Reference Ranges, Part 2
By David Plaut and Clarice Perry
May 2, 2012
Welcome back. In this installment we will tackle the establishment (as opposed to verification) of a reference range. (Establish: to create or begin; Verify: to make sure or demonstrate that (something) is true, accurate, or justified). We will review a bit from our last installment on verification. You will recall at that time we said that if 95% of your values are not within the expected range of the package insert, you need to establish your own reference range (RR).
The first step is to select the data from your population. It is recommended (CLSI) that 120 samples be collected for each sex and then by decade of age (e.g., 20 - 29 years, 30 - 39 years, etc.) For example, if we wish to establish an RR for BNP for men and women from age 50 to 80 we have 2 sexes times 3 decades times 120, or 720 samples. Where do we find those 720 people? Perhaps we could use the data from a health fair, or blood donors. However, suppose that your study analyte is something like D-dimer, BNP, or TnI. These are tests that are normally run in the Emergency Department (ED) on patients with chest pain. Thus, your population should reflect ED patients with chest pain, but not VTE, CHF, or MI. Granted, this is a special, but not unique, situation. The point is we need selection criteria. In other words, can we "select" all the people at the health fair, or do we need to ask them if they have been diagnosed with thyroid disease, or diabetes, and so on? Or, do we need to run a CBC, urinalysis, and a BMP on all of them? Before you answer yes, consider the time and expense this would require. Then, if you still say yes, let's imagine that the blood sugar from the BMP comes back high. Does this mean that all of the values from the CBC, UA and BMP are excluded, or can we use all the other values but exclude the blood sugar? Data collection is no small task, but essential in establishing an appropriate reference range. As examples of how knotty this can get, consider the following exclusion criteria others have used. For example:
• Therapeutic drugs with influence on serum and plasma enzyme concentration (e.g. warfarin, antiepileptics, diphenylhydantoin, aminopyrin, antidepressants, analgesics).
• BMI >30
• Heavy exercise in the previous days
• Alcohol > 30 grams per day
To us, this is almost possible. That is, the number of samples that would be discarded is probably not excessive. But, the risk is including unhealthy people. For another example, look at these exclusion criteria:
• Diabetes mellitus type 1 and type 2
• Burns and muscle traumas
• Chronic nephropathies
• Acute and chronic infection
• Hepato-biliary diseases
Without a doubt, these exclusion criteria will net us a more "healthy" set of data, but a more difficult data set to obtain. The next example uses laboratory results versus diagnoses to exclude patients.
• FBS > 126 mg/dL
Creatinine > 0.2 mg/dL above URL
• CK > 300 U/L
• CRP > 12 mg/L
• UA > 8.0 mg/dL
• TG > 200 mg/dL
• Chol > 260 mg/dL
• Albumin < 3.2 mg/dL
• RBC (males) < 4.0 or > 5.5 mil/µL; (females) < 3.4 or > 5.2 mil/µL
• Hb < 13 g/dL (males); < 11 g/dL (females)
• WBC < 3000/µL or > 12000 /µL
• PLT < 100 /nL
• HCT (males), < 42 or > 52%; (females) < 37 or > 47%
• MCV < 80 or > 96 fL
• HB surface antigen Positive
• IgG antibodies Positive
• antiHB core antigen Positive
• AntiHCV antibodies Positive
Once you have determined the appropriate population, set your exclusion criteria, and collected your data, you are ready to establish the reference range. Discard the upper and lower 2.5% of your data. By taking the central 95% of your data, you eliminate the pitfalls of using the standard deviation to calculate the range. Consider the next example to illustrate a possible scenario where ±2SDs would not be appropriate. Below is a data set where N = 120, but as the histogram is skewed to the right, it would not be appropriate to use the standard deviation to set the reference range. Keep in mind our discussion of histograms where we showed you histograms from varying Ns. It is much easier to see skewness with 120 than 40 as the histograms below show.
Using the middle 95% of the data avoids the SD pitfall; avoids the pitfall in using a logarithmic transformation; or any number of other mathematical approaches, useful as they may be. If a nearby hospital is doing a health fair or has the same instrument as you, it might be a good idea to share resources if you find it wise or necessary to establish an RR. As we said in our last installment, a poorly selected sample and inappropriate selection criteria yields data that can be more confusing to the doctor than no value.
Our next installment begins our discussion of inferential statistics with the confidence of the mean. As always, we encourage you to send comments or suggestions to email@example.com
Interpreting Statistics Archives