Univariate analyses with incomplete data set I'm dealing with a fairly large ecological data set (field data, not collected by me) of approximately 60 attributes measured for about 400 individual trees. Very few trees (~20%) have complete data (have a value for every attribute), the rest are missing one or more data points, due to field-research limitations and error. 
I'm working on univariate analyses--basically just generating a bunch of summary statistics--and I'm not sure what is the most responsible approach to handling the missing values?
The conservative approach would be to restrict my analyses to the 20% of individuals with complete data, but that looses a lot of critical information--the data is batched by field sites, and only a few sites have complete information, so I'd have to throw out a bunch of sites, which would kill the overall research objectives. 
Alternatively, I could just use all the data, and use a somewhat different set of individuals for each attribute--but my concern with that approach is that it makes comparison of different attributes invalid, because the data isn't paired. This is important, because some of the data is time-series (values for the same attribute in multiple years) and I want to compare different years to each other. For example, I want to look at how "attribute A" changes from year to year, and eventually I intend to correlate that change with other data, such as local temperature and rainfall.
I'm leaning toward the laborious hybrid-approach of using a different data sub-set for each comparison--for example when I look at "attribute A" over multiple years, I'll include only trees that were measured in all relevant sample periods, and exclude any with missing data. But if I want to do the same thing with "attribute B", I'll use a different set of trees, which is complete for that attribute. Before I dig too far into that approach though, I would love any advice folks have about how to handle this issue. I've tried poking around this forum and some text books, but I'm not finding any good answers.
 A: Your hybrid approach sounds a little bit odd to me, especially if you consider analysing time series data. This could lead to situations where you are comparing three points in time, which have no cases in common. 
Example: 


*

*t0 case1 case2 case3 case4 (rest is missing)

*t1 case3 case4 case5 case6 (rest is missing)

*t1 case5 case6 case7 case8 (rest is missing)


t0 & t1 = compare case3 and case4
t1 & t2 = compare case5 and case6
One possible way to deal with missing values without loosing cases, would be using multiple imputation. In multiple imputation you substitute each missing value by several plausible values based on a pre-specified regression model for each variable. 
An alternative would be estimating the model with the full information maximum likelihood approach, which is used in many structural equation modeling software packages. 
For univariate statictics both approaches could be a bit over the top, but I would give it a chance. 
Software: 


*

*http://www.jstatsoft.org/v45/i03/paper

*http://sites.stat.psu.edu/~jls/misoftwa.html
Sources:


*

*Graham, John W. "Missing data analysis: Making it work in the real world." Annual review of psychology 60 (2009): 549-576.

*Schafer, Joseph L., and John W. Graham. "Missing data: our view of the state of the art." Psychological methods 7.2 (2002): 147.
