Professional Documents
Culture Documents
1994,41
Correspondence and requests for reprints should be sent to Murray R. Barrick, De-
partment of Management and Organizations, Universityof Iowa, Iowa City, IA 52242
COPYRIGHT 0 1994 PERSONNEL PSYCHOLOGY. INC.
515
516 PERSONNELPSYCHOLOGY
Method
Sample
the same period for the sales representatives. It should be noted that
performance data for later periods were unavailable to the researchers.
GMA. GMA was assessed by the Wonderlic Personnel Test, Form 5.
Across forms, test-retest reliabilities reported in the test manual range
from .82 to .94. Alternate form reliabilities range from .73 to .95 while
other measures of internal consistency (e.g., a,KR-20) reported range
from .88 to .94 (see WonderlicPersonnel Test Manual, 1983).
Conrcientiousness. Conscientiousness was measured with the PCI
(for a more thorough description of item content and construct validity
evidence for the PCI, see Barrick & Mount, 1993, and Barrick et al.,
1993). Coefficient alpha was .85 for this scale, and in a previous study a
test-retest reliability of .84 was obtained.
Job involvement. Job involvement was measured with the 6-item
version of the job involvement scale developed by Lodahl and Kejner
(1965). Examples of items are: “The major satisfaction in my life comes
from my job” and “The most important things that happen to me involve
my work.” Corrected odd-even reliability coefficients for this scale have
ranged from .72 to .89 (Goodman, Furcon, & Rose, 1969; Lodahl &
Kejner, 1965), and in this study the alpha coefficient was .78. A 5-point
Likert scale anchored by “strongly agree” and “strongly disagree” was
used, with negative items recorded so that a high score on the scale is
indicative of high job involvement.
Job performance ratings. An 11-dimension measure of job perfor-
mance was developed by the researchers based on an analysis of the
sales jobs. The dimensions were job knowledge, quality of work, quan-
tity of work, initiative, customer communications, organizational com-
mitment, job commitment, planning and allocation, interpersonal ori-
entation, self-development7and account management. Each dimension
was defined by a one-sentence description, followed by three interpreta-
tive examples illustrating important facets of that dimension. The sub-
j e c t ~supervisors
~ rated their performance on a 5-point Likert scale from
“Consistently exceeds job requirements” to “Somewhat below job re-
quirements.” Raters were informed that ratings were being collected
for research purposes. Overall performance was the sum of the rankings
across all dimensions. The alpha coefficient was .86. Interrater agree-
ment could not be assessed but is assumed to be comparable to the mean
value (SO) reported in Rothstein (1990).
Sales data. The measure of sales performance was a z-score com-
puted in the following way. The mean sales of the geographic region
over the second half of 1991 was subtracted from the salesperson’s av-
erage monthly sales, divided by the standard deviation of that region. A
2-scorewas used in order to control for the wide variance in sales volume
in the 14 different geographical regions.
524 PERSONNEL PSYCHOLOGY
Results
TABLE 1
IntercorrelationMatrix
Variables
Variable Mean (SD) 1 2 3 4 5 6 7
1.GMA 28.08 (5.90)
2. Conscientiousness 2.61 (0.23) -.08
3. Age 38.00 (9.58) -.07 .15*
4. Gender 0.14 (0.29) .06 -.07 -.19*
%Jobinvolvement 3.53 (0.63) -.19** .29** .16* -.19**
6. Salesvolume 0.00 (0.91) .14* .19**-.04 -.01 .08
7. Supervisor ratings 3.56 (0.69) .18** .25**-.07 .02 -.16* .20**
8. Involuntary turnover 0.299 (0.28) -.14* -.15* .12* -.13* -.16* -.21** -.35**
N=194
*p< .05, all tests one-tailed
**p<.O1, all tests one-tailed
.25, p < .05). These results for conscientiousnesscorrespond to those re-
ported for sales representatives in previous meta-analytic reviews (e.g.,
Barrick & Mount, 1991). In addition, job involvement was significantly
correlated with conscientiousness, GMA, gender, and supervisory rat-
ings of performance (T- ranged from .16 to 29, p < .05). Finally, as
expected, none of the other Big Five personality constructs were signifi-
cantly correlated with involuntary turnover or either of the performance
indices.
We assessed the proposed turnover model by using maximum likeli-
hood estimation as implemented in LISREL VII (Joreskog & Sorbom,
1989). All subsequent analyses were based on the covariance matrix. In
this model, all the variables were treated as single indicators of latent
variables. An alternative approach would have been to operationalize
all variables in the model as multiple indicator variables, although this
would have required a larger sample to derive stable estimates of pa-
rameters. Finally, to control for the effects of measurement errors, the
estimates were corrected for unreliability. The following reliability esti-
mates were used: GMA (.85), conscientiousness (.89), job tenure (.95),
gender (.95), job involvement (.78), sales volume (.67), supervisor rat-
ings ( S O ) , and turnover (.95).l
‘The three “objective” indices (tenure, gender, and turnover) were assumed to be very
reliable indicators, although the .95 estimate allowed for some measurement error. The
performance ratings reliability estimate is based on interrater reliability rather than coef-
ficient alpha because the most important source of error in performance ratings are those
attributable to differences between raters rather than to the content of the measurement
procedure. The average interrater reliability of a single supervisor’srating of .50 is based
on the average mean estimate reported by Rothstein (1990). The sales volume reliability
estimate is based on a sample-weighted reliability estimate for sales measures reported in
a recent meta-analysis for objective measures assessed over 26 weeks (Hunter, Schmidt,
& Judiesch, 1990).
526 PERSONNEL PSYCHOLOGY
TABLE 2
Fit Statistics ofHypothesized and Alternative Turnover Models
Models
Fit statistic Hypothesized Null A C
Xa 10.43 131.16* 8.15 45.41*
(df1 (7) (22) (5) (12)
Difference in x a over hypothesized model NA 120.73’ -2.280 34.98’
Goodness-of-fit index (GFI) ,987 .858 .990 .949
Adjusted GFI (AGFI) .933 .768 .926 .847
Root mean square residual (RMSR) ,079 ,079 .078 ,082
Tucker-Lewis index .91 NA .873 .439
Normed-fit index .92 NA ,938 .654
Note: Model A = the “next most likely” unconstrained model, and Model C = the
“next most likely” constrained model; NA = Not Applicable, the GFI, AGFI, and RSMR
are recommended by Joreskog and Sorbom (1989); the normed-fit index by Bentler and
Bonnett (1980); the ’hcker-Lewis index by Marsh, Balla, and McDonald (1988).
*p<.o1
I I
22’.
significantly linked to sales volume ( p = .08), and tenure was not signif-
icantly associated with job involvement ( p = .Ol), as hypothesized.
For the entire model, the multiple correlation for involuntary turn-
over was .66 (R2= .440). As this suggests, 44% of the variability in
involuntary turnover was accounted for by the relationships posited in
the model. To enhance our understanding of the relations among vari-
ables, the sum of the direct and indirect relations of each antecedent
variable with involuntary turnover was calculated, based on the pattern
of correlations reported in Figure 1. Involuntary turnover was most
highly related to supervisory ratings of job performance (-.56), sales
volume (-.29), conscientiousness (-.33), and job involvement (-.18).
A major portion of the relationship with involuntary turnover for the
two performance variables and job involvement were direct. Two other
variables, tenure and gender, were also primarily related directly with
involuntary turnover (.16 and -.24, respectively). In contrast, GMA
and conscientiousness were associated only indirectly with involuntary
turnover through job involvement and the two performance measures,
and for both variables, the major portion of these relations was through
supervisory ratings (- .13 and - .22, respectively). In sum, the associa-
tion of GMA and conscientiousness with involuntary turnover was in-
direct, while the association of the other five predictor variables with
involuntary turnover primarily was direct. In fact, only two variables
other than GMA and conscientiousness had a substantial indirect rela-
tionship with turnover. They were job involvement and sales volume,
which indirectly affected turnover through supervisory ratings (.20 and
- .14, respectively).
A nonlinear maximum likelihood logistic regression of turnover was
also conducted in this study in order to ensure that the results were not
influenced by the use of a dichotomous dependent variable. Recent ev-
idence (Huselid & Day, 1991) demonstrates that turnover researchers
often adopt inappropriate estimation procedures, particularly ordinary
least squares, which provides biased parameters when the dependent
variable is binary and errors are not normally distributed. In contrast,
probit or logit analysis models have none of these limitations when pre-
dicting binary variables. Furthermore, these nonlinear procedures as-
sume the underlying distribution of the dichotomous variable is a con-
tinuous latent variable, which corresponds to the theoretical assumption
implicit in turnover research (Campion, 1991). Table 3 summarizes the
results of the logistic regression analysis for all the variables included
in the path model. The table contains the regression coefficients (and
t-values) reflecting the change in the logarithmic odds of turnover for
unit changes in the antecedent variables and the model chi-square, which
represents a comparison between the null model that all coefficients are
MURRAY R. BARRICK ET AL. 529
TABLE 3
Summaly of Results of Logistic Regression of Turnover
zero and the estimated model. As shown in Table 2, the model chi-square
was significant (p < .05), indicating that the estimated model represents
a significant improvement over the null model. Furthermore, each of the
variables was significantly related to turnover, and similar to previous
results, the two best predictors of turnover were supervisory ratings of
job performance (p < ,005) and sales volume (p < .02).
In general, results from the two sets of analyses were very similar.
However, the relative magnitude of some of the regression coefficients
did differ. For example, the coefficient for supervisory ratings of perfor-
mance was larger than salesvolume in the path analyses while the reverse
was true for the logistic regression. There are two possible explanations
for this finding. First, these variables are scaled differently; that is, sales
volume is expressed in z-score form, whereas the other variables are not.
Second, the nonlinear nature of the logit analysis may account for these
differences. Nevertheless, taken together, these analyses demonstrate
that the pattern of relations depicted in Figure 1is supported irrespec-
tive of whether zero-order correlations, linear structural equations, or
a nonlinear logistic regression estimation technique was used to analyze
the data.
Discussion
REFERENCES