NCSS Statistical Software NCSS.

com

Chapter 300

Linear Regression and Correlation

Introduction
Linear Regression refers to a group of techniques for fitting and studying the straight-line relationship
between two variables. Linear regression estimates the regression coefficients 𝛽𝛽0 and 𝛽𝛽1 in the equation

𝑌𝑌𝑗𝑗 = 𝛽𝛽0 + 𝛽𝛽1 𝑋𝑋𝑗𝑗 + 𝜀𝜀𝑗𝑗

where X is the independent variable, Y is the dependent variable, 𝛽𝛽0 is the Y intercept, 𝛽𝛽1 is the slope, and 𝜀𝜀 is
the error.

In order to calculate confidence intervals and hypothesis tests, it is assumed that the errors are
independent and normally distributed with mean zero and variance 𝜎𝜎 2 .
Given a sample of N observations on X and Y, the method of least squares estimates 𝛽𝛽0 and 𝛽𝛽1 as well as
various other quantities that describe the precision of the estimates and the goodness-of-fit of the straight
line to the data. Since the estimated line will seldom fit the data exactly, a term for the discrepancy between
the actual and fitted data values must be added. The equation then becomes

𝑦𝑦𝑗𝑗 = 𝑏𝑏0 + 𝑏𝑏1 𝑥𝑥𝑗𝑗 + 𝑒𝑒𝑗𝑗

= 𝑦𝑦�𝑗𝑗 + 𝑒𝑒𝑗𝑗

300-1
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

where j is the observation (row) number, 𝑏𝑏0 estimates 𝛽𝛽0 , 𝑏𝑏1 estimates 𝛽𝛽1 , and 𝑒𝑒𝑗𝑗 is the discrepancy between
the actual data value 𝑦𝑦𝑗𝑗 and the fitted value given by the regression equation, which is often referred to as
𝑦𝑦�𝑗𝑗 . This discrepancy is usually referred to as the residual.
Note that the linear regression equation is a mathematical model describing the relationship between X and
Y. In most cases, we do not believe that the model defines the exact relationship between the two variables.
Rather, we use it as an approximation to the exact relationship. Part of the analysis will be to determine how
close the approximation is.
Also note that the equation predicts Y from X. The value of Y depends on the value of X. The influence of all
other variables on the value of Y is lumped into the residual.

Correlation
Once the intercept and slope have been estimated using least squares, various indices are studied to
determine the reliability of these estimates. One of the most popular of these reliability indices is the
correlation coefficient. The correlation coefficient, or simply the correlation, is an index that ranges from -1 to
1. When the value is near zero, there is no linear relationship. As the correlation gets closer to plus or minus
one, the relationship is stronger. A value of one (or negative one) indicates a perfect linear relationship
between two variables.
Actually, the strict interpretation of the correlation is different from that given in the last paragraph. The
correlation is a parameter of the bivariate normal distribution. This distribution is used to describe the
association between two variables. This association does not include a cause-and-effect statement. That is,
the variables are not labeled as dependent and independent. One does not depend on the other. Rather,
they are considered as two random variables that seem to vary together. The important point is that in
linear regression, Y is assumed to be a random variable and X is assumed to be a fixed variable. In
correlation analysis, both Y and X are assumed to be random variables.

Possible Uses of Linear Regression Analysis

Montgomery (1982) outlines the following four purposes for running a regression analysis.

Description
The analyst is seeking to find an equation that describes or summarizes the relationship between two
variables. This purpose makes the fewest assumptions.

Coefficient Estimation
This is a popular reason for doing regression analysis. The analyst may have a theoretical relationship in
mind, and the regression analysis will confirm this theory. Most likely, there is specific interest in the
magnitudes and signs of the coefficients. Frequently, this purpose for regression overlaps with others.

300-2
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Prediction
The prime concern here is to predict the response variable, such as sales, delivery time, efficiency,
occupancy rate in a hospital, reaction yield in some chemical process, or strength of some metal. These
predictions may be very crucial in planning, monitoring, or evaluating some process or system. There are
many assumptions and qualifications that must be made in this case. For instance, you must not extrapolate
beyond the range of the data. Also, interval estimates require that normality assumptions to hold.

Control
Regression models may be used for monitoring and controlling a system. For example, you might want to
calibrate a measurement system or keep a response variable within certain guidelines. When a regression
model is used for control purposes, the independent variable must be related to the dependent variable in
a causal way. Furthermore, this functional relationship must continue over time. If it does not, continual
modification of the model must occur.

Assumptions
The following assumptions must be considered when using linear regression analysis.

Linearity
Linear regression models the straight-line relationship between Y and X. Any curvilinear relationship is
ignored. This assumption is most easily evaluated by using a scatter plot. This should be done early on in
your analysis. Nonlinear patterns can also show up in residual plot. A lack of fit test is also provided.

Constant Variance
The variance of the residuals is assumed to be constant for all values of X. This assumption can be detected
by plotting the residuals versus the independent variable. If these residual plots show a rectangular shape,
we can assume constant variance. On the other hand, if a residual plot shows an increasing or decreasing
wedge or bowtie shape, nonconstant variance (heteroscedasticity) exists and must be corrected.
The corrective action for nonconstant variance is to use weighted linear regression or to transform either Y
or X in such a way that variance is more nearly constant. The most popular variance stabilizing transformation
is the to take the logarithm of Y.

Special Causes
It is assumed that all special causes, outliers due to one-time situations, have been removed from the data.
If not, they may cause nonconstant variance, nonnormality, or other problems with the regression model.
The existence of outliers is detected by considering scatter plots of Y and X as well as the residuals versus X.
Outliers show up as points that do not follow the general pattern.

Normality
When hypothesis tests and confidence limits are to be used, the residuals are assumed to follow the normal
distribution.

300-3
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Independence
The residuals are assumed to be uncorrelated with one another, which implies that the Y’s are also
uncorrelated. This assumption can be violated in two ways: model misspecification or time-sequenced data.
1. Model misspecification. If an important independent variable is omitted or if an incorrect functional
form is used, the residuals may not be independent. The solution to this dilemma is to find the
proper functional form or to include the proper independent variables and use multiple regression.
2. Time-sequenced data. Whenever regression analysis is performed on data taken over time, the
residuals may be correlated. This correlation among residuals is called serial correlation. Positive
serial correlation means that the residual in time period j tends to have the same sign as the
residual in time period (j - k), where k is the lag in time periods. On the other hand, negative serial
correlation means that the residual in time period j tends to have the opposite sign as the residual in
time period (j - k).
The presence of serial correlation among the residuals has several negative impacts.
1. The regression coefficients remain unbiased, but they are no longer efficient, i.e., minimum variance
estimates.
2. With positive serial correlation, the mean square error may be seriously underestimated. The impact
of this is that the standard errors are underestimated, the t-tests are inflated (show significance
when there is none), and the confidence intervals are shorter than they should be.
3. Any hypothesis tests or confidence limits that require the use of the t or F distribution are invalid.
You could try to identify these serial correlation patterns informally, with the residual plots versus time. A
better analytical way would be to use the Durbin-Watson test to assess the amount of serial correlation.

Technical Details

Regression Analysis
This section presents the technical details of least squares regression analysis using a mixture of summation
and matrix notation. Because this module also calculates weighted linear regression, the formulas will
include the weights, 𝑤𝑤𝑗𝑗 . When weights are not used, the 𝑤𝑤𝑗𝑗 are set to one.

Define the following vectors and matrices.

𝑦𝑦1 1 𝑥𝑥1 𝑒𝑒1 1

⎡ ⋮ ⎤ ⎡⋮ ⋮ ⎤ ⎡⋮⎤ ⎡⋮⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ 𝑏𝑏
𝒀𝒀 = ⎢ 𝑦𝑦𝑗𝑗 ⎥ , 𝑿𝑿 = ⎢1 𝑥𝑥𝑗𝑗 ⎥ , 𝒆𝒆 = ⎢ 𝑒𝑒𝑗𝑗 ⎥ , 𝟏𝟏 = ⎢1⎥ , 𝒃𝒃 = � 0 �
𝑏𝑏1
⎢ ⋮ ⎥ ⎢⋮ ⋮ ⎥ ⎢⋮⎥ ⎢⋮⎥
⎣𝑦𝑦𝑁𝑁 ⎦ ⎣1 𝑥𝑥𝑁𝑁 ⎦ ⎣𝑒𝑒𝑁𝑁 ⎦ ⎣1⎦

𝑤𝑤1 0 0 ⋯ 0
⎡0 ⋱ 0 0 ⋮ ⎤
⎢ ⎥
𝑾𝑾 = ⎢ 0 0 𝑤𝑤𝑗𝑗 0 0⎥
⎢ ⋮ 0 0 ⋱ 0⎥
⎣0 ⋯ 0 0 𝑤𝑤𝑁𝑁 ⎦

300-4
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Least Squares
Using this notation, the least squares estimates are found using the equation.

𝒃𝒃 = (𝑿𝑿′ 𝑾𝑾𝑾𝑾)−𝟏𝟏 𝑿𝑿′ 𝑾𝑾𝑾𝑾

Note that when the weights are not used, this reduces to

𝒃𝒃 = (𝑿𝑿′ 𝑿𝑿)−𝟏𝟏 𝑿𝑿′ 𝒀𝒀

The predicted values of the dependent variable are given by

� = 𝒃𝒃′ 𝑿𝑿
𝒀𝒀

The residuals are calculated using

�
𝒆𝒆 = 𝒀𝒀 − 𝒀𝒀

Estimated Variances
An estimate of the variance of the residuals is computed using

𝒆𝒆′ 𝑾𝑾𝑾𝑾
𝑠𝑠 2 =
𝑁𝑁 − 2

An estimate of the variance of the regression coefficients is calculated using

𝑏𝑏0 𝑠𝑠𝑏𝑏20 𝑠𝑠𝑏𝑏0 𝑏𝑏1

𝑉𝑉 � � = � �
𝑏𝑏1 𝑠𝑠𝑏𝑏0 𝑏𝑏1 𝑠𝑠𝑏𝑏21

= 𝑠𝑠 2 (𝑿𝑿′ 𝑾𝑾𝑾𝑾)−1

An estimate of the variance of the predicted mean of Y at a specific value of X, say 𝑋𝑋0 , is given by

1
𝑠𝑠𝑌𝑌2𝑚𝑚|𝑋𝑋0 = 𝑠𝑠 2 (1, 𝑋𝑋0 )(𝑿𝑿′ 𝑾𝑾𝑾𝑾)−1 � �
𝑋𝑋0

An estimate of the variance of the predicted value of Y for an individual for a specific value of X, say 𝑋𝑋0 , is
given by

𝑠𝑠𝑌𝑌2𝐼𝐼 |𝑋𝑋0 = 𝑠𝑠 2 + 𝑠𝑠𝑌𝑌2𝑚𝑚|𝑋𝑋0

300-5
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Hypothesis Tests of the Intercept and Slope

Using these variance estimates and assuming the residuals are normally distributed, hypothesis tests may
be constructed using the Student’s t distribution with N - 2 degrees of freedom using

𝑏𝑏0 − 𝐵𝐵0
𝑡𝑡𝑏𝑏0 =
𝑠𝑠𝑏𝑏0
and
𝑏𝑏1 − 𝐵𝐵1
𝑡𝑡𝑏𝑏1 =
𝑠𝑠𝑏𝑏1

Usually, the hypothesized values of 𝐵𝐵0 and 𝐵𝐵1 are zero, but this does not have to be the case.

Confidence Intervals of the Intercept and Slope

A 100(1 − 𝛼𝛼)% confidence interval for the intercept, 𝛽𝛽0 , is given by

𝑏𝑏0 ± 𝑡𝑡1−𝛼𝛼/2,𝑁𝑁−2 𝑠𝑠𝑏𝑏0

A 100(1 − 𝛼𝛼)% confidence interval for the slope, 𝛽𝛽1 , is given by

𝑏𝑏1 ± 𝑡𝑡1−𝛼𝛼/2,𝑁𝑁−2 𝑠𝑠𝑏𝑏1

Confidence Interval of Y for Given X

A 100(1 − 𝛼𝛼)% confidence interval for the mean of Y at a specific value of X, say 𝑋𝑋0 , is given by

𝑏𝑏0 + 𝑏𝑏1 𝑋𝑋0 ± 𝑡𝑡1−𝛼𝛼/2,𝑁𝑁−2 𝑠𝑠𝑌𝑌𝑚𝑚|𝑋𝑋0

Note that this confidence interval assumes that the sample size at X is N.
A 100(1 − 𝛼𝛼)% prediction interval for the value of Y for an individual at a specific value of X, say 𝑋𝑋0 , is given
by
𝑏𝑏0 + 𝑏𝑏1 𝑋𝑋0 ± 𝑡𝑡1−𝛼𝛼/2,𝑁𝑁−2 𝑠𝑠𝑌𝑌𝐼𝐼 |𝑋𝑋0

Working-Hotelling Confidence Band for the Mean of Y

A 100(1 − 𝛼𝛼)% simultaneous confidence band for the mean of Y at all values of X is given by

𝑏𝑏0 + 𝑏𝑏1 𝑋𝑋 ± 𝑠𝑠𝑌𝑌𝑚𝑚|𝑋𝑋 �2𝐹𝐹1−𝛼𝛼,2,𝑁𝑁−2

This confidence band applies to all possible values of X. The confidence coefficient, 100(1 − 𝛼𝛼)%, is the
percent of a long series of samples for which this band covers the entire line for all values of X from
negativity infinity to positive infinity.

300-6
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Confidence Interval of X for Given Y

This type of analysis is called inverse prediction or calibration. A 100(1 − 𝛼𝛼)% confidence interval for the
mean value of X for a given value of Y is calculated as follows. First, calculate X from Y using

𝑌𝑌 − 𝑏𝑏0
𝑋𝑋� =
𝑏𝑏1

Then, calculate the interval using

2
(1 − 𝑔𝑔) �𝑋𝑋� − 𝑋𝑋��
�𝑋𝑋� − 𝑔𝑔𝑋𝑋�� ± 𝐴𝐴� 𝑁𝑁 +
∑𝑁𝑁 �
𝑗𝑗=1 𝑤𝑤𝑗𝑗 �𝑋𝑋𝑗𝑗 − 𝑋𝑋�
1 − 𝑔𝑔
where

𝑡𝑡1−α/2,𝑁𝑁−2 𝑠𝑠
𝐴𝐴 =
𝑏𝑏1

𝐴𝐴2
𝑔𝑔 =
∑𝑁𝑁 �
𝑗𝑗=1 𝑤𝑤𝑗𝑗 �𝑋𝑋𝑗𝑗 − 𝑋𝑋�

A 100(1 − 𝛼𝛼)% confidence interval for an individual value of X for a given value of Y is

2
(N + 1)(1 − 𝑔𝑔) �𝑋𝑋� − 𝑋𝑋��
�𝑋𝑋� − 𝑔𝑔𝑋𝑋�� ± 𝐴𝐴� 𝑁𝑁 +
∑𝑁𝑁 �
𝑗𝑗=1 𝑤𝑤𝑗𝑗 �𝑋𝑋𝑗𝑗 − 𝑋𝑋 �
1 − 𝑔𝑔

R-Squared (Percent of Variation Explained)

Several measures of the goodness-of-fit of the regression model to the data have been proposed, but by far
the most popular is 𝑅𝑅 2. 𝑅𝑅 2 is the square of the correlation coefficient. It is the proportion of the variation in
Y that is accounted by the variation in X. 𝑅𝑅 2 varies between zero (no linear relationship) and one (perfect
linear relationship).
𝑅𝑅 2, officially known as the coefficient of determination, is defined as the sum of squares due to the
regression divided by the adjusted total sum of squares of Y. The formula for 𝑅𝑅 2 is

𝒆𝒆′ 𝑾𝑾𝑾𝑾
𝑅𝑅 2 = 1 − � �
(𝟏𝟏′ 𝑾𝑾𝑾𝑾)𝟐𝟐
𝒀𝒀′ 𝑾𝑾𝑾𝑾 − 𝟏𝟏′ 𝑾𝑾𝑾𝑾

𝑆𝑆𝑆𝑆𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀
=
𝑆𝑆𝑆𝑆𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇

𝑅𝑅 2 is probably the most popular measure of how well a regression model fits the data. 𝑅𝑅 2 may be defined
either as a ratio or a percentage. Since we use the ratio form, its values range from zero to one. A value of
𝑅𝑅 2 near zero indicates no linear relationship, while a value near one indicates a perfect linear fit. Although

300-7
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

popular, 𝑅𝑅 2 should not be used indiscriminately or interpreted without scatter plot support. Following are
some qualifications on its interpretation:
1. Additional independent variables. It is possible to increase 𝑅𝑅 2 by adding more independent variables,
but the additional independent variables may actually cause an increase in the mean square error,
an unfavorable situation. This usually happens when the sample size is small.
2. Range of the independent variable. 𝑅𝑅 2 is influenced by the range of the independent variable. 𝑅𝑅 2
increases as the range of X increases and decreases as the range of the X decreases.
3. Slope magnitudes. 𝑅𝑅 2 does not measure the magnitude of the slopes.
4. Linearity. 𝑅𝑅 2 does not measure the appropriateness of a linear model. It measures the strength of
the linear component of the model. Suppose the relationship between X and Y was a perfect circle.
Although there is a perfect relationship between the variables, the 𝑅𝑅 2 value would be zero.
5. Predictability. A large 𝑅𝑅 2 does not necessarily mean high predictability, nor does a low 𝑅𝑅 2 necessarily
mean poor predictability.
6. No-intercept model. The definition of 𝑅𝑅 2 assumes that there is an intercept in the regression model.
When the intercept is left out of the model, the definition of 𝑅𝑅 2 changes dramatically. The fact that
your 𝑅𝑅 2 value increases when you remove the intercept from the regression model does not reflect
an increase in the goodness of fit. Rather, it reflects a change in the underlying definition of 𝑅𝑅 2.
7. Sample size. 𝑅𝑅 2 is highly sensitive to the number of observations. The smaller the sample size, the
larger its value.

Rbar-Squared (Adjusted R-Squared)

𝑅𝑅 2 varies directly with N, the sample size. In fact, when N = 2, 𝑅𝑅 2 = 1. Because 𝑅𝑅 2 is so closely tied to the
sample size, an adjusted 𝑅𝑅 2 value, called 𝑅𝑅� 2, has been developed. 𝑅𝑅� 2 was developed to minimize the impact
of sample size. The formula for 𝑅𝑅 � 2 is

�𝑁𝑁 − (𝑝𝑝 − 1)�(1 − 𝑅𝑅 2 )

𝑅𝑅� 2 = 1 − � �
𝑁𝑁 − 𝑝𝑝

where p is 2 if the intercept is included in the model and 1 if not.

Probability Ellipse
When both variables are random variables and they follow the bivariate normal distribution, it is possible to
construct a probability ellipse for them (see Jackson (1991) page 342). The equation of the 100(1 − α)%
probability ellipse is given by those values of X and Y that are solutions of

2
𝑠𝑠𝑌𝑌𝑌𝑌 𝑠𝑠𝑋𝑋𝑋𝑋 (𝑋𝑋 − 𝑋𝑋�)2 (𝑌𝑌 − 𝑌𝑌�)2 2𝑠𝑠𝑋𝑋𝑋𝑋 (𝑋𝑋 − 𝑋𝑋�)(𝑌𝑌 − 𝑌𝑌�)
𝑇𝑇2,𝑁𝑁−2,𝛼𝛼 = 2 � + − �
𝑠𝑠𝑌𝑌𝑌𝑌 𝑠𝑠𝑋𝑋𝑋𝑋 − 𝑠𝑠𝑋𝑋𝑋𝑋 𝑠𝑠𝑋𝑋𝑋𝑋 𝑠𝑠𝑌𝑌𝑌𝑌 𝑠𝑠𝑋𝑋𝑋𝑋 𝑠𝑠𝑌𝑌𝑌𝑌

300-8
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Orthogonal Regression Line

The least squares estimates discussed above minimize the sum of the squared distances between the Y’s
and their predicted values. In some situations, both variables are random variables and it is arbitrary which
is designated as the dependent variable and which is the independent variable. When the choice of which
variable is the dependent variable is arbitrary, you may want to use the orthogonal regression line rather than
the least squares regression line. The orthogonal regression line minimizes the sum of the squared
perpendicular distances from each observation to the regression line. The orthogonal regression line is the
first principal component when a principal components analysis is run on the two variables.
Jackson (1991) page 343 gives a formula for computing the orthogonal regression line without computing a
principal components analysis. The slope is given by

2
𝑠𝑠𝑌𝑌𝑌𝑌 − 𝑠𝑠𝑋𝑋𝑋𝑋 + �𝑠𝑠𝑌𝑌𝑌𝑌 − 𝑠𝑠𝑋𝑋𝑋𝑋 + 4𝑠𝑠𝑋𝑋𝑋𝑋
𝑏𝑏𝑜𝑜𝑜𝑜𝑜𝑜ℎ𝑜𝑜,1 =
2𝑠𝑠𝑋𝑋𝑋𝑋
where

∑𝑁𝑁 � �
𝑗𝑗=1 𝑤𝑤𝑗𝑗 �𝑋𝑋𝑗𝑗 − 𝑋𝑋 ��𝑌𝑌𝑗𝑗 − 𝑌𝑌�
𝑠𝑠𝑋𝑋𝑋𝑋 =
𝑁𝑁 − 1

The estimate of the intercept is then computed using

𝑏𝑏𝑜𝑜𝑜𝑜𝑜𝑜ℎ𝑜𝑜,𝑦𝑦 = 𝑌𝑌� − 𝑏𝑏𝑜𝑜𝑜𝑜𝑜𝑜ℎ𝑜𝑜,1 𝑋𝑋�

Although Jackson gives formulas for a confidence interval on the slope and intercept, we do not provide
them in NCSS because their properties are not well understood, and they require certain bivariate normal
assumptions. Instead, NCSS provides bootstrap confidence intervals for the slope and intercept.

The Correlation Coefficient

The correlation coefficient can be interpreted in several ways. Here are some of the interpretations.
1. If both Y and X are standardized by subtracting their means and dividing by their standard
deviations, the correlation is the slope of the regression of the standardized Y on the standardized X.
2. The correlation is the standardized covariance between Y and X.
3. The correlation is the geometric average of the slopes of the regressions of Y on X and of X on Y.
4. The correlation is the square root of R-squared, using the sign from the slope of the regression of Y
on X.

300-9
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

The corresponding formulas for the calculation of the correlation coefficient are

∑𝑁𝑁 � �
𝑗𝑗=1 𝑤𝑤𝑗𝑗 �𝑋𝑋𝑗𝑗 − 𝑋𝑋��𝑌𝑌𝑗𝑗 − 𝑌𝑌 �
𝑟𝑟 =
2 2
��∑𝑁𝑁 � 𝑁𝑁 �
𝑗𝑗=1 𝑤𝑤𝑗𝑗 �𝑋𝑋𝑗𝑗 − 𝑋𝑋 � � �∑𝑗𝑗=1 𝑤𝑤𝑗𝑗 �𝑌𝑌𝑗𝑗 − 𝑌𝑌� �

𝑠𝑠𝑋𝑋𝑋𝑋
=
√𝑠𝑠𝑋𝑋𝑋𝑋 𝑠𝑠𝑌𝑌𝑌𝑌

= ±�𝑏𝑏𝑌𝑌𝑌𝑌 𝑏𝑏𝑋𝑋𝑋𝑋

= sign(𝑏𝑏𝑌𝑌𝑌𝑌 )�𝑅𝑅2

where 𝑠𝑠𝑋𝑋𝑋𝑋 is the covariance between X and Y, 𝑏𝑏𝑋𝑋𝑋𝑋 is the slope from the regression of X on Y, and 𝑏𝑏𝑌𝑌𝑌𝑌 is the
slope from the regression of Y on X. 𝑠𝑠𝑋𝑋𝑋𝑋 is calculated using the formula

∑𝑁𝑁 � �
𝑗𝑗=1 𝑤𝑤𝑗𝑗 �𝑋𝑋𝑗𝑗 − 𝑋𝑋��𝑌𝑌𝑗𝑗 − 𝑌𝑌�
𝑠𝑠𝑋𝑋𝑋𝑋 =
𝑁𝑁 − 1

The population correlation coefficient, 𝜌𝜌, is defined for two random variables, U and W, as follows

𝜎𝜎𝑈𝑈𝑈𝑈
𝜌𝜌 =
�𝜎𝜎𝑈𝑈2 𝜎𝜎𝑊𝑊
2

𝐸𝐸[(𝑈𝑈 − 𝜇𝜇𝑈𝑈 )(𝑊𝑊 − 𝜇𝜇𝑊𝑊 )]

=
�𝑉𝑉𝑉𝑉𝑉𝑉(𝑈𝑈)𝑉𝑉𝑉𝑉𝑉𝑉(𝑊𝑊)

Note that this definition does not refer to one variable as dependent and the other as independent. Rather,
it simply refers to two random variables.

Facts about the Correlation Coefficient

The correlation coefficient has the following characteristics.
1. The range of r is between -1 and 1, inclusive.
2. If r = 1, the observations fall on a straight line with positive slope.
3. If r = -1, the observations fall on a straight line with negative slope.
4. If r = 0, there is no linear relationship between the two variables.
5. r is a measure of the linear (straight-line) association between two variables.
6. The value of r is unchanged if either X or Y is multiplied by a constant or if a constant is added.
7. The physical meaning of r is mathematically abstract and may not be very helpful. However, we
provide it for completeness. The correlation is the cosine of the angle formed by the intersection of
two vectors in N-dimensional space. The components of the first vector are the values of X while the
components of the second vector are the corresponding values of Y. These components are
arranged so that the first dimension corresponds to the first observation, the second dimension
corresponds to the second observation, and so on.

300-10
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Hypothesis Tests for the Correlation

You may be interested in testing hypotheses about the population correlation coefficient, such as ρ = ρ0 .
When ρ0 = 0, the test is identical to the t-test used to test the hypothesis that the slope is zero. The test
statistic is calculated using

𝑟𝑟
𝑡𝑡𝑁𝑁−2 =
2
�1 − 𝑟𝑟
𝑁𝑁 − 2

However, when 𝜌𝜌0 ≠ 0, the test is different from the corresponding test that the slope is a specified,
nonzero, value.
NCSS provides two methods for testing whether the correlation is equal to a specified, nonzero, value.
Method 1. This method uses the distribution of the correlation coefficient. Under the null hypothesis that
𝜌𝜌 = 𝜌𝜌0 and using the distribution of the sample correlation coefficient, the likelihood of obtaining the
sample correlation coefficient, r, can be computed. This likelihood is the statistical significance of the test.
This method requires the assumption that the two variables follow the bivariate normal distribution.
Method 2. This method uses the fact that Fisher’s z transformation, given by

1 1 + 𝑟𝑟
𝐹𝐹(𝑟𝑟) = ln � �
2 1 − 𝑟𝑟

is closely approximated by a normal distribution with mean

1 1 + 𝜌𝜌
ln � �
2 1 − 𝜌𝜌
and variance
1
𝑁𝑁 − 3

To test the hypothesis that 𝜌𝜌 = 𝜌𝜌0 , you calculate z using

𝐹𝐹(𝑟𝑟) − 𝐹𝐹(ρ0 )
𝑧𝑧 =
� 1
𝑁𝑁 − 3
1 + 𝑟𝑟 1 − 𝜌𝜌
ln �1 − 𝑟𝑟� − ln �1 − 𝜌𝜌0 �
0
=
1
2�𝑁𝑁 − 3

and use the fact that z is approximately distributed as the standard normal distribution with mean equal to
zero and variance equal to one. This method requires two assumptions. First, that the two variables follow
the bivariate normal distribution. Second, that the distribution of z is approximated by the standard normal
distribution.
This method has become popular because it uses the commonly available normal distribution rather than
the obscure correlation distribution. However, because it makes an additional assumption, it is not as
accurate as is method 1. In fact, we have included in for completeness, but recommend the use of Method
1.

300-11
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Confidence Intervals for the Correlation

A 100(1 − 𝛼𝛼)% confidence interval for 𝜌𝜌 may be constructed using either of the two hypothesis methods
described above. The confidence interval is calculated by finding, either directly using Method 2 or by a
search using Method 1, all those values of 𝜌𝜌0 for which the hypothesis test is not rejected. This set of values
becomes the confidence interval.
Be careful not to make the common mistake in assuming that this confidence interval is related to a
transformation of the confidence interval on the slope 𝛽𝛽1 . The two confidence intervals are not simple
transformations of each other.

Spearman Rank Correlation Coefficient

The Spearman rank correlation coefficient is a popular nonparametric analog of the usual correlation
coefficient. This statistic is calculated by replacing the data values with their ranks and calculating the
correlation coefficient of the ranks. Tied values are replaced with the average rank of the ties. This
coefficient is really a measure of association rather than correlation, since the ranks are unchanged by a
monotonic transformation of the original data.
When N is greater than 10, the distribution of the Spearman rank correlation coefficient can be
approximated by the distribution of the regular correlation coefficient.
Note that when weights are specified, the calculation of the Spearman rank correlation coefficient uses the
weights.

Smoothing with Loess

The loess (locally weighted regression scatter plot smoothing) method is used to obtain a smooth curve
representing the relationship between X and Y. Unlike linear regression, loess does not have a simple
mathematical model. Rather, it is an algorithm that, given a value of X, computes an appropriate value of Y.
The algorithm was designed so that the loess curve travels through the middle of the data, summarizing the
relationship between X and Y.
The loess algorithm works as follows.
1. Select a value for X. Call it X0.
2. Select a neighborhood of points close to X0.
3. Fit a weighted regression of Y on X using only the points in this neighborhood. In the regression, the
weights are inversely proportional to the distance between X and X0.
4. To make the procedure robust to outliers, a robust regression may be substituted for the weighted
regression in step 3. This robust procedure modifies the weights so that observations with large
residuals receive smaller weights.
5. Use the regression coefficients from the weighted regression in step 3 to obtain a predicted value
for Y at X0.
6. Repeat steps 1 - 5 for a set of X’s between the minimum and maximum of X.

300-12
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Mathematical Details of Loess

This section presents the mathematical details of the loess method of scatter plot smoothing. Note that
implicit in the discussion below is the assumption that Y is the dependent variable and X is the independent
variable.
Loess gives the value of Y for a given value of X, say X0. For each observation, define the distance between X
and X0 as

𝑑𝑑𝑗𝑗 = �𝑋𝑋𝑗𝑗 − 𝑋𝑋0�

Let q be the number of observations in the neighborhood of X0. Define q as [fN] where f is the user-supplied
fraction of the sample. Here, [Z] is the largest integer in Z. Often f = 0.40 is a good choice. The neighborhood
is defined as the observations with the q smallest values of 𝑑𝑑𝑗𝑗 . Define 𝑑𝑑𝑞𝑞 as the largest distance in the
neighborhood of observations close to X0.
The tricube weight function is defined as

(1 − |𝑢𝑢|3 )3 |𝑢𝑢| < 1

𝑇𝑇(𝑢𝑢) = �
0 |𝑢𝑢| ≥ 1

The weight for each observation is defined as

�𝑋𝑋𝑗𝑗 − 𝑋𝑋0�
𝑤𝑤𝑗𝑗 = 𝑇𝑇 � �
𝑑𝑑𝑞𝑞

The weighted regression for X0 is defined by the value of b0, b1, and b2 that minimize the sum of squares

𝑁𝑁
𝑋𝑋𝑗𝑗 − 𝑋𝑋0 2 2
� 𝑇𝑇 � � �𝑌𝑌𝑗𝑗 − 𝑏𝑏0 − 𝑏𝑏1�𝑋𝑋𝑗𝑗 � − 𝑏𝑏2�𝑋𝑋𝑗𝑗 � �
𝑑𝑑𝑞𝑞
𝑗𝑗=1

Note the if b2 is zero, a linear regression is fit. Otherwise, a quadratic regression is fit. The choice of linear or
quadratic is an option in the procedure. The linear option is quicker, while the quadratic option fits peaks
and valleys better. In most cases, there is little difference except at the extremes in the X space.
Once b0, b1, and b2 have be estimated using weighted least squares, the loess value is computed using

𝑌𝑌�𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 (𝑋𝑋0) = 𝑏𝑏0 − 𝑏𝑏1(𝑋𝑋0) − 𝑏𝑏2(𝑋𝑋0)2

Note that a separate weighted regression must be run for each value of X0.

300-13
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Robust Loess
Outliers often have a large impact on least squares impact. A robust weighted regression procedure may be
used to lessen the influence of outliers on the loess curve. This is done as follows.
The q loess residuals are computed using the loess regression coefficients using the formula

𝑟𝑟𝑗𝑗 = 𝑌𝑌𝑗𝑗 − 𝑌𝑌�𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 �𝑋𝑋𝑗𝑗 �

New weights are defined as

�𝑟𝑟𝑗𝑗 �
𝑤𝑤𝑗𝑗 = 𝑤𝑤𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙,𝑗𝑗 𝐵𝐵 � �
6𝑀𝑀

where 𝑤𝑤𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙,𝑗𝑗 is the previous weight for this observation, 𝑀𝑀 is the median of the q absolute values of the
residuals, and 𝐵𝐵(𝑢𝑢) is the bisquare weight function defined as

(1 − 𝑢𝑢2 )2 |𝑢𝑢| < 1

𝐵𝐵(𝑢𝑢) = �
0 |𝑢𝑢| ≥ 1

This robust procedure may be iterated up to five items, but we have seen little difference in the appearance
of the loess curve after two iterations.
Note that it is not always necessary to create the robust weights. If you are not going to remove the outliers
from your final results, you probably should not remove them from the loess curve by setting the number of
robust iterations to zero.

Testing Assumptions Using Residual Diagnostics

Evaluating the amount of departure in your data from each linear regression assumption is necessary to see
if any remedial action is necessary before the fitted results can be used. First, the types of plots and
statistical analyses the are used to evaluate each assumption will be given. Second, each of the diagnostic
values will be defined.

Notation – Use of (j) and p

Several of these residual diagnostic statistics are based on the concept of studying what happens to various
aspects of the regression analysis when each row is removed from the analysis. In what follows, we use the
notation (j) to mean that observation j has been omitted from the analysis. Thus, b(j) means the value of b
calculated without using observation j.

Some of the formulas depend on whether the intercept is fitted or not. We use p to indicate the number of
regression parameters. When the intercept is fit, p will be two. Otherwise, p will be one.

300-14
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

1 – No Outliers
Outliers are observations that are poorly fit by the regression model. If outliers are influential, they will
cause serious distortions in the regression calculations. Once an observation has been determined to be an
outlier, it must be checked to see if it resulted from a mistake. If so, it must be corrected or omitted.
However, if no mistake can be found, the outlier should not be discarded just because it is an outlier. Many
scientific discoveries have been made because outliers, data points that were different from the norm, were
studied more closely. Besides being caused by simple data-entry mistakes, outliers often suggest the
presence of an important independent variable that has been ignored.
Outliers are easy to spot on bar charts or box plots of the residuals and RStudent. RStudent is the preferred
statistic for finding outliers because each observation is omitted from the calculation making it less likely
that the outlier can mask its presence. Scatter plots of the residuals and RStudent against the X variable are
also helpful because they may show other problems as well.

2 – Linear Regression Function - No Curvature

The relationship between Y and X is assumed to be linear (straight-line). No mechanism for curvature is
included in the model. Although a scatter plot of Y versus X can show curvature in the relationship, the best
diagnostic tool is the scatter plot of the residual versus X. If curvature is detected, the model must be
modified to account for the curvature. This may mean adding quadratic terms, taking logarithms of Y or X, or
some other appropriate transformation.

Loess Curve
A loess curve should be plotted between X and Y to see if any curvature is present.

Lack of Fit Test

When the data include repeat observations at one or more X values (replicates), the adequacy of the linear
model can be evaluated numerically by performing a lack of fit test. This test procedure detects
nonlinearities.
The lack of fit test is constructed as follows. First, the sum of squares for error is partitioned into two
quantities: lack of fit and pure error. The pure error sum of squares is found by considering only those
observations that are replicates. The X values are treated as the levels of the factor in a one-way analysis of
variance. The sum of squares error from this analysis measures the underlying variation in Y that occurs
when the value of X is held constant. Thus, it is called pure error. When the pure error sum of squares is
subtracted from the error sum of squares of the linear regression, the result is measure of the amount of
nonlinearity in the data. An F-ratio can be constructed from these two values that will test the statistical
significance of the lack of fit. The F-ratio is constructed using the following equation.

𝑆𝑆𝑆𝑆𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 𝑜𝑜𝑜𝑜 𝑓𝑓𝑓𝑓𝑓𝑓

𝐹𝐹𝐷𝐷𝐷𝐷1,𝐷𝐷𝐷𝐷2 = 𝐷𝐷𝐷𝐷1
𝑆𝑆𝑆𝑆𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸
𝐷𝐷𝐷𝐷2

where DF2 is the degrees of freedom for the error term in the one-way analysis of variance and DF1 is N -
DF2 - 2.

300-15
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

3 – Constant Variance
The errors are assumed to have constant variance across all values of X. If there are a lot of data (N > 100),
nonconstant variance can be detected on a scatter plot of the residuals versus X. However, the most direct
diagnostic tool to evaluate this assumption is a scatter plot of the absolute values of the residuals versus X.
Often, the assumption is violated because the variance increases with X. This will show up as a ‘megaphone’
pattern to this plot.
When nonconstant variance is detected, a variance-stabilizing transformation such as the square-root or
logarithm may be used. However, the best solution is probably to use weighted regression, with weights
inversely proportional to the magnitude of the residuals.

Modified Levene Test

The modified Levene test can be used to evaluate the validity of the assumption of constant variance. It has
been shown to be reliable even when the residuals do not follow a normal distribution.
The test is constructed by grouping the residuals according to the values of X. The number of groups is
arbitrary, but usually, two groups are used. In this case, the absolute residuals of observations with low
values of X are compared against those with high values of X. If the variability is constant, the variability in
these two groups of residuals should be equal. The test is computed using the formula

𝑑𝑑1̅ − 𝑑𝑑̅2
𝐿𝐿 =
1 1
𝑠𝑠𝐿𝐿 � +
𝑛𝑛1 𝑛𝑛2
where

∑𝑛𝑛𝑗𝑗=1
1 𝑛𝑛2
�𝑑𝑑𝑗𝑗1 − 𝑑𝑑1̅ � + ∑𝑗𝑗=1�𝑑𝑑𝑗𝑗2 − 𝑑𝑑̅2 �
𝑠𝑠𝐿𝐿 = �
𝑛𝑛1 + 𝑛𝑛2 − 2

𝑑𝑑𝑗𝑗1 = �𝑒𝑒𝑗𝑗1 − 𝑒𝑒̃1 �

𝑑𝑑𝑗𝑗2 = �𝑒𝑒𝑗𝑗2 − 𝑒𝑒̃2 �

and 𝑒𝑒̃1 is the median of the group of residuals for low values of X and 𝑒𝑒̃2 is the median of the group of
residuals for high values of X. The test statistic L is approximately distributed as a t statistic with N - 2
degrees of freedom.

300-16
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

4 – Independent Errors
The Y’s, and thus the errors, are assumed to be independent. This assumption is usually ignored unless
there is a reason to think that it has been violated, such as when the observations were taken across time.
An easy way to evaluate this assumption is a scatter plot of the residuals versus their sequence number
(assuming that the data are arranged in time sequence order). This plot should show a relative random
pattern.
The Durbin-Watson statistic is used as a formal test for the presence of first-order serial correlation. A more
comprehensive method of evaluation is to look at the autocorrelations of the residuals at various lags. Large
autocorrelations are found by testing each using Fisher’s z transformation. Although Fisher’s z
transformation is only approximate in the case of autocorrelations, it does provide a reasonable measuring
stick with which to judge the size of the autocorrelations.
If independence is violated, confidence intervals and hypothesis tests are erroneous. Some remedial
method that accounts for the lack of independence must be adopted, such as using first differences or the
Cochrane-Orcutt procedure.

Durbin-Watson Test
The Durbin-Watson test is often used to test for positive or negative, first-order, serial correlation. It is
calculated as follows
2
∑𝑁𝑁
𝑗𝑗=2�𝑒𝑒𝑗𝑗 − 𝑒𝑒𝑗𝑗−1 �
𝐷𝐷𝐷𝐷 =
∑𝑁𝑁 2
𝑗𝑗=1 𝑒𝑒𝑗𝑗

The distribution of this test is difficult because it involves the X values. Originally, Durbin-Watson (1950,
1951) gave a pair of bounds to be used. However, there is a large range of ‘inclusion’ found when using
these bounds. Instead of using these bounds, we calculate the exact probability using the beta distribution
approximation suggested by Durbin-Watson (1951). This approximation has been shown to be accurate to
three decimal places in most cases which is all that are needed for practical work.

5 – Normality of Residuals
The residuals are assumed to follow the normal probability distribution with zero mean and constant
variance. This can be evaluated using a normal probability plot of the residuals. Also, normality tests are
used to evaluate this assumption. The most popular of the five normality tests provided is the Shapiro-Wilk
test.
Unfortunately, a breakdown in any of the other assumptions results in a departure from this assumption as
well. Hence, you should investigate the other assumptions first, leaving this assumption until last.

300-17
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Influential Observations
Part of the evaluation of the assumptions includes an analysis to determine if any of the observations have
an extra-large influence on the estimated regression coefficients, on the fit of the model, or on the value of
Cook’s distance. By looking at how much removing an observation changes the results, an observation’s
influence can be determined.
Five statistics are used to investigate influence. These are Hat diagonal, DFFITS, DFBETAS, Cook’s D, and
COVARATIO.

Definitions Used in Residual Diagnostics

Residual
The residual is the difference between the actual Y value and the Y value predicted by the estimated
regression model. It is also called the error, the deviate, or the discrepancy.

𝑒𝑒𝑗𝑗 = 𝑦𝑦𝑗𝑗 − 𝑦𝑦�𝑗𝑗

Although the true errors, 𝜀𝜀𝑗𝑗 , are assumed to be independent, the computed residuals, 𝑒𝑒𝑗𝑗 , are not. Although
the lack of independence among the residuals is a concern in developing theoretical tests, it is not a concern
on the plots and graphs.
The variance of the ε𝑗𝑗 is 𝜎𝜎 2 . However, the variance of the 𝑒𝑒𝑗𝑗 is not 𝜎𝜎 2 . In vector notation, the covariance
matrix of e is given by

𝟏𝟏 𝟏𝟏
𝑉𝑉(𝒆𝒆) = 𝜎𝜎 2 �𝑰𝑰 − 𝑾𝑾𝟐𝟐 𝑿𝑿(𝑿𝑿′ 𝑾𝑾𝑾𝑾)−𝟏𝟏 𝑿𝑿′ 𝑾𝑾𝟐𝟐 �

= 𝜎𝜎 2 (𝑰𝑰 − 𝑯𝑯)

The matrix H is called the hat matrix since it puts the ‘hat’ on y as is shown in the unweighted case.

𝑌𝑌� = 𝑿𝑿𝑿𝑿

= 𝑿𝑿(𝑿𝑿′ 𝑿𝑿)−𝟏𝟏 𝑿𝑿′ 𝒀𝒀

= 𝑯𝑯𝑯𝑯

Hence, the variance of 𝑒𝑒𝑗𝑗 is given by

𝑉𝑉�𝑒𝑒𝑗𝑗 � = 𝜎𝜎 2 �1 − ℎ𝑗𝑗𝑗𝑗 �

where ℎ𝑗𝑗𝑗𝑗 is the jth diagonal element of H. This variance is estimated using

𝑉𝑉� �𝑒𝑒𝑗𝑗 � = 𝑠𝑠 2 �1 − ℎ𝑗𝑗𝑗𝑗 �

300-18
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Hat Diagonal
The hat diagonal, ℎ𝑗𝑗𝑗𝑗 , is the jth diagonal element of the hat matrix, H, where

𝟏𝟏 𝟏𝟏
𝑯𝑯 = 𝑾𝑾𝟐𝟐 𝑿𝑿(𝑿𝑿′ 𝑾𝑾𝑾𝑾)−𝟏𝟏 𝑿𝑿′ 𝑾𝑾𝟐𝟐

H captures an observation’s remoteness in the X-space. Some authors refer to the hat diagonal as a
measure of leverage in the X-space. As a rule of thumb, hat diagonals greater than 4/N are considered
influential and are called high-leverage observations.
Note that a high-leverage observation is not a bad observation. Rather, high-leverage observations exert
extra influence on the final results, so care should be taken to ensure that they are correct. You should not
delete an observation just because it has a high-influence. However, when you interpret the regression
equation, you should bear in mind that the results may be due to a few, high-leverage observations.

Standardized Residual
As shown above, the variance of the observed residuals is not constant. This makes comparisons among the
residuals difficult. One solution is to standardize the residuals by dividing them by their standard deviations.
This will give a set of residuals with constant variance.
The formula for this residual is
𝑒𝑒𝑗𝑗
𝑟𝑟𝑗𝑗 =
𝑠𝑠�1 − ℎ𝑗𝑗𝑗𝑗

s(j) or MSEi
This is the value of the mean squared error calculated without observation j. The formula for s(j) is given by

𝑁𝑁
2
1
𝑠𝑠(𝑗𝑗) = � 𝑤𝑤𝑖𝑖 �𝑦𝑦𝑖𝑖 − 𝒙𝒙𝒊𝒊 𝒃𝒃(𝑗𝑗)�
𝑁𝑁 − 𝑝𝑝 − 1
𝑖𝑖=1,𝑖𝑖≠𝑗𝑗

𝑤𝑤𝑗𝑗 𝑒𝑒𝑗𝑗2
(𝑁𝑁 − 𝑝𝑝)𝑠𝑠 2 −
1 − ℎ𝑗𝑗𝑗𝑗
=
𝑁𝑁 − 𝑝𝑝 − 1

RStudent
Rstudent is similar to the studentized residual. The difference is the s(j) is used rather than s in the
denominator. The quantity s(j) is calculated using the same formula as s, except that observation j is
omitted. The hope is that be excluding this observation, a better estimate of σ2 will be obtained. Some
statisticians refer to these as the studentized deleted residuals.

𝑒𝑒𝑗𝑗
𝑡𝑡𝑗𝑗 =
𝑠𝑠(𝑗𝑗)�1 − ℎ𝑗𝑗𝑗𝑗

If the regression assumptions of normality are valid, a single value of the RStudent has a t distribution with
N - 2 degrees of freedom. It is reasonable to consider |RStudent| > 2 as outliers.

300-19
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

DFFITS
DFFITS is the standardized difference between the predicted value with and without that observation. The
formula for DFFITS is
𝑦𝑦�𝑗𝑗 − 𝑦𝑦�𝑗𝑗 (𝑗𝑗)
𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝑆𝑆𝑗𝑗 =
𝑠𝑠(𝑗𝑗)�ℎ𝑗𝑗𝑗𝑗

ℎ𝑗𝑗𝑗𝑗
= 𝑡𝑡𝑗𝑗 �
1 − ℎ𝑗𝑗𝑗𝑗

The values of𝑦𝑦�𝑗𝑗 (𝑗𝑗) and 𝑠𝑠 2 (𝑗𝑗) are found by removing observation j before the doing the calculations. It
represents the number of estimated standard errors that the fitted value changes if the jth observation is
omitted from the data set. If |DFFITS| > 1, the observation should be considered to be influential with
regards to prediction.

Cook’s D
The DFFITS statistic attempts to measure the influence of a single observation on its fitted value. Cook’s
distance (Cook’s D) attempts to measure the influence each observation on all N fitted values. The formula
for Cook’s D is
2
∑𝑁𝑁 �𝑗𝑗 − 𝑦𝑦�𝑗𝑗 (𝑖𝑖)�
𝑖𝑖=1 𝑤𝑤𝑗𝑗 �𝑦𝑦
𝐷𝐷𝑗𝑗 =
𝑝𝑝𝑠𝑠 2

The 𝑦𝑦�𝑗𝑗 (𝑖𝑖) are found by removing observation i before the calculations. Rather than go to all the time of
recalculating the regression coefficients N times, we use the following approximation

𝑤𝑤𝑗𝑗 𝑒𝑒𝑗𝑗2 ℎ𝑗𝑗𝑗𝑗

𝐷𝐷𝑗𝑗 = 2
𝑝𝑝𝑠𝑠 2 �1 − ℎ𝑗𝑗𝑗𝑗 �

This approximation is exact when no weight variable is used.

A Cook’s D value greater than one indicates an observation that has large influence. Some statisticians have
suggested that a better cutoff value is 4 / (N - 2).

300-20
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

CovRatio
This diagnostic flags observations that have a major impact on the generalized variance of the regression
coefficients. A value exceeding 1.0 implies that the ith observation provides an improvement, i.e., a reduction
in the generalized variance of the coefficients. A value of CovRatio less than 1.0 flags an observation that
increases the estimated generalized variance. This is not a favorable condition.
The general formula for the CovRatio is

−1
det �𝑠𝑠(𝑗𝑗)2 �𝑿𝑿(𝒋𝒋)′ 𝑾𝑾𝑾𝑾(𝒋𝒋)� �
𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝑜𝑜𝑗𝑗 =
det[𝑠𝑠 2 (𝑿𝑿′ 𝑾𝑾𝑾𝑾)−1 ]
𝑝𝑝
1 𝑠𝑠(𝑗𝑗)2
= � 2 �
1 − ℎ𝑗𝑗𝑗𝑗 𝑠𝑠

where p = 2 if the intercept is fit or 1 if not.

Belsley, Kuh, and Welsch (1980) give the following guidelines for the CovRatio:
• If CovRatio > 1 + 3p / N then omitting this observation significantly damages the precision of at least
some of the regression estimates.
• If CovRatio < 1 - 3p / N then omitting this observation significantly improves the precision of at least
some of the regression estimates.

DFBETAS
The DFBETAS criterion measures the standardized change in a regression coefficient when an observation is
omitted. The formula for this criterion is

𝑏𝑏𝑘𝑘 − 𝑏𝑏𝑘𝑘 (𝑗𝑗)

𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝑆𝑆𝑘𝑘𝑘𝑘 =
𝑠𝑠(𝑗𝑗)�𝑐𝑐𝑘𝑘𝑘𝑘

where 𝑐𝑐𝑘𝑘𝑘𝑘 is a diagonal element of the inverse matrix (𝑿𝑿′ 𝑾𝑾𝑾𝑾)−𝟏𝟏 .

Belsley, Kuh, and Welsch (1980) recommend using a cutoff of 2/√𝑁𝑁 when N is greater than 100. When N is
less than 100, others have suggested using a cutoff of 1.0 or 2.0 for the absolute value of DFBETAS.

Press Value
PRESS is an acronym for prediction sum of squares. It was developed for use in variable selection to validate
a regression model. To calculate PRESS, each observation is individually omitted. The remaining N - 1
observations are used to calculate a regression and estimate the value of the omitted observation. This is
done N times, once for each observation. The difference between the actual Y value and the predicted Y with
the observation deleted is called the prediction error or PRESS residual. The sum of the squared prediction
errors is the PRESS value. The smaller PRESS is, the better the predictability of the model.
The formula for PRESS is
𝑁𝑁
2
𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 = � 𝑤𝑤𝑗𝑗 �𝑦𝑦𝑗𝑗 − 𝑦𝑦�𝑗𝑗 (𝑗𝑗)�
𝑗𝑗=1

300-21
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Press R-Squared
The PRESS value above can be used to compute an 𝑅𝑅 2-like statistic, called R2Predict, which reflects the
prediction ability of the model. This is a good way to validate the prediction of a regression model without
selecting another sample or splitting your data. It is very possible to have a high 𝑅𝑅 2 and a very low R2Predict.
When this occurs, it implies that the fitted model is data dependent. This R2Predict ranges from below zero
to above one. When outside the range of zero to one, it is truncated to stay within this range.

2 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃
𝑅𝑅𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 =1−
𝑆𝑆𝑆𝑆𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇

Sum |Press residuals|

This is the sum of the absolute value of the PRESS residuals or prediction errors. If a large value for the PRESS
is due to one or a few large PRESS residuals, this statistic may be a more accurate way to evaluate
predictability. This quantity is computed as

𝑁𝑁

�|𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃| = � 𝑤𝑤𝑗𝑗 �𝑦𝑦𝑗𝑗 − 𝑦𝑦�𝑗𝑗 (𝑗𝑗)�

𝑗𝑗=1

Bootstrapping
Bootstrapping was developed to provide standard errors and confidence intervals for regression coefficients
and predicted values in situations in which the standard assumptions are not valid. In these nonstandard
situations, bootstrapping is a viable alternative to the corrective action suggested earlier. The method is
simple in concept, but it requires extensive computation time.
The bootstrap is simple to describe. You assume that your sample is actually the population and you draw B
samples (B is over 1000) of size N from your original sample with replacement. With replacement means
that each observation may be selected more than once. For each bootstrap sample, the regression results
are computed and stored.
Suppose that you want the standard error and a confidence interval of the slope. The bootstrap sampling
process has provided B estimates of the slope. The standard deviation of these B estimates of the slope is
the bootstrap estimate of the standard error of the slope. The bootstrap confidence interval is found by
arranging the B values in sorted order and selecting the appropriate percentiles from the list. For example, a
90% bootstrap confidence interval for the slope is given by fifth and ninety-fifth percentiles of the bootstrap
slope values. The bootstrap method can be applied to many of the statistics that are computed in
regression analysis.
The main assumption made when using the bootstrap method is that your sample approximates the
population fairly well. Because of this assumption, bootstrapping does not work well for small samples in
which there is little likelihood that the sample is representative of the population. Bootstrapping should only
be used in medium to large samples.
When applied to linear regression, there are two types of bootstrapping that can be used. See Neter, Kutner,
Nachtsheim, Wasserman (1996) page 430.

300-22
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Modified Residuals
Davison and Hinkley (1999) page 279 recommend the use of a special rescaling of the residuals when
bootstrapping to keep results unbiased. These modified residuals are calculated using

𝑒𝑒𝑗𝑗
𝑒𝑒𝑗𝑗∗ = − 𝑒𝑒̅ ∗
1 − ℎ𝑗𝑗𝑗𝑗
�
𝑤𝑤𝑗𝑗

where

∑𝑁𝑁 ∗
𝑗𝑗=1 𝑤𝑤𝑗𝑗 𝑒𝑒𝑗𝑗
∗
𝑒𝑒̅ =
∑𝑁𝑁
𝑗𝑗=1 𝑤𝑤𝑗𝑗

Bootstrap the Observations

The bootstrap samples are selected from the original sample of X and Y pairs. This method is appropriate
for data in which both X and Y have been selected at random. That is, the X values were not predetermined,
but came in as measurements just as the Y values.
An example of this situation would be if a population of individuals is sampled and both Y and X are
measured on those individuals only after the sample is selected. That is, the value of X was not used in the
selection of the sample.

Bootstrap the Residuals

The bootstrap samples are constructed using the modified residuals. In each bootstrap sample, the
randomly sampled modified residuals are added to the original fitted values forming new values of Y. This
method forces the original structure of the X values to be retained in every bootstrap sample.
This method is appropriate for data obtained from a designed experiment in which the values of X are
preset by the experimental design.
Because the residuals are sampled and added back at random, the method must assume that the variance
of the residuals is constant. If the sizes of the residuals are proportional to X, this method should not
be used.

Bootstrap Prediction Intervals

Bootstrap confidence intervals for the mean of Y given X are generated from the bootstrap sample in the
usual way. To calculate prediction intervals for the predicted value (not the mean) of Y given X requires a
modification to the predicted value of Y to account for the variation of Y about its mean. This modification of
the predicted Y values in the bootstrap sample, suggested by Davison and Hinkley, is as follows.
𝑦𝑦�+ = 𝑦𝑦� − 𝑥𝑥(𝑏𝑏1∗ − 𝑏𝑏1 ) + 𝑒𝑒+∗

where 𝑒𝑒+∗ is a randomly selected modified residual. By adding the randomly sampled residual we have
added an appropriate amount of variation to represent the variance of individual Y’s about their mean
value.

300-23
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Randomization Test
Because of the strict assumptions that must be made when using this procedure to test hypotheses
about the slope, NCSS also includes a randomization test as outlined by Edgington (1987). Randomization
tests are becoming more and more popular as the speed of computers allows them to be computed in
seconds rather than hours.
A randomization test is conducted by enumerating all possible permutations of the dependent variable
while leaving the independent variable in the original order. The slope is calculated for each permutation
and the number of permutations that result in a slope with a magnitude greater than or equal to the
actual slope is counted. Dividing this count by the number of permutations tried gives the significance
level of the test.
For even moderate sample sizes, the total number of permutations is in the trillions, so a Monte Carlo
approach is used in which the permutations are found by random selection rather than complete
enumeration. Edgington suggests that at least 1,000 permutations be selected. We suggest that this be
increased to 10,000.

Data Structure
The data are entered as two variables. If weights or frequencies are available, they are entered separately in
other variables. An example of data appropriate for this procedure is shown below. These data are the
heights and weights of twenty individuals. The data are contained in the LINREG1 database. We suggest that
you open this database now so that you can follow along with the examples.

LinReg1 Dataset (Subset)

Height Weight
64 159
63 155
67 157
60 125
52 103
58 122
56 101
52 82
79 228
76 199
73 195

Missing Values
Rows with missing values in the variables being analyzed are ignored. If data are present on a row for all but
the dependent variable, a predicted value and confidence limits are generated for that row.

300-24
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Example 1 – Running a Linear Regression Analysis

This section presents an example of how to run a linear regression analysis of the data in the LinReg1
dataset. In this example, we will run a regression of Height on Weight. Predicted values of Height are wanted
at Weight values equal to 90, 100, 150, 200, and 250.
This regression program outputs over thirty different reports and plots, many of which contain duplicate
information. For the purposes of annotating the output, we will output all the reports. (Normally, you would
only select a few of these reports.)

Setup
To run this example, complete the following steps:

1 Open the LinReg1 example dataset

• From the File menu of the NCSS Data window, select Open Example Data.
• Select LinReg1 and click OK.

2 Specify the Linear Regression and Correlation procedure options

• Find and open the Linear Regression and Correlation procedure using the menus or the Procedure
Navigator.
• The settings for this example are listed below and are stored in the Example 1 settings file. To load
these settings to the procedure window, click Open Example Settings File in the Help Center or File
menu.

Variables Tab
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Y: Dependent Variable(s) ....................................... Height

X: Independent Variable ......................................... Weight

Reports Tab
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Alphas, Confidence Level, Power, and Notes

___________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Compute Power ...................................................... Checked

Show Notes ............................................................ Checked

Select Reports
___________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

All Available Reports............................................... Checked (click the Check All button)

Predict Y at these X Values .................................... 90 100 150 200 250

Resampling
___________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Random Seed ......................................................... 3118927 (for reproducibility)

Perform Randomization Tests ................................ Checked
Monte Carlo Samples ........................................... 1000
Calculate Bootstrap Confidence Intervals for .......... Checked
Regression Estimates and Predicted Values

Bootstrap Calculation Options

Samples .................................................................. 1000

300-25
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Plots Tab
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

All Available Plots ................................................... Checked (click the Check All button)

3 Run the procedure

• Click the Run button to perform the calculations and generate the output.

Y vs X Linear Regression Plot

Y vs X Linear Regression Plot

─────────────────────────────────────────────────────────────────────────

The plot shows the data and the linear regression line. This plot is very useful for finding outliers and
nonlinearities. It gives you a good feel for how well the linear regression model fits the data.

300-26
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Run Summary

Run Summary
─────────────────────────────────────────────────────────────────────────
Item Value Rows Value
─────────────────────────────────────────────────────────────────────────────────────────────────────────
Dependent Variable (Y) Height Rows Processed 26
Independent Variable (X) Weight Rows Used in Estimation 20
Frequency Variable None Rows with X Missing 3
Weight Variable None Rows with Y Missing 3
Intercept 35.1337
Slope 0.1932
R² 0.9738
Correlation 0.9868
Coefficient of Variation 0.0226
Mean Square Error (MSE) 1.970176
Square Root of MSE 1.40363
─────────────────────────────────────────────────────────────────────────

This report summarizes the linear regression results. It presents the variables used, the number of rows
used, and the basic least squares results. These values are repeated later in specific reports, so they will not
be discussed further here.

Coefficient of Variation
The coefficient of variation is a relative measure of dispersion, computed by dividing the square root of the
mean square error by the mean of Y. By itself, it has little value, but it can be useful in comparative studies.

√𝑀𝑀𝑀𝑀𝑀𝑀
𝐶𝐶𝐶𝐶 =
𝑌𝑌�

Summary Statement

Summary Statement
─────────────────────────────────────────────────────────────────────────
The equation of the straight line relating Height and Weight is estimated as: Height = (35.1337) + (0.1932) * Weight
using the 20 observations in this dataset. The y-intercept, the estimated value of Height when Weight is zero, is
35.1337 with a standard error of 1.0887. The slope, the estimated change in Height per unit change in Weight, is
0.1932 with a standard error of 0.0075. The value of R², the proportion of the variation in Height that can be
accounted for by variation in Weight, is 0.9738. The correlation between Height and Weight is 0.9868.

A significance test that the slope is zero resulted in a t-value of 25.8679. The significance level of this t-test is
0.0000. Since 0.0000 < 0.05, the hypothesis that the slope is zero is rejected.

The estimated slope is 0.1932. The lower limit of the 95% confidence interval for the slope is 0.1775 and the upper
limit is 0.2089. The estimated intercept is 35.1337. The lower limit of the 95% confidence interval for the intercept is
32.8464 and the upper limit is 37.4209.
─────────────────────────────────────────────────────────────────────────

This report explains the results in text format.

300-27
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Descriptive Statistics

Descriptive Statistics
─────────────────────────────────────────────────────────────────────────
Model Variable
───────────────────
Parameter Dependent Independent
─────────────────────────────────────────────────────────────────
Variable Name Height Weight
Count 20 20
Mean 62.1 139.6
Standard Deviation 8.441128 43.1221
Minimum 51 82
Maximum 79 228
─────────────────────────────────────────────────────────────────────────

This report presents the mean, standard deviation, minimum, and maximum of the two variables. It is
particularly useful for checking that the correct variables were selected.

Regression Estimation

Regression Estimation
─────────────────────────────────────────────────────────────────────────
Intercept Slope
Parameter B(0) B(1)
──────────────────────────────────────────────────────────────────────────
Regression Coefficients 35.1337 0.1932
Lower 95% Confidence Limit 32.8464 0.1775
Upper 95% Confidence Limit 37.4209 0.2089
Standard Error 1.0887 0.0075
Standardized Coefficient 0.0000 0.9868

T-Statistic 32.2716 25.8679

P-Value (T-Test) 0.0000 0.0000
P-Value (Randomization Test*) 0.0010
Reject H0 (Alpha = 0.05) Yes Yes
Power† 1.0000 1.0000

Regression of Y on X 35.1337 0.1932

Inverse Regression from X on Y 34.4083 0.1984
Orthogonal Regression of Y and X 35.1076 0.1934
──────────────────────────────────────────────────────────────────────────

Estimated Model
────────────────────────────────────────────────────────────────
Height =
(35.1336680743148) + (0.193168566802902) * (Weight)
────────────────────────────────────────────────────────────────
─────────────────────────────────────────────────────────────────────────
* Number of Monte Carlo Samples = 1000, User-Entered Random Seed = 3118927.
† Power was calculated using the observed T-Statistic as the population effect size with a significance level of Alpha = 0.05.

Notes:
The above report shows the least-squares estimates of the intercept and slope followed by the corresponding standard errors,
confidence intervals, and hypothesis tests. Note that these results are based on several assumptions that should be validated
before they are used.

This section reports the values and significance tests of the regression coefficients. Before using this report,
check that the assumptions are reasonable by looking at the tests of assumptions report.

300-28
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Regression Coefficients
The regression coefficients are the least-squares estimates of the Y-intercept and the slope. The slope
indicates how much of a change in Y occurs for a one-unit change in X.

Lower and Upper 95% Confidence Limits

These are the lower and upper values of a 100(1 - α)% interval estimate for 𝛽𝛽𝑗𝑗 based on a t-distribution with
N - 2 degrees of freedom. This interval estimate assumes that the residuals for the regression model are
normally distributed.
The formulas for the lower and upper confidence limits are

𝑏𝑏𝑗𝑗 ± 𝑡𝑡1−𝛼𝛼/2,𝑛𝑛−2 𝑠𝑠𝑏𝑏𝑗𝑗

Standard Error
The standard error of the regression coefficient, 𝑠𝑠𝑏𝑏𝑗𝑗 , is the standard deviation of the estimate. It provides a
measure of the precision of the estimated regression coefficient. It is used in hypothesis tests or confidence
limits.

Standardized Coefficient
Standardized regression coefficients are the coefficients that would be obtained if you standardized both
variables. Here standardizing is defined as subtracting the mean and dividing by the standard deviation of a
variable. A regression analysis on these standardized variables would yield these standardized coefficients.
The formula for the standardized regression coefficient is:

𝑠𝑠𝑋𝑋
𝑏𝑏1,𝑠𝑠𝑠𝑠𝑠𝑠 = 𝑏𝑏1 � �
𝑠𝑠𝑌𝑌

where 𝑠𝑠𝑌𝑌 and 𝑠𝑠𝑋𝑋 are the standard deviations for the dependent and independent variables, respectively.
Note that in the case of linear regression, the standardized coefficient is equal to the correlation between
the two variables.

T-Statistic
These are the t-test values for testing the hypotheses that the intercept and the slope are zero versus the
alternative that they are nonzero. These t-values have N - 2 degrees of freedom.
To test that the slope is equal to a hypothesized value other than zero, inspect the confidence limits. If the
hypothesized value is outside the confidence limits, the hypothesis is rejected. Otherwise, it is not rejected.

P-Value (T-Test)
This is the two-sided p-value for the significance test of the regression coefficient. The p-value is the
probability that this t-statistic will take on a value at least as extreme as the actually observed value,
assuming that the null hypothesis is true (i.e., the regression estimate is equal to zero). If the p-value is less
than alpha, say 0.05, the null hypothesis is rejected.

300-29
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

P-Value (Randomization Test)

This is the two-sided p-value for the randomization test of whether the slope is zero. Since this value is
based on a randomization test, it does not require all of the assumptions that the t-test does. The number
of Monte Carlo samples of the permutation distribution of the slope is shown in parentheses.

Reject H0 (Alpha = 0.05)

This value indicates whether the null hypothesis was rejected. Note that the level of significance was
specified as the value of Alpha.

Power
Power is the probability of rejecting the null hypothesis that the regression coefficient is zero when in truth,
the regression coefficient is some value other than zero. The power is calculated for the case when the
estimated coefficient is the actual coefficient, the estimate variance is the true variance, and Alpha is the
given value.
High power is desirable. High power means that there is a high probability of rejecting the null hypothesis
when the null hypothesis is false. This is a critical measure of sensitivity in hypothesis testing. This estimate
of power is based upon the assumption that the residuals are normally distributed.

Regression of Y on X
These are the usual least squares estimates of the intercept and slope from a linear regression of Y on X.
These quantities were given earlier and are reproduced here to allow easy comparisons.

Regression of X on Y
These are the estimated intercept and slope derived from the coefficients of linear regression of X on Y.
These quantities may be useful in calibration and inverse prediction.

Orthogonal Regression of Y and X

These are the estimates of the intercept and slope from an orthogonal regression of Y on X. This equation
minimizes the sum of the squared perpendicular distances between the points and the regression line.

Estimated Model
This is the least squares regression line presented in double precision. Besides showing the regression
model in long form, it may be used as a transformation by copying and pasting it into the Transformation
portion of the spreadsheet.

300-30
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Bootstrap Confidence Intervals for Regression Coefficient Estimates

Bootstrap Confidence Intervals for Regression Coefficient Estimates

─────────────────────────────────────────────────────────────────────────
Estimation Results Bootstrap Confidence Interval Limits
─────────────────────── ────────────────────────────
Parameter Estimate | Confidence Level Lower Upper
────────────────────────────────────────────────────────────────────────────────────────────────────

Intercept
Original Value 35.1337 | 90% 33.4868 36.7965
Bootstrap Mean 35.1470 | 95% 33.1653 37.3351
Bias (BM - OV) 0.0133 | 99% 32.6453 38.1047
Bias Corrected Value 35.1204
Standard Error 1.0327

Slope
Original Value 0.1932 | 90% 0.1815 0.2049
Bootstrap Mean 0.1931 | 95% 0.1782 0.2075
Bias (BM - OV) -0.0001 | 99% 0.1703 0.2108
Bias Corrected Value 0.1932
Standard Error 0.0071

Correlation
Original Value 0.9868 | 90% 0.9797 0.9970
Bootstrap Mean 0.9867 | 95% 0.9788 1.0000
Bias (BM - OV) -0.0001 | 99% 0.9766 1.0000
Bias Corrected Value 0.9869
Standard Error 0.0054

R²
Original Value 0.9738 | 90% 0.9597 0.9937
Bootstrap Mean 0.9736 | 95% 0.9580 1.0000
Bias (BM - OV) -0.0002 | 99% 0.9535 1.0000
Bias Corrected Value 0.9740
Standard Error 0.0105

Standard Error of Estimate

Original Value 1.4036 | 90% 1.1608 1.8326
Bootstrap Mean 1.3204 | 95% 1.1275 1.9126
Bias (BM - OV) -0.0833 | 99% 1.0015 2.0661
Bias Corrected Value 1.4869
Standard Error 0.2051

Orthogonal Regression Intercept

Original Value 35.1076 | 90% 33.4494 36.7878
Bootstrap Mean 35.1205 | 95% 33.1306 37.3240
Bias (BM - OV) 0.0129 | 99% 32.6325 38.1117
Bias Corrected Value 35.0946
Standard Error 1.0383

300-31
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Orthogonal Regression Slope

Original Value 0.1934 | 90% 0.1817 0.2051
Bootstrap Mean 0.1933 | 95% 0.1785 0.2078
Bias (BM - OV) -0.0001 | 99% 0.1703 0.2111
Bias Corrected Value 0.1934
Standard Error 0.0072
─────────────────────────────────────────────────────────────────────────
Number of Bootstrap Samples = 1000, Sampling Method = Observations, Confidence Interval Method = Reflection,
User-Entered Random Seed = 3118927.

Notes:
The main purpose of this report is to present the bootstrap confidence intervals of various parameters. All gross outliers should
have been removed. The sample size should be at least 50 and the sample should be "representative" of the population from
which it was drawn.

This report provides bootstrap estimates of the slope and intercept of the least squares regression line and
the orthogonal regression line, the correlation coefficient, and other linear regression quantities. Details of
the bootstrap method were presented earlier in this chapter.

Original Value
This is the parameter estimate obtained from the complete sample without bootstrapping.

Bootstrap Mean
This is the average of the parameter estimates of the bootstrap samples.

Bias (BM - OV)

This is an estimate of the bias in the original estimate. It is computed by subtracting the original value from
the bootstrap mean.

Bias Corrected Value

This is an estimated of the parameter that has been corrected for its bias. The correction is made by
subtracting the estimated bias from the original parameter estimate.

Standard Error
This is the bootstrap method’s estimate of the standard error of the parameter estimate. It is simply the
standard deviation of the parameter estimate computed from the bootstrap estimates.

Conf. Level
This is the confidence coefficient of the bootstrap confidence interval given to the right.

Bootstrap Confidence Limits (Lower and Upper)

These are the limits of the bootstrap confidence interval with the confidence coefficient given to the left.
These limits are computed using the confidence interval method (percentile or reflection) designated on the
Bootstrap panel.
Note that to be accurate, these intervals must be based on over a thousand bootstrap samples and the
original sample must be representative of the population.

300-32
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Bootstrap Histograms of Regression Coefficient Estimates

Bootstrap Histograms of Regression Coefficient Estimates

─────────────────────────────────────────────────────────────────────────

(3 more histograms are shown)

Each histogram shows the distribution of the corresponding parameter estimate.

300-33
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Correlation and R²

Correlation and R²
─────────────────────────────────────────────────────────────────────────
Spearman
Pearson Rank
Correlation Correlation
Parameter Coefficient R² Coefficient
─────────────────────────────────────────────────────────────────────────────────────────────────
Estimated Value 0.9868 0.9738 0.9759
Lower 95% Conf. Limit (r distribution) 0.9646
Upper 95% Conf. Limit (r distribution) 0.9945
Lower 95% Conf. Limit (Fisher's z) 0.9662 0.9387
Upper 95% Conf. Limit (Fisher's z) 0.9949 0.9906
Adjusted (Rbar) 0.9723

T-Statistic for Testing H0: Rho = 0 25.8679 25.8679 18.9539

P-Value for Testing H0: Rho = 0 0.0000 0.0000 0.0000
P-Value of Randomization Test* 0.0010
─────────────────────────────────────────────────────────────────────────
* Number of Monte Carlo Samples = 1000, User-Entered Random Seed = 3118927.

Notes:
The confidence interval for the Pearson correlation assumes that X and Y follow the bivariate normal distribution. This is a
different assumption from linear regression which assumes that X is fixed and Y is normally distributed.

Two confidence intervals are given. The first is based on the exact distribution of Pearson's correlation. The second is based on
Fisher's z transformation which approximates the exact distribution using the normal distribution. Why are both provided?
Because most books only mention Fisher's approximate method, it will often be needed to do homework. However, the exact
methods should be used whenever possible.

The confidence limits can be used to test hypotheses about the correlation. To test the hypothesis that rho is a specific value, say
r0, check to see if r0 is between the confidence limits. If it is, the null hypothesis that rho = r0 is not rejected. If r0 is outside the
limits, the null hypothesis is rejected.

Spearman's Rank correlation is calculated by replacing the original data with their ranks. This correlation is used when some of
the assumptions may be invalid.

This report provides results about Pearson’s correlation, R2, and Spearman’s rank correlation.

Pearson Correlation Coefficient

Details of the calculation of this value were given earlier in the chapter. Remember that this value is an
index of the strength of the linear association between X and Y. The range of values is from -1 to 1. Strong
association occurs when the magnitude of the correlation is close to one. Low correlations are those near
zero.
Two sets of confidence limits are given. The first is a set of exact limits computed from the distribution of
the correlation coefficient. These limits assume that X and Y follow the bivariate normal distribution. The
second set of limits are limits developed by R. A. Fisher as an approximation to the exact limits. The
approximation is quite good as you can see by comparing the two sets of limits. The second set is provided
because they are often found in statistics books. In most cases, you should use the first set based on the r
distribution because they are exact. You may want to compare these limits with those found for the
correlation in the Bootstrap report.
The two-sided hypothesis test and probability level are for testing whether the correlation is zero.

300-34
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

P-Value of Randomization Test

This is the two-sided p-value for the randomization test of whether the slope is zero. This probability value
may also be used to test whether the Pearson correlation is zero. Since this value is based on a
randomization test, it does not require all of the assumptions that the parametric test does. The number of
Monte Carlo samples of the permutation distribution of the slope is shown in parentheses.

Spearman Rank Correlation Coefficient

The Spearman’s rank correlation is simply the Pearson correlation computed on the ranks of X and Y rather
than on the actual data. By using the ranks, some of the assumptions may be relaxed. However, the
interpretation of the correlation is much more difficult.
The confidence interval for this correlation is calculated using the Fisher’s z transformation of the rank
correlation.
The two-sided hypothesis test and probability level are for testing whether the rank correlation is zero.

R-Squared
𝑅𝑅 2, officially known as the coefficient of determination, is defined as

𝑆𝑆𝑆𝑆𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀
𝑅𝑅 2 =
𝑆𝑆𝑆𝑆𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇

𝑅𝑅 2 is probably the most popular statistical measure of how well the regression model fits the data. 𝑅𝑅 2 may
be defined either as a ratio or a percentage. Since we use the ratio form, its values range from zero to one.
A value of 𝑅𝑅 2 near zero indicates no linear relationship between the Y and X, while a value near one
indicates a perfect linear fit. Although popular, 𝑅𝑅 2 should not be used indiscriminately or interpreted
without scatter plot support. Following are some qualifications on its interpretation:
1. Linearity. 𝑅𝑅 2 does not measure the appropriateness of a linear model. It measures the strength of
the linear component of the model. Suppose the relationship between X and Y was a perfect circle.
The 𝑅𝑅 2 value of this relationship would be zero.
2. Predictability. A large 𝑅𝑅 2 does not necessarily mean high predictability, nor does a low 𝑅𝑅 2 necessarily
mean poor predictability.
3. No-intercept model. The definition of 𝑅𝑅 2 assumes that there is an intercept in the regression model.
When the intercept is left out of the model, the definition of 𝑅𝑅 2 changes dramatically. The fact that
your R 𝑅𝑅 2 value increases when you remove the intercept from the regression model does not
2

reflect an increase in the goodness of fit. Rather, it reflects a change in the underlying meaning of
𝑅𝑅 2.
4. Sample size. 𝑅𝑅 2 is highly sensitive to the number of observations. The smaller the sample size, the
larger its value.

Adjusted R-Squared
This is an adjusted version of 𝑅𝑅 2. The adjustment seeks to remove the distortion due to a small sample size.

2 𝑁𝑁 − 1
𝑅𝑅𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = 1 − (1 − 𝑅𝑅 2 ) � �
𝑁𝑁 − 2

300-35
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Analysis of Variance

Analysis of Variance
─────────────────────────────────────────────────────────────────────────
Sum of Mean
Source DF Squares Square F-Ratio P-Value Power*
─────────────────────────────────────────────────────────────────────────────────────────────────────────
Intercept 1 77128.2 77128.2
Slope 1 1318.337 1318.337 669.1468 0.0000 1.0000
Error 18 35.46317 1.970176
Lack of Fit 16 34.96317 2.185198 8.7408 0.1074
Pure Error 2 0.5 0.25
Adjusted Total 19 1353.8 71.25263
Total 20 78482
─────────────────────────────────────────────────────────────────────────────────────────────────────────

Standard Deviation of Residuals

────────────────────────────────────────────
s = Square Root(1.970176) = 1.40363
────────────────────────────────────────────
─────────────────────────────────────────────────────────────────────────
* Power was calculated using the observed F-Ratio as the population effect size with a significance level of Alpha = 0.05.

Notes:
The above report shows the F-Ratio for testing whether the slope is zero, the degrees of freedom, and the mean square error.
The mean square error, which estimates the variance of the residuals, is used extensively in the calculation of hypothesis tests
and confidence intervals.

An analysis of variance (ANOVA) table summarizes the information related to the sources of variation in
data.

Source
This represents the partitions of the variation in Y. There are four sources of variation listed: intercept, slope,
error, and total (adjusted for the mean).

DF
The degrees of freedom are the number of dimensions associated with this term. Note that each
observation can be interpreted as a dimension in N-dimensional space. The degrees of freedom for the
intercept, model, error, and adjusted total are 1, 1, N - 2, and N - 1, respectively.

Sum of Squares
These are the sums of squares associated with the corresponding sources of variation. Note that these
values are in terms of the dependent variable, Y. The formulas for each are

𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 = 𝑁𝑁𝑌𝑌� 2

2
𝑆𝑆𝑆𝑆𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 = Σ�𝑌𝑌� − 𝑌𝑌��

2
𝑆𝑆𝑆𝑆𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 = Σ�𝑌𝑌 − 𝑌𝑌��

𝑆𝑆𝑆𝑆𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 = Σ(𝑌𝑌 − 𝑌𝑌�)2

Note that the lack of fit and pure error values are provided if there are observations with identical values of
the independent variable.

300-36
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Mean Square
The mean square is the sum of squares divided by the degrees of freedom. This mean square is an
estimated variance. For example, the mean square error is the estimated variance of the residuals (the
residuals are sometimes called the errors).

F-Ratio
This is the F statistic for testing the null hypothesis that the slope equals zero. This F-statistic has 1 degree of
freedom for the numerator variance and N - 2 degrees of freedom for the denominator variance.

P-Value
This is the p-value for the above F test. The p-value is the probability that the test statistic will take on a value
at least as extreme as the observed value, assuming that the null hypothesis is true. If the p-value is less
than alpha, say 0.05, the null hypothesis is rejected. If the p-value is greater than alpha, the null hypothesis
is accepted.

Power
Power is the probability of rejecting the null hypothesis that the slope is zero when it is not.

Standard Deviation of Residuals

s is the square root of the mean square error. It is an estimate of the standard deviation of the residuals.

Summary Matrices

Summary Matrices
─────────────────────────────────────────────────────────────────────────
Calculation Matrix
──────────────────────────────────────────────────────────────────────────────────────────────────
X'X X'X X'Y X'X Inverse X'X Inverse
Index 0 1 2 0 1
──────────────────────────────────────────────────────────────────────────────────────────────────
0 20 2792 1242 0.6015912 -0.003951227
1 2792 425094 180208 -0.003951227 2.830392E-05
2 (Y'Y) 78482
Determinant 706616 1.415196E-06
──────────────────────────────────────────────────────────────────────────────────────────────────

Variance-Covariance Matrix of Regression Coefficients

───────────────────────────────────────────────────
VC(b) VC(b)
Index 0 1
───────────────────────────────────────────────────
0 1.185241 -0.007784612
1 -0.007784612 5.576369E-05
───────────────────────────────────────────────────

─────────────────────────────────────────────────────────────────────────

This section provides the matrices from which the least square regression values are calculated and the
variance-covariance matrix of regression coefficients. Occasionally, these values may be useful in hand
calculations.

300-37
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Tests of Assumptions

Tests of Assumptions
─────────────────────────────────────────────────────────────────────────
Test Is the Assumption
Statistic Reasonable at the 0.2
Assumption/Test Value P-Value Level of Significance?
──────────────────────────────────────────────────────────────────────────────────────────────────────

Residuals Follow a Normal Distribution?

Shapiro-Wilk 0.9728 0.8129 Yes
Anderson-Darling 0.2751 0.6609 Yes
D'Agostino Skewness -0.9590 0.3375 Yes
D'Agostino Kurtosis 0.1205 0.9041 Yes
D'Agostino Omnibus 0.9343 0.6268 Yes

Constant Residual Variance?

Modified Levene Test 0.0946 0.7620 Yes

Relationship is a Straight Line?

Lack of Linear Fit F(16, 2) Test 8.7408 0.1074 No

No Serial Correlation?
Evaluate the Serial-Correlation report and the Durbin-Watson test if you have equal-spaced time series data.
─────────────────────────────────────────────────────────────────────────
Notes:
A "Yes" means there is not enough evidence to make this assumption seem unreasonable. This lack of evidence may be
because the sample size is too small, the assumptions of the test itself are not met, or the assumption is valid.
A "No" means the that the assumption is not reasonable. However, since these tests are related to sample size, you should
assess the role of sample size in the tests by also evaluating the appropriate plots and graphs. A large dataset (say N > 500) will
often fail at least one of the normality tests because it is hard to find a large dataset that is perfectly normal.

Normality and Constant Residual Variance:

Possible remedies for the failure of these assumptions include using a transformation of Y such as the log or square root,
correcting data-recording errors found by looking into outliers, adding additional independent variables, using robust regression,
or using bootstrap methods.

Straight-Line:
Possible remedies for the failure of this assumption include using nonlinear regression or polynomial regression.

This report presents numeric tests of some of the assumptions made when using linear regression. The
results of these tests should be compared to appropriate plots to determine if the assumptions are valid or
not.
Note that a ‘Yes’ means that there is not enough evidence to reject the assumption. This lack of assumption
test rejection may be because the sample size is too small, or the assumptions of the test were no met. It
does not necessarily mean that the data met assumption. Likewise, a ‘No’ may occur because the sample
size is very large. It is almost always possible to fail a preliminary test given a large enough sample size. No
assumption is every fits perfectly. Bottom line, you should also investigate plots designed to check the
assumptions.

Residuals Follow a Normal Distribution?

This section displays the results of five normality tests of the residuals. The Shapiro-Wilk and Anderson-
Darling tests are usually considered as the best.
Unfortunately, these tests have small statistical power (probability of detecting nonnormal data) unless the
sample sizes are large, say over 300. Hence, if the decision is to reject normality, you can be reasonably
certain that the data are not normal. However, if the decision is to not reject, the situation is not as clear. If
you have a sample size of 300 or more, you can reasonably assume that the actual distribution is closely

300-38
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

approximated by the normal distribution. If your sample size is less than 300, all you know for sure is that
there was not enough evidence in your data to reject the normality of residuals assumption. In other words,
the data might be nonnormal, you just could not prove it. In this case, you must rely on the graphics to
justify the normality assumption.

Shapiro-Wilk W Test
This test for normality, developed by Shapiro and Wilk (1965), has been found to be the most powerful test
in most situations. It is the ratio of two estimates of the variance of a normal distribution based on a
random sample of N observations. The numerator is proportional to the square of the best linear estimator
of the standard deviation. The denominator is the sum of squares of the observations about the sample
mean. W may be written as the square of the Pearson correlation coefficient between the ordered
observations and a set of weights which are used to calculate the numerator. Since these weights are
asymptotically proportional to the corresponding expected normal order statistics, W is roughly a measure
of the straightness of the normal quantile-quantile plot. Hence, the closer W is to one, the more normal the
sample is.
The probability values for W are valid for samples in the range of 3 to 5000.
The test is not calculated when a frequency variable is specified.

Anderson-Darling Test
This test, developed by Anderson and Darling (1954), is based on EDF statistics. In some situations, it has
been found to be as powerful as the Shapiro-Wilk test.
The test is not calculated when a frequency variable is specified.

D’Agostino Skewness
D’Agostino (1990) proposed a normality test based on the skewness coefficient, �𝑏𝑏1. Because the normal
distribution is symmetrical, �𝑏𝑏1 is equal to zero for normal data. Hence, a test can be developed to
determine if the value of �𝑏𝑏1 is significantly different from zero. If it is, the data are obviously nonnormal.
The test statistic is, under the null hypothesis of normality, approximately normally distributed. The
computation of this statistic is restricted to sample sizes greater than 8. The formula and further details are
given in the Descriptive Statistics chapter.

D’Agostino Kurtosis
D’Agostino (1990) proposed a normality test based on the kurtosis coefficient, 𝑏𝑏2 . For the normal
distribution, the theoretical value of 𝑏𝑏2 is 3. Hence, a test can be developed to determine if the value of 𝑏𝑏2 is
significantly different from 3. If it is, the residuals are obviously nonnormal. The test statistic is, under the
null hypothesis of normality, approximately normally distributed for sample sizes N > 20. The formula and
further details are given in the Descriptive Statistics chapter.

D’Agostino Omnibus
D’Agostino (1990) proposed a normality test that combines the tests for skewness and kurtosis. The statistic,
𝐾𝐾 2 , is approximately distributed as a chi-square with two degrees of freedom.

300-39
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Constant Residual Variance?

Linear regression assumes that the residuals have constant variance. The validity of this assumption can be
checked by looking at a plot of the absolute values of the residuals versus the X variable. The modified
Levene test may be used when a numerical answer is needed.
If your data fail this test, you may want to use a logarithm transformation or a weighted regression.

Modified Levene Test

The modified Levene test can be used to evaluate the validity of the assumption of constant variance. It has
been shown to be reliable even when the residuals do not follow a normal distribution. The mathematical
details of the test were presented earlier in this chapter.

Relationship is a Straight Line?

Linear regression assumes that the relationship between X and Y is a straight line (linear). The validity of this
assumption can be checked by looking at the plot Y versus X and at the plot of the residuals versus X. The
lack of fit test may be used when a numerical answer is needed.
If your data fail this test, you may want to use a different model which accounts for the curvature. The
Growth and Other Models procedure in curve fitting is a good choice when curvature exists in your data.

Lack of Linear Fit Test

The lack-of-fit test is used to test for a departure from the linear fit. This test requires that there are multiple
observations for at least one X value. When such is the case, an estimate of pure error and lack of fit can be
found, and an F test created. The mathematical details of the test were presented earlier in this chapter.

Serial Correlation of Residuals and Durbin-Watson Test for Serial

Correlation

Serial Correlation of Residuals

─────────────────────────────────────────────────────────────────────────
Serial Serial Serial
Lag Correlation Lag Correlation Lag Correlation
────────────────────────────────────────────────────────────────────────────────────
1 0.1029 9 -0.2353 17
2 -0.4127* 10 -0.0827 18
3 0.0340 11 -0.0316 19
4 0.2171 12 -0.0481 20
5 -0.1968 13 0.0744 21
6 -0.0194 14 0.0073 22
7 0.2531 15 23
8 -0.0744 16 24
─────────────────────────────────────────────────────────────────────────
Notes:
Each serial correlation is the Pearson correlation calculated between the original series of residuals and the residuals lagged the
specified number of periods. This feature of residuals is only meaningful for data obtained sorted in time order. One of the
assumptions is that none of these serial correlations is significant. Starred correlations are those for which |Fisher's Z| > 1.645
which indicates whether the serial correlation is "large."

If serial correlation is detected in time series data, the remedy is to account for it either by replacing Y with first differences or by
fitting the serial pattern using a method such as that proposed by Cochrane and Orcutt.

300-40
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Durbin-Watson Test For Serial Correlation

─────────────────────────────────────────────────────────────────────────
Test Statistic
Value to Test Reject H0
Test Type H0: ρ(1) = 0 P-Value at α = 0.2?
──────────────────────────────────────────────────────────────────────────────────────────────
Positive Serial Correlation Test 1.6978 0.2366 No
Negative Serial Correlation Test 1.6978 0.7460 No
─────────────────────────────────────────────────────────────────────────
Notes:
The Durbin-Watson test was created to test for first-order serial correlation in regression data taken over time. If the rows of your
dataset do not represent successive time periods, you should ignore this test.

This report gives the probability of rejecting the null hypothesis of no first-order serial correlation. Possible remedies for serial
correlation were given in the Notes to the Serial Correlation report, above.

This section reports on the autocorrelation structure of the residuals. Of course, if your data were not taken
through time, this section should be ignored.

Lag
The lag, k, is the number of periods back.

Serial Correlation
The serial correlation reported here is the sample autocorrelation coefficient of lag k. It is computed as

∑ 𝑒𝑒𝑖𝑖−𝑘𝑘 𝑒𝑒𝑖𝑖
𝑟𝑟𝑘𝑘 = for 𝑘𝑘 = 1, 2, … , 24
∑ 𝑒𝑒𝑖𝑖2

The distribution of these autocorrelations may be approximated by the distribution of the regular
correlation coefficient. Using this fact, Fisher’s Z transformation may be used to find large autocorrelations.
If the Fisher’s Z transformation of the autocorrelation is greater than 1.645, the autocorrelation is assumed
to be large and the observation is starred.

Durbin-Watson Test Statistic

The Durbin-Watson test is often used to test for positive or negative, first-order, serial correlation. It is
calculated as follows
2
∑𝑁𝑁
𝑗𝑗=2�𝑒𝑒𝑗𝑗 − 𝑒𝑒𝑗𝑗−1 �
𝐷𝐷𝐷𝐷 =
∑𝑁𝑁 2
𝑗𝑗=1 𝑒𝑒𝑗𝑗

The distribution of this test is mathematically difficult because it involves the X values. Originally, Durbin-Watson
(1950, 1951) gave a pair of bounds to be used. However, there is a large range of indecision that can be found
when using these bounds. Instead of using these bounds, NCSS calculates the exact probability using the beta
distribution approximation suggested by Durbin-Watson (1951). This approximation has been shown to be
accurate to three decimal places in most cases.

300-41
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

PRESS Statistics

PRESS Statistics
─────────────────────────────────────────────────────────────────────────
From From
PRESS Regular
Parameter Residuals Residuals
────────────────────────────────────────────────────────────────────
Sum of Squared Residuals 43.15799 35.46317
Sum of |Residuals| 24.27421 22.02947
R² 0.9681 0.9738
─────────────────────────────────────────────────────────────────────────
Notes:
A PRESS residual is found by estimating the regression equation without the observation, predicting the dependent variable, and
subtracting the predicted value from the actual value. The PRESS values are calculated from these PRESS residuals. The
Regular values are the corresponding calculations based on the regular residuals.

The PRESS values are often used to compare models in a multiple-regression variable selection. They show how well the model
predicts observations that were not used in the estimation.

This section reports on the PRESS statistics. The regular statistics, computed on all of the data, are provided
to the side to make comparison between corresponding values easier.

Sum of Squared PRESS Residuals

PRESS is an acronym for prediction sum of squares. It was developed for use in variable selection to validate
a regression model. To calculate PRESS, each observation is individually omitted. The remaining N - 1
observations are used to calculate a regression and estimate the value of the omitted observation. This is
done N times, once for each observation. The difference between the actual Y value and the predicted Y with
the observation deleted is called the prediction error or PRESS residual. The sum of the squared prediction
errors is the PRESS value. The smaller PRESS is, the better the predictability of the model.

Sum of |PRESS Residuals|

This is the sum of the absolute value of the PRESS residuals or prediction errors. If a large value for the
PRESS is due to one or a few large PRESS residuals, this statistic may be a more accurate way to evaluate
predictability.

PRESS R²
The PRESS value above can be used to compute an 𝑅𝑅 2-like statistic, called R2Predict, which reflects the
prediction ability of the model. This is a good way to validate the prediction of a regression model without
selecting another sample or splitting your data. It is very possible to have a high 𝑅𝑅 2 and a very low R2Predict.
When this occurs, it implies that the fitted model is data dependent. This R2Predict ranges from below zero
to above one. When outside the range of zero to one, it is truncated to stay within this range.

300-42
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Predicted Values and Confidence Intervals of Y Means at Specific X Values

Predicted Values and Confidence Intervals for Y Means at Specific X Values

─────────────────────────────────────────────────────────────────────────
95% Confidence Interval
Predicted Standard Limits for Y Mean|X
Weight Height Error of ─────────────────
(X) (Yhat|X) Yhat Mean Lower Upper
─────────────────────────────────────────────────────────────────────────────────
90 52.51884 0.4855 51.49887 53.53881
100 54.45052 0.4312 53.54456 55.35649
150 64.10896 0.3233 63.42967 64.78824
200 73.76738 0.5495 72.61294 74.92183
250 83.42581 0.8821 81.57251 85.27911
─────────────────────────────────────────────────────────────────────────
The confidence interval estimates the mean of the Y values in a large sample of individuals with the stated value of X. The
interval is only accurate if all of the linear regression assumptions are valid.

The predicted values and confidence intervals of the mean response of Y given X are provided here. The
values of X used here are specified in the Predict Y at these X Values option on the Reports tab.
It is important to note that violations of any regression assumptions will invalidate this interval estimate.

X
This is the value of X at which the prediction is made.

Predicted Y (Yhat|X)
The predicted value of Y for the value of X indicated.

Standard Error of Yhat Mean

This is the estimated standard deviation of the predicted value.

95% Confidence Interval Limits for Y Mean|X (Lower and Upper)

These are the lower and upper limits of a 95% confidence interval estimate of the mean of Y at this value of
X. Note that you set the confidence interval alpha on the Reports tab of the procedure input window.

300-43
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Bootstrap Confidence Intervals and Histograms for Predicted Y Means at

Specific X Values

Bootstrap Confidence Intervals for Predicted Y Means at Specific X Values

─────────────────────────────────────────────────────────────────────────
Estimation Results Bootstrap Confidence Interval Limits
─────────────────────── ──────────────────────────────
Parameter Estimate | Confidence Level Lower Upper
───────────────────────────────────────────────────────────────────────────────────────────────────────

Predicted Mean of Height when Weight = 90

Original Value 52.51884 | 90% 51.80128 53.32835
Bootstrap Mean 52.5252 | 95% 51.69855 53.51639
Bias (BM - OV) 0.0064 | 99% 51.49534 53.82327
Bias Corrected Value 52.51248
Standard Error 0.4618

Predicted Mean of Height when Weight = 100

Original Value 54.45052 | 90% 53.82153 55.16396
Bootstrap Mean 54.45611 | 95% 53.72358 55.34012
Bias (BM - OV) 0.0056 | 99% 53.49232 55.54081
Bias Corrected Value 54.44493
Standard Error 0.4116

Predicted Mean of Height when Weight = 150

Original Value 64.10896 | 90% 63.59029 64.66124
Bootstrap Mean 64.11069 | 95% 63.46907 64.77720
Bias (BM - OV) 0.0017 | 99% 63.26657 64.88721
Bias Corrected Value 64.10722
Standard Error 0.3214

Predicted Mean of Height when Weight = 200

Original Value 73.76738 | 90% 72.84132 74.65286
Bootstrap Mean 73.76526 | 95% 72.69940 74.84036
Bias (BM - OV) -0.0021 | 99% 72.16286 75.20052
Bias Corrected Value 73.76951
Standard Error 0.5405

Predicted Mean of Height when Weight = 250

Original Value 83.42581 | 90% 81.98016 84.81046
Bootstrap Mean 83.41983 | 95% 81.65707 85.16362
Bias (BM - OV) -0.0060 | 99% 81.02856 85.72516
Bias Corrected Value 83.43179
Standard Error 0.8579
─────────────────────────────────────────────────────────────────────────
Number of Bootstrap Samples = 1000, Sampling Method = Observations, Confidence Interval Method = Reflection,
User-Entered Random Seed = 3118927.

Notes:
The main purpose of this report is to present the bootstrap confidence intervals of various parameters. All gross outliers should
have been removed. The sample size should be at least 50 and the sample should be "representative" of the population from
which it was drawn.

300-44
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Bootstrap Histograms of Predicted Y Means at Specific X Values

─────────────────────────────────────────────────────────────────────────

(3 more histograms are shown)

This report provides bootstrap estimates of the predicted Y means at each user-entered X value. Details of
the bootstrap method were presented earlier in this chapter.

Predicted Values and Prediction Intervals of Y Individuals at Specific X

Values

Predicted Values and Prediction Intervals for Y Individuals at Specific X Values

─────────────────────────────────────────────────────────────────────────
95% Prediction Interval
Predicted Standard Limits for Y|X
Weight Height Error of ────────────────
(X) (Yhat|X) Yhat Lower Upper
──────────────────────────────────────────────────────────────────────────────
90 52.51884 1.4852 49.39851 55.63917
100 54.45052 1.4684 51.36558 57.53547
150 64.10896 1.4404 61.08281 67.13509
200 73.76738 1.5074 70.60055 76.93422
250 83.42581 1.6578 79.94288 86.90874
─────────────────────────────────────────────────────────────────────────
The prediction interval estimates the predicted value of Y for a single individual with the stated value of X. The interval is only
accurate if all of the linear regression assumptions are valid.

The predicted values and prediction intervals of the response of Y given X are provided here. The values of X
used here are specified in the Predict Y at these X Values option on the Reports tab.
It is important to note that violations of any regression assumptions will invalidate this interval estimate.

X
This is the value of X at which the prediction is made.

300-45
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Predicted Y (Yhat|X)
The predicted value of Y for the value of X indicated.

Standard Error of Yhat

This is the estimated standard deviation of the predicted value.

95% Prediction Limits of Y|X (Lower and Upper)

These are the lower and upper limits of a 95% prediction interval estimate of an individual Y at this value of
X. Note that you set the prediction interval alpha on the Reports tab of the procedure input window.

Bootstrap Confidence Intervals and Histograms for Predicted Y Individuals

at Specific X Values

Bootstrap Confidence Intervals for Predicted Y Individuals at Specific X Values

─────────────────────────────────────────────────────────────────────────
Estimation Results Bootstrap Confidence Interval Limits
─────────────────────── ──────────────────────────────
Parameter Estimate | Confidence Level Lower Upper
───────────────────────────────────────────────────────────────────────────────────────────────────────

Predicted Height when Weight = 90

Original Value 52.51884 | 90% 49.89190 55.64400
Bootstrap Mean 52.49085 | 95% 49.58742 56.09112
Bias (BM - OV) -0.0280 | 99% 47.99874 57.09388
Bias Corrected Value 52.54683
Standard Error 1.7275

Predicted Height when Weight = 100

Original Value 54.45052 | 90% 51.79865 57.36475
Bootstrap Mean 54.41591 | 95% 51.48851 57.84039
Bias (BM - OV) -0.0346 | 99% 50.33467 59.90700
Bias Corrected Value 54.48514
Standard Error 1.7085

Predicted Height when Weight = 150

Original Value 64.10896 | 90% 61.36381 67.21675
Bootstrap Mean 64.12646 | 95% 61.10781 67.58835
Bias (BM - OV) 0.0175 | 99% 59.74065 68.87939
Bias Corrected Value 64.09145
Standard Error 1.7576

Predicted Height when Weight = 200

Original Value 73.76738 | 90% 70.94756 76.70295
Bootstrap Mean 73.76675 | 95% 70.53024 77.44383
Bias (BM - OV) -0.0006 | 99% 69.79328 78.23773
Bias Corrected Value 73.76801
Standard Error 1.7475

300-46
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Predicted Height when Weight = 250

Original Value 83.42581 | 90% 80.68179 86.51836
Bootstrap Mean 83.40162 | 95% 80.23896 87.15035
Bias (BM - OV) -0.0242 | 99% 79.26602 87.89960
Bias Corrected Value 83.45
Standard Error 1.7781
─────────────────────────────────────────────────────────────────────────
Number of Bootstrap Samples = 1000, Sampling Method = Observations, Confidence Interval Method = Reflection,
User-Entered Random Seed = 3118927.

Notes:
The main purpose of this report is to present the bootstrap confidence intervals of various parameters. All gross outliers should
have been removed. The sample size should be at least 50 and the sample should be "representative" of the population from
which it was drawn.

Bootstrap Histograms of Predicted Y Individuals at Specific X Values

─────────────────────────────────────────────────────────────────────────

(3 more histograms are shown)

This report provides bootstrap estimates of the predicted Y individual values at each user-entered X value.
Details of the bootstrap method were presented earlier in this chapter.

300-47
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Residual Plots
The residuals can be graphically analyzed in numerous ways. For certain, the regression analyst should
examine all of the basic residual graphs: the histogram, the density trace, the normal probability plot, the
serial correlation plots (for time series data), the scatter plot of the residuals versus the sequence of the
observations (for time series data), and the scatter plot of the residuals versus the independent variable.
For the scatter plots of residuals versus either the predicted values of Y or the independent variables,
Hoaglin (1983) explains that there are several patterns to look for. You should note that these patterns are
very difficult, if not impossible, to recognize for small data sets.

Point Cloud
A point cloud, basically in the shape of a rectangle or a horizontal band, would indicate no relationship
between the residuals and the variable plotted against them. This is the preferred condition.

Wedge
An increasing or decreasing wedge would be evidence that there is increasing or decreasing (nonconstant)
variation. A transformation of Y may correct the problem or weighted least squares may be needed.

Bowtie
This is similar to the wedge above in that the residual plot shows a decreasing wedge in one direction while
simultaneously having an increasing wedge in the other direction. A transformation of Y may correct the
problem or weighted least squares may be needed.

Sloping Band
This kind of residual plot suggests adding a linear version of the independent variable to the model.

Curved Band
This kind of residual plot may be indicative of a nonlinear relationship between Y and the independent
variable that was not accounted for. The solution might be to use a transformation on Y to create a linear
relationship with X. Another possibility might be to add quadratic or cubic terms of a particular independent
variable.

Curved Band with Increasing or Decreasing Variability

This residual plot is really a combination of the wedge and the curved band. It too must be avoided.

300-48
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Residual Distribution Plots

Histogram
─────────────────────────────────────────────────────────────────────────

The purpose of the histogram of the residuals is to evaluate whether they are normally distributed. Unless
you have a large sample size, it is best not to rely on the histogram for visually evaluating normality of the
residuals. The better choice would be the normal probability plot.

Normal Probability Plot

─────────────────────────────────────────────────────────────────────────

If the residuals are normally distributed, the data points of the normal probability plot will fall along a
straight line. Major deviations from this ideal picture reflect departures from normality. Stragglers at either
end of the normal probability plot indicate outliers. Curvature at both ends of the plot indicates long or

300-49
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

short distributional tails. Convex, or concave, curvature indicates a lack of symmetry. Gaps, plateaus, or
segmentation indicate clustering and may require a closer examination of the data or model. Of course, use
of this graphic tool with very small sample sizes is unwise.
If the residuals are not normally distributed, the t-tests on regression coefficients, the F-tests, and the
interval estimates are not valid. This is a critical assumption to check.

Residuals Plots

Residuals vs X Plots
─────────────────────────────────────────────────────────────────────────

This plot is useful for showing nonlinear patterns and outliers. The preferred pattern is a rectangular shape
or point cloud. Any other nonrandom pattern may require a redefining of the regression model.

300-50
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

|Residuals| vs X Plot
─────────────────────────────────────────────────────────────────────────

This plot is useful for showing nonconstant variance in the residuals. The preferred pattern is a rectangular
shape or point cloud. The most common type of nonconstant variance occurs when the variance is
proportion to X. This is shown by a funnel shape. Remedies for nonconstant variances were discussed
earlier.

RStudent vs X Plot
─────────────────────────────────────────────────────────────────────────

This is a scatter plot of the RStudent residuals versus the independent variable. The preferred pattern is a
rectangular shape or point cloud. This plot is helpful in identifying any outliers.

300-51
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Sequence Plot: Residuals vs Row Number

─────────────────────────────────────────────────────────────────────────

Sequence plots may be useful in finding variables that are not accounted for by the regression equation.
They are especially useful if the data were taken over time.

Serial Correlation Plot: Residuals vs Lagged Residuals

─────────────────────────────────────────────────────────────────────────

This is a scatter plot of the ith residual versus the ith-1 residual. It is only useful for time series data where the
order of the rows on the database is important.
The purpose of this plot is to check for first-order autocorrelation. You would like to see a random pattern,
i.e., a rectangular or uniform distribution of the points. A strong positive or negative trend indicates a need
to redefine the model with some type of autocorrelation component.

300-52
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Positive autocorrelation or serial correlation means that the residual in time period t tends to have the same
sign as the residual in time period (t - 1). On the other hand, a strong negative autocorrelation means that
the residual in time period t tends to have the opposite sign as the residual in time period (t - 1).
Be sure to check the Durbin-Watson statistic.

Predicted Y Values and Residuals

Predicted Y Values and Residuals

─────────────────────────────────────────────────────────────────────────
Predicted
Weight Height Height
Row (X) (Y) (Yhat|X) Residual
────────────────────────────────────────────────────────────────────
1 159 64 65.84747 -1.8475
2 155 63 65.07480 -2.0748
3 157 67 65.46114 1.5389
4 125 60 59.27974 0.7203
5 103 52 55.03003 -3.0300
6 122 58 58.70023 -0.7002
7 101 56 54.64369 1.3563
8 82 52 50.97349 1.0265
9 228 79 79.17610 -0.1761
10 199 76 73.57421 2.4258
11 195 73 72.80154 0.1985
12 110 56 56.38221 -0.3822
13 191 71 72.02886 -1.0289
14 151 65 64.30212 0.6979
15 119 59 58.12073 0.8793
16 119 59 58.12073 0.8793
17 112 58 56.76855 1.2315
18 87 51 51.93933 -0.9393
19 190 71 71.83569 -0.8357
20 87 52 51.93933 0.0607
21 100 54.45052
22 150 64.10896
23 200 73.76738
24 50
25 60
26 70
─────────────────────────────────────────────────────────────────────────
This report provides a data list that may be used to verify whether the correct variables were selected.

This report lists the values of X, Y, the predicted values of Y, and the residuals.

300-53
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Predicted Values and Confidence Intervals for Y Means

Predicted Values and Confidence Intervals for Y Means

─────────────────────────────────────────────────────────────────────────
95% Confidence Interval
Predicted Standard Limits for Y Mean|X
Weight Height Height Error of ─────────────────
Row (X) (Y) (Yhat|X) Yhat Mean Lower Upper
──────────────────────────────────────────────────────────────────────────────────────────────────────────
1 159 64 65.84747 0.3457 65.12122 66.57372
2 155 63 65.07480 0.3343 64.37253 65.77706
3 157 67 65.46114 0.3397 64.74746 66.17480
4 125 60 59.27974 0.3323 58.58169 59.97779
5 103 52 55.03003 0.4162 54.15566 55.90440
6 122 58 58.70023 0.3403 57.98536 59.41511
7 101 56 54.64369 0.4261 53.74841 55.53898
8 82 52 50.97349 0.5325 49.85482 52.09216
9 228 79 79.17610 0.7309 77.64045 80.71175
10 199 76 73.57421 0.5434 72.43261 74.71581
. . . . . . .
. . . . . . .
. . . . . . .
─────────────────────────────────────────────────────────────────────────
The confidence interval estimates the mean of the Y values in a large sample of individuals with the stated value of X. The
interval is only accurate if all of the linear regression assumptions are valid.

The predicted values and confidence intervals of the mean response of Y given X are given for each
observation.

X
This is the value of X at which the prediction is made.

Y
This is the actual value of Y.

Predicted Y (Yhat|X)
The predicted value of Y for the value of X indicated.

Standard Error of Yhat Mean

This is the estimated standard deviation of the predicted mean value.

95% Confidence Interval Limits for Y Mean|X (Lower and Upper)

These are the lower and upper limits of a 95% confidence interval estimate of the mean of Y at this value of
X. Note that you set the confidence interval alpha on the Reports tab of the procedure input window.

300-54
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Predicted Values and Prediction Intervals for Y Individuals

Predicted Values and Prediction Intervals for Y Individuals

─────────────────────────────────────────────────────────────────────────
95% Prediction Interval
Predicted Standard Limits for Y|X
Weight Height Height Error of ────────────────
Row (X) (Y) (Yhat|X) Yhat Lower Upper
──────────────────────────────────────────────────────────────────────────────────────────────────────
1 159 64 65.84747 1.4456 62.81044 68.88450
2 155 63 65.07480 1.4429 62.04341 68.10618
3 157 67 65.46114 1.4441 62.42709 68.49518
4 125 60 59.27974 1.4424 56.24933 62.31015
5 103 52 55.03003 1.4640 51.95422 58.10584
6 122 58 58.70023 1.4443 55.66590 61.73456
7 101 56 54.64369 1.4669 51.56187 57.72552
8 82 52 50.97349 1.5012 47.81952 54.12746
9 228 79 79.17610 1.5825 75.85130 82.50091
10 199 76 73.57421 1.5051 70.41203 76.73639
. . . . . . .
. . . . . . .
. . . . . . .
─────────────────────────────────────────────────────────────────────────
The prediction interval estimates the predicted value of Y for a single individual with the stated value of X. The interval is only
accurate if all of the linear regression assumptions are valid.

The predicted values and prediction intervals of individual Y response values given X are given for each
observation.

X
This is the value of X at which the prediction is made.

Y
This is the actual value of Y.

Predicted Y (Yhat|X)
The predicted value of Y for the value of X indicated.

Standard Error of Yhat

This is the estimated standard deviation of the predicted value suitable for creating a prediction limit for an
individual.

95% Prediction Interval Limits for Y|X (Lower and Upper)

These are the lower and upper limits of a 95% prediction interval estimate of an individual Y at this value of
X. Note that you set the prediction interval alpha on the Reports tab of the procedure input window.

300-55
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Working-Hotelling Simultaneous Confidence Bands

Working-Hotelling Simultaneous Confidence Bands

─────────────────────────────────────────────────────────────────────────
95% Simultaneous Confidence
Predicted Standard Band Limits for Y Mean|X
Weight Height Height Error of ─────────────────────
Row (X) (Y) (Yhat|X) Yhat Mean Lower Upper
─────────────────────────────────────────────────────────────────────────────────────────────────────────────────
1 159 64 65.84747 0.3457 64.80357 66.89137
2 155 63 65.07480 0.3343 64.06537 66.08422
3 157 67 65.46114 0.3397 64.43532 66.48695
4 125 60 59.27974 0.3323 58.27638 60.28310
5 103 52 55.03003 0.4162 53.77323 56.28683
6 122 58 58.70023 0.3403 57.67268 59.72778
7 101 56 54.64369 0.4261 53.35683 55.93056
8 82 52 50.97349 0.5325 49.36554 52.58144
9 228 79 79.17610 0.7309 76.96878 81.38342
10 199 76 73.57421 0.5434 71.93330 75.21513
. . . . . . .
. . . . . . .
. . . . . . .
─────────────────────────────────────────────────────────────────────────
This is a confidence band for the regression line for all possible values of X from -infinity to + infinity. The confidence coefficient is
the proportion of times that this procedure yields a band that includes the true regression line when a large number of samples
are taken using the same X values as in this sample.

The predicted values and confidence band of the mean response function are given for each observation.
Note that this is a confidence band for all possible values of X along the real number line. The confidence
coefficient is the proportion of time that this procedure yields a band that includes the true regression line
when a large number of samples are taken using the X values as in this sample.

X
This is the value of X at which the prediction is made.

Y
This is the actual value of Y.

Predicted Y (Yhat|X)
The predicted value of Y for the value of X indicated.

Standard Error of Yhat Mean

This is the estimated standard deviation of the predicted mean value.

95% Simultaneous Confidence Band Limits for Y Mean|X (Lower and Upper)
This is the lower and upper limits of the 95% simultaneous confidence band for the value of Y at this X. Note
that you set the confidence band alpha on the Reports tab of the procedure input window.

300-56
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Residuals

Residuals
─────────────────────────────────────────────────────────────────────────
Predicted Percent
Weight Height Height Standardized Absolute
Row (X) (Y) (Yhat|X) Residual Residual Error
──────────────────────────────────────────────────────────────────────────────────────────────────────────
1 159 64 65.84747 -1.8475 -1.3580 2.8867
2 155 63 65.07480 -2.0748 -1.5220 3.2933
3 157 67 65.46114 1.5389 1.1299 2.2968
4 125 60 59.27974 0.7203 0.5282 1.2004
5 103 52 55.03003 -3.0300 -2.2604 5.8270
6 122 58 58.70023 -0.7002 -0.5142 1.2073
7 101 56 54.64369 1.3563 1.0142 2.4220
8 82 52 50.97349 1.0265 0.7904 1.9741
9 228 79 79.17610 -0.1761 -0.1470 0.2229
10 199 76 73.57421 2.4258 1.8744 3.1918
. . . . . . .
. . . . . . .
. . . . . . .
─────────────────────────────────────────────────────────────────────────
The residual is the difference between the actual and the predicted Y values. The formula is Residual = Y - Yhat. The Percent
Absolute Error is the 100 |Residual| / Y.

This is a report showing the value of the residual at each observation.

X
This is the value of X at which the prediction is made.

Y
This is the actual value of Y.

Predicted Y (Yhat|X)
The predicted value of Y for the value of X indicated.

Residual
This is the difference between the actual and predicted values of Y.

Standardized Residual
The variance of the observed residuals is not constant. This makes comparisons among the residuals
difficult. One solution is to standardize the residuals by dividing by their standard deviations. This gives a set
of residuals with constant variance.
The formula for this residual is
𝑒𝑒𝑗𝑗
𝑟𝑟𝑗𝑗 =
𝑠𝑠�1 − ℎ𝑗𝑗𝑗𝑗

Percent Absolute Error

The percent is the absolute value of the Residual divided by the Actual value. Scrutinize observations with the
large percent errors.

300-57
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Residual Diagnostics

Residual Diagnostics
─────────────────────────────────────────────────────────────────────────
Weight Hat
Row (X) Residual RStudent Diagonal Cook's D MSEi
─────────────────────────────────────────────────────────────────────────────────────────────────────
1 159 -1.8475 -1.3931 0.0607 0.0595 1.8723
2 155 -2.0748 -1.5845 0.0567 0.0696 1.8176
3 157 1.5389 1.1392 0.0586 0.0397 1.9381
4 125 0.7203 0.5173 0.0560 0.0083 2.0537
5 103 -3.0300 * -2.5957 0.0879 0.2462 1.4939
6 122 -0.7002 -0.5034 0.0588 0.0083 2.0554
7 101 1.3563 1.0150 0.0922 0.0522 1.9669
8 82 1.0265 0.7818 0.1439 0.0525 2.0137
9 228 -0.1761 * -0.1429 0.2712 0.0040 2.0836
10 199 2.4258 * 2.0305 0.1499 0.3097 1.6789
. . . . . . .
. . . . . . .
. . . . . . .
─────────────────────────────────────────────────────────────────────────
Outliers are rows that are separated from the rest of the data. Influential rows are those whose omission results in a relatively
large change in the results. This report lets you see both.

An outlier may be defined as a row in which |RStudent| > 2. A moderately influential row is one with a CooksD > 0.5. A heavily
influential row is one with a CooksD > 1.

MSEi is the value of the Mean Square Error (the average of the sum of squared residuals) calculated with each row omitted.

This is a report gives residual diagnostics for each observation. These were discussed earlier in the technical
of this chapter and we refer you to that section for the technical details.

X
This is the value of X at which the prediction is made.

Residual
This is the difference between the actual and predicted values of Y.

RStudent
Sometimes called the externally studentized residual, RStudent is a standardized residual that has the
impact of a single observation removed from the mean square error. If the regression assumption of
normality is valid, a single value of the RStudent has a t distribution with N - 2 degrees of freedom.
An observation is starred as an outlier if the absolute value of RStudent is greater than 2.

Hat Diagonal
The hat diagonal captures an observation’s remoteness in the X-space. Some authors refer to the hat
diagonal as a measure of leverage in the X-space.
Hat diagonals greater than 4 / N are considered influential. However, an influential observation is not a bad
observation. An influential observation should be checked to determine if it is also an outlier.

300-58
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Cook’s D
Cook’s D attempts to measure the influence the observation on all N fitted values. The formula for Cook’s D
is
2
∑𝑁𝑁 �𝑗𝑗 − 𝑦𝑦�𝑗𝑗 (𝑖𝑖)�
𝑖𝑖=1 𝑤𝑤𝑗𝑗 �𝑦𝑦
𝐷𝐷𝑗𝑗 =
𝑝𝑝𝑠𝑠 2

The 𝑦𝑦�𝑗𝑗 (𝑖𝑖) are found by removing observation i before the calculations. A Cook’s D value greater than one
indicates an observation that has large influence. Some statisticians have suggested that a better cutoff
value is 4 / (N - 2).

MSEi
This is the value of the mean squared error calculated without observation j.

Leave One Row Out

Leave One Row Out

─────────────────────────────────────────────────────────────────────────
Row RStudent DFFITS Cook's D CovRatio DFBETAS(0) DFBETAS(1)
──────────────────────────────────────────────────────────────────────────────────────────────────────────────────
1 -1.3931 -0.3540 0.0595 0.9615 0.0494 -0.1483
2 -1.5845 -0.3885 0.0696 0.9023 0.0228 -0.1337
3 1.1392 0.2842 0.0397 1.0279 -0.0284 0.1087
4 0.5173 0.1260 0.0083 1.1511 0.0739 -0.0414
5 * -2.5957 -0.8059 0.2462 0.6304 -0.6820 0.5292
6 -0.5034 -0.1258 0.0083 1.1564 -0.0800 0.0486
7 1.0150 0.3234 0.0522 1.0978 0.2781 -0.2188
8 0.7818 0.3205 0.0525 1.2202 0.3024 -0.2589
9 -0.1429 -0.0872 0.0040 * 1.5346 0.0646 -0.0787
10 * 2.0305 0.8525 0.3097 0.8542 -0.5244 0.6959
. . . . . . .
. . . . . . .
. . . . . . .
─────────────────────────────────────────────────────────────────────────
Each column gives the impact on some aspect of the linear regression of omitting that row.

RStudent represents the size of the residual. DFFITS represents the change in the fitted value of a row. Cook's D summarizes the
change in the fitted values of all rows. CovRatio represents the amount of change in the determinant of the covariance matrix.
DFBETAS(0) and DFBETAS(1) give the amount of change in the intercept and slope.

Each column gives the impact on some aspect of the linear regression of omitting that row.

RStudent
Sometimes called the externally studentized residual, RStudent is a standardized residual that has the
impact of a single observation removed from the mean square error. If the regression assumption of
normality is valid, a single value of the RStudent has a t distribution with N - 2 degrees of freedom.
An observation is starred as an outlier if the absolute value of RStudent is greater than 2.

Dffits
Dffits is the standardized difference between the predicted value of Y with and without observation j. It
represents the number of estimated standard errors that the predicted value changes if that observation is
omitted. Dffits > 1 would flag observations as being influential in prediction.

300-59
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Cook’s D
Cook’s D attempts to measure the influence of the observation on all N fitted values. The formula for Cook’s
D is
2
∑𝑁𝑁 �𝑗𝑗 − 𝑦𝑦�𝑗𝑗 (𝑖𝑖)�
𝑖𝑖=1 𝑤𝑤𝑗𝑗 �𝑦𝑦
𝐷𝐷𝑗𝑗 =
𝑝𝑝𝑠𝑠 2

The 𝑦𝑦�𝑗𝑗 (𝑖𝑖) are found by removing observation i before the calculations. A Cook’s D value greater than one
indicates an observation that has large influence. Some statisticians have suggested that a better cutoff
value is 4 / (N - 2).

CovRatio
This diagnostic flags observations that have a major impact on the generalized variance of the regression
coefficients. A value exceeding 1.0 implies that the observation provides an improvement, i.e., a reduction in
the generalized variance of the coefficients. A value of CovRatio less than 1.0 flags an observation that
increases the estimated generalized variance. This is not a favorable condition.

DFBETAS(0) and DFBETAS(1)

DFBETAS(0) and DFBETAS(1) are the standardized change in the intercept and slope when an observation is
omitted from the analysis. Belsley, Kuh, and Welsch (1980) recommend using a cutoff of 2/√𝑁𝑁 when N is
greater than 100. When N is less than 100, others have suggested using a cutoff of 1.0 or 2.0 for the
absolute value of DFBETAS.

Outlier Detection Chart

Outlier Detection Chart

─────────────────────────────────────────────────────────────────────────
Weight Standardized
Row (X) Residual Residual RStudent
─────────────────────────────────────────────────────────────────────────────────────────────────────────────────
1 159 -1.8475 |||||||||...... -1.3580 ||||||||....... -1.3931 |||||||........
2 155 -2.0748 ||||||||||..... -1.5220 |||||||||...... -1.5845 |||||||||......
3 157 1.5389 |||||||........ 1.1299 |||||||........ 1.1392 ||||||.........
4 125 0.7203 |||............ 0.5282 |||............ 0.5173 ||.............
5 103 -3.0300 ||||||||||||||| -2.2604 ||||||||||||||| * -2.5957 |||||||||||||||
6 122 -0.7002 |||............ -0.5142 |||............ -0.5034 ||.............
7 101 1.3563 ||||||......... 1.0142 ||||||......... 1.0150 |||||..........
8 82 1.0265 ||||........... 0.7904 |||||.......... 0.7818 ||||...........
9 228 -0.1761 |.............. -0.1470 |.............. -0.1429 |..............
. . . . . . . .
. . . . . . . .
. . . . . . . .
─────────────────────────────────────────────────────────────────────────
Outliers are rows that are separated from the rest of the data. Since outliers can have dramatic effects on the results, corrective
action, such as elimination, must be carefully considered. Outlying rows should not automatically be removed unless a good
reason for their removal can be given.

An outlier may be defined as a row in which |RStudent| > 2. Rows with this characteristic have been starred.

Outliers are rows that are far removed from the rest of the data. Since outliers can have dramatic effects on
the results, corrective action, such as elimination, must be carefully considered. Outlying rows should not be
removed unless a good reason for their removal can be given.

300-60
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

An outlier may be defined as a row in which |RStudent| > 2. Rows with this characteristic have been starred.

X
This is the value of X.

Residual
This is the difference between the actual and predicted values of Y.

Standardized Residual
The variance of the observed residuals is not constant. This makes comparisons among the residuals
difficult. One solution is to standardize the residuals by dividing by their standard deviations. This gives a set
of residuals with constant variance.

RStudent
Sometimes called the externally studentized residual, RStudent is a standardized residual that has the
impact of a single observation removed from the mean square error. If the regression assumption of
normality is valid, a single value of the RStudent has a t distribution with N - 2 degrees of freedom.
An observation is starred as an outlier if the absolute value of RStudent is greater than 2.

Influence Detection Chart

Influence Detection Chart

─────────────────────────────────────────────────────────────────────────
Weight
Row (X) DFFITS Cook's D DFBETAS(1)
──────────────────────────────────────────────────────────────────────────────────────────────────────────────
1 159 -0.3540 ||||||......... 0.0595 ||............. -0.1483 ||.............
2 155 -0.3885 ||||||......... 0.0696 |||............ -0.1337 ||.............
3 157 0.2842 ||||........... 0.0397 |.............. 0.1087 ||.............
4 125 0.1260 |.............. 0.0083 |.............. -0.0414 |..............
5 103 -0.8059 ||||||||||||||. 0.2462 |||||||||||.... 0.5292 |||||||||||....
6 122 -0.1258 |.............. 0.0083 |.............. 0.0486 |..............
7 101 0.3234 |||||.......... 0.0522 ||............. -0.2188 ||||...........
8 82 0.3205 |||||.......... 0.0525 ||............. -0.2589 |||||..........
9 228 -0.0872 |.............. 0.0040 |.............. -0.0787 |..............
. . . . . . . .
. . . . . . . .
. . . . . . . .
─────────────────────────────────────────────────────────────────────────
Influential rows are those whose omission results in a relatively large change in the results. They are not necessarily harmful.
However, they will distort the results if they are also outliers. The impact of influential rows should be studied very carefully. Their
accuracy should be double-checked.
DFFITS is the standardized change in Yhat when the row is omitted. A row is influential when DFFITS > 1 for small datasets (N <
30) or when DFFITS > 2*SQR(1/N) for medium to large datasets.

Cook's D gives the influence of each row on the Yhats of all the rows. Cook suggests investigating all rows having a Cook's D >
0.5. Rows in which Cook's D > 1.0 are very influential.

DFBETAS(1) is the standardized change in the slope when this row is omitted. DFBETAS(1) > 1 for small datasets (N < 30) and
DFBETAS(1) > 2/SQR(N) for medium and large datasets are indicative of influential rows.

Influential rows are those whose omission results in a relatively large change in the results. They are not
necessarily harmful. However, they will distort the results if they are also outliers. The impact of influential
rows should be studied very carefully. The accuracy of the data values should be double-checked.

300-61
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

X
This is the value of X.

Dffits
Dffits is the standardized difference between the predicted value of Y with and without observation j. It
represents the number of estimated standard errors that the predicted value changes if that observation is
omitted. Dffits > 1 would flag observations as being influential in prediction.

Cook’s D
Cook’s D attempts to measure the influence of the observation on all N fitted values. The formula for Cook’s
D is
2
∑𝑁𝑁 �𝑗𝑗 − 𝑦𝑦�𝑗𝑗 (𝑖𝑖)�
𝑖𝑖=1 𝑤𝑤𝑗𝑗 �𝑦𝑦
𝐷𝐷𝑗𝑗 =
𝑝𝑝𝑠𝑠 2
The 𝑦𝑦�𝑗𝑗 (𝑖𝑖) are found by removing observation i before the calculations. A Cook’s D value greater than one
indicates an observation that has large influence. Some statisticians have suggested that a better cutoff
value is 4 / (N - 2).

DFBETAS(1)
DFBETAS(1) is the standardized change in the slope when an observation is omitted from the analysis.
Belsley, Kuh, and Welsch (1980) recommend using a cutoff of 2/√𝑁𝑁 when N is greater than 100. When N is
less than 100, others have suggested using a cutoff of 1.0 or 2.0 for the absolute value of DFBETAS.

Outlier and Influence Chart

Outlier and Influence Chart

─────────────────────────────────────────────────────────────────────────
Hat
Weight RStudent Cooks D Diagonal
Row (X) (Outlier) (Influence) (Leverage)
────────────────────────────────────────────────────────────────────────────────────────────────────────────────
1 159 -1.3931 |||||||........ 0.0595 ||............. 0.0607 |..............
2 155 -1.5845 |||||||||...... 0.0696 |||............ 0.0567 |..............
3 157 1.1392 ||||||......... 0.0397 |.............. 0.0586 |..............
4 125 0.5173 ||............. 0.0083 |.............. 0.0560 |..............
5 103 * -2.5957 ||||||||||||||| 0.2462 |||||||||||.... 0.0879 |..............
6 122 -0.5034 ||............. 0.0083 |.............. 0.0588 |..............
7 101 1.0150 |||||.......... 0.0522 ||............. 0.0922 |..............
8 82 0.7818 ||||........... 0.0525 ||............. 0.1439 |||............
. . . . . . . .
. . . . . . . .
. . . . . . . .
─────────────────────────────────────────────────────────────────────────
Outliers are rows that are separated from the rest of the data. Influential rows are those whose omission results in a relatively
large change in the results. This report lets you see both.

An outlier may be defined as a row in which |RStudent| > 2. A moderately influential row is one with a CooksD > 0.5. A heavily
influential row is one with a CooksD > 1.

This report provides diagnostics about whether a row is an outlier, influential, and has high leverage.
Outliers are rows that are removed from the rest of the data. Influential rows are those whose omission
results in a relatively large change in the results. This report lets you see both.

300-62
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

X
This is the value of X.

RStudent (Outlier)
RStudent is a standardized residual that has the impact of a single observation removed from the mean
square error. If the regression assumption of normality is valid, a single value of the RStudent has a t
distribution with N - 2 degrees of freedom.
An observation is starred as an outlier if the absolute value of RStudent is greater than 2.

Cook’s D (Influence)
Cook’s D attempts to measure the influence the observation on all N fitted values. The formula for Cook’s D
is
2
∑𝑁𝑁 �𝑗𝑗 − 𝑦𝑦�𝑗𝑗 (𝑖𝑖)�
𝑖𝑖=1 𝑤𝑤𝑗𝑗 �𝑦𝑦
𝐷𝐷𝑗𝑗 =
𝑝𝑝𝑠𝑠 2

The 𝑦𝑦�𝑗𝑗 (𝑖𝑖) are found by removing observation i before the calculations. A Cook’s D value greater than one
indicates an observation that has large influence. Some statisticians have suggested that a better cutoff
value is 4 / (N - 2).

Hat Diagonal (Leverage)

The hat diagonal captures an observation’s remoteness in the X-space. Some authors refer to the hat
diagonal as a measure of leverage in the X-space.
Hat diagonals greater than 4 / N are considered influential. However, an influential observation is not a bad
observation. An influential observation should be checked to determine if it is also an outlier.

Inverse Predicted Values and Confidence Intervals for X Means

Inverse Predicted Values and Confidence Intervals for X Means

─────────────────────────────────────────────────────────────────────────
95% Confidence Interval
Predicted Limits for X Mean|Y
Height Weight Weight ─────────────────
Row (Y) (X) (Xhat|Y) X-Xhat|Y Lower Upper
────────────────────────────────────────────────────────────────────────────────────────────────────────────
1 64 159 149.43600 9.5640320 145.98320 153.01930
2 63 155 144.25910 10.7408600 140.84410 147.73610
3 67 157 164.96640 -7.9664460 161.13100 169.13870
4 60 125 128.72870 -3.7286660 125.11810 132.19480
5 52 103 87.31406 15.6859400 81.48936 92.44440
6 58 122 118.37500 3.6249850 114.39470 122.07350
7 56 101 108.02140 -7.0213630 103.52270 112.10070
8 52 82 87.31406 -5.3140610 81.48936 92.44440
9 79 228 227.08830 0.9116458 219.73880 235.59970
10 76 199 211.55790 -12.5578800 205.22830 218.84300
. . . . . . .
. . . . . . .
. . . . . . .
─────────────────────────────────────────────────────────────────────────
This confidence interval estimates the mean of X in a large sample of individuals with the stated value of Y. This method of
inverse prediction is also called "calibration."

300-63
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

This report provides inverse prediction or calibration results. Although a regression of Y on X has been fit,
our interest here is predicting the value of X from the value of Y. This report provides both a point estimate
and an interval estimate of the predicted mean of X given Y.

Y
This is the actual value of Y.

X
This is the value of X at which the prediction is made.

Predicted X (Xhat|Y)
The predicted value of X for the value of Y indicated.

X-Xhat|Y
This is the difference between the actual value of X and the predicted value of X at this value of Y.

95% Confidence Interval Limits for X Mean|Y (Lower and Upper)

These are the lower and upper limits of a 95% confidence interval estimate of the mean of X at this value of
Y. Note that you set the confidence interval alpha on the Reports tab of the procedure input window.

Inverse Predicted Values and Prediction Intervals of X Individuals

Inverse Predicted Values and Prediction Intervals for X Individuals

─────────────────────────────────────────────────────────────────────────
95% Prediction Interval
Predicted Limits for X|Y
Height Weight Weight ─────────────────
Row (Y) (X) (Xhat|Y) X-Xhat|Y Lower Upper
────────────────────────────────────────────────────────────────────────────────────────────────────────────
1 64 159 149.43600 9.5640320 133.78580 165.21670
2 63 155 144.25910 10.7408600 128.59060 159.98960
3 67 157 164.96640 -7.9664460 149.30360 180.96620
4 60 125 128.72870 -3.7286660 112.93650 144.37650
5 52 103 87.31406 15.6859400 70.70028 103.23350
6 58 122 118.37500 3.6249850 102.44360 134.02460
7 56 101 108.02140 -7.0213630 91.90588 123.71750
8 52 82 87.31406 -5.3140610 70.70028 103.23350
9 79 228 227.08830 0.9116458 210.42140 244.91720
10 76 199 211.55790 -12.5578800 195.27440 228.79690
. . . . . . .
. . . . . . .
. . . . . . .
─────────────────────────────────────────────────────────────────────────
This prediction interval estimates the predicted value of X for a single individual with the stated value of Y. This method of inverse
prediction is also called "calibration."

This report provides inverse prediction or calibration results. Although a regression of Y on X has been fit,
our interest here is predicting the value of X from the value of Y. This report provides both a point estimate
and an interval estimate of the predicted value of X given Y.

300-64
© NCSS, LLC. All Rights Reserved.
NCSS Statistical Software NCSS.com

Linear Regression and Correlation

Y
This is the actual value of Y.

X
This is the value of X at which the prediction is made.

Predicted X (Xhat|Y)
The predicted value of X for the value of Y indicated.

X-Xhat|Y
This is the difference between the actual value of X and the predicted value of X at this value of Y.

95% Prediction Interval Limits for X|Y (Lower and Upper)

These are the lower and upper limits of a 95% prediction interval estimate of X at this value of Y. Note that
you set the prediction interval alpha on the Reports tab of the procedure input window.

300-65
© NCSS, LLC. All Rights Reserved.