of about $100,000. A five-variable discriminant function realized disappointing re-
sults, however. Only 64% of the original sample of 36 failed and 36 nonfailed firms
and 54% of the test sample of a like number of firms were correctly classified. He
concluded that the discriminant analysis procedure was not successful. Knight did
combine firms in many different industries, including manufacturing, service, retail,
and construction and this will contribute to estimation problems, especially if the data
are not adjusted to take into consideration industry differences and/or accounting dif-
ferences, for instance, lease capitalization. We discuss this industry effect at length in
the Australian situation.
(b) Altman and Lavallee (1981). The results of Altman and Lavallee (1981) were
more accurate when manufacturing and retailing firms are combined but they do not
advocate a single model for both sectors. Indeed, the holdout tests of this study indi-
cate that nonmanufacturers cannot be confidently measured when the model contains
variables which are industry sensitive.
The Altman and Lavallee (A&L) study was based on a sample of 54 publicly
traded firms, half failed and half continuing entities. The failures took place during
the ten years 1970–1979 and the average tangible asset size of these 27 failures was
$12.6 million at one statement date prior to failure (average lag was 16 months).
Manufacturers and retailer-wholesalers were combined although the data did not en-
able them to adjust assets and liabilities for lease capitalization. The continuing firms
were stratified by industry, size, and data period and had average assets of $15.6 mil-
lion. One can observe, therefore, that the Canadian model for the 1970s decade con-
sisted of firms with asset sizes similar to those of the previously reported U.S. mod-
els (e.g., Altman, 1968) constructed from the 1950s and l960s data period.
A&L examined just 11 ratios, and their resulting model contained five based on a
forward stepwise selection procedure. The model for Canada (ZC) is
where
(c) Classification Results. The overall classification accuracy of the Canadian Z
model on the original 54-firm sample was 83.3%, which is quite high, although not
as impressive as that reported in some of the other economic environments discussed
in this international review article. Practically speaking, classification criteria are
based on a zero cutoff score with positive scores indicating a nonfailed classification
and negative scores a failed assignment. Reliability, or holdout tests, included
Lachenbruch (1967) test replications, the original sample broken into randomly cho-
sen classification and test samples, and testing the model on prior years data, for ex-
ample years 2 through 4 before failure. The Lachenbruch and replication holdout re-
sults showed accuracies very similar to those of the original sample results and the
prior year accuracies were 73% (Year 2), 53% (Year 3), and only 30% (Year 4).
X 5 rate of growth of equityrate of asset growthX 4 net profits after tax>total debtX 3 current assets>current liabilitiesX 2 total debt>total assetsX 1 sales>total assetsZCCanadian Z-scoreZC1.6260.234 1 X 12 0.531 1 X 22 1.002 1 X 32 0.972 1 X 42 0.612 1 X 5210 • 16 BUSINESS FAILURE CLASSIFICATION MODELS