Surfaced while measuring agriculture-module coverage for GH #603; verified at source, so this is cache-independent.
Measured
Read directly from the raw .dta via get_dataframe() on origin/development:
| wave |
ag file |
households in ag file |
in cover page |
in ag but NOT in cover |
| 2018-19 |
sect3_pp_w4.dta |
2,879 |
6,770 |
120 |
| 2021-22 |
sect3_pp_w5.dta |
2,019 |
4,959 |
40 |
Key is household_id in both files.
Why it matters
The cover page is the household frame — it is where Rural, region, and the sampling variables live. A household present in the agriculture modules but absent from the cover page:
- has no urban/rural classification, so it is invisible to any coverage or representativeness analysis (this is exactly how it was found);
- cannot be joined to
sample, so it gets no weight and no v;
- will be silently dropped by any inner join against the frame, or will arrive carrying NaN for every frame-derived column.
160 households across two waves is small in absolute terms, but they are not a random 160: they are households that farmed and were enumerated for it, yet are missing from the roster of the survey that enumerated them.
Not determined
Whether these are (a) a genuine frame/module mismatch in the released data, (b) households dropped from the cover page during anonymisation — Ethiopias releases do strip identifiers, and the 2018-19/2021-22 waves are the ones with the heaviest editing — or (c) an artifact of which cover-page file the library reads. Deciding needs the questionnaire and the release notes, not more code.
Worth checking whether the same pattern exists in ESS1–ESS3, which were not measured.
Refs #603.
Surfaced while measuring agriculture-module coverage for GH #603; verified at source, so this is cache-independent.
Measured
Read directly from the raw
.dtaviaget_dataframe()onorigin/development:sect3_pp_w4.dtasect3_pp_w5.dtaKey is
household_idin both files.Why it matters
The cover page is the household frame — it is where
Rural, region, and the sampling variables live. A household present in the agriculture modules but absent from the cover page:sample, so it gets no weight and nov;160 households across two waves is small in absolute terms, but they are not a random 160: they are households that farmed and were enumerated for it, yet are missing from the roster of the survey that enumerated them.
Not determined
Whether these are (a) a genuine frame/module mismatch in the released data, (b) households dropped from the cover page during anonymisation — Ethiopias releases do strip identifiers, and the 2018-19/2021-22 waves are the ones with the heaviest editing — or (c) an artifact of which cover-page file the library reads. Deciding needs the questionnaire and the release notes, not more code.
Worth checking whether the same pattern exists in ESS1–ESS3, which were not measured.
Refs #603.