cm cutoffs for outliers of MUAC measurements?

I think you could use a method like that used in SMART/ENA which uses the collected data to define outliers. SMART uses the mean and standard deviation of anthropometric indices calculated from collected data to decide which observations are more likely to be outliers than true values. In this scheme, we exclude (flag) that lie outside of +/- 3 z-scores from the mean z-score so that values of (e.g.) WHZ that are below:

    mean(WHZ) - 3

and:

    means(WHZ) + 3

are flagged as as implausible values.

I think this approach may be problematic as it assumes that SD(WHZ) = 1 and it assumes the distribution of WHZ is normal. Work done by members of the SMART team shows that these assumptions are reasonable with WHZ usually normally distributed (with only slight deviations from normality) and SD(WHZ) tending to be limited to between 0.8 and 1.2 z-scores and centered around 1 z-score.

This does not, however, help us much with your problem as we tend to use raw measurment values rather than z-scores with MUAC. An analogous approach woud be to flag MUAC values that lie outside of mean(MUAC) +/- 3 * SD(MUAC). The problem here is that both the mean and the SD are strongly influenced by outlier values. It might be better, therefore, to use more robust summary measures such as the median (as an alternative to using the mean) and the median absolute deviation (multiplied by 1.4826) as an alternative to using the SD).

A common alternative is to use the method used with boxplots. The "box" on boxplots ranges between the upper and lower quartiles of the data.
The “whiskers” on boxplots usually extend to 1.5 times the interquartile range from the ends of the box (i.e. the lower and upper quartiles). This is known as the "inner fence".  Data points that are outside the inner fence are considered to be mild outliers. The "1.5" can be increased to 3.0. This defines the "outer fence" which can be used to identify extreme outliers. I tend to think of mild outliers as possible outliers and severe outliers as probable outliers.

See these sections:

Box Plot 

What are outliers in the data? 

Detection of Outliers

of the NIST/SEMATECH Handbook of Statistical Methods. Links on theses pages lead to tests that can help identify outliers.

You could use the SMART method with MUAC/A z-scores.

I hope this is of some use.

Hi Emily,

No argument with the previous posts on data cleaning your MUAC measurements. From a practical point of view you need to consider the context too. In extemely malnourished children (e.g. in some famine situations) with very low MUAC and very little body fat it becomes very difficult to measure MUAC without folding the skin, thus affecting accuracy.

In a normal context, accurately measured MUACs in children 6-59 months old < 9cm should be very infrequent (and tend toward poor outcomes). Accurately measured MUACs less than 8cm are rare. Most examples of MUACs between 7-8cm that I have personally verified in the field were false readings (MUAC tape too tight). I have measured a child with a MUAC of about 7.5cm+ in an emergency situation but it was difficult to get an accurate reading because of the skin folding. 

If this was survey data and the enumerators were standardised take a look at the data for the standardisation exercise and see if the accuracy was within the stated parameters. If this is raw field data you might consider the context and the reliability of the source of measurement and make a judgement call. 

Hi Emily

In our experience cutoffs are more straightforward for surveys of a relatively normal population. For an expected abnormal population  we use the same method suggested by Mark Myatt of calculating standard deviation (sd) within our sample population to look for outliers. For MUAC there is some change with age so the ideal is to use MUACz scorea and calculate the mean and sd for those values. For undernourished infants and children many of the low values are true on rechecking so we generally apply a more liberal cutoff on the low side e.g. -5. We apply the same method to changes between two time points in longditudinal studies.

Please let me know if you need any further information.

Cheers

Jay

I think Jay makes a good point. The flagging rules for SMART and the WHO flagging rules used in DHS / MICS are suited to survey data. Thay are not suited to data from (e.g.) CMAM program admissions as these will likely include a great many extreme low values ... even below WHO biologically plausible values ... as we are treating children with very high risk of near term death.