Wilcoxon test could be a good test for your work

Wilcoxon isn't the best first choice here. Cohen's kappa is a measure of agreement, intended to evaluate two raters evaluating the same x-ray, ice cream flavor, what have you. These are parallele groups of measurements. Why not do a chi-square test, with a Cochran-Armitage test for trend?

Thanks for the help, ill have a look at these tests.

You may construct square contingency tables, and model them by using symmetry models. This way may provide you rich inferences.

I think chi-square test could be a good choice, however I still think that Wilcoxon test, which is the non parametric equivalent of t-test, is the better choice. Moreover, a correlation analysis (i.e. Spearman Correlation) may be performed among variables.

For the purpose of deciding if one method is better than the other on a particular characteristic (e.g. organelles easily recognizable), I would use a chi-square test. However, the chi-square test is based on the assumption that the attributed scores are "correct" (no misclassification). In order to asses how objective/subjective is the scoring, I would then have two independent technicians (blind to the method used to prepare the slides) assign scores to a sub-sample of the slides and I would then calculate the kappa statistic on this. I would then present both results in my report. In any case, in order to avoid bias, it is important that the person(s) doing the scoring be "blind" to the method used to prepare the slides.

Thanks for the help everyone.

Mann-Whitney U test ( Wilcoxon rank sum test) is a non parametric equivalent to the unpaired t test and it looks to be correct in your situation to use. You can not use wilcoxon signed rank test basically because it is a non parametric equivalent for paired t test and your data is not paired.

If the scale you are using is perfectly ordinal and if your sample is sufficiently large you can even treat your data as on continuous scale and if it is following normal distribution ( which you can check by doing shapiro wilk test or klomogorov - smirnov test), you can even use unpaired t test.

Hi Everyone

If I get the pathologist to score 8 semi thin sections + ultra thin sections

out of 5 for both the old and new processing method,comparing them.

What would be the easiest statistal test to put into SPSS or excel?

I think this is a easier way of getting statistics than getting two pathologists to score subjectively.

Any help would be appreciated

Thanks

Tom

Hello Tim,

What are you scoring? Is it on categorical scale or continuous scale? How many groups are you comparing? What is the pattern of data distribution? Without knowing answers for these questions one can not decide about the statistical tests. If your data is on categorical scale and if you are comparing between only two groups you have to use Chi square test or if your data is on categorical scale and having more than two groups you have to use Kruskal wallis ANOVA. If your data is continuous and data shows normal distribution and if you are comparing between only two groups then simple two sample t test is suffice. If your data is on continuous scale showing normal distribution and if you are comparing between more than two groups you have to do ANOVA and if results are significant you need to follow it up with Suitable Post hoc. If your data is continuous but not following normal distribution and if you are comparing between only two groups you can use Man whitney U test . Your explanation looks to be insufficient to give specific answer.

Thanks

Nagesh Lakshminarayan

Wilcoxon (or MannWhitney or U-test and all othe synonyms) is not at all a choice here. The data is clearly not continuous, and there will be many many ties. Chisquare is ok, an ordered logistic model would give you more insights but this might be overdone. Forget Wilcoxon, use Chisquare.

I would also try to define continuous variables rather than scores...can you measure time to process rather than ease of processing score...etc. You can always generate scores from measurements, but you can't go the other way. Perhaps your scores are not sensitive to the differences you are trying to test?

I again wonder what people feel to be "poor content" in my answer. It is a clear advice, there is an argument given and it is not off-topic. If it is wrong, it would be good to get a correction.

I'm guessing someone meant to vote up...clearly nothing controversial here.

I think RG should revise the vote down to require that the respondent provide the rationale for the negative response which would be delivered and posted anonymously. This would provide opportunity to address what are more often than not mis-communications /understandings due to the short hand nature of these interactions.

Hi

My data looks like this-

Semi thins-

old processing method / New processing method

1) 3 4

2) 3 3

3) 4 4

4) 2 5

5) 3 4

6) 3 3

7) 4 4

8) 3 4

Ultra thins-

old processing method / New processing method

1) 3 5

2) 3 4

3) 3 4

4) 3 5

5) 2 4

6) 3 3

7) 4 3

8) 2 5

Semi thin scoring - 1-5

1) Very poor – Very poor quality of staining, preservation and nuclei/cell preservation.

2) Poor - Poor quality of staining, preservation, and nuclei/cell preservation.

3) Average– Average quality of staining, preservation, and nuclei/cell preservation.

4) Very good - Very good quality of staining, preservation, and nuclei/cell preservation.

5) Excellent – Excellent quality of staining, preservation, and nuclei/cell preservation.

Ultra-thin Scoring of muscle biopsy - 1-5

1) Very poor – Artefact present (All 4), very poor contrast, organelles not recognised, very poor resolution - grainy

2) Poor – Artefact present (up to 3), poor contrast, organelles not easily recognisable, poor resolution many areas of grainy background.

3) Average – Some artefact present (up to 2), average contrast, some most organelles recognisable, average resolution (areas of both good and grainy background).

4) Very Good – Minimal artefact present (up to 1), contrast very good, organelles recognisable, very good resolution of electron microscopy image

.

5) Excellent – No artefact present, excellent contrast, organelles easily recognisable, excellent resolution of electron microscopy image.

Thanks

Tom

Your data is organized as if the values are paired (each specimen analyzed twice: once with the old and once with the new method). My first guess was to look at the differences "New-Old", but at an ordinal scale such differences are not comparable (example: both 2-1 and 5-4 give +1, but the meaning can be entirely different in both cases).

To my knowledge, the only way to use thie pairing information was to make a simple comparison and record whether New>Old, New=Old, or New

In a previous comment, I said that Wilcoxon test could be used in this case (of course, if data are paired), I did not change my mind about this, but I agree with the chi-square test that you (Jochen Wilhelm) performed. However, I think the number of observations is too small for performing chi-square, so perhaps the logistic model could be more appropriate. Nevertheless, if is possible to increase the number of observations, I'd suggest this.

Regards

Jurandy, the logistic model has the same shortcomings like the chi-square test (the chi-square test *is* based on a simple logistic model). But I still have to insist that Wilcoxon is not a good choice because there are too many ties. It is one of the assumptions of this test that the variable is near continuous. See for instance: http://www.vassarstats.net/textbook/ch12a.html (end of this chapter).

Thanks for your help, I shall look into these tests

If im looking at current method vs new method for 8 specimens -

1)Semis vs Semis

2) Ultras vs Ultras

and instead i look at statistics to score individual parameters for each eg.

Semi thin sections- Score 1-3 -

Quality of staining – Strong or weak? Excellent to very poor

Preservation of staining – Artefacts? Scores?

Nuclei/cell preservation

Ultra thin sections Scores 1-5

Artefact – Staining contamination, dust, creases, scores?

Contrast – Black/white grey scale

Organelles recognisable? Mitochondria, lysosomes, nuclei, membranes, inclusions.

Resolution – Good/grainy background?

Eg for each specimen ( both curent+new method) the ultra thin section will get 4 scores from the parameters

eg

Ultra thin section scoring

Specimen 1

Current method

ARTEFACT 2

CONTRAST 3

ORGANELLES 4

RESOLUTION 3

Specimen 1

New method

ARTEFACT 2

CONTRAST 3

ORGANELLES 4

RESOLUTION 3

How do i test the null hypothesis of no difference?

Can I do a paired samples test?

Appreciate any help, as not sure how I can link the two together at the moment