Math2010 - Statistical Methods I
Chapter 2 - Simple Linear Regression
(Notes 5)
Stefanie Biedermann
2.16 Bodyfat Example
Knowledge of the fat content of the human body is
physiologically and medically important. The fat content may
influence susceptibility to disease, the outcome of disease, the
effectiveness of drugs (especially anaesthetics) and the ability to
withstand adverse conditions including exposure to cold and
starvation.
In practice, fat content is difficult to measure directly one way is
by measuring body density which requires subjects to be
weighed underwater! For this reason, it is useful to try to relate
simpler measures such as skinfold thicknesses (which are
readily measured using calipers) to body fat content and then
use these to estimate the body fat content.
Dr R. Telford collected skinfold (the sum of four skinfold
measurements) and percent body fat measurements on 102 elite
male athletes training at the Australian Institute of Sport.
The scatterplot of the data shows that the percent body fat
increases linearly with the skinfold measurement
From the plot we might also be concerned that the variation in
percent body fat seems to increases with the skinfold
measurement.
To take a closer look we can fit a simple linear regression model
and examine the diagnostic plots as well
From the scatter plot there do not seem to be any unusual data
points and there are no other obvious patterns to note.
20
18
16
14
Fat
12
10
8
6
40
60
80
100
Skinfold
Figure: Scatterplot of bodyfat against skinfold
We can see in the residual plot that there is a fan or funnel shape
evident, giving more evidence to our concern of heteroscedacity
The normal probability plot of the residuals also seems to
indicates that the normality assumption is not appropriate in the
tails of the distribution
2
1
0
Sample Quantiles
-2
-1
0
-2
-1
Residuals
Normal Q-Q Plot
10
12
Fitted values
14
16
18
-2
-1
Theoretical Quantiles
Figure: Anscombe and normal probability plot from a linear regression of
bodyfat on skinfold
It is often the case that we can simplify the relationships
exhibited by data by transforming one or more of the variables.
In the example, we want to preserve the simple linear
relationship for the conditional mean but stabilise the variability
(i.e. make it constant) so that it is easier to describe. We
therefore try transforming both variables.
There is no theoretical argument to suggest a transformation
here but empirical experience suggests that we try the log
transformation.
The scatterplot on the log scale shows that the log
transformation of both variables preserves the linear conditional
location relationship and stabilises the conditional variability.
After examining the diagnostics, we conclude that the linear
regression model provides a good approximation to the
underlying population.
Anscombe plot
1.8
-0.2
2.0
-0.1
0.0
Residuals
2.4
2.2
log Fat
2.6
2.8
0.1
3.0
Scatterplot
3.4
3.6
3.8
4.0
4.2
4.4
4.6
1.8
2.0
log Skinfold
2.2
2.4
2.6
2.8
Fitted values
0.0
-0.2
-0.1
Sample Quantiles
0.1
Normal Q-Q Plot
-2
-1
Theoretical Quantiles
Figure: Scatterplot (top left) Anscombe plot (top right) and normal probability
plot (bottom) on the transformed scale
Let Yi denote the log percent body fat measurement and Xi the
log skinfold measurement on athlete i, where i = 1, . . . , 102.
Then we have
E(Yi |xi ) = 0 + 1 xi and Var (Yi |xi ) = 2
Fitting the model to the data results in estimates of the
parameters so that we can write:
= 1.25 + 0.88x
Y
(0.097)(0.025)
where the standard errors of the intercepts and slope are the
numbers in parentheses.
We also obtain a variance estimate S 2 = 0.0067 on 100 degrees
of freedom.
Parameter Estimates
Term
Intercept
logskin
Analysis of Variance
Source
Estimate
Std Error
t Ratio
Prob>|t|
1.250
0.882
0.097
0.024
12.940
35.590
0.000
0.000
DF
SS
MS
Prob>F
logskin
Residuals
1 8.5240 8.5240 1266.3
100 0.6731 0.0067
0.0000
Total
101
9.1971
A 95% confidence interval for the population slope parameter 1 is
(0.83, 0.93).
A common test of interest is the test of the hypothesis that
1 = 0. i.e. that skinfold is not useful for predicting bodyfat.
The 95% confidence interval for 1 does not contain zero so we
can conclude that there is evidence against the hypothesis and
that skinfold is a useful variable in the model for predicting body
fat.
A formal hypothesis test can be carried out by computing the
t-ratio 35.59 which has a p-value of zero (to 4 decimal places).
We conclude that 1 is significantly different from zero.
A test of the hypothesis 1 = 1 is also of interest. This is
because:
A 95% confidence interval for the mean log body fat percentage
of all individuals with a skinfold of 70 is (2.474, 2.522).
A 95% prediction interval for the log body fat percentage of a
single male athlete with a skinfold of 70 is (2.334, 2.663).
Although in our body fat example, we have worked on the simpler
log scale, it may be useful to make predictions on the raw scale.
We can also plot confidence bands across the range of the
skinfold measurements taken for the mean log body fat
percentage at any given skinfold measurement
We can also plot prediction bands across the range of the
skinfold measurements taken for the log body fat percentage of a
single male athlete at any given skinfold measurement
The bands are widests at the extremes and narrowest at the
centre
Also note that the prediction band is always wider than the
confidence band
3.0
Scatter plot with the fitted line and 95% confidence/prediction bands
2.4
1.8
2.0
2.2
log Fat
2.6
2.8
Fitted line
95% CI for mean
95% PI for observation
3.4
3.6
3.8
4.0
4.2
4.4
4.6
log Skinfold
Figure: Scatterplot of log bodyfat against log skinfold with fitted regression
line and 95% confidence bands and prediction bands
The validity of the inferences depends on the representativeness
of the sample and the validity of the model.
It is vital when collecting data to ensure that it is representative of
the population of interest and before making inferences to ensure
that the model is appropriate.
Strictly speaking, model assessment is a form of informal
inference and has an impact on other inferences but it is not
sensible to simply hope that a model is valid when empirical
verification is available.