Unformatted text preview: ISSN 1183-1057 Department of Economics
Discussion Papers 02-3
Oh No! I Got The Wrong
Sign! What Should I Do?
S O M M E S PR SIMON FRASER UNIVERSITY Oh No! I Got the Wrong Sign! What Should I Do?
Professor of Economics,
Dept. of Economics
Simon Fraser University
Canada V5A 1S6
Email: [email protected]
Getting a “wrong” sign in empirical work is a common phenomenon. Remarkably,
econometrics textbooks provide very little information to practitioners on how this
problem can arise. This paper exposits a long list of ways in which a “wrong” sign can
occur, and how it might be corrected. Oh No! I Got the Wrong Sign! What Should I Do?
We have all experienced, far too frequently, the frustration caused by finding that the
estimated sign on our favorite variable is the opposite of what we anticipated it would be.
This is probably the most alarming thing "that gives rise to that almost inevitable
disappointment one feels when confronted with a straightforward estimation of one's
preferred structural model." (Smith and Brainard, 1976, p.1299). To address this problem,
we might naturally seek help from applied econometrics texts, looking for a section
entitled "How to deal with the wrong sign." Remarkably, a perusal of existing texts does
not turn up sections devoted to this common problem. Most texts mention this
phenomenon, but provide few examples of different ways in which it might occur.1 This
is unfortunate, because expositing examples of how this problem can arise, and what to
do about it, can be an eye-opener for students, as well as a great help to practitioners
struggling with this problem. The purpose of this paper is to fill this void in our textbook
literature by gathering together several possible reasons for obtaining the "wrong" sign,
and suggesting how corrections might be undertaken.
A wrong sign can be considered a blessing, not a disaster. Getting a wrong sign is
a friendly message that some detective work needs to be done – there is undoubtedly
some shortcoming in the researcher’s theory, data, specification, or estimation
procedure. If the “correct” signs had been obtained, odds are that the analysis would
not be double-checked. The following examples provide a checklist for this doublechecking task, many illustrating substantive improvements in specification. 1. Bad Economic Theory. Suppose you are regressing the demand for Ceylonese
tea on income, the price of Ceylonese tea and the price of Brazilian coffee. To
your surprise you get a positive sign on the price of Ceylonese tea. This dilemma
is resolved by recognizing that it is the price of other tea, such as Indian tea, that
is the relevant substitute here. Rao and Miller (1971, p.38-9) provide this
example. Gylfason (1981) refers to many studies which obtained “wrong” signs
because they used the nominal rather than real interest rate when explaining
1 Wooldridge (2000) is an exception; several examples of wrong signs are scattered throughout this text. 2. Omitted Variable. Suppose you are running an hedonic regression of automobile
prices on a variety of auto characteristics such as horsepower, automatic
transmission, and fuel economy, but keep discovering that the estimated sign on
fuel economy is negative. Ceteris paribus, people should be willing to pay more,
not less, for a car that has higher fuel economy, so this is a “wrong” sign. An
omitted explanatory variable may be the culprit. In this case, we should look for
an omitted characteristic that is likely to have a positive coefficient in the hedonic
regression, but which is negatively correlated with fuel economy. Curbweight is a
possibility, for example. (Alternatively, we could look for an omitted
characteristic which has a negative coefficient in the hedonic regression and is
positively correlated with fuel economy.) Here is another example, in the context
of a probit regression. Suppose you are using a sample of females who have been
asked whether they smoke, and then are resampled twenty years later. You run a
probit on whether they are still alive after twenty years, using the smoking
dummy as the explanatory variable, and find to your surprise that the smokers are
more likely to be alive! This could happen if the non-smokers in the sample were
mostly older, and the smokers mostly younger, reflecting Simpson’s paradox.
Adding age as an explanatory variable solves this problem, as noted by Appleton,
French, and Vanderpump (1996).
3. High Variances. Suppose you are estimating a demand curve by regressing
quantity of coffee on the price of coffee and the price of tea, using time series
data, and to your surprise find that the estimated coefficient on the price of coffee
is positive. This could happen because over time the prices of coffee and tea are
highly collinear, resulting in estimated coefficients with high variances – their
sampling distributions will be widely spread, and may straddle zero, implying that
it is quite possible that a draw from this distribution will produce a “wrong” sign.
Indeed, one of the casual indicators of multicollinearity is the presence of
“wrong” signs. In this example, a reasonable solution to this problem is to
introduce additional information by using the ratio of the two prices as the
explanatory variable, rather than their levels. This example is one in which the
wrong sign problem is solved by incorporating additional information to reduce high variances. Multicollinearity is not the only source of high variances,
however; they could result from a small sample size, or minimal variation in the
explanatory variables. Leamer (1978, p.8) presents another example of how
additional information can solve a wrong sign problem. Suppose you regress
household demand for oranges on total expenditure E, the price po of oranges, and
the price pg of grapefruit (all variables logged), and are surprised to find wrong
signs on the two price variables. Impose homogeneity, so that if prices and
expenditure double, the quantity of oranges purchased should not change; this
implies that the sum of the coefficients of E, po, and pg is zero. This extra
information reverses the price signs.
4. Selection Bias. Suppose you are regressing academic performance, as measured
by SAT scores (the scholastic aptitude test is taken by many students to enhance
their chances of admission to the college of their choice) on per student
expenditures on education, using aggregate data on states, and discover that the
more money the government spends, the less students learn! This “wrong” sign
may be due to the fact that the observations included in the data were not obtained
randomly – not all students took the SAT. In states with high education
expenditures, a larger fraction of students may take the test. A consequence of this
is that the overall ability of the students taking the test may not be as high as in
states with lower education expenditure and a lower fraction of students taking the
test. Some kind of correction for this selection bias is necessary. In this example,
putting in the fraction of students taking the test as an extra explanatory variable
should work. This example is taken from Guber (1999). Currie and Cole (1993)
exposit another good example of selection bias. Suppose you are regressing the
birthweight of children on several family and background characteristics,
including a dummy for participation in AFDC (aid for families with dependent
children), hoping to show that the AFDC program is successful in reducing low
birthweights. To your consternation the slope estimate on the AFDC dummy is
negative! This probably happened because mothers self-selected themselves into
this program – mothers believing they were at risk for delivering a low
birthweight child may have been more likely to participate in AFDC. This could be dealt with by using the Heckman two-stage correction for selection bias or an
appropriate maximum likelihood procedure. A possible alternative solution is to
confine the sample to mothers with two children, for only one of which the
mother participated in the AFDC program. A panel data method such as fixed
effects (or differences) could be used to control for the unobservables that are
causing the problem.
5. Data Definitions/Measurement Error. Suppose you are regressing stock price
changes on a dummy for bad weather, in the belief that bad weather depresses
traders and they tend to sell, so you expect a negative sign. But you get a positive
sign. Rethinking this, you change your definition of bad weather from 100 percent
cloud cover plus relative humidity above 70 percent, to cloud cover more than
80% or relative humidity outside the range 25 to 75 percent. Magically, the
estimated sign changes. This example illustrates more than the role of variable
definitions/measurement in affecting coefficient signs – it illustrates the dangers
of data mining and underlines the need for sensitivity analysis. This example
appears in Kramer and Runde (1997). This is not the only way in which
measurement problems can contribute to generating a wrong sign. It is not
uncommon to regress the crime rate on the per capita number of police and obtain
a positive coefficient, suggesting that more police engender more crime. One
possible reason for this is that having extra police causes more crime to be
reported. Another reason for how measurement error can cause a wrong sign is
exposited by Bound, Brown, and Mathiowetz (2001). They document that often
measurement errors are correlated with the true value of the variable being
measured (contrary to the usual econometric assumption) and show how this can
create extra bias sufficient to change a coefficient’s sign.
6. Outliers. Suppose you are regressing infant mortality on doctors per thousand
population, using data on the 50 US states plus the District of Columbia, but find
that the sign on doctors is positive. This could happen because the District of
Columbia is an outlier – relative to other observations, it has large numbers of
doctors, and pockets of extreme poverty. Removing the outlier should solve the sign dilemma. This example appears in Wooldridge (2000, p.303-4). Rowthorn
(1975) points out that a nice OECD cross-section regression confirming Kaldor’s
law resulted from a random scatter of points and an outlier, Japan.
7. Simultaneity/Lack of Identification. Suppose you are regressing quantity of an
agricultural product on price, hoping to get a positive coefficient because you are
interpreting it as a supply curve. Historically, such regressions produced negative
coefficients and were interpreted as demand curves – the exogenous variable
“weather” affected supply but not demand, rendering this regression an identified
demand curve. Estimating an unidentified equation would produce estimates of an
arbitrary combination of the supply and demand equation coefficients, and so
could be of arbitrary sign. The lesson here is check for identification. A classic
example here is Moore (1914) who regressed quantity of pig iron on price,
obtained a positive coefficient and announced a new economic discovery – an
upward-sloping demand curve. He was quickly rebuked for confusing supply and
demand curves. Morgan (1990, chapter 5) discusses historical confusion on this
issue. The generic problem here is simultaneity. More policemen may serve to
reduce crime, for example, but higher crime will cause municipalities to increase
their police force, so when crime is regressed on police, it is possible to get a
positive coefficient estimate. Identification is achieved by finding a suitable
instrumental variable. This suggests yet another reason for a wrong sign – using a
8. Bad Instruments. Instrumental variable (IV) estimation is usually employed to
alleviate the bias caused by correlation between an explanatory variable and the
equation error. Suppose you are regressing incidence of violent crime on
percentage of population owning guns, using data on U.S. cities. Because you
believe that gun ownership is endogenous (i.e., higher crime causes people to
obtain guns), you use gun magazine subscriptions as an instrumental variable for
gun ownership and estimate using two-stage least squares. You have been careful
to ensure identification, and check that the correlation between gun ownership and
gun magazine subscriptions is substantive, so are very surprised to find that the IV slope estimate is negative, the reverse of the sign obtained using ordinary least
squares. This was caused by negative correlation between gun subscriptions and
crime. The instrumental variable gun subscriptions was representing gun
ownership which is culturally patterned, linked with a rural hunting subculture,
and so did not represent gun ownership by individuals residing in urban areas,
who own guns primarily for self-protection.2 Another problem with IV estimation
is that if the IV is only weakly correlated with the endogenous variable for which
it is serving as an instrument, the IV estimate is not reliable and so a wrong sign
9. Specification Error. Suppose you have student scores on a pretest and a posttest
and are regressing their learning, measured as the difference in these scores, on
the pretest score (as a measure of student ability), a treatment dummy (for some
students having had an innovative teaching program) and other student
characteristics. To your surprise the coefficient on pretest is negative, suggesting
that better students learn less! Becker and Salemi (1977) spell out several ways in
which specification bias could cause this. One example is that the true
specification may be that the posttest score depends on the pretest score with a
coefficient less than unity. Subtracting pretest from both sides of this relationship
produces a negative coefficient on pretest in the relationship connecting the score
difference to the pretest score. Measurement error could also be playing a role
here. A positive measurement error in pretest appears negatively in the score
difference, creating a negative correlation between the pretest explanatory
variable and the equation error term, creating bias.
10. Ceteris Paribus Confusion. Suppose you have regressed house price on square
feet, number of bathrooms, number of bedrooms, and a dummy for a family room,
and are surprised to find the family room coefficient has a negative sign. The
coefficient on the family room dummy tells us the change in the house price if a
family room is added, holding constant the other regressor values, in particular
holding constant square feet. So adding a family room under this constraint must
2 I am indebted to Tomislav Kovandzic for this example. entail a reduction in square footage elsewhere, such as smaller bedrooms or loss
of a dining room, which will entail a loss in house value. In this case the net effect
on price is negative. This problem is solved by asking what will happen to price
if, for example, a 600 square foot family room is added, so that the proper
calculation of the value of the family room involves a contribution from both the
square feet regressor coefficient and the family room dummy coefficient. As
another example, suppose you are regressing yearling (racehorse) auction prices
on various characteristics of the yearling, plus information on its sire (father) and
dam (mother). To your surprise you find that although the estimated coefficient
on dam dollar winnings is positive, the coefficient on number of dam wins is
negative, suggesting that yearlings from dams with more race wins are worth less.
This wrong sign problem is resolved by recognizing that the sign is
misinterpreted. In this case, the negative sign means that holding dam dollar
winnings constant, a yearling is worth less if its dam required more wins to earn
those dollars. Although proper interpretation solves the sign dilemma, in this case
an adjustment to the specification seems appropriate: replace the two dam
variables with a new variable, earnings per win. This example is taken from
Robbins and Kennedy (2001).
11. Interaction Terms. Suppose you are regressing economics exam scores on grade
point average (GPA) and an interaction term which is the product of GPA and
ATTEND, percentage of classes attended. The interaction term is included to
capture your belief that attendance benefits better students more than poorer
students. Although the estimated coefficient on the interaction term is positive, as
you expected, to your surprise the estimated coefficient on GPA is negative,
suggesting that students with higher ability, as measured by GPA, have lower
exam scores. This dilemma is easily explained – the partial derivative of exam
scores with respect to GPA is the coefficient on GPA plus the coefficient on the
interaction term times ATTEND. The second term probably outweighs the first
for all ATTEND observations in the data, so the influence of GPA on exam scores
is positive, as expected. Wooldridge (2000, p.190-1) presents this example. 12. Regression to the Mean. Suppose you are testing the convergence hypothesis by
regressing average annual growth over the period 1950-1979 on GDP per work
hour in 1950. Now suppose there is substantive measurement error in GDP. Large
underestimates of GDP in 1950 will result in low GDP per work hour, and at the
same time produce a higher annual growth rate over the subsequent period
(because the 1979 GDP measure will likely not have a similar large
underestimate). Large overestimates will have an opposite effect. As a
consequence, your regression is likely to find convergence, even when none
exists. This is a type of “wrong” sign, in this case produced by the regression to
the mean phenomenon. For more on this example, see Friedman (1992). A similar
example is identified by Hotelling (1933). Suppose you have selected a set of
firms with high business-to-sales ratios and have regressed this measure against
time, finding a negative relationship i.e., over time the average ratio declines. This
result is likely due to the reversion to the mean phenomenon – the firms chosen
probably had high ratios by chance, and in subsequent years reverted to a more
13. Nonstationarity. Regressing a random walk on an independent random walk
should produce a slope coefficient insignificantly different from zero, but far too
frequently does not, as is now well-known. This spurious correlation represents a
“wrong” sign – the sign should not be significantly positive or negative. This is a
very old problem, identified by Yule (1926) in an article entitled “Why do we
sometimes get nonsense correlations between time series?”
14. Common Trends. A common trend could swamp what would otherwise be a
negative relationship between two variables; omitting the common trend would
give rise to the wrong sign.
15. Functional Form Approximation. Suppose you are running an hedonic
regression of house prices on several characteristics of houses, including number
of rooms and the square of the number of rooms. Although you get a positive
coefficient on the square of number of rooms, to your surprise you get a negative coefficient on number of rooms, suggesting that for a small number of rooms
more rooms decreases price. This could happen because in your data there are no
(or few) observations with a small number of rooms, so the quadratic term
dominates the linear term throughout the range of the data. The negative sign on
the linear term comes about because it provides the best approximation to the
data. Wooldridge (2000, p.188) provides this example.
16. Dynamic Confusion. Suppose you have regressed income on lagged income and
investment spending. You are interpreting the coefficient on investment as the
multiplier and are surprised to find that it is less than unity, a type of “wrong
sign.” Calculating the long-run impact on income this implies, however, resolves
this dilemma. This example appears in Rao and Miller (1971, p.44-5). Suppose
you have panel data on the US states and are estimating the impact of public
capital stock (in addition to private capital stock and labor input) on state output.
You estimate using fixed effects and to your surprise obtain a negative sign on the
public capital stock coefficient estimate. Baltagi and Pinnoi (1995) note that this
could be because fixed effects estimates the short-run reaction; pooled OLS, the
“between” estimator, and random effects all produce the expected positive sign,
suggesting that the long-run impact is positive. Suppose you believe that x affects
y positively but there is a lag involved. You regress yt on xt and xt-1 and are
surprised to find a negative coefficient on xt-1. The explanation for this is that the
long-run impact of x is smaller than its short-run impact.
17. Reversed Measure. Suppose you are regressing consumption on a consumer
confidence measure, among other variables, and unexpectedly obtain a negative
sign. This could happen because you didn’t realize that small numbers for the
consumer confidence measure correspond to high consumer confidence. It has
been known3 for an economist to present an entire seminar trying to explain a
wrong sign only to discover afterwards that it resulted from his software reversing
the coding on his logit analysis. 3 I am indebted to Marie Rekkas for this anecdote. 18. Heteroskedasticity. Suppose you are estimating a probit model, with the latent
equation a linear function of x, namely y* = α + βx + ε, but the error ε is
heteroskedastic, with variance σ2 proportional to the square of x. Probit estimates
β/σ, not β, because the likelihood function is based on the cumulative standard
normal density. So the operative latent equation is proportional to α/x + β, in
which the influence of x is reversed in sign. See Wooldridge (2001, p.479) for
19. Underestimated Variances. If the variance of a coefficient estimate is
underestimated, an irrelevant variable could be statistically “significant,” of either
sign. The Poisson model assumes that the variance of the counts is equal to its
expected value. Because of this Poisson estimation produces marked
underestimates of coefficient estimates’ variances in the typical case in which
there is overdispersion (the count variance is larger than its expected value).
Researchers often rely on asymptotic properties of test statistics which could be
misleading in small samples. A classic example appears in Laitinen (1978) who
showed that failure to use small-sample adjustments explained why demand
homogeneity had been rejected so frequently in the literature. What should be done if your double-checking can turn up no reasonable
explanation for the “wrong” sign? Try and get it published. Wrong sign puzzles, such
as the Leontief paradox, are a major stimulus to the development of our discipline.
For example, recent evidence suggests that there is a positive relationship between
import tariffs and growth across countries in the late 19th century, a “wrong” sign in
many economists’ view. Irwin (2002) extends the relevant economic theory to offer
an explanation for this.
There is no definitive list of ways in which “wrong” signs can be generated. In
general, any theoretical oversight, specification error, data problem, or inappropriate
estimating technique could give rise to a “wrong” sign. Observant readers might have
noted that many could be classified under a single heading: Researcher Foolishness. This serves to underline the importance of the first of Kennedy’s (2002) ten commandments of
applied econometrics: Use Common Sense. REFERENCES
Appleton, D. R., J. M. French, and M. P. J. Vanderpump. 1996. Ignoring a Covariate: An
Example of Simpson’s Paradox. American Statistician 50: 340-1.
Baltagi, B. H. and N. Pinnoi. 1995. Public Capital Stock and State Productivity Growth:
Further Evidence from an Error Components Model. Empirical Economics 20: 351-9.
Becker, W. E. and M. K. Salemi. 1977. The Learning and Cost Effectiveness of AVT
Supplemented Instruction: Specification of Learning Models. Journal of Economic
Education 8: 77-92.
Bound, J., C. Brown and N. Mathiowetz. 2001. Measurement Error in Survey Data. In J.
J. Heckman and E. E. Leamer (eds), Handbook of Econometrics, vol.V. Amsterdam:
North Holland, 3705-3843.
Currie, J. and N. Cole. 1993. Welfare and Child Health: The Link between AFDC
Participation and Birth Weight. American Economic Review 83: 971-83.
Friedman, M. 1992. Do Old Fallacies Ever Die? Journal of Economic Literature 30:
Guber, D. C. 1999. Getting What You Pay For: The Debate over Equity in Public School
Expenditures. Journal of Statistics Education 7(2). [Online] Available at:
Gylfason, H. 1981. Interest Rates, Inflation, and the Aggregate Consumption Function.
Review of Economics and Statistics 63: 233-45.
Hotelling, H. 1933. Review of The Triumph of Mediocrity in Business, by Horace Secrist.
Journal of the American Statistical Association 28: 463-5.
Irwin, D. A. 2002. Interpreting the Tariff-Growth Correlation of the Late 19th Century.
American Economic Review, Papers and Proceedings 92: 165-9.
Kennedy, P. E. 2002. Sinning in the Basement: What are the Rules? The Ten
Commandments of Applied Econometrics. Journal of Economic Surveys, 16: Kramer, W. and R. Runde. 1997. Stocks and the Weather: An Exercise in Data Mining or
Yet Another Capital Market Anomaly? Empirical Economics 22: 637-41.
Laitinen, K. (1978) Why is Demand Homogeneity So Often Rejected? Economics Letters
Leamer, E. E. 1978. Specification Searches: Ad Hoc Inference with Nonexperimental
Data. New York: John Wiley.
Morgan, M. S. (1990) The History of Econometric Ideas. Cambridge: Cambridge
Moore, H. L. (1914) Economic Cycles – Their Law and Cause. New York: Macmillan.
Rao, P. and R. Miller (1971) Applied Econometrics. Belmont, CA: Wadsworth.
Robbins, M. and P. E. Kennedy. 2001. Buyer Behaviour in a Regional Thoroughbred
Yearling Market, Applied Economics 33: 969-77.
Rowthorn, R. E. 1975. What remains of Kaldor’s Law? Economic Journal 85: 10-19.
Smith, G. and W. Brainard. 1976. The Value of A Priori Information in Estimating a
Financial Model. Journal of Finance 31: 1299–322.
Wooldridge, J. M. (2000) Introductory Econometrics. Cincinnati: South-Western.
Wooldridge, J. M. (2002) Econometric Analysis of Cross Section and Panel Data.
Cambridge, Mass.: MIT Press.
Yule, G. U. 1926. Why Do We Sometimes Get Nonsense Correlations Between TimeSeries? Journal of the Royal Statistical Society 89: 1-64. ...
View Full Document