We have found statistical differences with t-tests, but there must be more!

[MissyPausch]

Hello TS users,

The attached data set refers to patients that had an operation, a vast number of measurements were taken before and after.

We have found some evidence of improvement in questionaires for pain scale (VAS), ability of living (RMS & ODS) and walking distance, all tested using t-tests, comparing means before and after.

This is great but nothing new in terms of the effect the operation has on patients, we have about 20 other variables of various kinds. We are currently looking at the 'compare means' tab in SPSS to try and find alternative hypothesis.

Our supervisor has gone AWOL and we are rapidly nearing the end of our knowledge for statistics.

We have for sure tried, everything from correlations, multiple regressions, t-tests until the cows come home and various data mining techniques. We have been using Dytham C. 3rd edition to try and filter out any relevant information.

If anyone has any suggestions on where we can go next with this data, it would be truley wonderful. We have all our outputs, transformations, etc. to send if anyone has the time and the challenge in them to try and find something meaningful. We would love to read anyones comments on the data set.

With kind regards and crossed fingers, and oh course a massive thank you.

Samuel and Sibylle

PS in the data set k.A. denotes that no data is available.

[ledzep]

First of all, a very interesting topic.

Before jumping guns, the most important question: WHAT IS YOUR OBJECTIVE? What is the aim of your project?

Just had a quick look at your data.
I am never comfortable working out means for variables which has scale measurements. The reason being the uncertainity in uniformity of scale. For example, is the difference in pain between 2 and 3 same as the difference in pain between 7 and 8? (I don't know). If the scale doesn't represent the severity adequately, I would be more inclined towards simple frequency table. You can take the difference in pain score, and do the frequency table. A lot of negatives will suggest reduced pain.

# For your pain measurements (post-pre)
# Frequency counts (VAS)
Diff -10  -9  -8  -7  -6  -5  -4  -3  -2  -1   0   1   2   3   4 
Freq  6   2  11   9   6   8   9  13  10   7  16   1   2   2   1 

As can be seen a lot of negatives suggesting the pain relief post treatment.

If anyone has any suggestions on where we can go next with this data, it would be truley wonderful.

I understand it is a research work and you'll like to explore new things based on the data you have in hand. However, the objectives has to be set out before collecting the data (forget before analyzing). Changing objectives after seeing the data is not a very good idea.
The generic things to try are look at the side effects

We have for sure tried, everything from correlations, multiple regressions, t-tests until the cows come home and various data mining techniques. We have been using Dytham C. 3rd edition to try and filter out any relevant information.

Again its all great running a bunch of exploratory analysis but you have hold back and answer. Do they add something to answering the objectives of your research?

[Dason]

Yes with all the analysis that has been done without planning for it beforehand this sounds very exploratory to me. Which is fine but I do hope it's just exploratory and you're not hoping to make conclusive statements. If you find something interesting would you be doing another test at a later time to verify whether the finding was actually real or if it was an artifact of running tests until the cows came home?

[ledzep]

If you find something interesting with the data, then the best you could do is to "Generate Hypothesis" (no conclusions can be reached because your data comes from an experiment which was not designed to answer the interesting finding you may find with this data).
As Dason suggested, then you can propose this hypothesis and design a new experiment to test this hypothesis.

[MissyPausch]

Sorry, I am a little embarressed that we didn't outline what we were looking for.

The study had the sole aim to find out if the oporation corrected the condition long term, up to 3 years, using primarily the questionaire scores before and after. It may not be so clear in the data, sorry, follow-up occurred in 2009. Patients with a Year of Op. in 2006 had a follow up time of 3 years, 2 years for 2007 and 1 year for 2008.

The rest of the data was taken as it is standard practice to do this for the type of oporation. We wanted to check to see if any other patterns occurred within the additional data.

We thought about the following hypothesis when we realized we had the additional data.

Can walking distance predict long term operation success?
Does physio before and/or after the oporation play a role in recovery speed?
Does 'Doctors Pain Rank' reflect the self-assesment questionaires completed by the patient before the operation?
Does complication occurance effect recovery time?
Can 'period of pain' before the op. predict long term recovery?


TBH there are about another ten we came up with whilst imputting the data. I guess if it is not supposed to be done then I won't waste your time by doing so!

So is it a done thing to continue to analyse the data in the circumstances?

Hi Ledzep (great name!), thank you for your reply and interest.

is the difference in pain between 2 and 3 same as the difference in pain between 7 and 8?

Yes these are the same, or are to the extent that every persons opinion on their own pain is judged in the same way. The pain scale though is though of as having equal distance between increments!

Ledzep, regarding the other questions about adding something to our objectives, then I guess they would! This of course depends on if we are 'aloud' to analyze the additional hypothesis we possed during data input. It was pretty much a case of the fact that we didn't know this data existed until we were inputting the standard data for the patients (DoB for example).

Dason,
We thank you also for your comment. Unfortunately we will not be able to run further tests to confirm such effects. Is it possible to simply suggest that 'this data may suggest..' for example (regarding the additional hypothesis')?

Thank you again for your advice and comments.