Censoring values that are "extreme" can be done using several approaches. I'll go through a few: (1) Censor based on "arbitrary" thresholds. I use quotes around the "arbitrary" as these are not selected at random (e.g. we wont be picking 135 mm and 155 mm as our extremes in children). Sensible values would be (e.g.) 85 mm and 200 mm. No-one would argue that 80 mm in a 6 month old is extremely thin but 220 mm in a well-nourished 5 year old is not particularly extreme. In most settings where we (EN-NET people) will be using MUAC the 220 mmm will probably be an upper extreme. (2) Censor using thresholds derived from a reference (simple approach). In this approach you take a low extreme for a 6 month old and a high extreme for a 5 year old child. This is as (1) except that the thresholds are derived from a reference such as the WGS. You might choose the -5 z-scores and +5 z-scores. If (e.g.) the median MUAC for a 6 month old in the reference is 135 mm with a (negative) SD of 9 mm then the lower threshold will be 135 - (5 * 9) = 90 mm. (3) Censor using thresholds derived from a reference (full approach). Here you calculate MUAC/A or MUAC/H z-scores for every child using a reference and then censor those with z-scores of (e.g.) below -5 and above +5. Again, upper thresholds may be ridiculously high. (4) Trim the distribution. This involves stripping (e.g.) the top and bottom 0.5% of the values or, looked at the other way, keep the middle 99% of values. This approach will, in the absence of low extreme errors censor real cases of SAM. NOT A GOOD METHOD! (5) Censor using thresholds derived from the survey data. This approach uses the survey data, or probably better, the middle 99% of the survey data (as in (4) above). Using the survey data, calculate the mean and SD and use these to create your thresholds. For example, if you have a mean MUAC of 142 mm with and SD of 11 mm then your thresholds would be 142 - (5 * 11) = 87 mm and your upper threshold would be 142 + (5 * 11) = 197 mm. Method (1) is very common. Method (3) is the approach used in EpiInfo for WHZ. Method (5) is similar to the approach used by SMART (someone from SMART might like to confirm this). Sorry not to provide a definitive answer ... I'd go for (1) or (5). You could even use methods (1), (3), and (5) and take a "majority voting" approach ... if two of the three methods say censor then censor. The problem with (3) is that you need to measure age or height which might be OK if you are using both W/H and MUAC as survey indicators. Now to a digression ... why censor? I assume this is about error. Better to avoid error in the first place. You could use a colour-banded strap and record (e.g.) red, yellow, green. This has been shown to reduce error. You could have a survey rule that remeasured all below 120 mm. I hope this is of some use.
|