Forming a reference/normal range for laboratory test in SPSS

#1
Hi,

I'm doing a write up for my thesis and having trouble with forming a reference range in SPSS for medical laboratory FBC results. Without SPSS I used the formula I would use for normally distributed data: Mean +/- (1.96xStandard deviation)

For my normally distributed data I used:
Analyse - frequencies - choose variable - statistics - percentiles - 2.5 and 97.5.

I'm not sure if I should use the 2.5th and 97.5th percentiles or should I use the 5th and 95th percentiles? I'm not sure which is correct.


For my non parametric data I used the same as the normally distributed but also:
Bootstapping - 112 samples
Confidence intervals - 95%


Am I correct in doing this? Any advice would be greatly appreciated!!

Annmarie :)
 
#2
For a reference range of "normal" lab values, I don't think you want a confidence interval. I think people inappropriately do so because they don't realize you're not making an inference on some population parameter, and they confuse a 95% CI (inferential) with an interval that has 95% of observations contained within it (descriptive).

I think the percentiles is a better approach, unless you have data from a roughly normal distribution, then 1.96 SD on either side of the mean would be the cutoffs.

Ideally, you would have a large enough database of values from a spectrum of healthy patients (say 10,000, for example). You want to take the 2.5 and 97.5 percentile values as this is an interval to encompass ~95% of the individual normal values (5th and 95th covers 90% of cases). Values outside of these ranges may be flagged as unusual (which is where a suspect outlier of >|2| SD from the mean comes in a normally distributed variable, but you want to use percentiles because some variables are not normally distributed).

The caveat is that some labs may use different methods of establishing a "normal" range, so reading up on this for various tests across companies may be useful.
 

hlsmith

Omega Contributor
#3
If you go the bootstrap approach for the nonparametric, I would use it for constructing all NR regardless of normality.

I you just making a recommendment or you changing practices with them.
 
#4
If you go the bootstrap approach for the nonparametric, I would use it for constructing all NR regardless of normality.

I you just making a recommendment or you changing practices with them.
I personally agree with you in the sense that using percentiles is the way to go because that will align with 1.96 SD, more or less, in a roughly normal distribution, and it gets the job done appropriately for data with unknown or non-normal distributions.
 

hlsmith

Omega Contributor
#5
Also, establishing NR probably doesn't have to be based on 95% NR, probably more so on clinical risk, etc.

I thought I saw something from a hospital the other day mentioning the approval of NR per a local committee.
 
#6
Also, establishing NR probably doesn't have to be based on 95% NR, probably more so on clinical risk, etc.

I thought I saw something from a hospital the other day mentioning the approval of NR per a local committee.
This is what I was hinting at when I said there are different ways of establishing normal. The OP may be well served to see how other labs come up with normal for the kind of test he or she wants to create a range for. There are also instances where only a certain lab method of classification is acceptable because others fall short in utility, as you seem to allude to.