From: Damon Hart-Davis
Date: Thu, 19 Jan 2017 15:06:27 +0000
To: Paul Miller
Content-Transfer-Encoding: quoted-printable

Hi Paul,

> On 19 Jan 2017, at 14:44, Paul Miller <paul.miller@npl.co.uk> wrote:
>=20
> Hi Damon,
> =20
> The brief comments during our review of the your statistics were:
> =20
> How sensitive is the fit to the chosen baseline?

Emprically, with other data sets, I have not found it to make a huge =
difference, or at least not systematically.  I have typically tried a =
range of base temperature values to find an optimal fit (and that might =
be applicable to an update of this protocol for example) but have not =
found a consistent alternative baseline for (eg) separate time periods =
at the same property to fall out of the data.

In other words, the key slope results appear to remain relatively robust =
in my experience in the face of possibly non-optimal baseline values.

I think that the consistency of a fixed, standard, well-understood base =
temperature outweighs the small amount of noise that it probably =
introduces to the results.

> How significant is the =E2=80=93ve intercept to the calculated =
performance parameter?

Again, empirically, I haven=E2=80=99t found this to be a consistent sign =
of a problem unlike (say) a very low R^2 value.  I=E2=80=99d want to =
investigate further to understand better if I was consistently seeing =
such results.

In particular, when the heating fuel use is not separated from other =
uses of that fuel for metering purposes (eg gas being used for water as =
well as space heating), as will almost always be the case for the =
application of this protocol, noise in the intercept is to be expected; =
otherwise in general a -ve intercept would suggest that a higher =
baseline should be used.

> Do the responses to the two questions above include the impact on the =
reported measurement uncertainty?

That I=E2=80=99d like to take further advice on.

I will conduct a bit more desktop research to provide better answers to =
the above and could reanalyse some other data sets that I have to better =
answer these for my understanding and for audit purposes.  This is not a =
new analysis technique so there will be opinions available.

I note this commentary on such issue from my source of heating degree =
data:

    http://www.degreedays.net/regression-analysis

which is not *entirely* in line with my answers above, I note, but =
close.

> Will the 2nd set of test data help deal with these points?

Possibly.  In particular it may be possble to hand-tweak the baseline in =
the code and observe how sensitive the reported results are to it, even =
if that parameter is nominally fixed.

> As discussed, for the audit we would expect a written procedure (could =
be a single side of A4) that describes your decision making, perhaps =
with worked example(s).

Yes, thank you.

> Again, if you have any questions, let me know.
> =20
> Paul

Regards

Damon=