THE JOURNAL OF FINANCE * VOL XXXIV, NO.

3 * JUNE 1979

Determinants of Financial Structure: a New

Methodological Approach
MICHAEL G. FERRI and WESLEY H. JONES*

I. Introduction

THE ASSOCIATION BETWEEN A firm's financial structure and its operating char-
acteristics-its industrialclassificationand its size, among others-took on added
importance as a result of the debate, started by Modigliani and Miller [8],
regardingcost of capital and optimal structure. Impressive evidence that indus-
trial class influences financial structure has been marshalled by Scott [12] and
Scott and Martin [13]. Support for the view that size might shape a firm's debt-
equity mix may be found in Scott and Martin [13]. However, dissenting evidence
has been presented,most notably by Remmers,Stonehill, Wrightand Beekhuisen
[11] who argue that neither size nor industryis clearly a determinantof the firm's
use of debt.
The aim of this paper is to investigate the relationships between a firm's
financial structure and its industrial class, size, variability of income, and oper-
ating leverage. The methodology used in this paper is new to this area of inquiry
and promises superiorresults, because it avoids several measurement difficulties
encounteredin previous work. The resolution of these difficultiesoccurs through
the development, within this paper, of a taxonomy of firms that is based on the
firms' actual financial behavior. The taxonomicalstructurewill provide the basis
for an examination of associations between financial structure and industrial
class, size, variability of income, and operatingleverage.

II. Hypotheses
This paper's investigation of the determinants of financial structure will be
performedthrough the examinationof four specific hypotheses. The first hypoth-
esis is concerned with the relationship between industrial classification and
financial leverage. The view that industry class has an impact on financial
leverage has intuitive appeal. Firms in the same industry class should experience
similar amounts of business risk, because these firms produce similar products,
face similar costs for materials and skilled labor, and rely on similar technology.
Business risk-the uncertainty of future income streams-should substantially
determine the amount of debt the capital markets will provide. The markets set
interest rates and maximal debt loads by reference to the volatility of a firm's
income stream. Because this volatility should be related to the products of the
firm, there is reason to believe that financial structure is molded by a firm's
industrial classification. The evidence, cited earlier, that industry type does not
* Assistant Professorsof Business Administration,Universityof KentuckyLexington,Kentucky.

631
632 The Journal of Finance
influence financial structure thus runs contrary to intuition and deserves addi-
tional analysis.
The second hypothesis concerns the connection between size of the firm and
leverage, which has gained some attention in recent years [11, 13]. The rationale
for the belief that the firm's size is influential with respect to financial structure
lies in the evidence that larger firms may be more diversified [11], enjoy easier
access to the capital markets, receive higher credit ratings for their debt issues,
and pay lower interest rates on borrowedfunds [10]. Thus it is plausible that the
size of the firm is positively related to its use of debt capital, and this is the second
hypothesis that will be tested in this paper.
The third hypothesis concerns the impact of business risk (measureddirectly
and not through the proxy of industrialclass) on the firm's use of fixed payment
funding. The variability of the firm's future income is the chief factor in ex ante
estimates of its ability to meet fixed charges.As a result, one may anticipate that
business risk is negatively correlated with the percentage of debt in a firm's
financial structure. This expected negative relationshipwill be the third hypoth-
esis subjected to empirical examinationin this paper.
The fourth and final hypothesis pertainsto the suspected influenceof operating
leverage upon financial structure. Operatingleverage may be defined as the use
of fixed costs in the firm'sproductionscheme, but is generallyassociated with the
employment of fixed assets. The use of fixed assets can magnify the variabilityin
the firm's future income. Thus, this type of leverage should be negatively related
to the percentage of debt in the firm's financial structure; and this negative
relationship is the focus of the fourth hypothesis.

III. Data and variables

The data used in the present investigation were gathered from the Compustat
data tapes which contain the year-end balance sheet and income statement for
industrial, domestically headquartered,unregulated firms. The total sample of
firms selected (for consistency of fiscal year and availabilityof data) amounted to
233 firms. The distributionof selected firms by StandardIndustrialClassification
Code (SIC Code) and by generic industry groupingsis given in the appendix.
Data on the sample firms was gathered for two five year time spans:from 1969
to 1974and from 1971to 1976.Multiperiodvariables (averagesales, coefficient of
variation in operatingincome, etc.) are calculated on the basis of data from each
year in the five year spans. Single period variables (debt to total assets, current
sales, etc.) are computed on the basis of data from the terminal year in the two
time spans. The terminal years of these two spans were very differentin terms of
general business conditions, a point to be discussed later.

Dependent Variable
Financial structure (or financial leverage) was measured by the ratio of total
debt to total assets (D/TA), the leverage measure employed by Remmers et alia
[11]. Other measures of financial structure have been used. Scott [12] adopted
the percentageof common equity in the financialstructure,and Scott and Martin
 Financial Structure 633
[13] employed the ratio of common equity (inclusive of common stock at par
value plus surplus retained earnings and surplus reserves) to total assets at book
value as their indicators of leverage. In the present study it was decided to use
the ratio of total debt to total assets at book value (D/TA) for reasons of
conceptual simplicity and the variable's ability to more completely reflect a firm's
total reliance on borrowed funds.

Industry
Two different measures of industry type were used in tests of the first hypoth-
esis. The first of these is the conventional four digit SIC code (a measure defined
by a firm's dominant product line) as provided in the Compustat tapes. The
second indicator of industry type is a generic industry measure defined by
grouping firms with similar dominant product lines and SIC codes. The adopted
groupings are shown in the appendix.

Size
Four different measures of firm size were employed in tests of the second
hypothesis. Two of these are standard and have been used in previous research.
Remmers et alia [11] used total sales (SA) to represent the size of the firm, while
Scott and Martin [13] employed total assets at book value (TA). In addition to
these variables, this paper will use two long-term measures of size: a) the average
level of total assets, tt(TA), over the current and preceeding four years; and, b)
the average level of sales, tt(SA), over the same time interval. It was felt that
these average measures might give a truer indication of firm size than single
period measures.

Business Risk
Business risk-the expected variability in future income-will be represented
by a set of variables that measure the historical volatility in a firm's sales revenues
and in its pre-tax cash flows. Sales and cash flow were selected as the pertinent
revenues because of their obvious importance. Sales constitute the basis and
source of the bulk of a firm's income, and cash flow is that particular revenue
stream from which interest payments are met. The variables used in the present
study to measure business risk are: a) the coefficient of variation in sales, a(SA)/
tt(SA); b) the coefficient of variation in pre-tax cash flow, a(CF)/ti(CF); c) the
standard deviation of the standardized growth in sales, a(SGSA); and, d) the
standard deviation of the standardized growth in cash flow, a(SGCF).
As in the case of the size variables, the coefficients of variation were calculated
with data from the current and preceeding four years. The two standardized
growth variables were calculated with data from the current and preceeding five
years. The standard deviation of the standardized growth in X (whether sales or
cash flow) is the standard deviation of the dollar change in X divided by the
mean of X, or a(SGX) = a[Xt - Xt-_)/tt(X)], where t = 1, 2, .*,5. Each of the four
measures of business risk adjusts variability for the size of the firm. These
adjustments to the business risk measures were made in order to decouple
volatility from the sheer effects of size.
634 The Journal of Finance
Operating Leverage
A traditional measure of operating leverage is the ratio of the percentage
change in earnings to the percentage change in sales. The application of this sales
elasticity measure rests on important certeris paribus assumptions and on hy-
pothetical intra-period comparisons of different levels of sales. One asks, in effect,
how earnings would change if sales were to change, all else being the same.
However, historical data often violate this external consistency requirement.
From one period to the next, a firm may have different levels of price per unit,
fixed costs, total assets and other important variables. Furthermore, available
longitudinal data bases force one to compute change across rather than within
periods. Thus, it is not possible to operationalize a pure measure of operating
leverage that is consistent with the conceptual definition of sales elasticity of
earnings. Several proxies are available. The first, and most important, is simply
the percentage change of earnings before interest and taxes, as a proportion of
percentage change in sales. This measure (to be called the degree of operating
leverage (DOL)) was computed for each firm as DOLt = [(Et -Et_1)1Et_11[(SAt
- SAt-1)/SAt-1], t = 1974 and 1976.
Two other measures, based on balance sheet data and to some extent flawed,
are also available and will be employed in the tests that follow. They are: a) the
ratio of net fixed assets to total book assets, (FA/TA); and b) the ratio of the
average of net fixed assets in the current and preceeding four years to the average
of total book assets over the same time period, g(FA)/,t(TA). Again, it was felt
that the average may be a better indicant of the firm's use of fixed assets than the
single period equivalent.

IV. METHOD
The classification methodology employed in this paper is quite different from
that used in earlier efforts to investigate the linkage between financial structure
and the characteristics of the firm. This paper's method can, however, be most
easily explained by a comparison with methods used in earlier studies. In those
studies (particularly [11], [12], and [13]) firms were divided into industry classes
according to their SIC codes. As is necessary, the choice of SIC codes was
designed to include truly distinct industrial classes. The means and variances of
selected measures of financial leverage were then evaluated to determine the
im5act that class differences exert on the degree of financial leverage.
Several problems with this approach are evident. First, despite appearances,
SIC codes are not isomorphic with respect to the industries in which coded firms
participate. Many firms are multiproduct enterprises. The reflection of this
diversification within firms eludes the simple categorizations of the SIC code
system. The lack of isomorphism not only confounds the examination of firms in
a single SIC code, but also (to the extent that diversification of product lines
varies across SIC code categories) confounds examinations based on comparisons
among multiple SIC code classes.
A second difficulty relates to the grouping of firms into generic (multiple SIC)
industry classes. While no specific objection can be made to the particular generic
 Financial Structure 635

groupings employed in earlier studies, they are subject to a certain degree of

arbitrariness. At times, firms with quite similar SIC codes have been excluded
from generic classes. The impact of these exclusions cannot be assessed and the
theoretical justification of such groupings, and associated findings, is open to
some question. And, third, since there is little certainty that chosen industry
classes do, indeed, span the range of distinct financial structures the generaliza-
bility of reported findings is in some doubt.
Alternatives to SIC code based groupings of firms are taxonomies based on a
partitioning of firms, ranked by some structural measure (e.g. leverage) into
equally sized structural categories, e.g., the approach of Remmers et al., [11].
However, such an approach yields groupings that are highly dependent on the
particular set of firms selected. In the general situation where leverage is not
distributed uniformly across the sample of firms (i.e., where leverage has one or
more modes) decile or other rank based assignments can cause firms with similar
leverage to be classified into different groups, and highly dissimilar firms to be
grouped together. This classification problem is likely to inflate the error terms
in subsequent analyses and reduce the power of associated statistici1 tests.
The approach employed in this paper avoids the methodological difficulties
discussed above. This paper does not assume that taxononies based on dominant
product lines (SIC codes and generic industry groupings) necessarily represent
sets of firms that are distinct in terms of leverage. In fact, this relationship will be
here subjected to empirical tests. Nor does this paper assume that leverage sets
are adequately characterized by arbitrary assignments based on firms' relative
leverage rankings in a sample. Rather, the current investigation employes an
algorithm (to be described below) that developes a taxonomy of firms according
to the similarity of their financial leverage. In other words, this algorithm assigns
each firm to one of a set of "leverage classes" on the basis of the firm's use of
debt. Thus, the similarity of firms' actual financial behavior, not judgements
concerning dominant product lines nor idiosyncratic rankings in a sample, estab-
lishes the grouping of firms. The taxonomy of firms developed by the algorithm
forms the basis of subsequent analysis, facilitating the investigation of associations
between attributes of firms and leverage classes.

Howard-Harris Algorithm
The leverage classes were derived by use of the Howard-Harris clustering
algorithm [2, 7], which created a taxonomy of firms based on measures of their
financial structure. Formally, this algorithm partitions a set of objects (X1, X2,
*--, X,4XiES), where each object is characterized by a multivariate vector of q
attributes, such that a p-fold partitioning P(S, p) of disjoint subsets, or groups, is
obtained. The algorithm partitions hierarchically in a 'step-down' manner (i.e., in
2-way, 3-way, * **, p-way order) such that at each step groups are derived that
are as internally similar and as externally dissimilar as possible. Thus, each of the
resulting leverage groups derived in the present application of the algorithm are
as distinct as computationally possible.
In the Howard-Harris algorithm dissimilarity is defined as the q-dimensional
squared Euclidean distance between objects. The implications of this choice of
636 The Journal of Finance
Table 1
Mean Debt to Total Assets by Leverage Groups
Leverage Groups
A/W
Low 2 3 4 5 High Statistic
1976
D/TA 0.198 0.310 0.409 0.501 0.593 0.717 800.00
Number 14 42 74 54 28 21
of firms
1974
D/TA 0.149 0.272 0.363 0.465 0.586 0.744 673.14
Number 10 23 51 78 55 16
of firms

dissimilarity measure are several. Most germain to the present research is the
fact that the resultant partitioning maximizes the ratio of variance among groups
to variance within groups. Since the firms were grouped according to their ratios
of debt to total assets (D/TA), the groups derived by the algorithm maximized
(in terms of analysis of variance) the 'explained sum of squares'.' As in analysis
of variance, the explained sum of squares can be interpreted as a measure of
group differences.
Various criteria have been used with hierarchical clustering algorithms in order
to determine the number of distinct groups that exist within some set of objects.
The criterion adopted in this research was the location of the point at which 82X/
Sg2 = 0; where, g = the number of groups, A = the ratio of the determinant of the
pooled within sums of squares matrix to the determinant of the total sums of
squares matrix. The application of this criterion rests on the notion that an initial
increase in the number of partitions brings about a decrease in the relative
amount of within group variance primarily because of the recognition of true
group structure differences. Additional partitioning yields, at some point, a
reduction of within group variance only because of the artifactual partitioning of
error variance. Since the later reduction is a function, principally, of the number
of partitions (recognized groups), the statistic &\/Sg approaches a constant.
Further decreases in within group error (under constant SX/8g) do not indicate
discovery of additional true group differences.

V. RESULTS

Before proceeding to a discussion of the substantive hypotheses regarding firm

characteristics and financial leverage, it is necessary to analyze the validity of the
conceptual device-the clustering of firms by their use of debt-that will underlie
the tests of the hypotheses. In this connection two aspects of the Howard-Harris
clustering technique's results are important. The first is the existence and stability
of the leverage classes produced by the clustering algorithm. As was pointed out

' In the present analysis each firm (object) was characterized by a single variable DITA; thus
potential local optimality difficulties related to correlated variables in the partitioning criterion were
not encountered.
 Financial Structure 637
Table 2
1974-1976 Leverage Group Contingency Table
1974 1976 Leverage Groups
Leverage
Group Low 2 3 4 5 High Total
Low 5 2 2 1 0 0 10
2 7 12 3 0 1 0 23
3 1 21 24 4 0 1 51
4 1 7 41 25 3 1 78
5 0 0 3 22 20 10 55
High 0 0 1 2 4 9 16
Total 14 42 74 54 28 21 233

in the section on method, various criteria can be employed in an effort to

determine the number of distinct groups. The lambda criterion employed here
suggests that there are six distinct debt groups in the 1974 and 1976 measures of
D/TA. The existence of six distinct leverage classes receives further support in
the form of a statistic, similar to the conventional F-statistic of analysis of
variance, which relates the sum of squares among groups to the sum of squares
within groups produced by the cluster algorithm.2 This statistic, which will be
labelled A/W is reported in Table 1, along with the mean values for the ratio of
debt to total assets for firms in each of the six groups. The high values of the A/
W ratio and the distinct kinks in the lambda curves strongly suggest the presence
of distinct leverage classes.
A second issue related to the use of the clustering technique and its resulting
taxonomy is the stability of class membership across time and differing financial
environments. Visual evidence of stability is available in Table 2, which is a table
relating the class of firms in 1976 to their class as determined from 1974 data.
Table 2 should be interpreted in this way: 5 of the 10 firms classed in group 1
in 1974 remained in that group in 1976; 2 of those assigned to group 1 in 1974
moved to group 3 in 1976; and 22 of the 55 members of leverage class 5 in 1974
were in group 4 in 1976. The general impression conveyed by Table 2 is that
there was some movement of firms from their 1974 leverage class, but that this
movement is not large, representing as a rule a shift of only one leverage class.3
It may be concluded that distinct leverage classes are apparent in each of the
two years; and, importantly, the classes identified by the clustering algorithm are
substantially stable through time.

Industry
The leverage classes will be used now in the evaluation of the four substantive
hypotheses that were discussed in an earlier section. The first of these is the
2 This statistic is presented for descriptive purposes only; it is not appropriate for use in classical
hypothesis tests of differences between clustered groups (see [6]).
'The stability of leverage class membership from 1974 to 1976 is summarized by Kendall's tau c
statistic (a measure of association between ordinal scales) which is analogous to Pearson's product
moment correlation coefficient. Tau c was found to be 0.64 (significant at less than the 0.01 level)
indicating that the ordinal leverage class taxonomies are highly congruent across the two years
investigated.
638 The Journal of Finance
Table 3
Leverage Groups by Generic Industries (1976 and 1974)
Generic Industries
Leverage
Groups 1 2 3 4 5 6 7 8 9 10
1976 Results
Low 1 6 2 1 0 1 0 0 2 1
2 3 4 8 5 2 5 3 5 4 3
3 4 7 9 12 16 1 7 9 7 2
4 1 6 5 4 4 7 4 8 8 7
5 0 6 2 2 0 8 0 3 4 3
High 1 4 4 0 0 1 1 4 3 3
1974 Results
Low 2 3 0 2 0 1 0 1 0 1
2 3 4 5 1 2 3 1 1 1 2
3 3 8 7 8 7 2 4 6 3 3
4 1 7 10 9 11 7 7 9 12 5
5 1 7 6 3 2 9 3 9 8 7
High 0 4 2 1 0 1 0 3 4 1
Totals 10 33 30 24 22 23 15 29 28 19
Note: Generic Industries; 1) mining-metal, 2) oil-gas, 3) textiles-apparel, 4) lumber, 5) chemicals, 6)
rubber-plastics, 7) glass 8) machinery, 9) electronics, 10) retail food.
X2= 72.73 (1976), d.f. = 45, a < .01
= 52.45 (1974), d.f. = 45, a > .20

much analyzed relationship between financial leverage and industrial classifica-

tion, which is the conventional typology based on the principal product of the
firm. This issue was addressed by cross-tabulating financial leverage classes, from
both 1974 and 1976 clustering, with the firms' SIC code and with its generic
industry classification (this grouping of firms into generic industries is described
in the appendix). The relationship between financial leverage class and both the
generic industry group and individual SIC codes was examined from an infor-
mation theoretic perspective. The asymmetric uncertainty coefficient [5] was
used to indicate the degree to which knowledge of the independent variable (SIC
code or industry class) reduces error in predicting the value of the dependent
variable (leverage class). The findings of this analysis with respect to generic
industry class are given in Table 3.
-The data in Table 3 suggest that there is a slight statistical relationship between
relative debt structure class and generic industry class. A test of the hypothesis
that the relationship between leverage class and generic industry is perfectly
random was rejected, at the 0.05 level, for 1976, but could not be rejected in 1974.
It should be emphasized, that rejecting a single null hypothesis of randomness is
far from demonstrating an isomorphic relationship. As Table 3 shows, the
members of every industrial classification are widely dispersed across leverage
classes in both 1976 and 1974. While industrial classes 4, 5, 6, and 9 (lumber,
chemicals, rubber and electronics) seem to be more concentrated in a set of
leverage classes than any other industrial grouping, the degree of concentration
is not large. The dispersion of the industrial class members across the leverage
classes can be further documented by the value of the asymetric uncertainty
 Financial Structure 639
coefficient (a statistic bounded by 0 and 1) which took on values of 0.076 in 1974
and 0.082 in 1976.These values indicate that knowledgeof industry class reduces
error in predictingleverage class by only 7.6%in 1974 and 8.2%in 1976,relative
to errors that would be made by simply predicting the most common (modal)
debt class for all firms in the sample.
A final point regardingTable 3 is worth noting. It seems that some industrial
groups are less likely to vary with respect to leverage classes across time than
others. For example, industry 2 (oil/gas) has virtually the same distribution
across the six leverage classes in 1974 as it has in 1976. On the other hand,
industry 9 (electronics) shows great variation in the distributionacross leverage
classes in the two years, and firms of industry 8 (machinery)and of industry 4
(lumber)shift easily from one leverage class to another.
An analysis of the relationship between leverage class membership and SIC
code designation was performedin a manner similar to that used for industry
classification.The twenty-five SIC codes were cross-tabulatedwith leverage class
for both 1974and 1976.The bulk involved in displayinga 6 x 25 cross-tabulation
table precludes presenting it here. The results can be simply reported, however.
Since it was not possible to satisfy the conditionthat expected cell frequenciesbe
greater than or equal to five, a valid, comparable,test of the random association
hypothesis could not be performed.4However, the asymetric uncertainty coeffi-
cient could be calculated,and was found to take on values of 0.15 (1974)and 0.20
(1976). These statistics indicate that leverage class can be better predicted with
SIC codes than with generic industry groupings. This result, however, is not
surprising,because there were nearly three times as many distinct SIC codes as
generic industry groups in the predictor variable sets. Thus the three-fold im-
provement in prediction performance can be explained without recourse to
theoretically substantive arguments.

Size
The second, third, and fourth hypotheses concern the relationships between
leverage classes and firm size, business risk, and operatingleverage, respectively.
In each case, the hypotheses were tested by means of multivariate discriminant
analysis [14]. A summary of the results of these tests is displayed in Table 4.
Insignificantdiscriminatorshave been omitted from Table 4.
The first of the discriminantanalyses tested the null hypothesis of no differ-
ences across leverage groups with respect to the following measures of size:
current and average sales, and current and average total assets. In 1976 the
historical averages of sales and total assets, as well as current sales, were found
to differ across the leverage groups. The most efficient discriminant function
involved, however, only the historical averages of sales and total assets; the high
degree of correlation between current and historical sales lead to the former
measure's exclusion from the discriminantfunctions. Wilks' lambda (a measure
of group differences[8]) was found to be 0.90 in 1976;the associated multivariate
F statistic (F = 2.42, d.f. = 10,452) was significant at less than the 0.01 alpha
level. In 1974 a more complicated pattern was found to emerge. None of the
4 The computed values for x2were found to be 143.40 (1976) and 140.96 (1974), d.f. = 120, a > .07.
640 The Journal of Finance
Table 4
Mean Characteristics by Leverage Groups (1976 and 1974)
LeverageGroups Univariate
F
Characteristics Low 2 3 4 5 High (d.f. = 5,227)
1976 Results
Size
SA 374.03 559.54 951.64 1137.35 448.61 167.79 2.61a
A(SA) 285.57 454.92 765.63 922.29 350.91 127.40 2.76a
A(TA) 360.11 422.54 712.07 595.85 246.85 86.49 2.66a
Business Risk
u(CF)/A(CF) -9.58 0.37 0.31 0.42 0.04 0.49 3.30a
ur(SGCF) 20.60 0.41 0.37 0.59 0.55 0.69 3.18a
Operating Leverage
A(FA)IA(TA) 0.46 0.43 0.44 0.37 0.38 0.34 2.13b
DOL 3.23 3.35 9.98 5.05 -9.63 -74.10 2.09b

1974 Results
Size
SA 137.04 599.51 651.79 835.71 646.56 138.19 1.27
A(SA) 104.33 405.08 478.88 640.65 449.35 135.60 1.50
A(TA) 130.99 447.97 502.58 466.08 379.22 81.39 1.30
Business Risk
u(CF)/A(CF) 0.38 0.26 0.44 0.36 1.32 0.50 1.15
ur(SGCF) 0.41 0.56 0.55 0.37 2.60 1.16 1.00
Operating Leverage
A(FA)/j(TA) 0.53 0.43 0.46 0.41 0.35 0.37 3.66a
DOL 2.49 34.97 0.23 -0.18 0.87 26.09 1.59
Note: only variables which contributed to significant discriminant functions have been reported.
ap < .05
bp < .10

univariate tests of size differences were found to be significant. However, linear

combinations of current and historical sales, together with historical total asset
average, were derived that significantly discriminated among the leverage groups.
The significant tri-variate discriminant functions for 1974 yielded a Wilks' lambda
of 0.90, and the associated multivariate F statistics (F = 1.64, d.f. = 15,622) was
significant at the 0.06 level.
The significant size effects noted in both 1974 and 1976 indicate that size can
account for differences in financial structure. Nonetheless, the relationship is not
straightforward. In 1974, a recessionary period, a complex combination of current
ind historical performance (i.e. levels of sales and assets) is required to predict
differences in leverage. In 1976, an expansionary period, a firm's ability to aquire
debt could be predicted on the basis of either its historical or its current
performance. This difference between the recessionary and expansionary periods
is consistent with the results of other studies [1, 4]. These studies have indicated
that in expansions even marginal firms have ready access to debt capital, while
in recessions the established firms that have both a record of past success and
relatively good current performance obtain a large percentage of the new debt.

Business Risk
Similar analyses were performed, for both 1976 and 1974, using measures of
business risk: the standard deviation of standardized growth in sales, the standard
 Financial Structure 641
deviation of standardized growth in pre-tax cash flow, the coefficient of variation
in sales and the coefficient of variation in cash flow. The null hypothesis of no
difference in risk among the different leverage groups was rejected in 1976.
Univariate tests of the coefficient of variation and the standard deviation of
standardized growth in pre-tax cash flow, and a multivariate test of the discrim-
inant functions involving these two variables, were significant. The tests suggest
a linkage between income volatility and leverage class. However, the noted
differences are principally due to differences between the lowest debt to total
asset group (leverage group 1) and the other five groups. A closer examination of
group 1 indicated that the standard deviation of the risk measures within this
group were nearly ten times as large as the other five groups; the heterogeneity
of within group variance violates a necessary condition for valid discriminant
tests, suggesting that the null hypothesis in this case may have been falsely
rejected.
Leverage class 1 is dominated by firms in the extractive industries (as shown in
Table 3), especially those firms which provide fossil fuels. In the 1970's, partly in
response to the perceived difficulties of energy supply, a new wave of legislation
affecting these firms was passed. It is likely that these laws, which changed
depreciation, depletion, and valuation standards of energy related firms, had a
radical impact on reported income figures. Furthermore, any such mid-period
change would substantially inflate the variance in the present longitudinal mea-
sures of business risk.
In 1974 the leverage classes were not found to differ significantly on any of the
measures of risk. Wilks' lambda of 0.96 in 1974 (F = 1.68, a > 0.10) is less
significant than the value obtained in 1976, 0.87 (F = 3.26, a < 0.01). Thus the
statistical tests in 1974 do not support the suspected relationship between income
variation and a firm's use of debt.

Operating Leverage
The hypothesized relationship between operating leverage and financial struc-
ture was tested by use of three variables. The first, an income statement variable,
is the percentage change in EBIT to the percentage change in sales, across the
years 1973 to 1974 and 1975 to 1976. The remaining variables are based on balance
sheet data: fixed assets to total assets in the years of 1974 and 1976 and the ratio
of average fixed assets to average total assets for the five year intervals of 1970 to
1974 and 1972 to 1976. The null hypothesis of no connection between operating
leverage and financial structure was rejected for both 1976 and 1974. In each case
the ratio of average fixed assets to average total assets (over the current and
preceeding five years) varied more significantly over the leverage groups than
other measures of operating leverage. In 1976, the sales elasticity of earnings (a
traditional measure of the impact of operating leverage) was found to vary
substantially among the six leverage classes. This variable, however, was not
related to differences in debt load classes in 1974, a time when income statements
of many firms were in a deteriorating position. It should be noted (and will be
discussed later) that the relationship between operating leverage, measured by
balance sheet data, and financial leverage is nearly a perfectly negative monotonic
relationship.
 642 The Journal of Finance
VI. Discussion of hypotheses
The first hypothesis of this paper concerns the association between a firm's
financial structure and its industrial class. This association, which is theoretically
and intuitively plausible, has received empirical support in previous studies [11,
13]. The current investigation offers additional evidence in favor of the relation-
ship. Table 3 and its surrounding discussion indicate that a firm's leverage class-
its position in the array of firms ranked by debt use-is linked to the firm's
industrial classification. It was possible to reject the hypothesis of a random
association between leverage class and type of industry.
Nonetheless, the results of the analysis suggest that the relationship is neither
strong nor straightforward. Table 3 reveals four aspects of the association between
leverage class and industry. First, members of every industry group are dispersed
across several leverage classes. Second, the degree of dispersion is quite large in
some industries (e.g. industries 1, 2, 10). Third, dispersion occurs both in time of
recession (1974) and in a period of expansion (1976), indicating that dispersion is
not peculiar to a particular set of financial conditions. Fourth, the leverage class
characteristic of an industry (its modal leverage class) may shift sharply (indus-
tries 4 and 8 are the most dramatic examples of such a shift). Thus, the only
justifiable conclusion regarding the first hypothesis must be that, although
industry and financial structure are not totally independent of each other, the
dependence is, at best, modest and indirect.
The tests in this paper support the relationship, noted in Scott and Martin
[13], between a firm's size and its use of financial leverage, the focus of the second
hypothesis. The relationship was noted in Table 4 where certain measures of size
are shown to be associated with leverage class. The measure of size that best
discriminated between leverage class is the average level of sales, ,u(SA), a variable
that has not previously been used to represent size. The conventional measures
of size, total assets (TA) and total sales (SA), are inferior to average total assets
and average sales as discriminators among leverage groups.
It should be noted that the association between size and financial leverage is
not as simple as previous work might seem to suggest. Table 4 reveals a nearly
curvilinear relationship between size and leverage class. Small firms (those with
low levels of average sales) are clustered in classes of high and low leverage, and
the larger firms seem to be concentrated in the classes of intermediate leverage.
At the very least, this fact precludes the monotonically positive association
between size and leverage that has previously been posited [13].
Although this paper found support for the anticipated relationships between
financial leverage and measures of firm size and operating leverage, the statistical
tests did not provide evidence for the posited relationship between financial
leverage and business risk. Business risk was proxied by four measures of the
historical volatility in sales and cash flow. These measures have intuitive appeal
and should adequately reflect the expected variability in a firm's future income.
Despite the apparent appropriateness of the variables, significant differences in
business risk measures were not found among the debt classes in the present
samnple.
Finally, the results strongly confirm the expected negative relationship between
operating leverage and a firm's use of debt. As Table 4 shows for both 1976 and
 Financial Structure 643

1974,firms with a high proportionof fixed to total assets are concentratedin the
low leverage classes. The ratio of average fixed assets to average total assets,
IL(FA)/fL(TA), generallyfalls fromthe lower leverageclasses to the higherleverage
classes. This result is in agreement with the anticipatedrelationshipbetween the
two types of leverage. The measure of degree of operatingleverage, DOL, whose
limitations were discussed at length earlier, varied erratically across the six
leverage groups. The contrast between the consistent trend noted with the
balance sheet measure and the erratic performanceof DOL points to operational
limitations of this conceptually elegant measure.

VII. Summary and concluding remarks

This paper investigated possible linkages between a firm'sfinancialstructureand
its industry class, size, variability in income, and operating leverage. The study
developed a taxonomy of firms that avoided methodological and conceptual
difficultiesassociated with schemes based solely on SIC codes or on firms'relative
rankingsin a sample. The results of the study's effort to relate firm characteristics
to leverage class can be summarizedin this way: a) industry class is linked to a
firm's leverage, but in a less pronounced and direct manner than has been
previously suggested; b) a firm's use of debt is related to its size, but the
relationship does not conform to the positive, linear scheme that has been
indicated in other work;c) variation in income, measured in several ways, could
not be shown to be associated with a firm's leverage; and d) operating leverage
does influence the percentage of debt in a firm's financial structure and the
relationship between these two types of leverage is quite similar to the negative,
linear form which financial theory would suggest.

APPENDIX
Description of Sample
Component Firms in Sample
Compustat
IndustryClass Codes Components Total
Mining-Metal 1000,1021,1031 7, 2, 1 10
Oil-Gas 1311,1381, 1382 26, 5, 2 33
Textiles-Apparel 2200, 2300 13, 17,- 30
Lumber 2400,2600 8, 16,- 24
Chemicals 2800,2810,2820 15, 4, 3 22
Rubber-Plastics 3000,3069,3079 12, 4, 7 23
Glass 3210,3221 5, 10,- 15
Machinery 3550,3560 13, 16,- 29
Electronics 3662, 3670,3679 8, 7, 13 28
Retail Food 5411,5812 10, 9,- 19
25 distinct codes Total Firms 233

REFERENCES
1. T. R. Atkinson. Trends in CorporateBond Quality (New York:National Bureau of Economic
Research, 1967).
644 The Journal of Finance
2. J. D. Carroll."Modelsand Algorithmsfor MultidimensionalScaling ConjointMeasurementand
Related Techniques."MultiattributeDecisions in Marketing,A MeasurementApproach, by
Paul E. Greenand YoramWind. (Hinsdale,Illinois:Dryden Press, 1973).
3. R. A. Eisenbeis. "Pitfalls in the Applicationof DiscriminantAnalysis in Business, Finance and
Economics."Journal of Finance, Volume32 (June 1977).
4. M. G. Ferri and C. G. Martin. "The Cyclical Pattern in the Quality of New CorporateBonds."
Journal of Portfolio Management (forthcoming).
5. L. A. Goodmanand W. H. Kruskal."Measuresof Associationfor Cross Classification."Journal
of the American Statistical Association, Volume49 (December1954).
6. P. E. Green and D. S. Tull. Research for Marketing Decisions, Fourth Edition. (Englewood
Cliffs:Prentice-Hall,1978).
7. N. Howardand B. Harris."A HierarchicalGroupingRouting,IBM 360/65 FortranIV Program."
Universityof PennsylvaniaComputerCenter, 1966.
8. F. Modiglianiand M. H. Miller. "The Cost of Capital,CorporationFinance, and the Theory of
Investment."American Economic Review, Volume48 (June 1958).
9. D. F. Morrison.Multivariate Statistical Methods (New York:McGraw-Hill,1967).
10. G. E. Pinches and K. A. Mingo. "A MultivariateAnalysis of IndustrialBond Ratings."Journal
of Finance, Volume 28 (March1973).
11. L. Remmers, A. Stonehill, R. Wright and T. Beekhuisen. "Industryand Size as Debt Ratio
Determinants in ManufacturingInternationally."Financial Management, Volume 3 (Spring
1975).
12. D. F. Scott, Jr. "Evidenceon the Importanceof FinancialStructure."Financial Management,
Volume 1 (Summer1972).
13. D. F. Scott, Jr. and J. D. Martin. "Industry Influence on Financial Structure." Financial
Management,Volume4 (Spring1976).
14. M. M. Tatsuoka. Multivariate Analysis: Techniques for Educational and Psychological Re-
search (New York:John Wiley and Sons, 1971).