0892-6638/87/0001-0365/$01.50.

FASEB
365
Fitting curves to data using nonlinear regression:
a practical and nonmathematical review
HARVEY J. MOTULSKY AND LENNART A. RANSNAS
Department of Pharmacology, University of Calfornia, San Diego, La Jolla, California 92093, USA
ABSTRACT
Many types of data are best analyzed by fitting a curve
using nonlinear regression, and computer programs
that perform these calculations are readily available.
Like every scientific technique, however, a nonlinear
regression program can produce misleading results
when used inappropriately. This article reviews the use
of nonlinear regression in a practical and nonmathe-
matical manner to answer the following questions: Why
is nonlinear regression superior to linear regression of
transformed data? How does nonlinear regression
differ from polynomial regression and cubic spline?
How do nonlinear regression programs work? What
choices must an investigator make before performing
nonlinear regression? What do the final results mean?
How can two sets of data or two fits to one set of data
be compared? What problems can cause the results to
be wrong? This review is designed to demystify non-
linear regression so that both its power and its limita-
tions will be appreciated. - MOTULSKY, H. J.;
RANSNAS, L. A. Fitting curves to data using nonlinear
regression: a practical and nonmathematical review.
FASEBJ 1: 365-374; 1987.
Key Words: nonlinearregressioncomputerprograms - curve
fitting data analysis
MANY TYPES OF DATA ARE BEST ANALYZED by fitting a
curve using nonlinear regression, and computer pro-
grams that can perform these calculations are readily
available. However, it is difficult to learn about non-
linear regression because it is a topic virtually ignored
by most statistical textbooks, and because many articles
on the subject assume an advanced level of mathemati-
cal background. We therefore prepared this review to
explain simply the theory, use, and pitfalls of nonlinear
regression. Our goal is to present a practical approach
to the problem, and we largely cite review articles.
More mathematical details, as well as citations to the
primary literature, can be found in several reviews (see
refs 1-5).
TYPES OF CURVE FITTING
Nonlinear regression is a powerful tool for fitting data
to an equation to determine the values of one or more
parameters. Before discussing nonlinear regression,
however, we will first review the other methods used for
fitting curves to data.
Linear regression of transformed data
Linear regression is familiar to all scientists. Data are
graphed so that the x axis represents the independent
variable and they axis represents the dependent variable.
The line drawn by the linear regression procedure is
chosen to minimize the sum of the squares of the verti-
cal distances of the points from that line.
An easy method for dealing with curved relation-
ships is to transform the data into a straight line and
then perform linear regression. An example is shown
in Fig. 1. The data follow an exponential decay
(y = Ae). By taking the logarithm of each y value,
the data points form a straight line [ln(y) = ln(A) - BXI.
Other common transformations are used to convert
saturation radioligand-binding isotherms into Scatchard
plots, and enzyme kinetic data into Lineweaver-Burke
plots. Linear regression of the transformed data yields
a slope and intercept that can be used to determine the
parameters of interest.
Performing linear regression on transformed data
has several advantages: It is intuitively straightforward;
it does not require a computer; and the result may seem
easy to evaluate statistically. However, the results are
not statistically optimal. Linear regression calculations
are valid only when the experimental uncertainty of
replicate y values is not related to the values of x or y.
This assumption is usually not valid after data have
been transformed. For example, Fig. 1 shows an expo-
nential decay curve, and a linear plot constructed by
taking the logarithm of ally values. However, this trans-
formation enhanced the errors associated with the
points with a small y value. Thus those points will be
emphasized by linear regression, and the points with
a low x value will be relatively ignored. Accordingly,
the resulting slope will not be the best estimate of the
rate constant. Linearizing transformations are not op-
timal because they distort the experimental errors. This
distortion is especially severe in transformations that
combine x and y values, for example, Scatchard plots
used to analyze radioligand-binding data.
Polynomial regression
Polynomial regression is used to fit data points to the
following equation:
-Cubic Spline
---Polynomial
a-
4J
Ia
0
a-
I
>
#{149}3
0
C
-J
366
MOTULSKY AND RANSNAS
Y=A + Bx+ Cx2 +Dx3 + Ex4
The goal of polynomial regression is to determine
values for the parameters (A, B, C,. . -) that make the
curve best fit the data points. When using a polynomial
regression program, you must specify the order of the
equation - the number of parameters to be fit. When
only A and B are to be fit, the equation describes a line,
and polynomial regression is identical to linear regres-
sion. With three or more parameters, the equation
describes a curve; more parameters create a more flex-
ible curve. In this equation, y is linear with (propor-
tional to) each of the parameters (when x and the other
parameters are held constant). Accordingly, a unique
solution can be obtained without the use of the iterative
procedures described below. The methods used for
fitting polynomial equations are extensions of linear
regression.
Polynomial regression is often used to create a
generic curve through the data points; in such cases the
mathematical form of the equation is irrelevant. See,
for example, Fig. 2. Polynomial regression is not fre-
quently used for data analysis in biology.
Cubic spline
A cubic spline is a curve that goes through every point.
Cubic spline is useful for plotting a smooth curve
through a set of data points (Fig. 2). This technique
may also be used for interpolating between data points
on a standard curve. Unlike nonlinear regression, cubic
spline is not a tool useful for analyzing data. Different
cubic spline algorithms may find slightly different
curves, especially near the first and last data points.
The idea behind a cubic spline curve is straight-
forward. Although a pair of data points can be fit only
to a single straight line, they can be fit to many cubic
TIME TIME
Figure 1. Curve fitting by linear transformation. The left panel
shows computer generated data points after an exponential dissoci-
ation curve. The replicate values are equally spaced, representing
experimental uncertainty that is not related to the values of x ory.
The right panel shows the same data after eachy value is converted
to its logarithm. After this transformation, the relationship is linear
and, accordingly, linear regression can be used to analyze the data.
Note, however, that the experimental uncertainty of the points now
is related to the values of x andy. Linear regression on the trans-
formed data would therefore overemphasize the points with large x
values.
Figure 2. Cubic spline vs. cubic polynomial. The solid curve fit via
a cubic spline procedure goes through every data point. The dashed
curve was determined by fitting the data to the cubic polynomial
equation y = A + Bx2 + C,3. Depending on the circumstances,
both of these procedures are useful for drawing generic curves
through data points. These procedures should be considered the
electronic equivalent of a French curve or a flexible ruler; they are
not often useful for data analysis.
(third-degree polynomial) curves. The cubic spline
technique finds a different cubic equation for every pair
of adjacent points, and selects these equations so that
the overall curve is smooth. It does this by ensuring that
the first and second derivatives of each curve segment
match those of the adjacent segments.
Note the distinction between fitting data to a cubic
polynomial equation and using the cubic spline method.
Equation 1 is a cubic equation when E and all variables
beyond are set to zero. A polynomial fit with a cubic
equation fits one equation to all the data. A cubic
spline, in contrast, uses a different cubic equation for
each adjacent pair of data points. The cubic spline
curve goes through each data point; a polynomial fit
usually does not.
THEORY OF NONLINEAR REGRESSION
Nomenclature
We confine our discussion to the common situation in
which there is a single independent variable x, and a single
dependent variable y. The independent variable is con-
trolled by the experimenter; the dependent variable is
measured. The relationship between the variables x and
y can be described by an equation that includes one or
more parameters, which we call A, B, C, etc.
Goal
Nonlinear regression is a procedure for fitting data to
any selected equation. As with linear regression, non-
linear regression procedures determine values of the
parameters that minimize the sum of the squares of the
distances of the data points to the curve. If they value
of each data point is called ydata and they value of the
curve is the goal is to minimize the residual
sum of squares (SS):
ss
Best
FITflNG CURVES TO DATA USING NONLINEAR REGRESSION 367
SS = sum[(ydata - ycu,-ve)2]
Because this criterion minimizes the sum of the square
of the distances, it is called a least-squares method; such
methods are appropriate when the experimental uncer-
tainty is Gaussian (normally distributed), and not
related to the values of x or y.
Unlike linear or polynomial regression, a nonlinear
regression problem cannot be solved in one step. Instead
the problem must be solved iteratively. An initial esti-
mate (first guess) of the value of each parameter must
be provided. The nonlinear regression procedure then
adjusts these values to improve the fit of the curve to the
data. It then adjusts those new values to improve the fit
again. These iterations continue until negligible, if any,
improvement occurs. Although iterative procedures are
not required, iterative nonlinear regression programs
may also be used with equations (such as polynomial
equations) in which y is linear with each parameter.
In the discussion that follows, we first consider only
equations that contain two parameters to be fit, A and
B. Accordingly, nonlinear regression can be viewed in
a three-dimensional topographical analogy (Fig. 3). In
this model, the horizontal plane represents A and B,
and the vertical axis represents the sum of squares.
Thus every pair of possible values for A and B is
associated with a single sum of squares value z. De-
pending on the data and the equation chosen, this sur-
face may be simple and symmetrical, or it may contain
numerous peaks and valleys. The goal of nonlinear
regression - to find the pair of values for A and B that
minimize the sum of squares - is to find the deepest val-
ley. Note that the data points do not individually
appear in this graph; instead goodness of fit (incorpo-
Initial
Value
A
Figure 3. A topographical analogy of nonlinear regression. This
three-dimensional graph plots the sum of squares SS (a measure of
goodness of fit) as a function of possible values for the two
parameters A and B. The purpose of nonlinear regression is to find
the values for A and B that minimize the sum of squares. This
occurs at the bottom of the valley. In this example the topographical
surface is simple; in other examples the surface may contain several
peaks and valleys. Note that the individual data points do not
appear on this type of plot.
rating all the data) is graphed as a function of each pair
of possible values for the parameters A and B.
When the equation has more than two parameters,
the topographical analogy contains more than three
dimensions and cannot be visualized. Nonetheless, the
mathematical principles are the same.
Algorithms
All iterative techniques must be given a starting point
on the surface - an initial estimate of the parameters
obtained by calculation or intelligent guessing. The
nonlinear regression procedure then iteratively moves
along the surface by altering the values of the param-
eters to improve the fit. Several methods can be used to
calculate these iterations. In pharmacological and bio-
chemical research, the Marquardt method is most com-
monly used. This method is a hybrid between two older
algorithms, the method of steepest descent and the
method of Gauss-Newton.
The method of steepest descent moves along the
direction of steepest descent with an arbitrary step
length. Then the slope is calculated in the new spot and
the procedure is repeated. Essentially this method finds
the minimum by repeatedly moving downhill. Although
the initial iterations rapidly advance toward the goal,
later iterations often zigzag, especially if the step length
is large. Many iterations are often required before the
method converges on a solution. In some implementa-
tions of this method, the step length is varied to hasten
the process.
The Gauss-Newton algorithm utilizes another prin-
ciple. With the topological analogy introduced above,
equations in which y is linearly dependent on A and B
always generate a single smooth ellipsoid crater. Other
equations generate an asymmetrical surface. The
Gauss-Newton method approximates the equation so
that it does generate a symmetrical ellipsoid surface.
From the initial position on that surface, the algorithm
can project the entire ellipsoid. It then alters the values
of A and B to jump straight to the minimum. With
functions linear with the parameters, therefore, the
Gauss-Newton method immediately finds the solution
without need for further iterations. Nonlinear func-
tions do not generate an ellipsoid surface, but that
simplifying assumption usually (but not always) leads
to an improvement of the fit. With further iterations,
the fit improves. The Gauss-Newton method sometimes
works poorly in the initial iterations, and it may move
in the wrong direction altogether, making the fit worse.
However, this method works well when close to the
minimum, where the surface usually can be well ap-
proximated by an ellipsoid.
Marquardt designed a method that combines the ad-
vantages of both the steepest descent and Gauss-Newton
methods, while avoiding their limitations (6). The
method of steepest descent works best in initial itera-
tions, and the Gauss-Newton method works best in later
iterations. Marquardts algorithm uses a blend of the
two methods. The steepest descent method is empha-
sized in initial steps, and this method is used repeatedly
368 MOTULSKY AND RANSNAS
as long as the residual sum of squares decreases con-
siderably. When the sum of squares is no longer de-
creasing, Marquardts method gradually switches over
to the Gauss-Newton principle. This approach has
been found to be useful for fitting many types of data
to various types of equations.
The simplex algorithm (reviewed in refs 7 and 8) is
an alternative method for performing nonlinear regres-
sion analysis. One must provide both an initial value
and an initial increment for each parameter. From these
values, the algorithm generates N + 1 starting points,
where N is the number of parameters to be fit. Each of
these starting points, called a vertex, consists of a pos-
sible value for each parameter. The goodness of fit of
each vertex is evaluated. The worst vertex is rejected,
and a new vertex is generated by blending the best of
the others. This algorithm repeatedly rejects the worst
vertex and generates a new one until the vertices
coalesce to values within a specified tolerance. On our
topological analogy, the simplex method starts with
three starting values that form a triangle. With each
iteration the algorithm rejects the worst of the points,
and creates a new one. This can be visualized as a tri-
angular amoeba oozing down the surface. Eventually
all three points merge together at the minimum. With
more parameters, the vertices form an N-dimensional
shape called a simplex. The simplex algorithm for non-
linear regression bears little relationship to the simplex
algorithm used for linear programming.
The simplex method has three advantages over the
more commonly used methods discussed above: 1) It is
fast. It does not require calculating derivatives. 2) It
rarely converges at a local minimum (see below). 3) It can
be used with noncontinuous functions. The disadvan-
tage of the simplex method is that it does not estimate
the standard error of each parameter. It is also some-
what more difficult to use, inasmuch as you must pro-
vide starting increments for each parameter.
CHOICES TO BE MADE WHEN USING A
NONLINEAR REGRESSION PROGRAM
Scaling the data
In theoretical mathematics, the units used to express
data are irrelevant. When using a computer to solve a
numerical problem, however, it is important to choose
appropriate units. Computers keep track of only a cer-
tain number of significant digits (which varies between
programs), and all computer calculations involving
floating point (noninteger) values are therefore subject
to round-off error. It is important, therefore, to express
data in reasonable units to avoid very large or very
small numbers. For example, round-off errors may be
severe if concentrations were entered as 0_1014 M, but
would be trivial if those data were expressed as 0-10
ftvl.
Equation
A nonlinear regression program cannot find the best
curve through a set of data points; it can only optimize
the parameters in a specified equation. The equation
must calculate y as a function of x and one or more
parameters. The function should be selected because it
describes a hypothetical physical or molecular model.
Instead of entering a function, some programs allow
one to enter a system of differential equations.
In providing the equation, one must make the distinc-
tion between constants (fixed by the experimenter), and
parameters that are to be fit by the program. This dis-
tinction may greatly affect the final results.
All nonlinear regression procedures (except simplex)
repeatedly calculate the derivative (slope) of y with
respect to each parameter. Some programs require one
to determine the derivatives using calculus and to enter
the resulting equations. Other programs calculate the
derivatives numerically by evaluating the equation be-
fore and after altering the value of a parameter by a
small amount , and then dividing the change in y by
i. Although the calculations are slower, this latter
method is of more general utility.
When writing an equation to express a physical
model, parameters may be entered in various ways.
Thus acidity may enter an equation as [H] or as pH.
Similarly, a binding affinity may enter an equation as
either an equilibrium dissociation constant (KD, ex-
pressed in concentration units), the logarithm of the
KD, or the reciprocal of the KD (an equilibrium associ-
ation constant, expressed in reciprocal concentration
units). Switching between alternative forms is called
reparameterization. Note that this is quite different
from the linear transformations mentioned earlier; in
those cases the variables x andy were transformed. In
reparameterization, the parameters (A, B, . - .) are
transformed without altering the x and y values.
Reparameterization is useful because many of the
statistical inferences that can be made from nonlinear
regression are strictly valid only for equations in which
y is linear with respect to the parameters. Reparameter-
izing a nonlinear equation can make it behave more
nearly linear, and thus can improve the validity of
statistical values. Unfortunately it is quite difficult to
measure the extent of nonlinearity of an equation (9).
This makes it difficult to ascertain the effectiveness of
reparameterization.
Initial estimates
Initial values for each parameter in the equation must
be specified, based on previous experience, on prelimi-
nary analyses based on linear transformations, or on a
hunch. Although it is impossible to give any general
guidelines, it is usually not difficult to estimate the
parameters if one understands the physical model that
the equation represents, and understands the meaning
of each parameter. A poor selection of initial values
may have several consequences: 1) In a well-behaved
system, the amount of computer time required to reach
a solution will be increased, but the solution will be cor-
rect. 2) In other situations, poorly selected initial values
can lead the nonlinear regression program to go in the
wrong direction and never converge on a solution. 3) It
FITTING CURVES TO DATA USING NONLINEAR REGRESSION 369
is also possible that poorly selected initial values can
cause the program to converge on the wrong solution,
a local minimum (see below). The choice of initial
values is less important with equations containing only
one or a few parameters than with equations containing
many parameters.
Weighting scheme
Most common nonlinear regression schemes, like
linear regression, minimize the sum of the square of the
vertical distances of the points from the curve. This ap-
proach is statistically valid when the experimental un-
certainties do not relate in a systematic way to the
values of x ory. Often this is not true. For example, in
many experimental situations the experimental uncer-
tainty is (on the average) a constant fraction of the
value of y. In these situations, the usual regression
methods are not optimal, inasmuch as the points with
largey values tend to be further from the curve than are
points with small y values. In minimizing the sum of
squares, the program would therefore tend to relatively
emphasize those points with large y values, and ignore
points with smally values. To circumvent this problem,
the procedure for quantitating goodness of fit can be
altered, so that the deviation of each point from the
curve is divided by they value of that point, and then
squared. With this modification, the algorithm will not
minimize the square of the distances (expressed in the
units of they variable) of the points from the curve, but
instead will sum the square of the relative distances (ex-
pressed as fractions). This option, commonly available
in regression programs, is termed weighting the data
points. In fact this procedure reduces the inappropriate
weight given the points with largey values, and should
be thought of as unweighting.
The concept of differential weighting of data points
can be somewhat confusing. It may help to consider
instead the criteria used to measure how far a point is
from the curve. Normally this is measured as a distance,
expressed in units of the y axis. With the weighting
scheme described above, the relative distance, expressed
as a fraction, is used instead.
Other choices for differentially weighting the data
points are commonly used. One possibility is for the
distance of the point from the curve to be divided by the
square root ofy, and then squared; this is a blend of the
two methods listed above. Alternatively, it is common
to weight data points according to the number of repli-
cate determinations. Thus points that are the mean of
five replicates will be weighted more than points that
are the mean of only two replicate determinations.
Another possibility is to inversely weight each point
according to the standard deviation of the replicate
values that went into determining that point. Thus data
points with tight replicates will be weighted more heav-
ily than points with a great deal of variability.
When in doubt, it is best to weight each point
equally. Use differential weighting only when the rela-
tionship between the experimental uncertainties and
the value ofy is clear. It is not appropriate to weight the
data points by l/y or 1/y2 when some of they values are
equal to, or nearly equal to, zero, as commonly occurs
when one defines y as an experimentally determined
value minus a baseline (or nonspecific) value.
Convergence criterion
As nonlinear regression calculations occur iteratively,
the computer must be told when to stop-when the cal-
culations have converged. Some programs have built-in
criteria; other programs are more flexible. For example,
some nonlinear regression programs terminate when
an iteration reduces the sum of squares by less than one
part in a thousand. Others terminate when an iteration
alters the value of each variable by less than one part
in a thousand. Too loose a criterion can cause a pro-
gram to stop before it has reached the best fit. Too tight
a criterion can consume a great deal of computer time.
INTERPRETING THE RESULTS
Nonlinear regression programs typically spew out
several pages of information about the final equation.
This information is designed to answer several ques-
tions: How well does the model fit the data? Does this
model fit the data better than an alternative model?
How much uncertainty is there in the values of the
parameters? Does the equation fit this set of data differ-
ently from another set of data?
Assessing goodness of fit
In assessing goodness of fit, it is essential to first examine
a graph of a curve superimposed on the data points.
Many potential problems are easiest to spot graphical-
ly. It is inappropriate to use the results of a nonlinear
regression program without first examining a graph of
the data together with the fit curve. In addition to view-
ing the graph, several statistical methods can be used
for quantitating goodness of fit.
The average deviation of the curve from the points is
the square root of SS/df, where SS is the sum of
squares, and df is the degrees of freedom. This is also
called the root mean square (RMS). If goodness of fit
was quantitated using the actual deviation of the points
from the curve (no weighting), then the RMS will be
expressed in the units used by the y values.
In a residual plot, the y value of each point is
replaced by the distance of that point from the curve.
Two examples are shown in Fig. 4. If the equation is ap-
propriate for the data, the residuals represent only ex-
perimental error. Accordingly, the residuals should not
be systematically related to the x values, and the
residual plot will have a random arrangement of posi-
tive and negative residuals. If the equation is in-
appropriate, the positive residuals may tend to cluster
together at some parts of the graph, whereas negative
residuals cluster together at other parts. Such clustering
indicates that the data points differ systematically (not
just randomly) from the predictions of the curve.
0.5
-0.5
0.5
0 5 10
-0.5
15 0 5
l0
370
MOTULSKY AND RANSNAS
HyprboIo Rasiduals
Expon2ritIa Rasiduals
Figure 4. Residual plots. The solid line shows the results of non-
linear regression using the equation describing a rectangular hyper-
bola, y = Ax/(B + x). The dotted line is the result of nonlinear
regression using an equation describing an exponential association,
y = A[l - exp(-Bx)1. Which equation better describes the data? It
is difficult to answer that question by examining the curves on the
left. The residual plots on the right, however, make the answer clear,
The residual plots show how far each point is from the curve. The
residuals from the rectangular hyperbola are randomly above and
below zero. The residuals from the exponential equation, however,
show a systematic pattern, with positive residuals for the first and
last few points and negative residuals for the middle points. Such
systematic deviations indicate that the data are not well-described
by that equation. If these were real data, the choice between the two
equations would be based not only on goodness of fit, but also on
the physical meaning of the two equations.
The runs test is a simple and robust method for
determining whether data differ systematically from
a theoretical curve. A run is a series of consecutive
points with a residual of the same sign (positive
or negative). The runs test statistic is calculated
from the number of runs associated with a particular fit
of the data. For example the residuals of fit of the ex-
ponential equation in Fig. 4 have the following signs:
++ ++-++. Thus these 17 points
have only 5 runs. This is associated with a P value of
< 0.05 as determined from an appropriate table (9). A
low P value indicates that the curve deviates sys-
tematically from the points. In contrast the residuals
from the fit of those data to the hyperbolic equation
have 11 runs.
The r2 value (square of the correlation coefficient)
often accompanies the results of linear regression, and
most scientists have developed a good intuitive grasp of
its meaning. This value represents the fraction of the
overall variance of the y values that is reduced (or ex-
plained) by the line. The r2 value is traditionally
defined only for linear, and not for curved, relation-
ships. Nonetheless it can easily be calculated after non-
linear regression as the fraction of the variance of the
y values (from their overall mean) that is reduced (or
explained) by the curve. In a perfect fit, r2 = 1.00; in a
very poor fit, r2 = 0.00.
Uncertainty in the value of the parameters
After obtaining the converged values of the parameters,
one wants to know the reliability of those values. Non-
linear regression programs generally print out estimates
of the standard error of the parameters, but these values
should not be taken too seriously. In nonlinear functions,
errors are neither additive nor symmetrical, and exact
confidence limits cannot be calculated. The reported
standard error values are based on linearizing assump-
tions, and will always underestimate the true uncer-
tainty of any nonlinear equation. The extent to which
a particular error value is underestimated depends on
the particular equation and data being analyzed. As we
will discuss below, reparameterizing the equation may
improve the accuracy of the error estimates. In a non-
linear equation, the uncertainty of all the parameters
will be underestimated, even parameters that are linear
withy [for example, A in the equationy = Aexp(Bx)].
For the reasons given above, it is not appropriate to
use the standard error values printed by a nonlinear
regression program in further formal statistical calcula-
tions. They are quite useful, however, as an informal
measure of the goodness of a fit. Because the estimated
standard errors are always underestimates, large error
values point to a real problem. Large standard errors
will occur in several situations: when the data points
have a lot of scatter, when the parameters are highly
correlated or redundant (see below), or when data have
not been collected over a large enough range of x
values.
We have found another approach useful for qualita-
tively demonstrating the certainty of the values deter-
mined by a nonlinear regression program. Along with
the data points and the best-fit curve, one may super-
impose various other curves calculated by changing the
value of one of the parameters altered by a specified
amount (see Fig. 5).
Figure 5. A qualitative method for demonstrating the uncertainty
in the values of a variable determined by nonlinear regression. The
data points are the same as those in Fig. 4. The solid curve was
determined by nonlinear regression using an equation describing a
rectangular hyperbola, y = Ax/(B + x). The best-fit values deter-
mined by the nonlinear regression program (without weighting)
were A 10.00 0.12 and B = 2.00 0.10. Inasmuch as those
uncertainties are estimated standard errors, the 95% confidence
limits for the value of B is between 1.8 and 2.2. However, the error
values determined by nonlinear regression are not precise. To
understand intuitively the uncertainty of the value of B, the dotted
curves were computer generated leaving A = 10.0, but setting B to
either 1.7 (top) or 2.3 (bottom). Clearly these curves do not fit the
data well, and B must therefore be between 1.7 and 2.3.
FITTING CURVES TO DATA USING NONLINEAR REGRESSION 371
This approach can be extended systematically in the
Monte Carlo method. First the data are fitted using any
of the algorithms. Then a set of ideal data are created
using the same x values as the actual data, but replacing
they values with values predicted by the best-fit curve.
Next pseudoexperimental data are generated by adding
random error to the ideal data points. This is done by
adding to each point a random number calculated from
a Gaussian distribution with a standard deviation equal
to the mean standard deviation of replicate values in
the actual data. Multiple sets of pseudo data should be
generated in this manner, and each should be analyzed
by nonlinear regression. The standard deviation of esti-
mates of a parameter from multiple sets of pseudo data
is a reasonable estimate of the error of the value of that
parameter determined from analysis of the actual data.
This method is especially useful when used in conjunc-
tion with the simplex method, which does not otherwise
provide any estimate of the errors.
In both linear and nonlinear regression, the param-
eters are usually not entirely independent. Thus, alter-
ing one parameter will make the fit worse, but this can
be partially offset by adjusting the value of another
variable. The degree to which two parameters depend
on one another is shown in the covariance matrix,
which is often reported as correlation coefficients for
every pair of parameters. A correlation of zero means
that two parameters are completely independent; a
correlation of 1.0 or -1.0 means that the two param-
eters are redundant (see below). Parameters in most
models are correlated, and correlation coefficients as
large as 0.8 are commonly seen. Very high correlations
can indicate that the equation includes redundant vari-
ables, that the data points do not span a sufficiently
large range of x values to define a curve well, or that too
few data points are collected. The joint distribution of
the confidence limits of two parameters is sometimes
presented as a confidence ellipse, as shown in Fig. 6.
I-
0
U,
Value of A
Figure 6. A confidence ellipse. The values of the parameters
reported by a nonlinear regression program are not independent of
one another. Thus increasing the value of one parameter may
worsen the fit, but altering the value of another parameter may
partially offset this decrease. The joint distribution of two
parameters can be plotted as a confidence ellipse. In the example
shown the 95% confidence limits of A and B are plotted. Any pair
of values of A and B can be plotted as a single point, and we are
95% certain that the true values of A and B lie within the ellipse.
Comparing two fits to one set of data
The sum of squares value allows one to compare two
fits to the same data. Such a comparison is meaningful
only when the data have not been changed or trans-
formed between fits, and when the same weighting
scheme was used for each fit. This allows one to com-
pare two different equations (models) fit to the same
data. Clearly, the selection of a model to explain a par-
ticular set of data should not, be based entirely on
statistical measures. More important are the physical
plausibility of the model, and its consistency with other
types of data.
Comparing two models with the same number of
parameters is easy: the fit with the lower sum of squares
is superior, for its curve lies closer to the points. The
statistical significance is obtained by F = SS1/SS2,
where both numerator and denominator have N - V
degrees of freedom (SS = residual sum of squares of
each fit, N = number of data points; V = number of
parameters fit by the program).
Comparing two models with a different number of
parameters is less straightforward because increasing
the number of parameters gives more flexibility to the
curve-fitting procedure, and almost always leads to a
curve that is closer to the points. The question is
whether the improved fit is worth the cost (in lost
degrees of freedom) of the additional parameter or
parameters. This question is usually answered statisti-
cally by performing an F test with the following
equation:
F = (SS1 - SS2)/(df1 - df2)
SS2/df2
Here SS refers to the sum of squares, and df refers to
the number of degrees of freedom (number of data
points minus number of parameters). The subscript I
refers to the fit with fewer parameters, the more simple
model. A P value is obtained from the F value by con-
sulting a standard table using (df1 - df2) and df2
degrees of freedom. A small P value indicates that the
more complex model (with more parameters) fits the
data significantly better than the simpler model.
Comparing data sets
Usually one wishes to fit two sets of data to the same
general model, and use the results to determine whether
the two sets of data differ significantly. For example,
one may perform dissociation rate experiments under
two different conditions and ask whether the resulting
exponential dissociation rate constants differ signifi-
cantly. Despite the fact that this is a common problem,
there is no clear consensus for its solution. One intui-
tively straightforward method is not appropriate. As
discussed above, it is not proper to use the standard
error values reported by the nonlinear regression pro-
gram to compare two models, as those standard error
values are underestimates of the actual uncertainty.
372 MOTULSKY AND RANSNAS
One approach is to repeat the experiment several
times, and to compare the resulting parameters using
paired I tests. We prefer this method and use it in our
work inasmuch as it is straightforward to calculate and
easy to explain to others. However, this method does
not use all available information for each experiment;
it uses only the estimated value of each parameter
without taking into account its standard error. Accord-
ingly, the method is statistically conservative, and the
resulting P value may be too high. Thus small differ-
ences may be missed by this method, especially if the
experiment has been performed only two or three times.
A more powerful approach can also be used (10).
First the two sets of data are analyzed separately. The
overall values for the sum of squares for the two sets of
data analyzed separately are the sums of the individual
values from each fit. Similarly, the number of degrees
freedom is the sum of the values from each fit:
SS5 = SS1 + SS2 dfsep = df1 + df2
Next the two sets of data are pooled (combined) and
analyzed simultaneously. This pooled fit yields values
for SS001 and df01. The question is whether the sep-
arate fit is significantly better than the pooled fit. The
significance of the improvement is determined from the
F ratio calculated as
F = (SS001 - SSsep)/(dfpooi - df,)
SSsep/dtsep
To interpret the meaning of this F value, a statistical table
is used to convert to a P value. In using such a table,
the numerator has (df1 - dfsep) degrees of freedom;
the denominator has dfsep degrees of freedom. A large
F value (with corresponding low P value) indicates that
the separate fit is much better than the pooled fit - that
the two sets of data are not well fit by one curve.
The t test method mentioned above is used to deter-
mine one P value from a series of paired experiments.
With this method, N refers to the number of experi-
ments. With the F test method, a P value is calculated
from one paired experiment. With this method, N
refers to the number of data points in each half of that
experiment. Although more powerful, this method is
potentially misleading inasmuch as an extremely small
P value can come from a single experiment. Despite a
low P value, such results cannot be considered to be sig-
nificant until the experiment is repeated.
Linear assumptions applied to nonlinear models
The methods used for calculating standard errors of the
parameters and for comparing two fits are strictly valid
only when applied to equations in whichy is linear with
each of the parameters. With other equations, the un-
certainties do not follow a Gaussian distribution. Thus
the calculated confidence limits are only approximate.
Moreover, the sums of squares also do not follow a
Gaussian distribution, and statistical inferences based
on comparisons of sums of squares are also only ap-
proximate. In a practical sense, these problems rarely
have a substantial impact on conclusions based on non-
linear regression.
PROBLEMS
Nonconvergence
Sometimes the calculations terminate without con-
verging because of overflow or underfiow (a number
becomes too large or too small for the computer to
handle), or because the system becomes ill-conditioned
or a matrix becomes singular. This occurs in several
situations: 1) The data contain numbers that are too
large or too small. 2) The selected equation does not
reasonably fit the data. 3) The initial values are far
from correct. 4) The data points are quite scattered.
5) The data were not collected over a sufficiently wide
range of x values. 6) The computer calculations were
not sufficiently precise (not enough significant digits).
Slow convergence
Nonlinear regression analyses usually converge in fewer
than 5-10 iterations, even when the initial estimates are
far from the final values. Slow convergence indicates
that the program is having difficulty finding a solution.
This may occur for several reasons: 1) The convergence
criterion is too stringent. 2) The data are not adequate
to define the parameters; more data points or more
widely spaced points are required. 3) The equation
contains too many parameters; it may help to fix one
or more parameters as a constant. 4) The equation
needs to be reparameterized, as discussed above.
Redundant variables
Nonlinear regression fails when one tries to fit data to
an equation with too many parameters. For example, it
would be a mistake to try to fit exponential growth data
to the equationy = Aexp(Bx + C), as the parameters A
and C are essentially interchangeable, both representing
they intercept. {This can be seen by rewriting the equa-
tion asy = [Aexp(C)]exp(Bx); only a single parameter
is needed to represent they intercept, and a nonlinear
regression program has no basis for deciding how to
apportion that value between A and C.) With a compli-
cated equation, it can be difficult to detect such redun-
dancy. Consider this possibility if the reported standard
error values of the parameters are extremely large, or
if the nonlinear regression program consistently fails to
converge on a solution.
Local minima
In nonlinear regression, it is possible that the final (con-
verged) fit is not the best possible fit. Recall that non-
linear regression makes relatively small changes during
each iteration, and stops when making such small
ss
K Local Best
Minimum Fit
FITTING CURVES TO DATA USING NONLINEAR REGRESSION
373
changes no longer improves the fit of the equation to
the data. It is possible, however, that making much
larger changes in the values of the variables would lead
to a much better fit. This is usually not a problem when
only one or two variables are being fit; it is more likely
to occur when many variables are being fit. This
problem is known as finding a local minimum. In our
topographical analogy, the goal is to find the spot with
the lowest altitude. The nonlinear regression procedures
nearly always move downhill, so it is possible that it
converges at the bottom of one valley without finding a
much deeper valley just over a ridge (Fig. 7). To guard
against this problem, one can repeat nonlinear regres-
sion several times, using different starting values. More
important, it helps to have a good intuitive grasp of the
equation used.
MODIFICATIONS TO NONLINEAR
REGRESSION
Robust nonlinear regression
Most statistical analyses, including nonlinear regres-
sion, are based on the assumption that the experimen-
tal variation of each data point is the sum of many
small random inaccuracies. From this assumption
comes the notion that experimental data ought to be
distributed in a Gaussian (normal) fashion. Accord-
ingly, about 5% of the data points should lie more than
two standard deviations from the mean, and fewer than
1 in 10,000 data points should be more than four stan-
dard deviations from the mean. In real life, outliers
occur more commonly than that! This is because ex-
perimental variation results not only from the sum of
small inaccuracies, but also from larger mistakes. As
used by statisticians, the term experimental error does
Initial
Value
V
A
Figure 7. Local minimum. This figure is similar to Fig. 3. The goal
of nonlinear regression is to find the values of the parameters that
minimize the sum of squares. With reasonable starting values, non-
linear regression algorithms nearly always move stepwise downhill
until the minimum (best fit) is found, With an inappropriate start-
ing place, however, a local minimum might be found instead.
not include mistakes. For example, the Gaussian distri-
bution accounts for the imprecision of pipettes; it does
not account for the fact that the experimenter may ac-
cidentally include a large bubble, thus reducing the
volume of reagent.
Several approaches can be used to deal with outlier.
One approach is to let them be - to include all points
in the analysis. Although this seems like a compulsive
and conservative approach, it may lead to incorrect
results, as outlier can have a big influence on the result-
ing curve. The square of the distance of an outlying
point from the curve can be huge, and the nonlinear
regression will twist the curve considerably to reduce
that distance (see Fig. 8). A more common approach is
to delete outlying points, either on an ad hoc or a
systematic basis.
Another approach for dealing with outlier is to
modify the criteria used to assess goodness of fit to
reduce the weight given outlying points and thus make
the nonlinear regression process more robust. Several
such methods have been described. Each method is de-
signed to be identical to the sum of squares method for
most of the points, but to reduce the importance of out-
lying points. Each such method requires that you choose
a robustness constant c, a deviation (z, the absolute
value of the distance of a point from the curve) beyond
which a point is likely to be invalid. Wahrendorf pro-
posed that the sum of squares (z2) be replaced by
2Cz - C2, when z > C (11). Thus points that are close
to the curve (z < C) are treated as usual, but as the
deviations increase their contribution will be propor-
tional to the distance, not the square, of the distance.
Mosteller and Tukey proposed an alternate scheme
where points with z > C are ignored, and the other
points are used to minimize the sum of [1 - (z/C)2]2 (12).
Extended nonlinear regression
We discussed earlier the importance of choosing the
appropriate weighting scheme, as nonlinear regression
will converge on the correct solution only when the
points are appropriately weighted according to the ex-
pected experimental uncertainties. To account for com-
mon patterns of experimental uncertainty, the sum of
squares is often calculated as the sum of (D2/y),
where D is the distance of the point from the curve, y
is they value of the point, and Nis 0, 1, or 2. If n is zero,
this method is the same as ordinary sum of squares; it
sums the square of the distances of the points from the
curve. When N = 2, this equation reduces to (D/y)2,
which sums the square of the relative distance (as a
fraction) of the points from the curve. Setting N = 1 is
a blend between using absolute and relative distances.
In many experimental contexts it is difficult to know
what value to give N, and N does not even have to be
given an integer value. The idea of extended nonlinear
regression is to let the regression procedure fit the value
of N as well as the other parameters in the equation. In
two simulations of pharmacokinetics, extended non-
linear regression was found to work as well as ordinary
weighted nonlinear regression using the appropriate
374 MOTULSKY AND RANSNAS
Figure 8. Effects of an outlier. These graphs clearly show how a
single outlier point can greatly distort the curve determined by non-
linear regression. Outlier are more influential in data sets with few
points (right). Except for the outlier, the data are the same as in
Figs. 4 and 5.
value for N and much better than ordinary regression
given the wrong value for N (13, 14). Because this
method increases the number of parameters to be fit by
regression procedure, it will be less useful when the
number of data points is small.
PERSPECTIVE
Nonlinear regression is a powerful technique for data
analysis. Although nonlinear regression calculations
cannot be reasonably performed by hand, computer
programs that perform these calculations are available
for microcomputers found in many laboratories. Non-
linear curve-fitting analyses are easy, fast, and practical
for routine use. No sophisticated knowledge of com-
puters or mathematics is required to use such programs.
To use nonlinear regression properly, however, it is
necessary to have an intuitive feel for the selected equa-
tion and for the physical model it represents. Without
such an understanding, the use of nonlinear regression
will be frustrating and potentially misleading. Results
from a nonlinear regression program should be careful-
ly scrutinized and viewed graphically. Like every scien-
tific technique, a nonlinear regression program may
produce misleading results when used inappropriately.
Computers are associated with a powerful mystique,
and incorrect or incomplete computer analyses are
sometimes published and accepted uncritically. We
hope that this review will demystify nonlinear regres-
sion, so that both its power and limitations will be
appreciated.
This work was supported by a grant from the National Institutes
of Health and the Swedish Medical Research Council.
REFERENCES
1. LANDAW, E. M.; DISTEFANA, A. J. Multiexponential,
multicompartmental, and noncompartmental model-
ling. II. Data analysis and statistical considerations.
Am. J. Physiol. 246: R665-R677; 1984.
2. DUGGLEBY, R. G. A nonlinear regression program for
small computers. Anal. Biochem. 110: 9-18; 1981.
3. JENNRICH, R. I.; RALSTON, M. L. Fitting nonlinear
models to data. Annu. Rev. Biophys. Bioeng. 8: 195-238;
1979.
4. DRAPER, N. R.; SMITH, H. An introduction to non-
linear estimation. Applied regression analysis. New York:
Wiley; 1981: 458-517.
5. PRESS, W. H.; FLANNERY, B. P.; TEUKOLSKY, S. A.;
VErrERLING, W. T. Modelling of data. Numerical
recipes: the art of scien4fic computing. Cambridge: Cam-
bridge University Press; 1986: 498-546.
6. CACECI, M. S.; CACHERIS, W. P. Fitting curves to data.
The simplex algorithm is the answer. Byte 9: 340-362;
1984.
7. PRESS, W. H.; FLANNERY, B. P.; TEUKOLSKY, S. A.;
VETTERLING, W. T. Minimization or maximization of
functions. Numerical recipes: theart ofscien4/ic computing.
Cambridge: Cambridge University Press; 1986: 289-
293.
8. MARQUARDT, D. W. An algorithm for least-squares esti-
mation of nonlinear parameters. J. Soc. md. Appi.
Math. 11: 431-444; 1963.
9. BEYER, W. H. Handbook of tables for probability and statis-
tics. Cleveland: CRC Press; 1968: 397-398.
10. RATKOwSKY, D. Comparing parameter estimates from
more than one data set. Nonlinear regression modelling: a
unifiedand practical approach.New York: Dekker; 1983:
135-152.
11. WAHRENDORF, J. The application of robust nonlinear
regression methods for fitting hyperbolic scatchard
plots. mt. J. Bio-Med. Comput. 10: 75-87; 1979.
12. MOSTELLER, F.; TUKEY, J. W. Data analysis and regression.
New York: Addison-Wesley; 1977.
13. PECK, C. C.; BEAL, S. L.; SHEINER, L. B.; NICHOLS,
A. I. Extended least squares nonlinear regression: a
possible solution to the choice of weights problem in
analysis of individual pharmacokinetic data. J. Phar-
macokinet.Biopharm. 12: 545-558; 1984.
14. THOMSON, A. H.; KELMAN, A. W.; WHIrrING, B.
Evaluation of nonlinear regression with extended least
squares: simulation study. J. P/zarm. Sd. 74:
1327-1331; 1985.