A collection of what I've learned about the LTZ1000, mostly from the epic thread on the EEVBlog forum: Ultra Precision Reference LTZ1000.
From the datasheet:
Here's the same data, but stated as "relative impact":
That is, if R1 changes in value, 1% of that change will show up on the output voltage (if R1 changes by 100ppm, the output voltage will change by 1ppm).
However, members of the community have measured different results:
Here's the same data, but expressed as the importance of each resistor relative to the importance of the R4/R5 ratio (i.e. the data is normalized against R4/R5). E.g. according to zlymex, drift in R1 is only 15% as important as the drift in R4/R5.
So, spend your money on R4/R5 first, then R2, then R1, then R3.
References:
Note: the above data does not pertain to circuits which feature the "R60" resistor.
A circuit from a CERN presentation: