SMALL BUSINESS
INDICATORS OF MACROECONOMIC ACTIVITY
William C. Dunkelberg
Chief Economist, National Federation of Independent
Business
Professor of Economics,
Jonathan A. Scott
Professor of Finance,
William J. Dennis
National Federation of Independent Business
September, 2003
The
National Federation of Independent Business (NFIB) began quarterly economic
surveys of its membership (about 350,000 firms) in 1973.[1] Since that time, a virtually identical
three-page questionnaire has been mailed to a sample of NFIB’s small-business
owner members on a regular basis. A copy
of the current questionnaire is included in Appendix 2. From October of 1973 through 1985, a random
sample of the NFIB membership list was selected and survey form was mailed them
on the first day of every quarter. This
mailing was followed by a second about 10 days later. Since January 1986, the same procedure has
been followed monthly. Responses are
collected for about 25 days after which duplicate responses are
eliminated. The yield is 1,300 to 1,800
responses in the first month of each quarter and 500 to 700 responses in the
following two months. A monthly
report, Small Business Economic Trends, based on the findings from the survey is available
from NFIB in both electronic (nfib.com/research) and printed forms. The data are also presented in models and
used to forecast selected macroeconomic phenomena. The objective of this paper is to assess how
well the NFIB quarterly economic survey data anticipate changes in highly
visible measures of macroeconomic activity.
Because
small firms make up such a large fraction of the total economy, it is logical
to look to indicators of their collective economic health as reliable
indicators of the entire economy’s performance.[2] This argument is reinforced by the notion
that the same basic economic forces impact all firms, large or small. Federal Reserve policy, tax based fiscal
policy, shifts in consumer spending, for example, all affect businesses of
every size. Therefore, owners of small firms should experience the same
economic forces that are experienced by the managers of large firms.[3] The sectoral composition of “large” and
“small” firms does differ, with
manufacturing dominated by larger firms and construction and many services
dominated by small firms. Large firms
are heavily involved in international trade while small firms are domestically
focused. So, from time to time, the
economic fortunes of large and small firms may collectively diverge. But, if the same fundamental economic forces
are impacting all firms, large and small, these differences should not
seriously compromise the usefulness of small business based indicators for
economic analysis.
The
NFIB membership reasonably reflects the small-business population.[4] However, members tend to be somewhat older
and are over-represented in the Midwestern, Plains, and Mountain states. Nevertheless, indicators based on surveys of
non-representative samples of economic agents can provide reliable indicators
of economic activity. As long as sample
frame and response biases are stable over time, thereby not disturbing the
relationship between changes in the indicators and changes in the
economic activity measures of interest, the indicators will be reliable
predictors. Therefore, the practical
value of data from surveys are that: (1) they are not subject to revision since
they are taken at one point in time; (2) they can be empirically related to changes
in macroeconomic indicators that are historically correct (after all revisions
are in) and, (3) the sample of units included need not be representative of all
units in the population nor do all units in the sample need to respond to the
survey.[5] Although the industry composition of the NFIB
membership has shifted toward services as has the makeup of the
In
the following sections, the contribution of the NFIB indicators to the
prediction of dependent variables of interest is quantified using multiple
regression analysis. For simplicity,
standard errors are not presented in the equations below. All variables are significant at the 95%
level unless otherwise indicated.[6] No specific theory, such as that underlying
the consumption function in macroeconomics, lies behind the equations. Still, many of the NFIB indicators have
strong theoretical counterparts such as those described by the generalized
stock adjustment model. The analysis is
a quantification of the parameters of empirical relationships which would be
observed in the use of the NFIB measures as leading indicators. If the relationship between the economic
variable and the NFIB indicator is reasonably stable, it becomes possible to
make quantitative estimates or predictions of the values of the economic
variables based on the NFIB survey results.
The
quarterly survey data used in this analysis are collected in the first month of
each quarter, January, April, July and October. This timing builds in a small implicit
“lead” of at least one to two months into a forecast. For example, if the January survey is the
best predictor of the dependent variable of interest, say GDP growth, this
means that the January survey predicts
the value of the dependent variable for January, February and March. Thus, the NFIB survey data “leads” the GDP
growth number by two months.
Furthermore, the value of the dependent variable may not be officially
known (first estimate) for several weeks after the end of the quarter. If NFIB data lead one-quarter, then the
January survey anticipates the second quarter (April - June) value of the
dependent variable. So, the NFIB data
always lead the variable of interest, sometimes by several quarters.
Leads
are denoted by negative subscripts on the variables. A subscript of “-1” means that the NFIB data
lead by 1 quarter (plus the built-in two months discussed above), “-2” by two
quarters and so on. The forecasting
literature often refers to such a relationship as a “lag”, meaning that the
current value of the dependent variable to be predicted is empirically
determined by an earlier value of the NFIB variable. If the first quarter reading of the NFIB
variable determines the second quarter value of the change in GDP, this is
called a “lag” in the equation, %DGDP = a + b NFIB-1. The percentage change in GDP depends on the
value of the NFIB variable in the prior quarter. Thus, changes in GDP in the current period
depend on changes in the NFIB variables in earlier periods and the NFIB
variables lead changes in GDP. This lead
is represented by a “lag” of the NFIB variable in the equation, meaning that
values of the NFIB variables occur before the relevant value of the dependent
variable of interest.
1. Employment,
Unemployment and Labor Mark et Indicators
Because
small firms play such a critical role in the job creation process, the NFIB
employment measures should have a strong relationship to measures of aggregate
employment growth and other labor market indicators. Two survey measures are used to explain
variations in employment, the net percent of owners who report plans to expand
total employment at their firms, HIREPLN[7],
and the percent of owners who report at least one hard-to-fill job opening,
JOBOPEN.,[8] The larger the percent of owners planning to
expand total employment in the months following the survey, the larger the
expected employment growth in current and future periods.[9]
A
high level of unfilled job openings indicates disequilibrium between the
desired level of employment at the firm and its actual level. This disequilibrium takes time to resolve
(collecting applications, interviewing candidates, etc.). When the percent of owners reporting one or
more hard-to-fill job openings is high, owners have more difficulty getting
employees and employment will grow more slowly.
Job openings exhibit a slightly positive bivariate correlation with
employment change (.17), and a correlation with hiring plans of .75. Thus, after accounting for the effect of
hiring plans, the percent of owners with hard to fill job openings will likely
be negatively associated with the growth in employment.[10] When job openings are pervasive, it is hard
to hire and consequently employment grows more slowly.[11]
Two
commonly used measures of growth in total employment are the quarterly change
in total employment (%DEMPT)
and the quarterly change in private sector employment (%DEMPP). NFIB’s
hiring plans indicator (HIREPLN) performs as expected, with a higher net
percent of firms planning to expand employment associated with stronger growth
in employment over the survey period.
The higher the percent of owners with hard to fill openings (JOBOPEN),
the harder it is to grow employment.
But these indicators, individually and jointly, do not do a good job
predicting the change in the level of employment (Equation 1.1). The NFIB variables perform better when
predicting private sector employment changes as might be expected (Equation
1.2). Still, neither equation forecasts
changes in employment levels well, with or without leads. Small business owners simply fail to
anticipate the very large changes (volatility) in employment. The predicted change in total employment from
Equation 1.1 is plotted against the actual change in Exhibit 1.1.
[1.1]
%DEMPT = 1.57 + .178
HIREPLN - .08 JOBOPEN R2
= 13%
[1.2]
%DEMPP = 1.59 + .273
HIREPLN - .11 JOBOPEN R2 =
20%
a.
Unemployment
Although
the change in employment is not well anticipated by the NFIB labor market
measures, the national unemployment rate (UNE) as reported monthly by the
Bureau of Labor Statistics (BLS) is accurately predicted by the same two
variables. Apparently these variables capture shifts in the balance between
movements into and out of the labor force (changes in the labor force
participation rate) and the creation and destruction of jobs. The unemployment rate is the ratio of the
number of people who want a job (and searched in the prior 4 weeks) divided by
the sum of those who want a job and those who have a job (the labor
force). This gives the unemployment a
variance that is not perfectly related to changes in employment. The best fit is obtained with NFIB variables
leading by one-quarter. Strengthening
hiring plans produce an increase in employment but the unemployment rate will
depend additionally on net flows into or out of the labor force. Often, the unemployment rate and the growth
of employment move in the same, not opposite, directions. Rising reports of hard-to-fill job openings
indicate a reduced unemployment rate [Equation 1.3]. The plot of the predicted unemployment rate
based on Equation 1.3 against the actual unemployment rate is shown in Exhibit
1.2.
[1.3] UNE = 10.22
- .04 HIREPLN-1 - .17 JOBOPEN-1 R2 = 72%
The
percentage change in total (or private, non-farm) employment has a much higher
variance than the unemployment rate.
This volatility explains at least in part the difference in predictive
power of the independent variables.
Larger firms contribute substantially more to the volatility in total
employment over the business cycle than do small firms.[12] Thus, it is unlikely that the employment
decisions of small-business owners would anticipate percentage changes in total
employment. However, owners do sense “tightness” in their
local labor markets and, collectively, do a very solid job of anticipating
changes in the unemployment rate over the business cycle.
The Help Wanted Index (HWI) should be highly correlated
with the percent of owners reporting hard to fill job openings and the strength
of hiring plans. The NFIB variables do
carry the expected sign, but explain only 32
percent of the variation in the HWI.
If larger firms with more formal personnel departments are more likely
to use help wanted ads and have more cyclical employment, the correlation would
be expected to be weaker. Small
employers often use employee networks (hiring friends, family members etc.) to
locate, screen and qualify employees, not want ads and search agencies.[13]
[1.4]
HWI = 54.92 + .47 HIREPLN + .94 JOBOPEN
R2 = 32%
b. Labor Compensation
The
major cost incurred by small business is labor cost, although changes in labor
costs are less volatile than changes in energy or other business costs. Overall, the path of labor costs drives the
price level because firms that cannot cover labor costs will fail. Since April 1982, the NFIB survey has asked a
series of questions about past and planned labor cost changes in addition to
the indicators of the demand for labor.
WAGEUP[14]
is the net percent of owners reporting that they raised labor compensation in
the prior three-month period; PLNWAGE[15]
is the net percent of owners planning to increase compensation during the next
three-month period. Both are seasonally
adjusted. HIRED[16]
is the net percent of owners who reported increasing employment in the previous
3 months and HIREPLN is the net percent of owners who planned to increase the
total number of people working at the firm.
Again, both are seasonally adjusted.
JOBOPEN is the percent of owners reporting at least one hard-to-fill job
opening. JOBOPEN, HIREPLN and HIRED are
measures of the strength of the demand for labor and tightness in the labor
market. Both should be positively
related to measures of labor cost (wages and benefits).
WAGEUP and PLNWAGE are direct measures of actual and anticipated changes in labor compensation, or planned changes to wages and benefits. These variables should have positive relationship to macro measures of labor compensation, but they should occur prior to ( lead) actual events because implementation of compensation changes takes time (especially benefit changes).
The
level of the Bureau of Labor Statistic’s Employment Cost Index (ECI) is well
anticipated by the NFIB data. Hiring
plans and job openings explain 59 percent of the variation in the ECI with a
2-quarter lead (Equation 1.5).
Similarly, reports of past and planned changes in employee compensation
(WAGEUP and PLNWAGE) explain 58 percent of the variation in the ECI, but with a
three quarter lead (Equation 1.6).
Although the fit of the equations is fairly good, the ability of the
NFIB variables to predict the ECI outside of the sample period is compromised
by the fact that the ECI is a trend variable and continually rises, sometimes
faster, sometimes more slowly, while the NFIB variables are percentages and
move within a limited range. Equation
1.7 uses the percentage change in the ECI as the dependent variable [Exhibit
1.3]. The explanatory power of a
difference or percentage change equation is always lower since the advantage of
the presence of a time trend is eliminated.
Equation 1.7 does not capture the amplitude of the fluctuations in the
ECI but does a fair job of anticipating its directional changes (with a 3
quarter lead).
[1.5] ECI = 50.0 + .82 HIREPLN-2 + 2.64
JOBOPEN-2 R2 =
59%
[1.6] ECI = 88.39 + 4.72 WAGEUP-3 – 4.69
PLNWAGE-3 R2
= 58%
[1.7] %DECI
= 1.78 -.04 WAGEUP-3 – .17
PLNWAGE-3 R2
= 24%
(not significant at the 95% level)
(estimated 1984-2002)
Overall,
the NFIB labor market indicators are highly correlated with two of three
important macro labor market variables.
Movements in the ECI can be reasonably well anticipated with the NFIB
survey measures while the unemployment rate is very well anticipated by
them. The lead time in the “best fit” relationships make the
survey measures particularly useful indicators of future economic developments
as measured by popular government labor market statistics.
2. Inflation
Along
with “full employment,” inflation is the other major concern of economic
policy. Two survey questions address
this economic measure: reported changes
in average selling prices over the past 3 months (PASTP)[17]
and reported plans for raising selling prices in the next 3 months (PLANP).[18] The variable PASTP is the percent of owners
who report raising average selling prices less the percent who report lowering
prices (the net percent, seasonally adjusted). PASTP should impact current Consumer Price
Index (CPI) changes, since price changes implemented in the three months prior
to the survey will impact price measures in the current period. PLANP, the percent of owners planning to
increase average selling prices less the percent planning to reduce average
selling prices, should lead the CPI measures.
Plans from the prior quarter should show up as changes in prices in the
current period.
The
NFIB survey also asks for the actual magnitude of past and planned price
changes in categorical classifications.[19] During periods of rapid price changes,
movements in the tails of the distribution of reported price changes should
have an important impact on changes in the average price level. The
information in the tails of the distribution, that is to say, the incidence of
extremely high or extremely low reports of actual and planned selling price
changes adds substantial predictive content.
The findings also suggest that the process of inflation is fairly
gradual. The R2 statistics on
the equations with longer lead structures (not shown) do not deteriorate
substantially as the leads are lengthened.
Thus, plans to raise prices expressed several quarters earlier or
reports of actual changes in average prices in recent past quarters appear to
take some time to feed into the CPI.
PASTP>5 is the percent of firms reporting
price hikes of 5 percent or more in the past three months, and PLNP>5 is the
percent of firms planning to raise prices by 5 percent or more in the next 3 to
6 months (not seasonally adjusted). %DCPI is the annualized percentage change in the
headline Consumer Price Index. As shown
in Equation 2.1, the NFIB price measures anticipate most of the quarterly
variation in the CPI. Exhibit 2.1 plots
the predicted quarterly CPI inflation rate against the actual percentage change
in the CPI.
[2.1]%DCPI = -.20 + .07 PASTP + .08 PASTP>5 +.23
PLNP>5-1 R2 = 78%
Converting these
coefficients into Beta Coefficients (denoted b),[20]
the most important variable in the model is PASTP (b=.41), followed by PLAN>5 (b=.34), and finally, PASTP>5 (b=.21).
Reported past changes are major determinants of the percentage change in
the CPI for the current period. But
plans to change prices, leading one quarter, also exert a heavy influence on
the current period inflation measure.
Substituting
the core CPI, %DCORECPI,
for the %DCPI, to eliminate the
volatile energy and food components of the CPI, produces a similar relationship
(Equation 2.2). The NFIB variables were
not able to capture the decline in the CPI in the late 1980s due to energy
price declines, but did capture the tumble in the CPI in 2002. Excluding energy and food, the NFIB measures
perform well in the 1980s, but still predict a major deceleration in the
inflation rate in 2002 that does not appear in the Core CPI (see Exhibits 2.2
and 2.3).
[2.2] %DCORECPI = -1.15 + .09 PAST + .32*PAST>5 +
.06*PLAN>5-1 R2=79%
Two other important measures of inflation, based on the
GDP deflator (%DGDPDEFLATOR) and the Personal Consumption Expenditures deflator (%DPCEDEFLATOR),
are also well anticipated by NFIB owner reports of past and planned future
price changes. Equations 2.3 and 2.4
illustrate how well the NFIB survey measures anticipate changes in these
popular inflation measures.
[2.3] %DPCEDEFLATOR =
-1.71 + .07 PAST + .22
PAST>5
+.12 PLAN5-1 R2
= 80%
[2.4] %DGDPDEFLATOR = -1.74 + .10 PAST +.24 PAST>5
+.08 PLAN>5-1 R2 =
81%
When
the time period is shortened to exclude the volatile 1970s, the variables
reflecting the tails of the distribution become insignificant and PASTP and
lagged PLANP explain the same fraction of the variance in the CPI. Shortening the time period to exclude the
1970s (Equation 2.5) reduces the R2 to 75% and including only the
1990s (Equation 2.6) reduces the R2 to 41 percent. There is much less predictable variance in
the price measures in the 1990s and into 2000 compared to prior decades, while
the “noise” remains. The coefficients in
the three regressions covering the various sample periods are consistently
similar, suggesting that if inflationary forces were to reappear, the model
based on NFIB survey responses would quickly pick it up.
[2.5]
%DCPI = -.70 + .125 PASTP +.152 PLANP-1 [1980-2002]
R2 = 75%
[2.6]
%DCPI = -.14 + .095 PASTP +.126 PLANP-1 [1990-2002]
R2 = 41%
Overall,
small business owner reports of past and planned price changes do a very good
job of anticipating inflation. Reports
of actual price changes in prior months and plans to raise prices in the
current quarter expressed in the previous quarter both make substantial
contributions to the anticipation of changes in the various price indices.
3.
Business Inventories
Changes
in non-farm business inventories are notoriously difficult to predict. These changes are the direct result of owner
decisions to actively increase or decrease inventories, and of consumer
(customer) decisions to buy more or buy less in a given period of time. Mismatches between these two sets of
decisions can produce wide swings in business inventories. The basic model for examining inventory
investment in macroeconomics is the stock adjustment model: the desired stock of inventories depends on
expected sales in the future period, the cost of holding inventories, and the
ratio of inventory to sales that is desireable for that particular type of
business. Comparing the desired stock to
the stock on hand produces a gap that, if positive, must be closed by
additional inventory accumulation and, if negative, must be closed by reducing inventories.
The
net percent of owners characterizing their current stocks as “too low” (INVSAT)
is a direct proxy for the gap between desired and actual stocks. [21] The percent of owners planning to
intentionally add to inventory stocks (INVPLN) is driven directly by the
pervasiveness among small businesses of a gap between desired and actual
inventory stocks. [22]
Equation
3.1 relates the actual change in business inventories (DINV) as reported in the National Income and Product
Accounts to the NFIB survey measures of inventory satisfaction (INVSAT), the net percent of owners reporting
that current holdings are too low, and inventory plans (INVPLN), the net
percent of owners planning to intentionally increase inventory holdings. The best model incorporates a lead of one
quarter for the NFIB measures. A plot of
the actual and predicted values from Equation 3.1 is shown in Exhibit 3.1).
[3.1] DINV= 19.41 +2.41 INVSAT-1
+ 4.01 INVPLN-1 R2
= 32%
The
NFIB inventory model tracks changes in business inventory fairly well, except
for the period 1997-2002 where it consistently underestimates the buildup of
inventories and then misses the dramatic reduction in 2002. This poor performance in 1997-2002 may be the
result of inventory changes confined to sectors of the economy that are
dominated by large firms, such as manufacturing and telecommunications. There was certainly no gain to be expected
from inventory building in anticipation of rising prices as the economy has
been experiencing “disinflation” for two decades. Re-estimating the equation through 1997
provides virtually identical coefficients, but much higher explanatory power
(Equation 3.2):[23]
[3.2]
DINV = 20.5 +2.35 INVSAT-1 + 3.05 INVPLN-1 R2 = 42%
Economic
growth faltered substantially in the second half of 2000, signaling an end to
the frenetic growth that typified the last half of the decade, creating large
excess inventory holdings. Cash flow
also came under pressure and the gap between S&P reported operating profits
and NIPA profit measures diverged.[24] This helped trigger reductions in inventories
and, as of March 2003, NFIB owners more frequently reported reductions than
additions to inventory holdings for 24 straight months (there were 41
consecutive months of net reductions in the 1990-91 recession). This was preceded by 4 months beginning in
December, 2000 when the net percent of firms adding to inventory was either 1
percent or 0. Unable to raise prices,
the carrying costs of inventory (the nominal interest rate minus the inflation
rate) became positive and, in some industries, substantial. This triggered a record decumulation in
inventory no longer needed to support economic growth which slowed dramatically
in the second half of 2000. The
predictive power of the model was substantially degraded by events during this
boom and bust period that included the largest period of inventory liquidation
in modern economic history.
Beginning
in the fourth quarter of 1982, NFIB began asking about actual changes in
inventories in the preceding 3-month period.
ACTUAL is the seasonally adjusted net percent of firms reporting
an increase in inventory holdings over the past 3 months.[25] There is no lead for this variable as it
reports actual behavior in the months preceding the survey. Adding this variable to the inventory
satisfaction variable and the inventory plan variable yields an improved fit
(Equation 3.3):
[3.3] DINV = 19.0 +3.94 INVSAT-1 + 3.84 INVPLN-1 + 4.24
ACTUAL
R2
= 50% [1982:4
– 2002:4]
Considering
the volatility of inventory investment in the NIPA accounts, this equation
explains well the actual dollar amount of inventory investment. Since NFIB members cover all sectors of the
economy (NAICS), better measures of inventory behavior for predictive purposes
might be obtained by selecting firms to include in the creation of the survey
indicators that are from inventory intensive sectors (e.g. manufacturing,
construction, wholesale trades etc.).
However, sample size considerations prevent the creation of industry
based indicators.
4. Capital
Expenditures
Although
individual small firms rarely make massive capital outlays, the accumulation of
small outlays by six million small-business employers can have a substantial
impact on aggregate capital spending in the U.S. The median outlay for NFIB members is $20,000
(in the prior six months), typically reported by 55 percent to 70 percent of
the owners from quarter to quarter. One
percent of the owners typically report outlays in excess of $500,000 (in the
prior six months).
The
relationship between gross private domestic investment (NIPA) and the NFIB
indicators (CXPLAN),[26]
the percent of owners planning capital outlays in the next 3 to 6 months, and
the incidence of past capital spending (CXPAST)[27]
is not particularly strong (Equation 4.1).
Incorporating leads did not improve the R2. The results fit the predictions that would
follow from a typical capital stock adjustment model: higher levels of past spending (CXPAST) are
associated with lower levels of spending in the current period since the prior
expenditures brought the actual stock closer to that desired by owners. The higher the percent of owners reporting
that they planned outlays in future months (CXPLAN), the higher the actual
level of expenditures in the months following the survey, an indication that
the desired stock of capital must have risen relative to current capacity.[28]
Narrowing the definition of investment spending does
improve the fit, with the major gain coming from the shift from the gross
private domestic investment measure to a measure of private fixed investment or
equipment only (Equations 4.2, 4.3 and 4.4).
All equations are estimated beginning with 1979, the first year that
past expenditure data are available.
Constant CXPLAN CXPAST R2
[4.1] CAPX% 4.18 2.16 -1.22 12%
[4.2] PRIVFIXED% -10.17 1.74 - .73 31%
[4.3] NONRESPRIVFIX% -27.20 1.50 - .28 32%
[4.4] NONRESEQUIP% -16.61 1.80 - .62 28%
Profitability is an important determinant of capital
spending in most macroeconomic models.
Adding the net percent of owners reporting that earnings in the prior
quarter were higher than in the quarter before (EARN)[29],
adds some explanatory power, and leaves the other coefficients fundamentally
unchanged in most cases. Again, narrowing
the definition to fixed investment measures, with or without residential
structures (Equation 4.5 vs 4.6), improves the explanatory power of the
equation as does a focus on equipment (Equation 4.8). The predictions from Equation 4.6 are shown
in Exhibit 4.1.
Constant CXPLAN CXPAST EARN R2
[4.5] CAPX% 24.02 1.40 -1.01 .47 15%
[4.6] PRIVFIXED% -
4.21 1.51 - .67 .14
32%
[4.7] NONRESPRIVFIX% -20.09 1.21 - .20 .18 33%
[4.8] NONRESEQUIP% -
7.39 1.45 - .52 .21 31%
Restricting the estimation period to 1990-2002 provides
much the same result, indicating that the relationships are quite stable over
time. Although there was more
variability in capital spending in the 1970s than in later periods, the
stability of the equations suggest that the NFIB contribution to anticipating
changes in capital spending is fairly stable.
1990-2002 Constant
CXPLAN CXPAST EARN R2
[4.9] CAPX% 20.30 1.56
-1.06
.28 11%
[4.10] PRIVFIXED% -17.12 1.36
- .40 .01
29%
[4.11] NONRESPRIVFIX% -15.25 1.37
- .36 .24 39%
[4.12] NONRESEQUIP% -11.78 1.14
- .28 .18 22%
5. Capital Mark et Indicators
Although most small firms are financed primarily with the
entrepreneur’s own savings when started, the cost and availability of capital
is critical to the survival of small firms once the business is operating. NFIB asks about the ease of obtaining the
most recent loan relative to prior attempts (CREDHARD).[30] CREDHARD is the net percent of owners who
said that it was “harder” to get the loan on the last attempt. Aggregated over all business owners, this is
a proxy for how tight monetary policy is in the months prior to the
survey. The demand for credit varies
substantially over the period of analysis.
To compensate, CREDHARD normalized by the percent of firms reporting
that they borrow on a regular basis [BOR], which varies from a high of 53
percent in 1979 to a low of 29 percent in 1999.
Monetary policy is implemented through bond market
transactions by the Fed in New York.
Thus, the first banks to “feel” a change in policy should be the money
center banks. The Federal Reserve
surveys senior loan officers in about 50 money center banks and ascertains the
percent of these loan officers that report “tightening” or “loosening” the
lending standards for “small” firms, presumably in response to changes in
market conditions created by the Fed or permitted by Fed policy. These changes in credit market conditions
should be transmitted to the remaining 8,000 plus commercial banks, ultimately
being reflected in NFIB owner reports of “easier” or “harder” credit
conditions. The percent of money center
bank senior loan officers who report tightening lending conditions for small
businesses (net of those reporting easier terms) is used as a predictor to
explain variations in the NFIB data on “harder to get the last loan.” With little theoretical guidance other than
“long and variable lags” in monetary policy, lags of varying lengths are tested
looking for the highest R2.
The R2 peaks in both regressions with the lenders’
assessments demonstrating a 17 quarter lag (over 4 years!). The coefficients are more stable, but largest
when the R2 peaks (Exhibit 5.1).
[5.1] CREDHARD/BOR
= 9.40 + .18 FED-17 R2
= 34%
[5.2] CREDHARD
= 3.32 + .07 FED-17 R2 = 35%
Plots
of the predicted percent of owners reported credit “harder” to get are shown in
Exhibit 5.2 and Exhibit 5.3 (the period over which forecasts are available due
to the availability of Fed data and the 17 quarter lag). In simple terms, this relationship indicates
that changes in credit availability reported by senior loan officers of money
center banks takes 17 quarters to have its maximum impact on credit
availability for small business owners.
Since policy transmission effects are not one quarter events, a more
complex distributed lag model that incorporates many past quarterly
observations on loan officer reports may produce higher explanatory power, but
would not alter the conclusion that small business borrowers detect tighter or
easier borrowing conditions long after money center banks report implementing
such policy changes in response to market conditions. Unfortunately, Fed Loan Officer survey data
are not available during the volatile 1770s and early 1980s (see Exhibit
5.2). However, even in the “tame” credit
markets of the 1990s, the relationship between the Fed survey and CREDHARD with
a 17 quarter lag is fairly clear (Exhibit 5.3).
If this is any commentary on the span between the time money center banks
sense a change in market conditions and when this filters out to the smaller
and rural banks in the system, lags are long indeed, and variable, reaching the
most sophisticated banks first and the small banks last. Rising rates may not make borrowing “harder”,
and, for more infrequent borrowers with longer term loans, some considerable
time many pass before they notice a change in credit market conditions.
6. Real GDP Growth
The
most widely reported indicator derived from the NFIB survey is the Index of
Small Business Optimism (INDEX). The 10
questions included in the INDEX have been part of the questionnaire since
October 1974. They include:
·
Good Time for
Business Expansion (GTEX)[31]
·
Outlook For The
Economy: Better Or Worse (EXBUS)[32]
·
Net Earnings
Trends: Higher Or Lower (EARN)
·
Expected Real
Sales Volume: Higher or Lower (EXSALE)[33]
·
Plans To
Increase/Decrease Employment (HIREPLN)
·
Job Openings Not
Able To Fill (JOBOPEN)
·
Current Inventory
Satisfaction: Too High Or Low (INVSAT)
·
Planned
Inventory Change: Increase or Decrease (INVPLN)
·
Expected Change
In Credit
·
Planned Capital
Expenditures (CXPLAN)[35]
Most
of the questions used to construct the INDEX are symmetric, such as whether the
owner expects the economy to be “better” or “worse” in the next 6 months or
plans to “increase” or “decrease” the total number of people working for the
firm. For these questions, a balance
variable (or diffusion index) is formed. The percent of unfavorable responses
(“worse;” “reduce”) is subtracted from the favorable responses (“better;”
“increase”) to provide a net percent variable.
For the three other questions, only the percent of owners offering an
affirmative answer is included, e.g. the percent planning capital spending or
reporting that the current period is a good time for capital expansion. Some variables have strong seasonal patterns
such as hiring plans, though others have little or none, such as capital
spending plans or expected credit conditions.
All ten variables are seasonally adjusted. The INDEX is computed as the sum of the ten
seasonally adjusted components plus 100 to keep the INDEX from becoming
negative. Finally, the INDEX is based to
its average value in 1986 (1986=100), the middle of the 1980s expansion.
Perhaps
the most important indicator of an economy’s overall economic health is growth
in the Gross Domestic Product (GDP).
Two GDP measures are used, the percentage change in real GDP
quarter-to-quarter (annualized), %DGDP, and the quarter over quarter change in real GDP
(from 1st quarter to 1st quarter, etc.), %DGDPQQ.
The
quarterly percentage change in GDP is quite volatile (Exhibit 6.1) and the
INDEX does not do particularly well in explaining the actual magnitude of the
change (Equation 6.1). Equation 6.1’s
predicted values are plotted against the actual quarterly change in real GDP in
Exhibit 6.2. Note in Exhibit 6.2 how
poorly the NFIB indicator forecasts the wild ride of the late 1970s and early
1980s. Even the extremes of the more
docile1990s are not well anticipated, although the trend in changes is
reasonably well anticipated.
The INDEX does a better job of predicting %DGDPQQ because %DGDPQQ smoothes the volatility of GDP growth, and the
less volatile measure more accurately reflects the path of real economic
activity (Equation 6.2). However, it is
more difficult to interpret year over year equations in a quarterly forecasting
context. The best prediction for growth over any 4-quarter period is obtained
using the survey results from the middle of the period.[36] Thus, calendar year growth is best predicted
using the July survey (the “third quarter” survey in the NFIB sequence).
[6.1] %DGDP = -47.75 + .51 INDEX R2
= 37%
[6.2] %DGDPQQ = -36.11 + .40 INDEX -1 R2 = 51%
Final
domestic sales (%DFSALES)
as a dependent variable produces a better fitting model than GDP. The R2
in Equation 6.3 is substantially higher than in the GDP equation, 6.1. The coefficients on INDEX in the GDP and the
Final Sales equations are very close, even though the R2 is higher
in the Final Sales equations.
[6.3] %DFSALES = -45.27 + .49 * INDEX R2 = 43%
[6.4] %DFSALESQQ = -35.81+ .39 * INDEX-1 R2 = 54%
Using all 10 INDEX components raises the R2
(to 44%, 38% adjusted R2) compared to Equation 6.1, but all
coefficients are statistically insignificant and often carry theoretically incorrect
signs. This is a result of the
collinearity that exists among the components (see Appendix 1). Two of the variables, inventory investment
plans, INVPLN and expected changes in the real sales volume, EXSALES, perform
as well as the INDEX in predicting changes in GDP.
Bivariate Full Equation
Component Coeff. R2 Coefficient
CXPLAN .22 .09
-.01
INVPLN .51 .36 .29
INVSAT .72 .22 .10
GTEX .29 .26 .13
EXBUSCOND .04 .05 .00
EARN .17 .17 -.07
EXSALES .20
.39 .10
7. Other
Indicators
A
number of popular survey indicators used by analysts to anticipate changes in
economic activity include the University of Michigan Consumer Confidence Index
(MICH) and the Conference Board Consumer Confidence Index (CONF). Business owners are, of course, consumers and
as such, their views should be reflected in both indices (Equations 7.1 and
7.2).
Although
the NFIB Index is related to these indicators in a way that would be expected,
the relationships are not very strong.
[7.1]
MICH = -8.42 + 1.02 INDEX R2
= 13%
[7.2] CONF = -228.59 + 3.27 INDEX R2 = 32%
The
variance in the two consumer confidence indices is substantial relative to the
INDEX (see Exhibit 7.1]. Thus, it is not surprising that the NFIB
INDEX explains relatively little of the variance in these two measures.
The
Index of Leading Economic Indicators (LEI) is also a popular and widely used
forecasting measures. The LEI leads
changes in economic activity (GDP), but its components are, with few exceptions
(e.g. the University of Michigan Consumer Confidence Index), subject to
revisions that are often substantial.
Thus, one cannot be sure of the existence of a signal unless the change
in the LEI is substantial and persistent.
Its relationship to the INDEX is not strong, either on a concurrent
basis or with NFIB data leading.
(Equation 7.3).
[7.3] LEI = -3.12 + .96 INDEX R2 = 12%
The
LEI and the consumer sentiment measures receive considerable attention in the
press and their release often moves financial markets. One would expect, then, to observe a
substantial relationship between changes in GDP and these measures. However, the R2 is only 2 percent
between the LEI and %DGDP. The University of Michigan Index (MICH) and
the Conference Board Index (CONF) performances are somewhat better with R2
statistics of 19 percent and 12 percent respectively and neither performs
better with one or two quarter leads.
None of the three perform as well as the NFIB INDEX with its R2
of 37 percent.
[7.4] %DGDP = -7.01 + .11 MICH R2 = 19 %
[7.5] %DGDP
= -2.06 + .05 CONF R2 = 12%
[7.6] %DGDP = -1.49 + .05 LEI R2 = 2%
[7.7] %DGDP
= -47.75 + .51 INDEX R2
= 37%
A priori, it is not clear whether the level of these indices should be
used to anticipate changes in GDP or changes in these indices. Converting the predictors to percentage
changes, the R2 for the LEI rises to 25 percent, while MICH falls to
9 percent; CONF remains about the same, 13 percent, and INDEX falls to 4
percent. Percentage changes perform more
poorly for all except the LEI, but the INDEX level produces a better predictive
performance than all of the other three indicators.
9. Summary
Small
business produces half of the private sector GDP and accounts for an even
larger fraction of the private sector labor force and new jobs created. As a consequence, the collective actions of
small firm owners have a major impact on the U.S. economy. The economic indicators pioneered by NFIB are
shown to have strong empirical relationships to important economic measures
such as the growth in GDP, the inflation rate, the unemployment rate, inventory
investment and the Employment Cost Index.
For
the most part, the best models anticipate economic activity by 1 or 2 quarters,
making the NFIB measures useful indicators of future change. Because the NFIB survey measures are never
revised, their relationship to NIPA and other BLS and BEA data that are subject
to revision can provide a sound guide to the direction of the economy though
they may differ from the preliminary figures released by government
agencies. The NFIB indicators are good
predictors of changes in final BLS/BEA numbers and thus may be better
indicators of changes in these measures than the preliminary releases of their
values.
Though
survey data cannot be effectively used to produce long-term forecasts that
require knowing future values of the survey variables, the NFIB measures
contain useful information not captured by variables traditionally employed in
forecasting models. The NFIB data
provide helpful data for identifying the near-term path of the economy and
insight into which forecast scenario might be developing. Because most of the forecasting relationships
use NFIB measures to forecast subsequent events, their usefulness as predictors
of near-term future economic activity is enhanced.
Bibliography
Bram, Jason and Sydney
Ludvigson, “Does Consumer Confidence Forecast Household Expenditure? A
Sentiment Index Horse Race” Federal
Reserve Bank of New York Research Paper # 9708, printed in FRBNY Economic
Policy Review, June, 1998
Campbell,
Carroll, Christopher D.,
Jeffrey C. Fuhrer, and David W. Wilcox, “Does Consumer Sentiment Forecast
Household Spending? If So, Why?”, American Economic Review, 84,
no. 5, pp1397-1408, 1994.
Cashell, Brian W., “Measures
of Consumer Confidence: Are They
Useful?”, Congressional Research Service, Library of Congress, June, 2003.
Dennis, William J. Jr. and
William C. Dunkelberg, “Small Business Economic Trends: A Quarter Century
Longitudinal Data Base of Small Business Economic Activity”, J.A. Katz ed., Data
Bases for the Study of Entrepreneurship, New York, JAI Press, 2000.
Dunkelberg, William C.
and Jonathan A. Scott, “Report on
the Representativeness of the National Federation of Independent Business
Sample of Small Firms in the United States, office of Advocacy, U.S. Small
Business Administration, contract # SBA2A-0084-01, mimeo, 1983.
Howrey, Phillip, “The
Predictive Power of the Index of Consumer Sentiment”, Brookings Papers on
Economic Activity, #1, 2001.
Mishkin, Frederic S., “Consumer
Sentiment and Spending on Durable Goods”,
Brookings Papers on
Economic Activity, no. 1, pp217-32,
1978
Otoo, Maria Ward, “Consumer
Sentiment and the Stock
Popkin Joel and Company,Small Business During the Business
Cycle. Office of Advocacy, U.S. Small
Business Administration, contract # SBAHQ-01-R-0011, 2003.
Popkin Joel and Company ,
“Labor Shortages, Needs and Related Issues in Small and Large Businesses: Part A: Labor Shortages in Small Firms.” Office of Advocacy, U.S. Small Business
Administration contract # SBAHQ-98-C-00017, 1999.
Appendix 1: Principle Components Analysis
The notion of “optimism” is
a conceptual construct that cannot be directly measured. Consequently, subjects are asked a number of
questions that relate to dimensions of what might be part of “optimism”. These measures can then be combined in some
way to identify a construct that more closely approximates “optimism.”
As
the correlation matrix below shows, several single components have a high
correlation with GDP growth (REALSAL is correlated .62) and with other INDEX
components (EARN and GTEX are correlated .81).
Using all 10 of the INDEX components to predict the percentage change in
GDP does produce a higher R2 (44 percent vs. 37 percent for the
INDEX alone), but only one of the components is significant. This may be acceptable for an overall
forecast of GDP growth, but no partial analysis would be reliable (e.g. using
the change in HIREPLN, ceteris paribus, to assess the possible impact of the
change on GDP).
TABLE
1
|
|
HIRE |
OPEN |
EXCRED |
EBCOND |
REALSAL |
NEARN |
INVSAT |
INVPLN |
GTEX |
CAPX |
SALES |
|
|
HIRE |
1 |
|
|
|
|
|
|
|
|
|
|
|
|
OPEN |
0.775467 |
1 |
|
|
|
|
|
|
|
|
|
|
|
EXCRED |
0.420581 |
0.078659 |
1 |
|
|
|
|
|
|
|
|
|
|
EBCOND |
-0.2418 |
-0.629 |
0.400516 |
1 |
|
|
|
|
|
|
|
|
|
REALSAL |
0.425853 |
0.008994 |
0.471566 |
0.424375 |
1 |
|
|
|
|
|
|
|
|
NEARN |
0.632472 |
0.397864 |
0.403976 |
-0.11266 |
0.60178 |
1 |
|
|
|
|
|
|
|
INVSAT |
0.413706 |
0.065554 |
0.324099 |
0.09731 |
0.629813 |
0.586218 |
1 |
|
|
|
|
|
|
INVPLN |
0.683859 |
0.277986 |
0.520037 |
0.19057 |
0.780956 |
0.679639 |
0.703643 |
1 |
|
|
|
|
|
GTEX |
0.709636 |
0.436378 |
0.589673 |
0.016338 |
0.693577 |
0.809932 |
0.512649 |
0.673083 |
1 |
|
|
|
|
CAPX |
0.833069 |
0.610202 |
0.33429 |
-0.31626 |
0.340427 |
0.606001 |
0.385588 |
0.60225 |
0.590023 |
1 |
|
|
|
SALES |
0.318249 |
0.348849 |
-0.03113 |
-0.27245 |
0.513351 |
0.675439 |
0.384106 |
0.375165 |
0.547934 |
0.321874 |
1 |
|
|
%GDP |
0.372992 |
0.102144 |
0.386221 |
0.228409 |
0.623369 |
0.420866 |
0.466718 |
0.602545 |
0.516234 |
0.296841 |
0.346755 |
|
Principle
components, and related techniques such as factor analysis and cluster
analysis, can be employed when a number of measurements are taken that are
related to some underlying conceptual construct such as “optimism.” The results of the analysis are presented in
Tables 2 and 3. The information content
of the ten components of the INDEX can be reasonably represented by linear
combinations of the ten components combined into four new indices. These four independent constructs account for
88 percent of the variation contained in the ten question series.
Good
Time to Expand dominates the first component with strong contributions from
Hiring Plans, Inventory Plans and Expected Real Sales gains. A supporting role is played by reports of
improved earnings. The second component
is dominated by Expected Business Conditions.
This INDEX component is highest when the percent of firms with
hard-to-fill job openings is lowest (at or near the trough in the business
cycle). The third component is dominated
by Inventory Satisfaction and, to a lesser degree, by Planned Capital Outlays. However, the latter never loads more than .6
in any component. Component four
represents the Credit Outlook. This
variable appears to be independent of the other nine Index variables, although
it is not a significant factor.[37] Using the four strongest components in place
of the INDEX to explain changes in real GDP produces a somewhat higher R2
(just the use of 4 instead of 1 predictor will do this), but on an adjusted R2
basis, there is little gain and the constructs are more difficult to use and
interpret in a forecasting context.
|
|
|
|
TABLE 2 |
|
|
|
|
|
PRINCIPAL
COMPONENTS ANALYSIS: SMALL BUSINESS OPTIMISM INDEX |
|||||||
|
|
|
PERCENT
OF VARIANCE EXPLAINED |
|
|
|||
|
EIGENVALUE |
COMPONENT |
CUMULATIVE |
|
|
|||
|
1 |
4.9 |
49.0 |
|
49.0 |
|
|
|
|
2 |
2.1 |
21.0 |
|
70.0 |
|
|
|
|
3 |
1.2 |
12.0 |
|
82.0 |
|
|
|
|
4 |
0.6 |
5.6 |
|
87.6 |
|
|
|
|
5 |
0.4 |
4.1 |
|
91.7 |
|
|
|
|
6 |
0.4 |
3.7 |
|
95.4 |
|
|
|
|
7 |
0.2 |
1.8 |
|
97.2 |
|
|
|
|
8 |
0.2 |
1.5 |
|
98.7 |
|
|
|
|
9 |
0.1 |
0.8 |
|
99.5 |
|
|
|
|
10 |
0 |
0.5 |
|
100 |
|
|
|
|
|
|
|
|
TABLE 3 |
|
|
|
|
|
PRINCIPLE
COMPONENTS, SMALL BUSINESS OPTIMISM INDEX |
||||||
|
|
|
|
|
|
|
|
|
|
COMPONENT |
|
1 |
2 |
3 |
4 |
5 |
|
|
Hiring Plans |
|
0.85 |
-0.4 |
0.14 |
0 |
0.24 |
|
|
Job Openings |
|
0.45 |
-0.79 |
-0.13 |
0.25 |
0.15 |
|
|
Credit Outlook |
|
0.64 |
0.35 |
0.34 |
0.54 |
0 |
|
|
Expected Business Conditions |
0 |
0.91 |
0.25 |
0 |
0.15 |
||
|
Expected Real Sales |
0.77 |
0.46 |
0 |
-0.23 |
0 |
||
|
Net Earnings Change |
0.86 |
0 |
-0.14 |
-0.12 |
-0.39 |
||
|
Inventory Satisfaction |
0.52 |
0.23 |
-0.75 |
0 |
0 |
||
|
Inventory Plans |
|
0.88 |
0.18 |
-0.2 |
0 |
0.25 |
|
|
Good Time to Expand |
0.91 |
0 |
0 |
0 |
-0.28 |
||
|
Planned Capital Outlays |
0.61 |
-0.26 |
0.6 |
-0.34 |
0 |
||