I trained as a biologist so my knowledge of statistics is minimal

I've just managed to get some image analysis for HCS running on our cluster. I'd like to prove that the results for my HPC technique are statistically similar to using their existing software, what form of statistics should I run on it? The data is in the form:

I don't think i fully understand the question - but I think you are basially asking how you can prove that the numbers in the column for HPC and those in the column for current method are very similar??

If so, then can't you just find out the difference between the two in every case? Then you can have an average difference, minimum and maximum difference, and even a spread with pretty graphs etc showing every value, giving an idea of the distribution of values.

Please note, I've never studied statistics in my life - but that's how I would do it!

paired student T-test could be the answer.

I was going to suggest that next.

Excel can do t-tests if you load the Analysis toolpack.

If it were me I'd try to prove both difference and similarity.

T-test proves difference
Correlation proves similarity

Always best to avoid Excel for stats Minitab at the least.



T-test proves if the difference is significant according the parameters set out. You can have sets of results that are perfectly correlated but are different and the OP wants to prove the results are not different.

but i could be wong

Thanks for the help.

Here's a scatterplot for my data:



And the original file in csv format is here (22kb). My technique is image2_PixelIntensityMean and Average Outer Intensity is the existing technique.

The same tool which I'm doing the image analysis on has an inbuilt T test and it comes up with:



(For some reason it insists on renaming the fields treatment is my technique.)

Does the above table look sensible? I don't know how many decimal places the P value is to.

Telling biologists that their past few year's work maybe a pile of poo because they'd used Excel's own regression functions is a hobby of mine.



I'd go for R myself (if I knew enough stats).

Actually, from what I've read on the t test, it may not work as the data that I'm looking at are not normalised. I'm wondering if carrying out the same statistics as the scientist originally performed (in this case working out the efficacy of various drugs on cells) and then doing a T test on his results and my own.

mann whitney?

This is why I like the fact we have a dedicated statistician at work to go and ask

First off, can you clarify what the two columns of figures represent - are they individual results or are the A/B/C blocks replicates?

Second, are you trying to prove that the PixelIntensity method gives comparable results to the AOT method?

Each row in the file represents the same image, but with two different methods of analysis one in each column. The other column is actually a coordinate (rows are letters, columns are numbers eg B1 is 2nd from left in the top column).



Exactly.

Right.

So, summing all the results for each block (A, B, C etc), it is clear that the pixel method is consistently "over counting", which isn't surprising. However, if you plot the totals against each other, you find that there is an almost perfect linear relationship between the two methods. (AOI = 0.898xPixel - 78289) with an R-squared of 0.9913. For each individual data point, the equation is AOI = 0.8739xPixel - 2976.6 with an R squared of 0.9261.

So, while the two methods give quite different results, there is a very strong relationship between the two. So, if you use the pixel method and apply the above equation, you get reusults that are quite close to those from the AOI method.

Here's block A:

Pixel AOI calculated
14836.56298 10266.68164 9989.072388
13948.07763 9259.987305 9212.625041
6812.057559 1813.310059 2976.457101
8836.744484 4011.632568 4745.831005
5655.597539 1482.571533 1965.826689
15833.87909 10933.84375 10860.62694
5449.987808 1227.18811 1786.144345
15648.57683 10695.43164 10698.69129
6458.702376 2258.579102 2667.660006
15303.77094 10883.83203 10397.36542
9439.19292 4675.243164 5272.310693
13390.46867 10026.29004 8725.330571
11123.52801 5451.89209 6744.251128
14322.2248 9913.797852 9539.592253
12166.87358 6828.822754 7656.030822
15872.66113 11304.67285 10894.51856
13034.07752 8356.905273 8413.880345
13245.9075 9129.931641 8598.998564
13466.1725 8569.248047 8791.488148
13099.45097 8024.505859 8471.010203
15464.76456 9860.850586 10538.05775
12566.94026 8175.209473 8005.649093
15169.36717 10313.85254 10279.90997
6220.324126 6889.153809 2459.341254


Is that what you were after?

Thanks for that. The correlation is pretty good. I had a quick look at the outliers on my graph. Most of them are caused by some form of artefact in the field. The thresholding picks the artefact up rather than the cells. I need to look at the original technique to see how they get around that.

I'm still trying to get hold of a copy of the original data so I can compare IC50s between techniques.