Lecture 7

What we know now…

^ 0
^ β −β
T tests t = 1 1
^
s.e.( β 1 )

and confidence intervals

⎡^ . 05 / 2
^ ^
. 05 / 2
^ ⎤
Pr ⎢ β 1 − t N − k * s.e.( β 1) ≤ β1 ≤ β 1 + t N − k * s.e.( β 1)⎥ = 0.95
⎢⎣ ⎥⎦

are useful ways to check whether the OLS estimates are compatible
with our (economic) priors about how the world works
But they are not the only tests that we can use to judge the
performance of the model and our economic priors

In what follows look at a set if different tests (F tests) that can be used to
test different sorts of hypotheses and apply them to the issues of
forecasting and parameter stability
F test of Goodness of Fit of the Model

We know that the R2 gives an idea of the strength of association

between dependent and explanatory variable
 F test of Goodness of Fit of the Model

We know that the R2 gives an idea of the strength of association

between dependent and explanatory variable

Because of statistical variation unlikely that R2 would ever fall to zero in

absence of any (true) correlation between y and X variables
 F test of Goodness of Fit of the Model

We know that the R2 gives an idea of the strength of association

between dependent and explanatory variable

Because of statistical variation unlikely that R2 would ever fall to zero in

absence of any (true) correlation between y and X variables

So how do we know that the value for the R2 reflects a true relationship
rather than the result of chance?
 F test of Goodness of Fit of the Model

We know that the R2 gives an idea of the strength of association

between dependent and explanatory variable

Because of statistical variation unlikely that R2 would ever fall to zero in

absence of any (true) correlation between y and X variables

So how do we know that the value for the R2 reflects a true relationship
rather than the result of chance?

We know that
^
Var (Y ) = Var(Yˆ ) + Var(u )
 F test of Goodness of Fit of the Model

We know that the R2 gives an idea of the strength of association

between dependent and explanatory variable

Because of statistical variation unlikely that R2 would ever fall to zero in

absence of any (true) correlation between y and X variables

So how do we know that the value for the R2 reflects a true relationship
rather than the result of chance?

We know that
^
Var (Y ) = Var(Yˆ ) + Var(u )
and that this can be also written as
^2
2 2
∑ (Y − Y ) = ∑ (Yˆ − Y ) + ∑ u

TSS = ESS + RSS

Can show (don’t need to learn proof) that can use the value

ESS / k − 1
F= ~ F[k-1, N-k]
RSS / N − k

to test whether the R2 is statistically significantly different from zero –

allowing for the random variation that use of a sample entails
Can show (don’t need to learn proof) that can use the value

ESS / k − 1
F= ~ F[k-1, N-k]
RSS / N − k

to test whether the R2 is statistically significantly different from zero

allowing for the random variation that use of a sample entails

Intuitively if the model is a good fit then the ESS will be high relative to
the RSS.
Can show (don’t need to learn proof) that can use the value

ESS / k − 1
F= ~ F[k-1, N-k]
RSS / N − k

to test whether the R2 is statistically significantly different from zero

allowing for the random variation that use of a sample entails

Intuitively if the model is a good fit then the ESS will be high relative to
the RSS.

Just how high depends (like t values) on the sample size N and the
number of right hand side coefficients k (hence the corrections above)
Can show (don’t need to learn proof) that can use the value

ESS / k − 1
F= ~ F[k-1, N-k]
RSS / N − k

to test whether the R2 is statistically significantly different from zero

allowing for the random variation that use of a sample entails

Intuitively if the model is a good fit then the ESS will be high relative to
the RSS.

Just how high depends (like t values) on the sample size N and the
number of right hand side coefficients k (hence the corrections above)
The F statistic

(named after Ronald Fisher 1890-1962 – the same person who

developed the t test and the same person who wrote a book
attributing the decline of the British empire to the failure of its upper
classes to breed)

is the ratio of 2 chi-squared distributions divided by their respective

degrees of freedom
The F statistic

has its own set of degrees of freedom

F[k-1, N-k]
The F statistic

has its own set of degrees of freedom

F[k-1, N-k]

k-1 said to be the numerator

N-k said to be the denominator (for obvious reasons)
(In stata
rndf 1000 5 55
Now, unlike the t distribution, the F distribution because it is effectively
the t distribution “squared”) is bounded from below at zero.
Now, unlike the t distribution, the F distribution (because it is effectively
the t distribution “squared”) is bounded from below at zero.

This means that any hypothesis testing is based on a “1-tailed test” since
the rejection region now only lies at the far right-end tail of the
distribution
Now, unlike the t distribution, the F distribution because it is effectively
the t distribution “squared”) is bounded from below at zero.

This means that any hypothesis testing is based on a “1-tailed test” since
the rejection region now only lies at the far right-end tail of the
distribution

The basic rule is that

^
if F > Fα( k_−critical
1, N − k )
Now, unlike the t distribution, the F because it is effectively the t
distribution “squared”) is bounded from below at zero.

This means that any hypothesis testing is based on a “1-tailed test” since
the rejection region now only lies at the far right-end tail of the
distribution

The basic rule is that

^
if F > Fα( k_−critical
1, N − k )

then can be 1-α% confident that result is unlikely to have arisen by

chance (the ESS is high wrt the RSS)
Since the F-statistic can also be manipulated to give

ESS / k − 1 ( ESS / TSS ) / k − 1

F= =
RSS / N − k ( RSS / TSS ) / N − k
Since the F-statistic can also be manipulated to give

ESS / k − 1 ( ESS / TSS ) / k − 1

F= =
RSS / N − k ( RSS / TSS ) / N − k

R2 / k − 1
then F =
(1 − R 2 ) / N − k

which is an equivalent way of doing the test

Since the F-statistic can also be manipulated to give

ESS / k − 1 ( ESS / TSS ) / k − 1

F= =
RSS / N − k ( RSS / TSS ) / N − k

R2 / k − 1
then F =
(1 − R 2 ) / N − k

which is an equivalent way of doing the test

and still the rule is reject the null hypothesis if

^
F > Fα( k_−critical
1, N − k )
The F test in this example is 51.48
The p value is less than 0.05 which is a quick way of realising that this F
value is statistically significant from zero
Which means the chances of this R2 arising by chance are small (<5%)
Which means the model has some explanatory power
Forecasting Using OLS

One of the many uses of regression analysis is that can use the
coefficient estimates to make predictions about the behaviour of
economic agents (individuals, time periods) that may not be included
in the original data
Bank of England forecast of level of inflation (CPI) & GDP growth in
November 2007

November 2007 CPI Fan Chart November 2007 GDP Fan Chart

Source Bank of England Inflation report

Actual Outturn
Forecasting Using OLS
One of the many uses of regression analysis is that can use the
coefficient estimates to make predictions about the behaviour of
economic agents (individuals, time periods) that may not be included
in the original data

^ ^
Already seen that given a set of estimates β 0 and β 1
and a set of observations on the variable X can use this to predict a
value for y
Forecasting Using OLS
One of the many uses of regression analysis is that can use the
coefficient estimates to make predictions about the behaviour of
economic agents (individuals, time periods) that may not be included
in the original data

^ ^
Already seen that given a set of estimates β 0 and β 1
and a set of observations on the variable X can use this to predict a
value for y

^ ^ ^
y = β 0 + β1 X
Forecasting Using OLS
One of the many uses of regression analysis is that can use the
coefficient estimates to make predictions about the behaviour of
economic agents (individuals, time periods) that may not be included
in the original data
^ ^
Already seen that given a set of estimates β 0 and β 1
and a set of observations on the variable X can use this to predict a
value for y

^ ^ ^
y = β 0 + β1 X

and can do this for any X value even if the X value did not belong to
the original data set
Forecasting Using OLS
One of the many uses of regression analysis is that can use the
coefficient estimates to make predictions about the behaviour of
economic agents (individuals, time periods) that may not be included
in the original data
^ ^
Already seen that given a set of estimates β 0 and β 1
and a set of observations on the variable X can use this to predict a
value for y

^ ^ ^
y = β 0 + β1 X

and can do this for any X value even if the X value did not belong to
the original data set

So it is good practice when estimating equations to leave out some

observations (eg end few time periods or a random sub-sample of a
cross-section)
If the model is a good one it should predict well out-of-sample.
How do we know how good the prediction is ?
How do we know how good the prediction is ?

- use the idea of a forecast error

^
the closer is the estimate y to its actual value, the smaller will be the
forecast error
How do we know how good the prediction is ?

- use the idea of a forecast error

^
the closer is the estimate y to its actual value, the smaller will be the
forecast error which is just the difference between the prediction (the
forecast) and the actual (out-of-sample) value
How do we know how good the prediction is ?

- use the idea of a forecast error

^
the closer is the estimate y to its actual value, the smaller will be the
forecast error which is just the difference between the prediction (the
forecast) and the actual (out-of-sample) value
^ ^
u 0 = yo − yo (o is “out of sample”)
How do we know how good the prediction is ?

- use the idea of a forecast error

^
the closer is the estimate y to its actual value, the smaller will be the
forecast error which is just the difference between the prediction (the
forecast) and the actual (out-of-sample) value
^ ^
u 0 = yo − yo (o is “out of sample”)

and if we use OLS to generate the estimates to make the forecasts then
can show that the forecast error will on average be zero and have the
smallest variance of any other (linear unbiased) technique.
Proof: The expected value of the forecast error is
^ ^
E (u 0 ) = E ( yo − yo )
Proof: The expected value of the forecast error is
^ ^
E (u 0 ) = E ( yo − yo )

^ ^ ^
sub. in for y 0 = β 0 + β 1 X o and yO = β 0 + β 1 X o + uo
Proof: The expected value of the forecast error is
^ ^
E (u 0 ) = E ( yo − yo )

^ ^ ^
sub. in for y 0 = β 0 + β 1 X o and yO = β 0 + β 1 X o + uo
^ ^ ^
E (u 0 ) = E ( β 0 + β 1 X o − ( β 0 + β 1 X o + uo ))
Proof: The expected value of the forecast error is
^ ^
E (u 0 ) = E ( yo − yo )

^ ^ ^
sub. in for y 0 = β 0 + β 1 X o and yO = β 0 + β 1 X o + uo
^ ^ ^
E (u 0 ) = E ( β 0 + β 1 X o − ( β 0 + β 1 X o + uo ))

using what we know about the expected value of OLS estimates,

Proof: The expected value of the forecast error is
^ ^
E (u 0 ) = E ( yo − yo )

^ ^ ^
sub. in for y 0 = β 0 + β 1 X o and yO = β 0 + β 1 X o + uo
^ ^ ^
E (u 0 ) = E ( β 0 + β 1 X o − ( β 0 + β 1 X o + uo ))

using what we know about the expected value of OLS estimates,

^ ^
E ( β 0 ) = β0 and E ( β 1) = β1
Proof: The expected value of the forecast error is
^ ^
E (u 0 ) = E ( yo − yo )

^ ^ ^
sub. in for y 0 = β 0 + β 1 X o and yO = β 0 + β 1 X o + uo
^ ^ ^
E (u 0 ) = E ( β 0 + β 1 X o − ( β 0 + β 1 X o + uo ))

using what we know about the expected value of OLS estimates,

^ ^
E ( β 0 ) = β0 and E ( β 1) = β1
^
E (u 0 ) = β 0 + β 1 X o − β 0 − β 1 X o − E (u0 )
= 0 − E (u0 )
Proof: The expected value of the forecast error is
^ ^
E (u 0 ) = E ( yo − yo )

^ ^ ^
sub. in for y 0 = β 0 + β 1 X o and yO = β 0 + β 1 X o + uo
^ ^ ^
E (u 0 ) = E ( β 0 + β 1 X o − ( β 0 + β 1 X o + uo ))

using what we know about the expected value of OLS estimates,

^ ^
E ( β 0 ) = β0 and E ( β 1) = β1
^
E (u 0 ) = β 0 + β 1 X o − β 0 − β 1 X o − E (u0 )
= 0 − E (u0 )

and since Gauss-Markov assumption is that for any observation the expected value of
the true residual is zero, E(u0)=0
Proof: The expected value of the forecast error is
^ ^
E (u 0 ) = E ( yo − yo )

^ ^ ^
sub. in for y 0 = β 0 + β 1 X o and yO = β 0 + β 1 X o + uo
^ ^ ^
E (u 0 ) = E ( β 0 + β 1 X o − ( β 0 + β 1 X o + uo ))

using what we know about the expected value of OLS estimates,

^ ^
E ( β 0 ) = β0 and E ( β 1) = β1
^
E (u 0 ) = β 0 + β 1 X o − β 0 − β 1 X o − E (u0 )
= 0 − E (u0 )

and since Gauss-Markov assumption is that for any observation the expected value of
the true residual is zero, E(u0)=0

^
E (u 0 ) = 0 the mean value of any OLS forecast error is zero
The variance of each individual forecast error is given by

⎡ ⎤
⎢ _
2 ⎥
^ ^ −
Var (Y 0 − Yo ) = Var (u o ) = s 2 ⎢1 + + ⎥
1 ( X o X )
⎢ N N _ ⎥
⎢ ∑(Xi − X ) ⎥
2
⎢⎣ i =1 ⎥⎦
The variance of each individual forecast error is given by

⎡ ⎤
⎢ _
2 ⎥
^ ^ −
Var (Y 0 − Yo ) = Var (u o ) = s 2 ⎢1 + + ⎥
1 ( X o X )
⎢ N N _ ⎥
⎢ ∑(Xi − X ) ⎥
2
⎢⎣ i =1 ⎥⎦

where N is the number of in-sample observations,

s2 is the residual variance = RSS/N-k

The variance of each individual forecast error is given by

⎡ ⎤
⎢ _
2 ⎥
^ ^ −
Var (Y 0 − Yo ) = Var (u o ) = s 2 ⎢1 + + ⎥
1 ( X o X )
⎢ N N _ ⎥
⎢ ∑(Xi − X ) ⎥
2
⎢⎣ i =1 ⎥⎦

where N is the number of in-sample observations,

s2 is the residual variance = RSS/N-k

So precision of forecast is increased by

a) the closer is the out of sample observation Xo to mean value of in-sample X’s
The variance of each individual forecast error is given by

⎡ ⎤
⎢ _
2 ⎥
^ ^ −
Var (Y 0 − Yo ) = Var (u o ) = s 2 ⎢1 + + ⎥
1 ( X o X )
⎢ N N _ ⎥
⎢ ∑(Xi − X ) ⎥
2
⎢⎣ i =1 ⎥⎦

where N is the number of in-sample observations,

s2 is the residual variance = RSS/N-k

So precision of forecast is increased by

a) the closer is the out of sample observation Xo to mean value of in-sample X’s
b) the larger the sample size
The variance of each individual forecast error is given by

⎡ ⎤
⎢ _
2 ⎥
^ ^ −
Var (Y 0 − Yo ) = Var (u o ) = s 2 ⎢1 + + ⎥
1 ( X o X )
⎢ N N _ ⎥
⎢ ∑(Xi − X ) ⎥
2
⎢⎣ i =1 ⎥⎦

where N is the number of in-sample observations,

s2 is the residual variance = RSS/N-k

So precision of forecast is increased by

a) the closer is the out of sample observation Xo to mean value of in-sample X’s
b) the larger the sample size
(not expect a good forecast if only have little information)
The variance of each individual forecast error is given by

⎡ ⎤
⎢ _
2 ⎥
^ ^ −
Var (Y 0 − Yo ) = Var (u o ) = s 2 ⎢1 + + ⎥
1 ( X o X )
⎢ N N _ ⎥
⎢ ∑(Xi − X ) ⎥
2
⎢⎣ i =1 ⎥⎦

where N is the number of in-sample observations,

s2 is the residual variance = RSS/N-k

So precision of forecast is increased by

a) the closer is the out of sample observation Xo to mean value of in-sample X’s
b) Larger sample size
(not expect a good forecast if only have little information)
c) Larger variance of in-sample X’s
The variance of each individual forecast error is given by

⎡ ⎤
⎢ _
2 ⎥
^ ^ −
Var (Y 0 − Yo ) = Var (u o ) = s 2 ⎢1 + + ⎥
1 ( X o X )
⎢ N N _ ⎥
⎢ ∑(Xi − X ) ⎥
2
⎢⎣ i =1 ⎥⎦

where N is the number of in-sample observations,

s2 is the residual variance = RSS/N-k

So precision of forecast is increased by

a) the closer is the out of sample observation Xo to mean value of in-sample X’s
c) Larger sample size
(not expect a good forecast if only have little information)
c) Larger variance of in-sample X’s
(wider variance means more likely already encountered a big range of possible X
values to make a forecast with)
The variance of each individual forecast error is given by
⎡ ⎤
⎢ _
2 ⎥
^ ^ −
Var (Y 0 − Yo ) = Var (u o ) = s 2 ⎢1 + + ⎥
1 ( X o X )
⎢ N N _ ⎥
⎢ ∑(Xi − X ) ⎥
2
⎢⎣ i =1 ⎥⎦
where N is the number of in-sample observations,

s2 is the residual variance = RSS/N-k

So precision of forecast is increased by

a) the closer is the out of sample observation Xo to mean value of in-sample X’s
b) Larger sample size
(not expect a good forecast if only have little information)

c) Larger variance of in-sample X’s

(wider variance means more likely already encountered a big range of possible X
values to make a forecast with)

d) Better fitting model (implies smaller RSS and smaller s2. The better the model is at
predicting in-sample, the more accurate will be the prediction out-of-sample)
Follows that can construct α% confidence interval for each individual forecast
observation (just as with the confidence interval for an unknown true coefficient value)
using the square root of the above variance and substituting this into the formula for the
confidence interval
Follows that can construct α% confidence interval for each individual forecast
observation (just as with the confidence interval for an unknown true coefficient value)
using the square root of the above variance and substituting this into the formula for the
confidence interval
Just as
^
β 1 − β1
follows a t distribution
^
s.e.( β 1)
Follows that can construct α% confidence interval for each individual forecast
observation (just as with the confidence interval for an unknown true coefficient value)
using the square root of the above variance and substituting this into the formula for the
confidence interval
Just as
^
β 1 − β1
follows a t distribution
^
s.e.( β 1)
^ ^
using u 0 rather than β 1 then
Follows that can construct α% confidence interval for each individual forecast
observation (just as with the confidence interval for an unknown true coefficient value)
using the square root of the above variance and substituting this into the formula for the
confidence interval
Just as
^
β 1 − β1
follows a t distribution
^
s.e.( β 1)
^ ^
using u 0 rather than β 1 then
^ _
u0 − u0
also follows a t distribution
^
s.e.(u 0 )
Follows that can construct α% confidence interval for each individual forecast
observation (just as with the confidence interval for an unknown true coefficient value)
using the square root of the above variance and substituting this into the formula for the
confidence interval
Just as
^
β 1 − β1
follows a t distribution
^
s.e.( β 1)
^ ^
using u 0 rather than β 1 then
^ _
u0 − u0
also follows a t distribution
^
s.e.(u 0 )
_
and since u 0 = E (u0 ) = 0
Follows that can construct α% confidence interval for each individual forecast
observation (just as with the confidence interval for an unknown true coefficient value)
using the square root of the above variance and substituting this into the formula for the
confidence interval
Just as
^
β 1 − β1
follows a t distribution
^
s.e.( β 1)
^ ^
using u 0 rather than β 1 then
^ _
u0 − u0
also follows a t distribution
^
s.e.(u 0 )
_
and since u 0 = E (u0 ) = 0
^
u0
also follows a t distribution
^
s.e.(u 0 )
 ^
u0
Given
^
s.e.(u 0 )
^ ^
Sub. in u 0 = yo − yo
 ^
u0
Given
^
s.e.(u 0 )
^ ^
Sub. in u 0 = yo − yo
^
y − yo
becomes o
^
s.e.(u 0 )
 ^
u0
Given
^
s.e.(u 0 )
^ ^
Sub. in u 0 = yo − yo
^
y − yo
becomes o
^
s.e.(u 0 )

Since this follows a t distribution we know that 95% of observations will lie in the region

⎡ ^ ⎤
y −y
Pr ⎢− t .N05−/k2 ≤ o o ≤ +t .N05−/k2 ⎥ = 0.95
⎢ ^ ⎥
⎢⎣ s.e.(u 0 ) ⎥⎦
 ^
u0
Given
^
s.e.(u 0 )
^ ^
Sub. in u 0 = yo − yo
^
y − yo
becomes o
^
s.e.(u 0 )

Since this follows a t distribution we know that 95% of observations will lie in the region

⎡ ^ ⎤
y −y
Pr ⎢− t .N05−/k2 ≤ o o ≤ +t .N05−/k2 ⎥ = 0.95
⎢ ^ ⎥
⎢⎣ s.e.(u 0 ) ⎥⎦

and rearranging terms gives

^ ^ ^ ^
α /2 α /2
Pr[Y o − t N − k s.e.(u o ) ≤ Yo ≤ Y o + t N − k s.e.(u o )] = 0.95
Given

^ ^ ^ ^
Pr[Y o − tα /2
N −k s.e.( u o ) ≤ Yo ≤ Y o + t α / 2 s.e.(u )] = 0.95
N −k o

^ ^
= Y o ± tα / 2 SE (u o )

then we can be 95% confident that the true value will lie within this range

If it does not then the model does not forecast very well
 November 2007 CPI Fan Chart November 2007 GDP Fan Chart

Source Bank of England Inflation report

Example
reg cons income if year<90 /* use 1st 35 obs and save last 10 for forecasts*/

Source | SS df MS Number of obs = 35

---------+------------------------------ F( 1, 33) = 3190.74
Model | 1.5750e+11 1 1.5750e+11 Prob > F = 0.0000
Residual | 1.6289e+09 33 49361749.6 R-squared = 0.9898
---------+------------------------------ Adj R-squared = 0.9895
Total | 1.5913e+11 34 4.6803e+09 Root MSE = 7025.8

------------------------------------------------------------------------------
cons | Coef. Std. Err. t P>|t| [95% Conf. Interval]
---------+--------------------------------------------------------------------
income | .9467359 .0167604 56.487 0.000 .9126367 .9808351
_cons | 6366.214 4704.141 1.353 0.185 -3204.433 15936.86

predict chat /* gets fitted (predicted) values from regression */

predict forcse, stdf /* gets standard error around forecast value */

/* graph actual and fitted values, draw line through OLS predictions */

two (scatter cons income) (line chat income if year<90, xline(426000))

twoway (scatter cons income) (line chat income, xline(426000))
600000
500000
400000
300000
200000

200000 300000 400000 500000 600000

income

cons Fitted values

Vertical line denotes boundary between in and out of sample observations

 ^ ^
N −k
/* now calculate confidence intervals using cons = cons ± t ∂ / 2 * SE (u )
g minconf= chat - (2.04*forcse)
g maxconf= chat + (2.04*forcse)

/* graph predicted consumption values and confidence intervals */

twoway (scatter cons income if year>89) (line chat income if year>89, lstyle(dot)) (line maxconf income if year>89)
> (line minconf income if year>89)

550000
500000
450000
400000

440000 460000 480000 500000 520000 540000

income

cons Fitted values

maxconf minconf

So for most of the out of sample observations the actual value lies outside the
95% confidence interval. Hence the predictions of this particular model are not
that good.
Chow Test of Forecast Stability
If a model forecasts well out of sample then we would expect all the
out-of-sample residuals to be close to zero.
Chow Test of Forecast Stability
If a model forecasts well out of sample then we would expect all the
out-of-sample residuals to be close to zero.

It can be shown that the joint test of all the out-of-sample-residuals

being close to zero is given by:

RSS in +out − RSS in / N o

F= ~ F[ N o , N − k ]
RSS in / N − k

where No is the number of out-of-sample observations

N is the number of in-sample observations
k is the number of RHS coefficients
Chow Test of Forecast Stability
If a model forecasts well out of sample then we would expect all the
out-of-sample residuals to be close to zero.

It can be shown that the joint test of all the out-of-sample-residuals

being close to zero is given by:

RSS in +out − RSS in / N o

F= ~ F[ N o , N − k ]
RSS in / N − k

where No is the number of out-of-sample observations

N is the number of in-sample observations
k is the number of RHS coefficients

Intuitively if the model fits well the RSS from the combined regression
should be close to that from the in-sample regression.

A “large” difference suggest the RSS are different and so model does
not forecast well)
Given a null hypothesis that the model is stable out of sample (predicts
well) then if
^
F > Fcriticalα [ N o , N − k ]
reject null of model stability out-of-sample
Example:
Estimate of consumption function model including out-of-sample observations is
reg cons income if year<90

Source | SS df MS Number of obs = 35

-------------+------------------------------ F( 1, 33) = 3190.74
Model | 1.5750e+11 1 1.5750e+11 Prob > F = 0.0000
Residual | 1.6289e+09 33 49361749.6 R-squared = 0.9898
-------------+------------------------------ Adj R-squared = 0.9895
Total | 1.5913e+11 34 4.6803e+09 Root MSE = 7025.8
------------------------------------------------------------------------------
cons | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
income | .9467359 .0167604 56.49 0.000 .9126367 .9808351
_cons | 6366.214 4704.141 1.35 0.185 -3204.433 15936.86
------------------------------------------------------------------------------

. reg cons income

Source | SS df MS Number of obs = 45

-------------+------------------------------ F( 1, 43) = 5969.79
Model | 4.7072e+11 1 4.7072e+11 Prob > F = 0.0000
Residual | 3.3905e+09 43 78849774.6 R-squared = 0.9928
-------------+------------------------------ Adj R-squared = 0.9927
Total | 4.7411e+11 44 1.0775e+10 Root MSE = 8879.7
------------------------------------------------------------------------------
cons | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
income | .9172948 .0118722 77.26 0.000 .8933523 .9412372
_cons | 13496.16 4025.456 3.35 0.002 5378.05 21614.26
------------------------------------------------------------------------------
Comparing RSS from this with that above can calculate the F value
RSS in + out − RSS in / N o
F= ~ F[ N o , N − k ]
RSS in / N − k
di ((3.3905-1.6289)/10)/(1.6289/(35-2))

3.3905 − 1.6289 / 10
F=
~ F [10,35 − 2]
1.6289 / 35 − 2

= 3.57 ~ F[10,33]

From tables Fcritical.05[10,33] = 2.10

So F > F and therefore reject null that model predicts well out of
^
critical

sample.