diff --git a/assets/css/stack.css b/assets/css/stack.css index 4c13b8b..d44bd78 100644 --- a/assets/css/stack.css +++ b/assets/css/stack.css @@ -224,8 +224,14 @@ a { color: inherit; text-decoration: none; } .prose hr { border: none; border-top: 2px solid var(--black); margin: 2rem 0; } /* Descriptions in the compact source list. A visible line beats a hover title, - which no touch device shows and no screen reader announces reliably. */ + which no touch device shows and no screen reader announces reliably. + `.row span` above sets white-space: nowrap for the short language tag. A + description is also a span, so it inherited that and ran off the side of the + phone instead of wrapping. Scope the nowrap to the tag. */ .row { flex-direction: column; align-items: flex-start; gap: .25rem; } -.row .d { font-size: .8rem; opacity: .78; line-height: 1.35; } +.row .d { font-size: .8rem; opacity: .78; line-height: 1.35; + white-space: normal; overflow-wrap: anywhere; } .row .t { display: flex; justify-content: space-between; align-items: baseline; - gap: .6rem; width: 100%; } + gap: .6rem; width: 100%; white-space: normal; } +.row .t b { overflow-wrap: anywhere; } +.row .t span { white-space: nowrap; flex: none; } diff --git a/index.html b/index.html index c368acc..417e30e 100644 --- a/index.html +++ b/index.html @@ -26,9 +26,9 @@

EquityStack

-

Python for development sector data workflows: cleaning, exploration, modelling, - impact evaluation, and design-based estimates from complex surveys. Part of - OpenStacks.

+

The Python you reach for on a development research project: getting messy data into + shape, asking whether a programme worked, and getting the confidence intervals right + when the survey is clustered. Part of OpenStacks.

9Modules
34Python files
@@ -48,27 +48,26 @@

Survey estimation

survey_estimation/

Design-based standard errors

-

Weights are the easy half. A national household survey is stratified and clustered, so an - independence standard error is too small, often by a factor of two, and every interval - built on it errs in the direction that flatters the result. This implements Taylor - linearisation, the estimator behind Stata's svy: and R's survey.

+

Weighting is the half everyone remembers. A national household survey is also clustered, + so an interval that ignores that is too narrow, often by half, and it errs in the + direction that flatters your result. This does it properly, the same way Stata's + svy: and R's survey do.

Read the module →

Worked example

NFHS-5 stunting, reproduced

-

Rebuilds India's published stunting table by wealth quintile from the raw DHS children's - recode: 35.5 percent nationally, 46.1 in the poorest quintile falling to 22.9 in the - richest. Reproducing a published table is the only cheap proof a pipeline is right end - to end.

+

Rebuilds India's published stunting table from the raw NFHS-5 files: 35.5 percent + nationally, 46.1 in the poorest fifth falling to 22.9 in the richest. If your numbers + match the published ones, your pipeline is sound. If they do not, you know before you + publish rather than after.

Read the script →
-

Checked against R's survey package on the same data, agreeing to - twelve significant figures. Three behaviours that look like details and are not: subgroups - are domains rather than subsets, proportions get a logit interval so a small one cannot go - negative, and intervals use t on clusters minus strata and say which distribution produced - them.

+

Checked against R's survey package on the same data, to twelve significant + figures. Three things it gets right that a hand-rolled version usually does not: a subgroup + keeps the clusters it does not appear in, a small percentage never gets a negative lower + bound, and the interval widens correctly when you have few clusters.