I have tended to use demi-span to estimate height in older adults before using the estimated height to calculate BMI. You could use knee-height or arm-span. Whichever measurement you use, you will suffer from problems with estimation error because the estimate is squared when calculating BMI and this has the effect of magnifying the error in estimating height.
There is a problem with using "standard formulae" for turning arm-spans (or whatever you pick) into heights. This is because the relationship between limb length (or whatever you use) and height varies between populations ... a formula that works well in one setting may not work as well in another setting. I usually develop context-specific formulae using data from a small cross-sectional survey of people whose height and arm-span can both be measured. With this data you can perform a linear regression:
height = constant + B1 * arm-spans
and use the results to find height from arm-span. You may want to include age (probably in decades) in the model and fit models for males and females separately.
Sample sizes do not need be very large. I tend to use a corrected "subject to variable ratio" approach to sample size for this application. There are a number of options for this. I prefer that presented in "Thorndike RM, Correlation procedures for research, Gardner, New York, 1978 (p 184)" which is:
N = 10p + 50
Where "N" is the minimum required sample size and "p" is the number of predictors. In the case of a model such as:
height = constant + armspan + sex
you have two predictors (arm-span and age) so the minimum sample size is:
N = 10 * 2 + 50 = 70 + 50 = 70
This is your absolute minimum sample size. If you can get and afford to measure more than 70 subjects then measure as many as you can. A sample size of < 100 will probably be considered as "small". An alternative rule from "Nunnally JC, Psychometric theory (2nd Ed.), McGraw-Hill, New York, 1978" is:
N = 40p
which, with two predictors, give:
N = 40 * 2 = 80
These sample sizes are for finding estimating equations for men only or women only.
The Nunnally (1978) rule is considered safer than the Thorndike (1978) rule if you use stepwise techniques to build a model (see "Tabachnick BG, Fiddel LS, Using multivariate statistics (2nd Ed.), Harper Collins, New York, 1989 (p. 129)"). As usual, large sample sizes are better than smaller sample sizes.
Any sample size > 100 will not be considered "small" and may be easier to justify when it comes to publication.
BTW : Work done in in elderly adults in Africa and Asia found that CAMA, AMA, calf-muscle area, and MUAC were better predictors of function (as measured over several dimensions of function as well as by a validated "activities of daily life" score) than BMI. You may consider, therefore, using MUAC rather than BMI. I am not sure of the value of BMI in care of Ebola cases.
I hope this is of some use.