Docs/news prs 72 77

DESCRIPTION

@@ -1,7 +1,7 @@
 Package: LAGO
 Type: Package
 Title: LAGO
-Version: 1.0.12
+Version: 1.1.0
 Author: Ante Bing, Minh Bui, Jingyu Cui
 Maintainer: Ante Bing <abing@bu.edu>
 Description: Calculates LAGO recommended interventions for the next stage.

NEWS.md

@@ -1,4 +1,4 @@
-# LAGO 1.0.12
+# LAGO 1.1.0
 
 * Fixed `outcome_goal_intention = "minimize"` ignoring the outcome goal. The
   minimize direction is implemented by negating the fitted coefficients and
@@ -145,6 +145,94 @@
 * Fixed passing more than one center characteristic, which either failed or
   silently added a single recycled column of values to the confidence set
   instead of one column per characteristic.
+* Fixed the estimated-outcome confidence interval and confidence set being
+  computed on the logit scale for a binomial outcome model fitted with
+  `link = "identity"`. The interval branch applied `expit()` and the logit
+  delta-method factor from the outcome type alone, ignoring the link, so an
+  identity-link binomial fit was reported on the wrong scale, with bounds that
+  could exclude the point estimate and wrongly discard or populate the
+  confidence set. The bounds are now built on the link the model was fitted on.
+  Binary logit results are unchanged. (#74)
+* Fixed the confidence interval bounds for a binary outcome on the logit link
+  being reported outside `[0, 1]`. The point estimate is `expit()` of the
+  linear predictor and is a probability by construction, but the delta-method
+  bounds are symmetric on the probability scale and could fall below 0 or above
+  1, so a set reported a lower bound of -0.106 or an upper bound of 1.049 for a
+  probability. The bounds in `rec_int_ci` and in the confidence set are now
+  confined to `[0, 1]`, while confidence-set membership is still decided from
+  the unconfined interval, so an outcome goal of exactly 1 does not drop a
+  qualifying intervention. (#75)
+* Fixed the numerical optimizer failing with the base-R error "argument is of
+  length zero" when every restart of its maximum-achievable-outcome search
+  failed. Only the cost search guarded against this, so an all-failed outcome
+  search reached `which.max()` over all-`NA` results and died before the
+  package's own message advising the `grid_search` method could appear. Both
+  searches now raise that actionable message when no restart succeeds. (#73)
+* Fixed a rank-deficient outcome model being detected only after a full search
+  over every intervention. When `glm()` returns `NA` for a coefficient the
+  optimization reads (the intercept, an intervention component or interaction
+  term, a fixed center or time effect, or a center characteristic), the fit is
+  now refused up front with an error naming the aliased terms, before any
+  intervention is tried. Where such a fit is still reached, both the numerical
+  and the grid-search paths raise that same error, instead of the numerical one
+  advising `grid_search`, which fails on the same fit, and the grid one failing
+  with "missing value where TRUE/FALSE needed". (#74, #76)
+* Fixed a factor, logical or ordered additional covariate or center
+  characteristic whose name begins with "center" or "period" still being
+  counted as a fixed center or time effect on the path a model passed directly
+  to `get_confidence_set()` can take, where the fitted model's
+  term-to-coefficient mapping is unavailable. The names held back from the
+  fallback name search were column names, but a contrast-coded column's
+  coefficient is named after the level, so `center_grp` never held back
+  `center_grpb`, which was then taken for a center dummy and recycled the center
+  weights into a silently wrong outcome. The held-back names now come from the
+  coding the model recorded, and a covariate named exactly "center" or
+  "period", whose dummies cannot be told from the genuine fixed-effect dummies,
+  raises an error reporting the collision. (#73)
+* `center_weights_for_outcome_goal` must now be numeric, finite and
+  non-negative at both entry points. Only their type, length and sum were
+  checked before, so weights of -10 and 11 summed to 1 and were accepted,
+  reporting an outcome of 10.95 for a binary outcome, and a missing weight
+  surfaced as an opaque "missing value where TRUE/FALSE needed" error. A weight
+  of exactly 0, which excludes that center from the average, is still allowed.
+  The exported `get_confidence_set()`, which does not pass through
+  `validate_inputs()`, previously ran none of these checks and could report a
+  confidence bound above 1 for a binary outcome. (#73, #74)
+* Fixed the estimated outcome being scaled by center weights that did not sum
+  to 1. Weights whose sum was within a thousandth of 1 passed validation and
+  were multiplied into the per-center outcomes as supplied, so a set summing to
+  0.999 scaled every reported outcome, including the goal comparison the
+  recommendation is chosen against. `lago_optimization()` now renormalises the
+  weights it accepts to sum to 1, so the estimated outcome and the
+  recommendation shift for accepted weights that did not already sum to exactly
+  1, and are unchanged for weights that did. The exported `get_confidence_set()`
+  instead refuses weights that do not sum to 1, since it is handed the weights
+  an optimization ran with and renormalising them would move the interval away
+  from the value it was computed at. (#72, #75)
+* `lago_optimization()` now refuses an additional covariate whose column is
+  entirely `NA`, in `validate_inputs()`, with an error naming the offending
+  covariate(s). Such a column made `glm()`'s internal `na.omit` drop every row,
+  so the fit died with an opaque "nonempty numeric vector" error deep in model
+  fitting that never named the covariate. The check catches all-`NA` numeric,
+  factor and character columns, while a partially or fully observed covariate is
+  untouched. (#77)
+* Added a warning when an additional covariate is dropped by `glm()` as
+  collinear. Its coefficient is `NA` but the optimization never reads it, so the
+  run no longer stops or drops it silently: it warns naming the covariate and
+  continues, returning the recommendation the fit without that covariate gives.
+  (#76)
+* Added a warning from `lago_optimization()` when a binary outcome's estimated
+  outcome is reported outside `[0, 1]`, which happens when an identity-link
+  linear probability model is extrapolated to an intervention beyond the range
+  its components were fitted over. The estimate is not clamped, because it
+  drives the optimizer and the recommendation, so the warning names the
+  extrapolation and reports how many interval bounds are affected while every
+  returned value is unchanged. (#75)
+* Added a warning when a numeric additional covariate whose observed range
+  excludes 0 is held at 0 to compute the confidence set, naming each such
+  covariate and its observed range, since the reported outcome and interval are
+  then an extrapolation to a covariate value that never occurs in the data. The
+  covariate is still held at 0, so no returned value changes. (#75)
 * Added runnable `@examples` to every exported function that lacked them:
   `get_confidence_set()` and the `print()`, `summary()`, and `plot()` methods
   for `"lago"` objects. Every exported function now ships an example.