Transcript Multiple Linear Regression

y '  a  b1 x1  b2 x 2  bi x i
Nothing
explains
everythin
g
Multiple Linear Regression
Laurens Holmes, Jr.
Nemours/A.I.duPont Hospital for Children
What is MLR?
• Multiple Regression is a statistical method
for estimating the relationship between a
dependent variable and two or more
independent (or predictor) variables.
Multiple Linear Regression
• Simply, MLR is a method for studying the
relationship between a dependent variable
and two or more independent variables.
• Purposes:
– Prediction
– Explanation
– Theory building
Operation?
• Uses the ordinary least squares solution (as
does simple linear or bi-variable regression)
• Describes a line for which the (sum of
squared) differences between the predicted and
the actual values of the dependent variable are
at a minimum.
• Represents the “function” that minimizes the
sum of the squared errors.
• Ypred = a + b1X1 + B2X2 … + BnXn
Operation?
• MLR produces a model that identifies the
best weighted combination of independent
variables to predict the dependent (or
criterion) variable.
• Ypred = a + b1X1 + B2X2 … + BnXn
• MLR estimates the relative importance of several
hypothesized predictors.
• MLR assess the contribution of the combined
variables to change the dependent variable.
Design Requirements
• One dependent variable (criterion)
• Two or more independent variables
(predictor or explanatory variables).
• Sample size: >= 50 (at least 10 times as
many cases as independent variables)
Variations
Predictable variation by
the combination of
independent variables
Total Variation in Y
Unpredictable
Variation
MLR Model: Basic Assumptions
• Independence: The data of any particular subject are
independent of the data of all other subjects
• Normality: in the population, the data on the dependent
variable are normally distributed for each of the possible
combinations of the level of the X variables; each of the
variables is normally distributed
• Homoscedasticity: In the population, the variances of the
dependent variable for each of the possible combinations
of the levels of the X variables are equal.
• Linearity: In the population, the relation between the
dependent variable and the independent variable is linear
when all the other independent variables are held
constant.
Simple vs. Multiple Regression
• One dependent variable Y
predicted from one
independent variable X
• One regression coefficient
• r2: proportion of variation
in dependent variable Y
predictable from X
• One dependent variable Y
predicted from a set of
independent variables (X1,
X2 ….Xk)
• One regression coefficient
for each independent
variable
• R2: proportion of variation
in dependent variable Y
predictable by set of
independent variables
(X’s)
MLR Equation
• Ypred = a + b1X1 + B2X2 … + BnXn (pred=predicted,
1 and 2 are underscore)
•
•
•
•
Ypred = dependent variable or the variable to be
predicted.
X = the independent or predictor variables
a = “raw score equations” include a constant or Y
• Intercept ob Y axis, representing the value of Y when X = 0.
• b = b weights; or partial regression coefficients.
• The bs show the relative contribution of their
independent variable on the dependent variable
when controlling for the effects of the other
predictors
Variables in the model?
• One approach is to perform literature review and
examine theories to identify potential predictors , thus
building a “theoretical” variate, which may reflect the
biologic or clinical relevance of the variable.
• This is sometimes referred to as the “standard”
(simultaneous) regression method.
• A second approach is to examine statistics that show
the effects of each variable both within and out of
• the equation.
• The “statistical variate” is built based on those variables
• showing the most effect (significant at 0.25).
– These are sometimes called “Forward and Backward Stepwise
Regression
MLR Output
• The following notions are essential for the
understanding of MLR output: R2, adjusted R2,
constant, b coefficient, beta, F-test, t-test
• For MLR “R2” (the coefficient of multiple
determination) is used rather than “r” (Pearson’s
correlation coefficient) to assess the strength of
this more complex
relationship (as compared to a bivariate
correlation)
Adjusted R square and b coefficient
• The adjusted R2 adjusts for the inflation in R2
caused by the number of variables in the
equation. As the sample size increases above
20 cases per variable, adjustment is less
needed (and vice versa).
• b coefficient measures the amount of increase or
decrease in the dependent variable for a oneunit difference in the independent variable,
controlling for the other independent variable(s)
in the equation.
B coefficient
• Ideally, the independent variables are
uncorrelated.
• Consequently, controlling for one of them
will not affect the relationship between the
other independent variable and the
dependent variable
Intercorrelation or collinearlity
• If the two independent variables are
uncorrelated, we can uniquely partition the
amount of variance in Y due to X1 and X2 and
bias is avoided.
• Small intercorrelations between the independent
variables will not greatly biased the b
coefficients.
• However, large intercorrelations will biased the b
coefficients and for this reason other
mathematical procedures are needed
MRL Model Building
• Each predictor is taken in turn. That is, all other
predictors are first placed in the equation and
then the predictor of interest is entered.
• This allows us to determine the unique
(additional) contribution of the predictor variable.
• By repeating the procedure for each predictor
we can determine the unique contribution of
each independent variable.
Different Ways of Building
Regression Models
• Simultaneous: all independent variables
entered together
• Stepwise: independent variables entered
according to some order
– By size or correlation with dependent variable
– In order of significance
• Hierarchical: independent variables
entered in stages
Various Significance Tests
• Testing R2
– Test R2 through an F test
– Test of competing models (difference between R2)
through an F test of difference of R2s
• Testing b
– Test of each partial regression coefficient (b) by ttests
– Comparison of partial regression coefficients with
each other - t-test of difference between
standardized partial regression coefficients ()
F and t tests
• The F-test is used as a general indicator of
the probability that any of the predictor
variables contribute to the variance in the
dependent variable within the population.
• The null hypothesis is that the predictors’
weights are all effectively equal to zero.
• Implying that, none of the predictors
contribute to the variance in the dependent
variable in the population
F and t tests
• t-tests are used to test the significance of
each predictor in the equation.
• The null hypothesis is that a predictor’s
weight is effectively equal to zero when
the effects of the other predictors are
taken into account.
• That is, it does not contribute to the
variance in the dependent variable within
the population.
R Square
• When comparing the R2 of an original set of variables to the R2
after additional variables have been included, the researcher is able
to identify the unique variation explained by the additional set of
variables.
• Any co-variation between the original set of variables and the new
variables will be attributed to the original variables.
• R2 (multiple correlation squared) – variation in Y accounted for by
the set of predictors
• Adjusted R2 – sample variation around R2 can only lead to inflation
of the value.
• The adjustment takes into account the size of the sample and number of
predictors to adjust the value to be a better estimate of the population value.
• R2 is similar to η2 value but will be a little smaller because R2 only
looks at linear relationship while η2 will account for non-linear
relationships.
Vignette
• Suppose we wish to examine the factors that predict the
length of hospitalization following spinal surgery in
children with CP(dependent continuous variable).
• The available variables in the dataset are hematocrit,
estimated blood loss, cell saver, operating time, age at
surgery, and parked red blood cells.
• If the dependent and independent variables are
measured on continuous scale, what will be an
appropriate test statistic?
– Select appropriate variables (theory based and statistical
approach), and determine the effect of estimated blood loss
while controlling hematocrit and parked red blood cell, age at
surgery, cell saver, operating time (duration of surgery).
SPSS: 1) analyze, 2)
regression, 3) linear
SPSS Screen
SPSS Output
Interpret
the
coefficients
SPSS Output
Interpret
the r
square
What does
the ANOVA
result mean?
RM removes variability in
baseline prognostic factor
– ideal model !!!
Repeated Measure Analysis of Variance
(RM ANOVA)
Univariable (Univariate)
Repeated Measures ANOVA
• Between Subjects Design
– ANOVA in which each participant
participated in one of the three treatment
groups for example.
• Within Subjects or Repeated Measures
Design
– Participants participate in one treatment
and the outcome of the treatment is
measured in different time points for
example 3, (before treatment, immediately
after, and 6 months after treatment)
RM ANOVA Vs. Paired T test
• Repeated measures ANOVA, also known as
within-subjects ANOVA, are an extension of
Paired T-Tests.
• Like T-Tests, repeated measures ANOVA gives
us the statistic tools to determine whether or not
changed has occurred over time.
• T-Tests compare average scores at two different
time periods for a single group of subjects.
• Repeated measures ANOVA compared the
average score at multiple time periods for a
single group of subjects.
RM ANOVA: Understanding the
terms & analysis interpretation
• The first step in solving repeated measures ANOVA is to
combine the data from the multiple time periods into a
single time factor for analysis.
• The different time periods are analogous to the
categories of the independent variable is a one-way
analysis of variance.
• The time factor is then tested to see if the mean for the
dependent variable is different for some categories of the
time factor.
• If the time factor is statistically significant in the ANOVA
test, then Bonferroni pair wise comparisons are
computed to identify specific differences between time
periods.
RM ANOVA: Understanding the
terms & analysis interpretation
• The dependent variable is measured at
three time periods, there are three paired
comparisons:
• time 1 versus time 2 (preoperative or
before treatment measure)
• time 2 versus time 3 (immediate after
surgery/treatment measure)
• time 1 versus time 3 (Follow-up post
operative measure)
Statistical Assumptions of RM ANOVA
•
•
•
•
Independence
Normality
Homogeneity of within-treatment variances
Sphericity
RM is ideal in testing the
hypothesis on treatment
effectiveness when ethical
constraints restricts the
use of control subjects
Homogeneity of Variance
• In one-way ANOVA, we expect the
variances to be equal
– We also expect that the samples are not
related to one another (so no covariance or
correlation)
Sphericity and Compound Symmetry
• Extension of homogeneity of variance
assumption
• Compound Symmetry is stricter than
Sphericity (but maybe easier to explain)
– All variances are equal to each other
– All covariance are equal to each other
Sphericity and Compound Symmetry
• If we meet assumption of Compound
Symmetry than we meet assumption of
Sphericity
• Sphericity is less strict and is the only thing
we need to meet for RM ANOVA
• Sphericity is that the variance of the
differences are equal
– Variance of difference scores between time 1
and 2 is equal to the variance of difference
scores between time 2 and 3.
Spericity Assumption Violations
• A more conservative method of evaluating
the significance of the obtained F is
needed
• Greenhouse-Geisser (1958) correction
– Gives appropriate critical value for worst situation in
which assumptions are maximally violated
• Huynh-Feldt correction
– The Huynh-Feldt epsilon is an attempt to correct the
Greenhouse-Geisser epsilon, which tends to be overly
conservative, especially for small sample sizes
Sample Table for RM ANOVA
RM ANOVA
• All participants participate in all treatment
conditions, ex. surgery for spinal deformity
correction.
• Participant emerges as an independent
source of variance.
– In RM ANOVA there is no such variability.
• The other sources of variance include the
repeated measures treatment and the
Participant x treatment interaction
RM ANOVA Equation
y i   0 i   1i (Tim e)   it
0i   0
 1i   1
Vignette
• Suppose a spinal fusion was performed to
correct spinal deformities in Adolescent
Idiopathic Scoliosis (AIS). If the main cobb angle
was measured preoperatively, immediately after
surgery (first erect), and during two years of
follow-up, was the surgical procedure effective in
correcting the curve deformity and maintaining
correction after two years of follow-up?
• Hint: correction loss > 10 degrees in indicative of a clinically
significant loss of correction.
Sample variables on
preoperative,
immediate operative
and 2 year follow-up
Normality assumption
of the variables on the
three measuring points
of the cobb angle.
From the variables box
select accordingly 1, 2,
and 3rd measurement
points during the study
period.
SPSS Output
Click the
option box
and select
descriptive,
and Bon
multiple
comparison.
SPSS OUTPUT
SPSS Output
SPSS Output
48