Measuring work organization exposure over the life course with a job-exposure matrix

Measuring work organization exposure over the life course with a job-exposure matrix. Scand J Work Environ Health 1993;19:21---8. In most epidemiologic studies of occupational stress, workexposure is measured at only one point in time. This article presents a meth odology for measuring the intensity and duration of work organization exposure over a lifetime. A job-exposure matrix for work organization was developedfrom data on a random sample (N=12 084) of the Swedish labor force. The matrix consisted of mean exposureestimates for work control, social support, psychological and physical job demands, and job hazards for 261 occupations. Several vali dations of the matrix were undertaken. The matrix scores were found to be significantly correlated with individual self-reportscores, and a similar increase in chronicdisease prevalence with decreasing work control was found for both the matrix and the individual scores. The matrix was applied to oc cupational history data to construct lifetime exposure profiles. Potential problems in using attribution systems for exposure assessment are discussed.

Increasing evidence suggests that expo sure to adver se work organization characteristics (suc h as high performance demands coupled with low leve ls of work control and social support) place ind ividuals at increased risk of stress and chronic illnesses (1-1 2). A number of major methodological challenges remain, however, before a clear causal association between work organization and physical illness can be established (5,(13)(14)(15)(16)(17). One such challenge involves the question of how exposure to work organization is conceptualized and measured.
In population-based epidemiologic studies, work characteristics have been measured in two ways. Questionnaire responses of large samples about their work content have been used to characterize the exposure of individuals at the time of the interview (5,7,11,12), or, alternatively, national survey registers have been used to develop attribution systems by aggregating survey responses of individuals to arrive at occupationally representative scores for work environment ch aracteristics (18)(19)(20)(21). These occupational scores are then attributed to subjects (freq uent-ly in another data set) on the basis of the job held at a given point in time (19,20,22,23). Both of these measurement systems share methodological we aknesses in that the exposure status is usually assessed cro ss-sectionally and is based solely on the job held at the time of interview. Indeed, with a few exceptions (5,16,24,25) , the idea of exposure as a timedependent process exerting an incremental effect on the risk of disea se has not been thoroughly conceptualized or investigated in the occupational stres s field. In earlier population-based studies it was assumed that work organization either exerts an acute effect on disease risk or that the exposures for all past occupations were identical to tho se of the last occupation. The fir st assumption is inconsistent with recent evidence that cumulative exposure to adverse work organization is associated with an excess risk of chronic disease (7-12, 18-20, 22, 23). Given the propensity for job change and career mobility in modern society, the second assumption is not reasonable . Moreover, in such studies the potential effects of prior exp osure such as that motivating an individual to change occupations and the time-dependent relationship between exposure and disease risk cannot be evaluated. This paper presents a procedure for assessing exposure to three psychosocial and two physical work organization charac teristics retrospectively. Using methods developed in occupational epidemiology (26), we have developed a job-exposure matrix for work org anization (ie, a work organization matrix) to assign exposure scores retrospectively to occupational title groups and create an individual worklife time-exposure profile.

Material and methods
In occupational studies of chemical agents a detailed profile of exposure to a specific agent is often lacking, and, for diseases with long incubation and latency periods, data are typically insufficient to define exposure quantitatively on an individual basis, Exact approximations of a worker's exposure profile is only possible when the induction and latency periods are short or when monitoring data have been systematically collected over the period of employment of the study cohort. These challenges have motivated the use of surrogates of the subject's specific exposure profile, for example, the job-exposure matrix (26)(27)(28), which is used to approximate exposure intensities over time. The simplest type of jobexposure matrix consists of occupational groups or some other means of defining homogeneous exposure clusters (27)(28). Other dimensions are added to the matrix when exposure variation exists within occupations. We have applied this general approach to the construction of a work organization matrix.

Sample
The 1977 and 1979 years of the Survey of Living Conditions made by Statistics Sweden were used to construct the work organization matrix. The response rates were 81% in 1977 and 89% in 1979. The survey is performed annually and consists of systematic random samples of the entire Swedish population from the ages of 16 to 74 years. These two years were chosen because they contained the most extensive information on the work environment. The work organization matrix was based on the responses of 12 084 employed Swedish citizens between the ages of 25 and 74 years. Altogether 5658 subjects (46.8%) were women and 6426 (53.2%) were men. Non-Swedish citizens were excluded because of the likelihood that a significant portion of their work history had occurred in their native country. A 25-yearold age entry criterion was used because no data on occupational history were available before this age.

Measurement of work characteristics
The indicators for the work environment characteristics used in this study have been shown to predict cardiovascular disease (11-12) and psychosomatic strain. The questionnaire items have been published elsewhere (12,25,(29)(30)(31), and the reliability and validity of the indicators have been discussed in several publications (7,11,25,(29)(30)(31). All of the scales were multiplicatively transformed so that the scale scores ranged from 0 to 10.
The work-control scale is a l2-item indicator measuring decision authority and skill discretion on the job with a mean of 4,89 and a standard deviation of 2.01. Cronbach's alpha is 0.75 for this scale.
The work social-support index is a four-item measure of the availability of social interaction in the workplace and with co-workers. The mean of the index is 8.67, and the standard deviation is 2.14. No alpha has been computed for this index since it measures interaction opportunities rather than a latent construct.
The psychological job-demand indicator is a twoitem scale that measures time pressure and psychological demand. It has a mean of 4.70, a standard deviation of 3.94, and a reliability coefficient of 0.60.
The physical job-demand indicator consists of five items that measure the physical burden of the job. The mean of the scale is 4.70, its standard deviation is 3.94, and its Cronbach's alpha is 0.73.
The hazardous exposure scale is made up of seven items that address a series of exposure characteristics that are potentially hazardous. It has a mean of 2.65, a standard deviation of 2.23, and a Cronbach's alpha of 0.71.

Statistical methods
In the interview the subjects were asked to recall their complete work history and the duration of time each job was held. Information on only the current occupation was used to construct the work organization matrix. The Nordic version of the three-digit International Standard Classification of Occupations was used to form 261 occupational title groups (32). Exposure scores for each job characteristic were estimated by gender, occupation, age (two categories: 25---44 and~45 years), and the duration of time in the specific occupation (three categories: 1-5,6-19, and~20 years). Separate job-exposure matrices were developed for the men and women on the basis of the previous work of Hall (30), who reported that the majority of Swedish workers are employed in jobs that are highly gender segregated. In addition, Hall reported that there are significant gender differences in the level of work control for men and women holding the same job (32). These findings suggest that men and women in the Swedish labor force, and even in the same occupation, can be exposed to differing levels of work organization characteristics (33).
In its most stratified form the work organization matrix distinguishes between chronological age and duration of employment within a specific occupation. Previous job-characteristic attribution systems have stratified or adjusted for age (20)(21). Since age is more readily available than duration of employment, and in order to increase the general utility of the work organization matrix for other investigators, genderspecific occupational scores were stratified by age alone (the duration categories being collapsed). Gender-specific occupational means (collapsed across age and duration of employment) have also been included in the work organization matrix. The decision to use one form or the other is up to the investigator and would depend on his or her specific research question and the avail abilit y of information related to age and duration of employment.
Although age is often used as a proxy for duration of empl oyment , the introduction of a sepa:ate time dimension strengthen s the acc uracy of the Jobcharacteristic estimates in those situations when age and durat ion of employment are not the same. At an individual level, to the extent that a worke r moves in and out of the labor force or undergoes career mobility, chronological age wi ll n~t~deq~atel'y measure dur ation of employment. Thi s situano n IS the truest for women, who often leave the labor force dur ing child-rearing years, to return again later in life. A wom an, on reentry into the labor force, must often lite rally "beg in at the beginning" -eve n though she may be older. We have found that in certain jobs predomin ately performed by women (such as typist, secretary , cashier, and bookkeeper) women who start working in their forties report lower levels of work control than do women who start these jobs in the ir twenties.
. . . In an analysis of variance of both the indiv idual scores and the matrix-attributed scores age, duration of employment, and the age-duration interaction were found to be statistically significant.
The Stati stical Analysis System (34) was used to generate the input data set for the matr ix. This proced ure was performed separately for the men and women, and it created occ upatio nal means and variances for the five work characteristics. Accord ing to the criterion established in earlier occupational attribution system s, a minimum sample size of four was required for a gender, occupation, age, and duration of work specific category to estimate directly the mean sco re and stand ard deviation (20-2 1). If the sample size was less than four, then we estimated the cell-specific mean by borrowin g data from the nearest neighb or ce lls until the sample size for the estimate was itself greater than or equal to four. The orde r of borrowing was gender-spec ific, occ urred mainl y within the three -digit occupational code, and was decided on the basis of proximity of an age-and duration-specific cell to the cell of interest.
The coeffi cient of variation (CV) (standard deviatio n/mean) was used to exam ine the relati ve variability of the estimates of the work organiza tion matrix for the work-control variable as a function of the sample size of the cell. The CV values generally ranged from 0 to 0.85 for both the men and the women, the majority of estimates rangi.ng from 0. 10 .to 0.60. The average CV was approx imately 0.35, Irrespecti ve of cell size. Though the CV range was somewhat broader for the sma ller cell sizes, there was little indication of a meaningful associ ation between cell size and the magni tude of the CV. We would also note that when the work organization matrix was applied to individuals in the study sample, only 4% of the men and 2% of the women received estim ates derived from cell sizes of four. Altogether 85% of the men and 90% of the women re-Scand J Work Enviro n Health 1993, vol 19, no I ceive d estimates derived from a cell size of eight or more.
All of the technical and other details regarding the work organization matrix have been publi shed as a research monograph that is ava ilable upon request (25).: 1

Proportion of variance explained by the work organization matrix
The proporti on of the populat ion variance in~w?rk cha racteristic explain ed by the work orgaruzation matrix was assessed through a comparison of attributed and self-report scores associated with the occupation held at the time of the interview acco rding to the methods developed by Schwartz et al (2 I). The proportion of the total sample vari ance explained by the work organization matrix was estimated from the correl ation of the attributed score specific for gender, occ upation, age, and duration of employment with the self-report score. The Cronb ach ' s alpha reliabil ity coefficients reported in the Measurement of Work Char acteristics section were used to estim ate the upper bound of variance that can be reli ably measured for each sca le (2 1). The ratio of variance estimated from the attr ibuted score s to the total reliable variance, estima ted by Cronb ach ' s alpha , was derived for each sca le.
Although the prop ortion of varian ce explained by the work organization matrix (table I) was greater for the men than for the women, the pattern was similar for both genders. Of the five variables , the attributed score for physical jo b dem ands accounted for the greatest proport ion of explained variance for both genders, and the largest ratio of explained to reliable variance (ie, 73% for the men and 49.3% for the women). The work-control and job-hazard indices also explained a meaningful proportion of the individual score variance for the men and women. The ratio of expl ained to reliabl e variance ranged from 32.4 to 63.5 % for the women and men, respectively. The attributed score for psych ological job demand s explained the least reliabl e variance (29.3% for the men and 24% for the women ). Thi s find ing was not unexpected, give n that the few items comprising this index were more likely to reflect the individual percepti on of work conditions at the time of the interview in contrast to the structura l characteristics of the job . As such, there is likely to be greater heterogeneity of exposure within anyone occupational category. The social support indicator also explained only 25% of the sampl e variance. Thi s situation may be due to the considerable variation in  social support opportunities that occur within occupations (ie, variance that is not captured by the work organization matrix indicator) . By contrast, the items comprising the work-control and physical-demand indicators focused on characteristics that were more likely to vary markedly between occupations. Similar findings have been noted by Schwartz et al (21) regarding the psycholo gical-dem and and work-control indicators. It is possible that a proportion of the variance in the individual self-report scores explained by the scores of the work organization matrix may be due to the fact that the scores were initially derived from these same individual self-reports. In order to test whether this issue of circul arity accounted for the observed association s, we applied the work organization matrix to a different nationally representative, random sample (N = 6138) that was collected by Statistics Sweden in 1980 and included a physical job-

WOM Attributed Scores
Chronic disease prevalence We compared the self-report and attributed scores by examining their respecti ve odds ratios for the prevalence of cross-s ectional chronic disease morbidi ty. Work control was selected for this purpo se because of its known association with stress-related chronic disease conditions (9). The latter was defined as a dichotomous variable of having one or more of a subset of illnesses with a plau sible stress-related etiology, includ ing cardiovascular disease, digestive system disorder s, and back problems. For this analysis both the attributed and self-report work-control scores were divided into quartiles using the same score values for defining each category.
The odds ratios for both the attribut ed and selfreport scores exhibited a similar stepwise increase in the odds ratios for chronic illness with decreasing demand indicator. This check yielded a correlation of 0.61 for the men and 0.47 for the women between the scores of the work organization matrix and the individual self reports for physical job demands. Although the magnitudes of these associations are somewhat less than those previously obtained, they are sufficiently strong to suggest that circularity does not account for the preponderance of the associations reported in this article.
The attributed score for work control was regressed on the self-report score to examine the specific relation between these two methods of assigning scores to individuals, and the results are depicted in figure 1 for the men. The pattern obtained for the women was almost identical and hence has not been displayed. The broken line represents a simple linear or best fit of the attributed and self-report scores. The unbroken line is the actual observed regression line estimated from the observations for each gender. The regression line (unbroken line) was estimated accordin g to Cleveland' s smoothing procedure (35), a nonparametric method that fits the best line to a sequential moving subset of observations, rather than imposing a single linear model on the data. The actual observed trend was found to approximate the "ideal" regression line very closely, for control values in the mid to upper range. At the lower end of the scale, however, the attributed score tended to underestimate the self-report score.  Table 2, Gender-specific age-adjusted odds ratios and 95% confidence intervals for the attributed scores of the work organizat ion matrix and self-report scores of the Swedish men (N =5312) and women (N =5069 ) determined for the relation between chronic disease prevalence and work control by quartiles. (OR =odds ratio, 95% CI =confidence interval) levels of work control for the men (tab le 2). Among the men the expos ure-resp onse trend was more indicative of a stepwise pattern when the attributed score was compare d with the self-report score. Amon g the women no stepwise pattern was observed; rather a similarly elevated odds ratio for both measurement methods was found for the medium high and lowest control quartiles.

Linking the work organization matrix with occupational history
The application of the work organization matrix to an individual work history is iIIustrated in table 3 and depicted graphicaIly for work control, psychological job demands, and socia l support in figure 2. We created an exposure pro file by applying the scores of the work organizatio n matrix for motor vehicle operlator and miner to the subject's work history. At each stage of his or her work career, age-and durationspecific job scores were assigned to the individual 's work history as he or she aged, gained more exper ience, and changed jobs.

Discussion
Previous exposure assessment of work organization in the occupational stress field has been based primarily on single point assessment of the respondent's percepti on of the work situation at the time of interview. Although the inadequacy of this approach has been pointed out by other investigators (5,16), to date, the methodol ogy has not existed that would permit an examination of the relatio nship between duration and intensity of work orga nization exposure and chronic disease risk. The purpose of this paper was to begin to address this maj or method ological weakness in the psychosocial work-environment field thro ugh developing a way to measure the durat ion and intensity of work exposure ove r a lifetime. We do not cla im that this is the perfect or even the most preferable approac h; rather it is an initia l attempt to address some of the limitations in the field in order to promote the study of work orga nization factors with greater meth odological rigo r.
Th is paper illustrates the applica tion of a matrix for work organization exposure to self-reported job histories to develop individual exposure profiles. It was found that the elements of the work organiz ation matrix that measure more objec tive aspects of occ upational exposure, such as work control and physical job demands, expl ained a substantial proport ion of the total sampl e variance. In addition, a similar pattern of association was found for both the work -control scores of the matrix and the individual self-report scores in regard to cross -sectional chro nic disease prevalence. The methods outlined in this article , although developed with Swedish data, should be applica ble to other national settings in which similar data elements are avai lable, namely, an eq uivalent, nationally specific matri x in combination with occ upational history data.
As menti oned earlier, we are awa re that it would be desirable to validate the matrix on other data sets that contain occupationa l and work envi ronment characteristics. Thi s vali dation would be useful not only because of the information gained with respect to the potent ial difficult y of circularity (disc ussed earlier) but also because this is a new research tool using new approac hes to work organizatio n charac -teristics. In this rega rd, we should note that the work orga nization matrix has recentl y been applied in a study of job strain among Swedish men, and a significant association with ambulatory diastolic blood pressure was found (36).
The work organization matri x and other attribution systems are designed to explain the maximal proportion of variance that exists for exposure characteristics at the ind ividual level. This approac h works well when the total varia tion in the work organization variables within the strata defined by the occupational classific ation groups of the matrix is small relative to the variation that exists between strata. Using covariates to define the strata further helps to minimize the variation that might exist within occupational strata. If a work-exposure characteristic is not associated with the stratification variables that define the job-exposure matri x, the attribution system will be of no value for that work environment trait. For exa mple, the work organization matrix accounted for only a small proportion of individual variance in psychologi cal job dem ands but for a large proportion of individu al variance in work contro l. This result was probably due to the fact that control is a work charac teristic that varies across occ upations, wherea s psychological demands vary across indiv iduals and work sites within occ upations.
Another potential probl em with the work organizat ion matrix involves its application to retrospective exposure assignment. When the work orga nization matrix scores are used to estimate indiv idual expos ure charac teristics retro spectively, misclassificatio n error may occur because the job has changed at some point in the past. Although the mag nitude of the exposure estimates may change, it is less likely that the relative position of one occupation to another will change. Therefore the work organization matrix may still provide valid ordinal-level estimates of past expo sure.
Although every job-exposure matrix has its deficiencies , in some instances, attr ibution systems are superior to individually reported expos ure status when an individual' s self-report represent s a biased assess ment of the work environm ent. Attributed scores are partic ularly valuable when the dai ly variation in indiv idual self-reported exposure status is large due to changes in individual perception or actual changes in work practices. In this instance attribution syste ms ca n provide a more reliable and more valid esti mate of average exposure status than individual self-reports taken at one point in time.
The exposure estimates of attribution systems like the work orga nization matri x could be improve d in several ways. The inherent properti es (the reliability and construct validity) of survey-based measurement s predetermin e the validity of matrix-derived estimates. An impro ved job -exposure matrix could be constructe d based on more rigorous and lengthy questionnaires that focus specifically on objec tive aspects of work content. Further stratification vari-abies could also be added to the matrix to measure aspects which are not captured by the jo b title and which are strongly associated with exposure of a specific occupat ion. For example, job control may be strongly assoc iated with position in the organizational hierarchy, size of the work organization, type of ownership, and degree of work part icipation. To the extent that these factors are important , future surveys may benefit from the collection of add itional informat ion that would provide the basis for a development of future job-exposure matrices.