In my understanding, Excel 2003 had finally worked out most of the statistics bugs. For years, the trendline formula was (believe it or not) considered a world-class techniue for calculating poly fits, but 2003 finally brought this capability to the worksheet functions.

Excel 2007 broke the trendline formula for many cases, inappropriately changing coefficients within a certain range to zero due to an overzealous rounding error correction algorithm. SP1 made a few improvements, SP2 has made more. We all hope and wish that Excel 2010 will clean up all of the problems (and not just with statistics!).

You didn’t include your trendline, so I can’t comment directly. When I plot your data in Excel 2003, I get this closely fitting trendline:

y = -746.65×3 – 317.45×2 – 3.2524x + 273.16
R² = 0.9996

I have heard of problems with the trendline formula in Excel 2007 with some data sets. I repeated this in Excel 2007, and got this trendline:

y = -746.65×3 – 317.45×2 – 3.2524x + 273.16
R² = 0.9996

Exactly the same. What’s more, the LINEST results below agree with the trendline:

Here is the data and calculated Y values based on the LINEST coefficients:

Here is the chart with the trendline:

Here is the chart with the original data and the LINEST-fitted data:

On the off chance that the chart type was incorrect, I converted the chart to a line chart. This changes the nature of the curve from concave downward, like half of a parabola, to a sigmoidal curve, and the trendline doesn’t really fit:

y = 0.0847×3 – 1.1768×2 + 3.2385x + 270.87
R² = 0.9932

Here’s the line chart, which obviously looks amiss:

0 273.15
0.019477851 273
0.043981716 272.3
0.074296729 270.9
0.109878671 267.9
0.127178507 266.2
0.135243092 265

Any suggestions would be appreciated.

Thanks,

Adam

I once worked for a guy who didn’t understand this whole concept. We’d be talking for an hour about how we were going to process this lot of material, and I’d show test data that he didn’t follow. Finally, he’d ask me (yet again), “What’s the definition of Standard Deviation”? He was the Director of Engineering, but we all called him Duh-rector.

I want to share a little bug-a-boo that I have with Excel trendlines.

I am a quantitative researcher with 15-years of experience using SAS as a statistical programming language — and have recently moved from investment banking to a small startup.

People here at my new job use Excel to place trendlines on graphs and refer to the trends — with no exploration into whether the coefficient on the slope of the line is significant!

I was in an internal meeting once where a long, painful discussion took place to figure out how best to explain the unexpected negative slope of a trendline on an Excel graph to a client. I finally asked whether the negative slope was actually *that* important — thinking about whether it was strongly significant or just borderline.

Turns out — no-one had checked the significance of that coeffcient at all.

When they did, it was really FAR from being significant. In other words — even though there seemed to be a negative slope when you eyeballed the data — mathematically (with the amount of data and the amount of noise in the data) we could not say with confidence that the slope wasn’t zero. In actuality, we should be telling the client that there is no trend. That the data is constant over time.

This was a rude awakening to some people at my work and I just wanted pass along the joy!