Thread Subject: linear regression - inconsistent results

I've been noticing when using regress or polyfit that I'm getting inconsistent result when I switch the 1 independent variable with the dependent variable. Mathematically this does not make sense to me.

If y = mx + b

then

x = y/m - b/m

But I've found for certain datasets that when I flip the y and x around using either regress (and adding a column of ones to X) or polyfit(x,y,1) I get non-consistent results.

This is a short example to illustrate my problem.

>> d=[ 0.0074143 0.052035
    0.0076173 0.0014361
    0.0077408 -0.0041507
     0.013317 0.054487
     0.013289 0.061777
     0.013346 0.055137
      0.01397 -0.046578
     0.014114 -0.026229
     0.014658 0.042499
     0.020282 -0.0010642];
>> polyfit(d(:,2),d(:,1),1)

ans =

    -0.011263 0.012788

>> polyfit(d(:,1),d(:,2),1)

ans =

      -1.0641 0.032315

THESE ARE INCONSISTENT RESULTS AREN'T THEY?
e.g.,
1/-0.011263 ~= -1.0641

"Russ Scott" <robinandruss@gmail.com> wrote in message
news:gevnh1$bnn$1@fred.mathworks.com...
> I've been noticing when using regress or polyfit that I'm getting
> inconsistent result when I switch the 1 independent variable with the
> dependent variable. Mathematically this does not make sense to me.
>
> If y = mx + b
>
> then
>
> x = y/m - b/m
>
> But I've found for certain datasets that when I flip the y and x around
> using either regress (and adding a column of ones to X) or polyfit(x,y,1)
> I get non-consistent results.
>
> This is a short example to illustrate my problem.
>
>>> d=[ 0.0074143 0.052035
> 0.0076173 0.0014361
> 0.0077408 -0.0041507
> 0.013317 0.054487
> 0.013289 0.061777
> 0.013346 0.055137
> 0.01397 -0.046578
> 0.014114 -0.026229
> 0.014658 0.042499
> 0.020282 -0.0010642];

Let's take a look at your data:

x = d(:, 1);
y = d(:, 2);
plot(x, y, 'go')
hold on

Do the green circles look like they're arranged on a line? Does it make
sense to fit a line to this data set, given how the points are distributed?

Pictures are sometimes truly worth a thousand words.

>>> polyfit(d(:,2),d(:,1),1)
>
> ans =
>
> -0.011263 0.012788

mb1 = polyfit(y, x, 1);
plot(polyval(mb1, y), y, 'r+')

When we evaluate the red line for the y values you used to fit the line, it
seems to be a rough approximation to the vertical line around x = 0.014.

>>> polyfit(d(:,1),d(:,2),1)
>
> ans =
>
> -1.0641 0.032315

mb2 = polyfit(x, y, 1);
plot(x, polyval(mb2, x), 'kx')

These look to be a rough fit to the horizontal line "splitting the
difference" between the points around y = 0 and the points around y = 0.05.

> THESE ARE INCONSISTENT RESULTS AREN'T THEY?
> e.g.,
> 1/-0.011263 ~= -1.0641

If your data was more "linear", then you might expect the relationship you
proposed above to hold.


figure
x = 1:10; y = 7*(x+rand(size(x)));
plot(x, y, 'go');
hold on

mb1 = polyfit(y, x, 1); % x = mb1(1)*y + mb1(2)
plot(polyval(mb1, y), y, 'r+')

mb2 = polyfit(x, y,1); % y = mb2(1)*x + mb2(2)
plot(x, polyval(mb2, x), 'kx')

compareLinearCoeff = [mb1(1), 1./mb2(1)]
compareConstantCoeff = [mb1(2), -mb2(2)./mb2(1)]


On these graphs, the black x's and the red +'s are much closer to one
another and to the green circles. While the relationship you proposed above
doesn't exactly hold, in this case the data is much better fit by the lines
and so the relationship is much closer to being satisfied.

--
Steve Lord
slord@mathworks.com

"Russ Scott" <robinandruss@gmail.com> wrote in message <gevnh1$bnn$1@fred.mathworks.com>...
> I've been noticing when using regress or polyfit that I'm getting inconsistent result when I switch the 1 independent variable with the dependent variable. Mathematically this does not make sense to me.
>
> If y = mx + b
>
> then
>
> x = y/m - b/m
>
> But I've found for certain datasets that when I flip the y and x around using either regress (and adding a column of ones to X) or polyfit(x,y,1) I get non-consistent results.
>
> This is a short example to illustrate my problem.
>
> >> d=[ 0.0074143 0.052035
> 0.0076173 0.0014361
> 0.0077408 -0.0041507
> 0.013317 0.054487
> 0.013289 0.061777
> 0.013346 0.055137
> 0.01397 -0.046578
> 0.014114 -0.026229
> 0.014658 0.042499
> 0.020282 -0.0010642];
> >> polyfit(d(:,2),d(:,1),1)
>
> ans =
>
> -0.011263 0.012788
>
> >> polyfit(d(:,1),d(:,2),1)
>
> ans =
>
> -1.0641 0.032315
>
> THESE ARE INCONSISTENT RESULTS AREN'T THEY?
> e.g.,
> 1/-0.011263 ~= -1.0641


In addition to the points made by Steve, note that regression minimizes

sum((y_data-y_predicition).^2)

which doesn't have to be the same as

sum((x_data-x_prediction).^2)

so transposing your data and repeating the fit won't normally give related regression parameters.

Ken

"Russ Scott" <robinandruss@gmail.com> wrote in message <gevnh1$bnn$1@fred.mathworks.com>...
> I've been noticing when using regress or polyfit that I'm getting inconsistent result when I switch the 1 independent variable with the dependent variable. Mathematically this does not make sense to me.
>
> If y = mx + b
>
> then
>
> x = y/m - b/m
>
> But I've found for certain datasets that when I flip the y and x around using either regress (and adding a column of ones to X) or polyfit(x,y,1) I get non-consistent results.
>

(snip)

> THESE ARE INCONSISTENT RESULTS AREN'T THEY?

NO! Sorry, but they are not. I'll just expand on
the other comments by a bit. What model do
you assume when you fit a linear regression
model? Do you know? If not, then you should
expend the effort to learn, because this is the
cause of your confusion.

You may think that you are fitting the model

  y = a*x + b

but, you are not. That model is actually missing
a term. In fact, your true model is of the form

  y = a*x + b + E_i

where E_i is assumed to be normally (Gaussian)
distributed error. Thus for each data point,
we assume additive, zero mean errors, but
with an unknown variance. Those errors are
added to the value of y.

What happens when you swap the x and y
in your regression? In effect, the model is
suddenly in a different form.

  x = c*y + d + F_i

Here, the errors are assumed to be in the
x variable. This difference is the source
of your problem. The model is truly
different, so those regression parameters
will certainly be different too.

John

"Ken Campbell" <campbeks@gmail.com> wrote in
news:gevq1m$g80$1@fred.mathworks.com:

> In addition to the points made by Steve, note that regression
> minimizes
>
> sum((y_data-y_predicition).^2)
>
> which doesn't have to be the same as
>
> sum((x_data-x_prediction).^2)
>
> so transposing your data and repeating the fit won't normally give
> related regression parameters.
>
> Ken
>

Exactly. The independent variable is called the independent variable for
a reason. you know that this isn't where your errors are. If you have
two variables with errors, the correct optimization is an orthogonal
regression, or an "error in variables" model.

--
Scott
Reverse name to reply

THANKS ALL for your help and information. Got it.