Switching Models From Revealed Preferences and

Stated Intentions, Transportation Research A, 24A(6), 485-495.

Ben-Akiva, M., & Lerman, S.R. (1985), Discrete Choice Analysis: Theory and Application to Travel Demand,

MIT Press, Cambridge, MA.

Brownstone, D. & Train, K. (1999) Forecasting New Product Penetration with Flexible Substitution Patterns,

Journal of Econometrics, 89, 109-129.

Green, P. E., & Tull, D. S. (1978) Research for Marketing Decisions. 4 ed. Prentice-Hall, Englewood Cliffs,

NJ.

Green, P.E. and Wind, Y., 1971, Multiattribute Decisions in Marketing: A Measurement Approach, Dryden

Press, Hinsdale, IL.

Greene, W.H. (1993) Econometric Analysis. Prentice-Hall, Englewood Cliffs, NJ.

Hensher, D.A. & Bradley, M. (1993) Using Stated Response Data to Enrich Revealed Preference Discrete

Choice Models, Marketing Letters, 4, 39-152.

Hensher, D. A. & Johnson, L. (1981), Applied Discrete Choice Modelling, Croom Helm, London.

Hensher, D.A., Louviere, J.J. and Swait, J. D. (1999) Combining Sources of Preference Data, Journal of

Econometrics, 89, 197-221.

Jans, N. A. (1979) Work involvement and work satisfaction: an investigation of two indicators of work

adjustment in organisations. Unpublished PhD thesis, Department of Organisational Behaviour,

University of New South Wales.

Jans, N. A. (1982) The nature and measurement of work involvement, Journal of Occupational Psychology,

55, 57-67.

Jans, N. A. (1985) Organisational factors and work involvement, Organisational Behaviour and Human

Decision Processes, 35, 382-396.

AJHR paper final draft Mar 01 24 Word count 8600

Jans, N. A. (1988) Careers in Conflict: A Study of Services Officers' Careers and Families in Peacetime.

Canberra College of Advanced Education , Canberra Series in Administrative Studies, No. 10.

Jans, N. A. (1989a) Organisational commitment, career factors and career/life stage, Journal of

Organizational Behaviour, 10, 247-266

Jans, N. A. (1989b) Problems in military professionalism: a study of change in the Australian Defence Force,

Armed Forces & Society, 15, 181-192.

Jans, N. A. (1989c) The career of the military wife, Human Relations, 1989, 42, 337-351.

Jans, N. A. (2000) Options for ADF Human Resource Management: Part of the Solution or Part of the

Problem? in Ball, D. (Ed.) Maintaining the Strategic Edge: The Defence of Australia in 2015, Strategic

& Defence Studies Centre, Australian National University, 381-406.

Jans, N. A. & Frazer-Jans, J. M. (1989) Facing up to the future: proposals for career/personnel issues to assist

in staffing the ADF in the 1990s and beyond. Report to the Chief of the Defence Force.

Krantz, D.H. & Tversky, A. (1971) Conjoint-Measurement analysis of Composition Rules in Psychology,

Psychological Review, 78,151-169.

Lerman, S.R. & Louviere, J.J. (1978) On the Use of Direct Utility Assessment to Identify Functional Form in

Utility and Destination Choice Models, Transportation Research Record, 673, 78-86.

Louviere, J. J. (1988a) Conjoint Analysis Modeling of Stated Preferences: A Review of Theory, Methods,

Recent Developments and External Validity, Journal of Transport Economics and Policy, 22 (1), 93-

120.

Louviere, J. J. (1988b), Analyzing Decision-Making: Metric Conjoint Analysis, Sage Publications, Newbury

Park, CA.

Louviere, J. J. (1994), Conjoint Analysis, In R. Bagozzi (Ed.), Advanced Marketing Research, Cambridge,

MA: Blackwell Publishers.

Louviere, J. J., & Meyer, R. J. (1981) A Composite Attitude-Behavior Model of Traveller Decision

Making, Transportation Research, 158(5), 411-420.

Louviere, J. J., & Swait, J.D. (1993) The Role of the Scale Parameter in the Estimation and Comparison of

Multinomial Logit Models, Journal of Marketing Research, 30, 305-314.

AJHR paper final draft Mar 01 25 Word count 8600

Louviere, J. J. & Woodworth, G. (1983) Design and Analysis of Simulated Consumer Choice or Allocation

Experiments, Journal of Marketing Research, 20 (November) 350-367.

Louviere, J. J., Levin, I, Schepanski, A. & Norman, K. L., (1983) Validity Tests and Applications of

Laboratory Studies of Information Integration, Organizational Behavior and Human Performance,

31, 173-193.

Louviere, J. J., M. Fox, & W. Moore (1993) Cross-Task Validity Comparisons of Stated Preference

Models, Marketing Letters, 4, 205-213.

Louviere, J. J., Moore, W. & Fox, M. (1993) Cross-Task Validity Comparisons of Stated Preference Choice

Models, Marketing Letters, 4, 205-213.

Louviere, J. J., Swait, J. & Williams, M. (1994) A Sequential Approach to Exploiting the Combined

Strengths of SP and RP Data: Application to Freight Shipper Choice. Special Issue on The Practice of

Stated Preference Methods, Transportation, 21, 135-152.

Louviere, J.J. (1974) Predicting the Evaluation of Real Stimulus Objects from an Abstract Evaluation of

Their Attributes: The Case of Trout Streams, Journal of Applied Psychology, 59(5), 572-577.

Louviere, J.J. and Hensher, D.A. (2000) Combining Sources of Preference Data, Invited Resource Paper,

Workshop on Combining Sources of Preference Data, International Association of Travel Behaviour

Researchers Conference Proceedings, Gold Coast, Australia, July.

Louviere, J.J., Hensher, D.A, & Swait, J.D. (2000) Stated Choice Methods: Analysis and Applications,

Cambridge, UK: Cambridge University Press.

Louviere, J.J., Hensher, D.A. & Swait, J. D. (1999) Conjoint Analysis Methods in the Broader Context of

Preference Elicitation Methods, In Gustafson, A., Hermann, A. and F. Huber (Eds.), Conjoint

Measurement: Methods and Applications, Berlin: Springer-Verlag, 279-318.

Louviere, J.J., Meyer, R. J., Bunch, D.S., Carson, R., Dellaert, B., Hanemann, W.M., Hensher, D.A., and

Irwin, J. (1999) Combining sources of preference data for modeling complex decision processes,

Marketing Letters, 10 (3), 205-218.

McFadden, D. (1973) Conditional Logit Analysis of Qualitative Choice Behavior, in P. Zarembka (ed.),

Frontiers in Econometrics, Academic Press: New York, 105-142.

AJHR paper final draft Mar 01 26 Word count 8600

McFadden, D. (2000) Disaggregate Behavioral Travel Demand's RUM Side: A 30-Year Retrospective, Paper

presented to the International Association For Travel Behavior Research, Surfer's Paradise, Australia

(July).

Morikawa, T., M. Ben-Akiva, K. Yamada (1991) Forecasting Intercity Rail Ridership Using Revealed

Preference and Stated Preference Data, Transportation Research Record, 1328, 30-35.

Morikawa, T.M. (1989) Incorporating stated preference data in travel demand analysis, Ph.D. Dissertation,

Department of Civil Engineering, MIT.

Muchinsky, P. M (1990) Psychology Applied to Work: An Introduction to Industrial and Organizational

Psychology. Brooks/Cole, Pacific Grove, CA.

Rust, R.T., Zahorik, A.J. & Keiningham, T.L. (1995) Return on Quality (ROQ): Making Service Quality

Financially Accountable, Journal of Marketing, 59, 58-70.

Schmitt, N. W. & Klimoski, R. J. (1991) Research Methods in Human Resource Management. South-Western,

Cincinnati, OH.

Zavoina, R. & McElvey, W. (1975) A statistical model for the analysis of ordinal level dependent variables,

Journal of Mathematical Sociology, Summer 1975, 103-120.