1. Regression @UWE_JT9 @dave_lush
Scientific Practice
Regression
@UWE_JT9
@dave_lush

2. Where We Are/Where We Are Going
We have looked at how correlation shows how two things might be associated
eg between arm length and leg length
no causality implied
ie leg length is not responsible for arm length!
The correlation coefficient, r, assumes a linear ‘fit’ between the data and reflects how far away from that ‘line of best’ fit the data lie
Regression takes this one step further
describes fit mathematically
generally implies a causal link; eg increased BP causes increased mortality

3. Independent/Dependent Variables
As a causal link is implied, then…
the independent variable is the thing doing the influencing
the dependent variable is the one being influenced
By convention, if we were to plot this graphically…
the x-axis represents the independent variable
the y-axis represents the dependent variable
eg…
blood pressure on the x-axis
cardiovascular mortality on the y-axis

4. The Equation of the Straight Line
Any linear relationship can be described as…
y = mx + c
y can be calculated (predicted) for any x using…
c, the intercept
where the line crosses the y-axis when x=0
m, the slope of the line
change in y per x
non-zero if relationship

5. The Line of Best Fit
For a given set of data, linear regression derives the straight line equation that best describes the data
by minimising the overall distance of data points to that straight line

6. The Power of Linear Regression
Eg, data from a practical class
looked at relationship between latency of the Achilles Heel stretch reflex (ms) and height (cm)
the taller the person, the longer the nerve pathway mediating the reflex

7. The Power of Linear Regression
The class results (n=85)

8. The Power of Linear Regression
The line of best fit : y = x
the non-zero slope suggests a relationship

9. Testing the Line of Best Fit
But even if we used random data, the line of best fit would have a non-zero slope!
so how do we know if the slope is significantly different to zero?
The reported slope is the best estimate based on data that doesn’t perfectly fit a straight line
when we use a collection of data to estimate a single value, then that estimate comes with a standard error
eg data  mean +/- SE of the mean
in our case data  slope +/- SE of the slope
we can predict the 95% CI of the slope!
does the 95% CI encompass zero?

10. The Power of Linear Regression
In reality, Minitab will do the significance calculation for us
a t-test on the slope and its SE
Null Hypo is that slope is zero
The regression equation is
latency = height
Predictor Coef SE Coef T P
Constant
height
S = R-Sq = 11.1% R-Sq(adj) = 10.1%
p < 0.05, so the slope is sig diff to zero
reflex latency increases with height (0.134 ms/cm)

11. The Power of Linear Regression
The intercept also reported as a non-zero value
is it significantly different to zero? (the Null Hypo)
a t-test on the intercept and its SE
The regression equation is
latency = height
Predictor Coef SE Coef T P
Constant
height
S = R-Sq = 11.1% R-Sq(adj) = 10.1%
p > 0.05, so the intercept not sig diff to zero
predictive equation is latency=0.134 height

12. The Power of Linear Regression
The analysis also yields r-squared
this is the square of the correlation coefficient
proportion of variation in y-axis variable that can be explained by variation in x-axis variable
The regression equation is
latency = height
Predictor Coef SE Coef T P
Constant
height
S = R-Sq = 11.1% R-Sq(adj) = 10.1%
at 10%, this is very low
but it is still highly significant!

13. R-squared and Significance
Just because something has a low r-squared does not mean it is not significant
means it has a low predictive power
Eg the more clothes worn, the heavier is a person’s weight
clothes significantly influence weight
But it can only account for a small amount of variation in weight that we see
ie r-squared is small

14. Summary
Linear regression extends correlation by reporting the mathematical ‘line of best fit’
y = mx + c
The slope needs to be tested statistically to ‘prove’ the relationship is ‘real’
eg the Null Hypo is that m = 0
The intercept should also be tested to see if it is non-zero
The equation of the ‘line of best fit’ is predictive
ie given a value of x, you can predict y
the usefulness of this depends on r-squared
proportion of variation in y ‘explained’ by x (0-100%)