Strategy for diff indiff analysis

[Ellabluebird]

Need help- powerpoint due Monday.
Ok I’m comparing the difference between a couple pediatric risk of mortality scores. I want to take the difference and see if telemedicine or telephone consults made a difference in risk of mortality.

2 of the 4 scores are so skewed we wanted to log transform them. So I log transformed them, then standardized them, then took the difference, then ran a regression. How do I know whether this is adequate? Or whether that’s too many transformations and I’m making my data uninterpretable?
Should I just standardize the data and then get the difference and run a wilcoxon? It’s the the down side of wilcoxon that it’s no good for skewed data?

 

[hlsmith]

What is sample? Can you post histograms of these variables? If it is a score, what do scores look like?

 

[Ellabluebird]

Thanks. The scores are PRISA and RePEAT ( ER risk of mortality) and PRISM and PIM ( Peds ICU risk of mortality).

these are the raw scores, and the histograms.

what I need to do is standardize them, and then take the difference, ( PIM - Repeat, PIM - PRISA, PRISM - Repeat, PRISM - PRISA ) to see if my intervention has an impact on the change in risk of mortality scores from when the patient is in the ER to when they are in the PICU.

these deltas will become my dependent variable in a multiple linear regression I hope.

Here are the problems we ran into :

the log transformation doesn't really normalize the data well.

my team is worried that doing a log transformation of PIM and PRISM and then taking the difference will obscure the data. I tried to do the standardization first and then log transform, but you have negative numbers after standardization and taking the difference. so you can't take the log of a negative.

thanks for your help.

 

[hlsmith]

Doesn't feel like a good idea subtract two different scales. Any precedence.

What if the two scales are focused n different attributes and they change over time but this doesn't get registered.

Your first scale seems heavily zero bound.

 

[Ellabluebird]

precedent? is that what you mean? thats why I standardized them, to use Z scores.

they are different scales but measure the same thing basically- risk of mortality.

yep, the PRISM score has lots of zeroes. And PRISA has a few as well.

 

[hlsmith]

Does anybody actually die in your sample?

How do you select which scales to subtract?

What does a std look like on a bounded variable, is the problem. What does it look like after transforming it?

How valid are the scales and does that impact the differences?

 

[hlsmith]

Sorry for typos and grammar. I was talking go some one while typing

 

[Ellabluebird]

Does anybody actually die in your sample? only 5 of the 345 with complete data died.

How do you select which scales to subtract?

PRISA and Repeat are ER scores ( taken first) Repeat is a severity of illness score and PRISA is risk of admission .
I'm doing all 4 permutations ( both PICU scores - both ER scores) to see if there is a difference. I guess if pressed my hypothesis would be you would see the greatest difference in PRISM - Repeat because there's PRISM is measured at 12 hours after ICU admission and PIM is measured at 1 hour.

What does a std look like on a bounded variable, is the problem. What does it look like after transforming it?








How valid are the scales and does that impact the differences? all 4 are validated pediatric scales to measure mortality/ severity of illness.

 

[hlsmith]

Yeah given the other issues its seems like you will also need to address false discovery, given multiple comparisons.

Also, if a person has a high ED score are they are likely to get different treatment course than their counterpart?

Can you state the purpose of these analyses succinctly again, because there could be a better approach.