occu_choice_pf_2 0.6.1__tar.gz

occu_choice_pf_2-0.6.1/.gitignore

@@ -0,0 +1,19 @@
+.venv/
+__pycache__/
+*.pyc
+.pytest_cache/
+.ipynb_checkpoints/
+dist/
+build/
+*.egg-info/
+
+# example logs; the tables in examples/*/output are committed
+*.log
+
+# fake data drawn by simulation/<module>.py
+src/occu_choice_pf/simulation/data/
+
+# LaTeX build of tex/examples.tex
+tex/out_dir/
+
+misc/

occu_choice_pf_2-0.6.1/CHANGELOG.md

@@ -0,0 +1,437 @@
+# Changelog
+
+The server pins an exact version (`pip install occu-choice-pf==<version>`),
+so a new release changes nothing there until the pin does. Every change to
+how a function is called is listed under **Update a server script**, with the
+old and the new call.
+
+## 0.6.1 (2026-10-08)
+
+`underestimation_gap.gap_table`'s `share_column` also takes a list of both
+share columns, `['underestimator', 'zero_gap']`. The package is now also
+published as `occu_choice_pf_2`, a third name for the same code
+(`.github/workflows/publish_2.yml`), so that three versions can be installed
+on the server within 24 hours. No call changes.
+
+### Update a server script
+
+- `pip install occu-choice-pf==0.6.1` (or `pip install occu_choice_pf_1==0.6.1`,
+  or `pip install occu_choice_pf_2==0.6.1`).
+- Calls of 0.6.0 are unchanged, and give the same tables.
+
+### Added
+
+- `underestimation_gap.gap_table(..., share_column=['underestimator', 'zero_gap'])`:
+  both share columns, 7 and 8 in the order of the list, each with its
+  standard errors and p-values; the counts follow in 9 and 10.
+
+## 0.6.0 (2026-10-08)
+
+`underestimation_gap.gap_table` takes `share_column` for column 7:
+`'underestimator'` (the default), `'zero_gap'`, the new share of job pairs
+with no perceived wage gap (`d_Vij == 0`), or `None`. It replaces
+`underestimator`, a change of call.
+
+### Update a server script
+
+- `pip install occu-choice-pf==0.6.0` (or `pip install occu_choice_pf_1==0.6.0`).
+- `gap_table`'s `underestimator` is gone; `share_column` says what column 7
+  shows, its column and header those of `labels`:
+  - old: `ug.gap_table(rows, underestimator=None)`
+    new: `ug.gap_table(rows, share_column=None)`
+  - old: `ug.gap_table(rows, underestimator={'u': 'Header'})`
+    new: `ug.gap_table({label: df.with_columns(d_v_vi_max_notmaxV_rate=c.u) for label, df in rows.items()})`;
+    the header is `labels.UNDERESTIMATOR`'s
+  - a call without `underestimator` is unchanged, and gives the same table.
+- `run_tests` checks `d_Vij` too, the perceived wage gap in levels: a
+  number, exactly 0 for equal perceived wages, minus itself in the other
+  order of the pair, and a rate `d_Vj_Vij_rate` of 1 where it is 0.
+
+### Added
+
+- `underestimation_gap.gap_table(..., share_column='zero_gap')`: column 7
+  is the share of job pairs whose perceived wage gap is 0 (`d_Vij == 0`),
+  among those where it is not missing, in place of the share of
+  underestimators; the header is `labels.ZERO_GAP`, with
+  $\Delta \widetilde{V_{i,j,j'}}=0$. Pairs with `d_Vij` missing are logged.
+  `share_column='underestimator'` (the default) is the share of
+  underestimators, and `share_column=None` no column 7.
+- The simulated frame has `d_Vij`, and 10% of persons perceive two job
+  types as paying the same; the example tables are rerun on it, and
+  `tab_estimation_gaps_zero` (in `tex/examples.tex`) shows the new column.
+
+### Changed
+
+- `gap_table(..., underestimator={column: header} | None)` is now
+  `share_column='underestimator' | 'zero_gap' | None`; another column goes
+  in under the default name, as for the other columns.
+
+## 0.5.2 (2026-10-08)
+
+`underestimation_gap.gap_table` compares rows that are not next to each
+other, a row with several others; `p_at_end=True` puts the p-values in the
+last rows of the table, and `underestimator=None` leaves out the column of
+the share of underestimators. The row numbers of a table of ten rows or
+more are right-aligned, so that the labels line up.
+
+### Update a server script
+
+- `pip install occu-choice-pf==0.5.2` (or `pip install occu_choice_pf_1==0.5.2`).
+- Calls of 0.5.1 are unchanged, and give the same tables but for the row
+  numbers of a table of ten rows or more (see Changed).
+
+### Added
+
+- `underestimation_gap.gap_table`: `compare` takes rows anywhere in `rows`,
+  a row in any number of pairs (they had to be next to each other, a row in
+  one pair at most). The rows from the first to the last of a pair are a
+  block, joined with the blocks it overlaps; the p-values of a block go
+  under it, in the order of `compare`, followed by a line. Comparing a row
+  with itself, or two rows twice, raises.
+- `gap_table(..., p_at_end=True)`: all the p-values in the last rows of the
+  table, under a line, in the order of `compare`.
+- `gap_table(..., underestimator=None)`: the table without column 7, the
+  share of underestimators, and its standard error and p-value; the frames
+  need not have the column.
+- The example `tab_estimation_gaps_gender` compares men with women and with
+  women reweighted to men, rows (1)-(2) and (1)-(3), and
+  `tab_estimation_gaps_gender_end` is the same with `p_at_end=True`; both are
+  in `tex/examples.tex`.
+
+### Changed
+
+- `underestimation_gap.gap_table` right-aligns the row numbers in a table
+  of ten rows or more, (1) to (9) padded by `\phantom{0}`, so that the row
+  labels line up with that of (10). Calls are unchanged; the example tables
+  are rerun.
+
+## 0.5.1 (2026-10-07)
+
+A new module, `underestimation_gap`, with Table 3 of the paper, the
+distribution of wage gap misperceptions by subsample:
+`gap_table(rows)` takes one frame of the extra-long data per table row.
+`bootstrap=B` adds standard errors from B draws of persons;
+`compare=[(label, label)]` puts the p-value of the difference of two rows,
+next to each other, under them, followed by a line; `ps_group` reweights
+one group to the other by a propensity score model estimated anew in each
+draw, as Table C1 of the parental transmission paper. The rows are
+numbered, and a p-value row names the rows it compares.
+
+### Update a server script
+
+- `pip install occu-choice-pf==0.5.1` (or `pip install occu_choice_pf_1==0.5.1`).
+- Calls of 0.5.0 are unchanged.
+- `run_tests` takes the extra-long frame too, and skips the data tests of
+  a frame it is not given: `occu_choice_pf.run_tests({'long': long, 'elong': elong})`.
+
+### Added
+
+- `underestimation_gap.gap_table(rows, weights=None, bootstrap=0, compare=(), ps_group=None, ...)`:
+  Table 3 of the paper, its cells as in the author's script
+  `underestimation_gap.py`. `rows` is `{label: frame}`, each frame the
+  extra-long frame (`pnr` x `type_sizerank_job` x `type_sizerank_jobj`) or
+  a subsample of it made with polars; `weights={label: column}` weights a
+  row. Labels and headers break lines with `\n`. The columns are
+  `labels.GAP_RATE`, `GAP_LOG` and `UNDERESTIMATOR` (`{column: header}`).
+  Two departures from the script: persons whose underestimator value is
+  missing are left out of that share, with a warning in the log (the script
+  counted them as not underestimators); and the job pairs are unique
+  (person, {j, j'}), so (j, j') and (j', j) are one pair (the script
+  counted unique (person, j)). A row's frame with NaN raises, as a polars
+  filter such as `c.d_vj > 0.05` keeps NaN: make it missing in prep with
+  `elong.with_columns(pl.col(pl.Float64).fill_nan(None))`.
+- `row_numbers=True` numbers the rows, (1), (2), ..., and a p-value row
+  reads `$p$-value of difference (1)-(2)`; `row_numbers=False` leaves both
+  out, as Table 3 in the paper.
+- `bootstrap=B`: standard errors under each row, from B draws of persons
+  with replacement, the same draws for every row, drawn with `seed`; the estimate rows are
+  pulled up to them with `\vspace{-.4em}`, as in the paper's regression
+  tables.
+- `compare=[(label, label), ...]`: the bootstrap p-value of no difference
+  between two rows next to each other, in a row under them (twice the
+  smaller share of draws on either side of 0, as the script for Table C1),
+  then an `\hline`, except under the last row.
+- `ps_group`, `ps_xs`, `ps_fes`, `ps_limit`: adds the weights `ps_weight`
+  (each group reweighted to the other) and `ps_support` (1 for persons with
+  a weight) from a logit of the 0/1 `ps_group` on `ps_xs` with fixed effects
+  `ps_fes` (default `disced_6d`), estimated anew in each draw. As
+  `compute_weights` in the author's `p_scores_funcs`, with pyfixest's
+  `feglm` in place of fixest in R; with one fixed effect and no `ps_xs` the
+  model is saturated and the score is the group's share in the cell,
+  computed without pyfixest. `underestimation_gap.ps_weights` returns the
+  weights themselves.
+- `labels.JOBJ`, `GAP_RATE`, `GAP_LOG`, `UNDERESTIMATOR`, `P_LABEL`,
+  `PS_FES`, `PS_LIMIT`.
+- `simulation/underestimation_gap.py` (fake extra-long frame), the example
+  `examples/underestimation_gap`, tests and data tests
+  (`tests/test_underestimation_gap_data.py`: no NaN, one row per pair, each
+  pair in both orders, a person-level underestimator and major, finite gaps).
+
+### Changed
+
+- The data tests of a frame not given to `run_tests` are skipped, rather
+  than failing to load it.
+- `tex/examples.tex` sets the rows of a tabular 1.25 apart, as the paper.
+
+## 0.5.0 (2026-10-05)
+
+A term of `cols_control` wrapped in `reg.plain()` enters its column as it
+is, not times `x`, so a fixed effect can go in one column only. The labels
+of the education dummies say "Major".
+
+### Update a server script
+
+- `pip install occu-choice-pf==0.5.0` (or `pip install occu_choice_pf_1==0.5.0`).
+- Calls of 0.4.0 are unchanged. The check-mark rows of `labels.EDUC_FIELD`
+  and `labels.DISCED` change (see Changed).
+
+### Added
+
+- `regressions.plain(*terms)`: terms of `cols_control` that enter their
+  table column as they are, not times `x`. An `i.<column>` among them is a
+  fixed effect absorbed in that column only, e.g. a job-type fixed effect
+  in one column: `cols_control=[{}, {'Job type FE': reg.plain('i.' + JOB)}]`.
+  A list of terms may mix `plain()` with terms times `x`.
+- A table column with a fixed effect twice raises.
+- The example fits a second column with the job-type fixed effect from
+  `reg.plain()`.
+
+### Changed
+
+- The labels of `labels.EDUC_FIELD` and `labels.DISCED` are
+  `Major field × ...` and `Major × ...` (were `Education field × ...` and
+  `Education (DISCED) × ...`), and so are their check-mark rows in the
+  default table. The example tables are rerun with them.
+
+## 0.4.0 (2026-10-01)
+
+The table is `reg_control_table`, and the caller lists its columns:
+`cols_control` takes the controls of each table column, and
+`common_control` those of every column, in place of `base`, `powers`,
+`controls`, `dummies` and `fe`. A term may be `i.<column>` (dummies, or in
+`common_control` a fixed effect absorbed by pyfixest) or a polars
+expression. `verbose=True` logs every coefficient.
+
+### Update a server script
+
+- `pip install occu-choice-pf==0.4.0` (or `pip install occu_choice_pf_1==0.4.0`).
+- `regressions.regression_table` is `regressions.reg_control_table`, and
+  `regressions.estimate` is `regressions.reg_control_fits`. With the default
+  columns the tables are the same, except where `educ_field` or `disced_6d`
+  is null (see Changed):
+
+  ```python
+  # 0.3.0
+  save_tex(reg.regression_table(long, method='ols'), out / 'tab_ols.tex')
+  fits = reg.estimate(long)
+  # 0.4.0
+  save_tex(reg.reg_control_table(long, method='ols'), out / 'tab_ols.tex')
+  fits = reg.reg_control_fits(long)
+  ```
+
+- `powers`, `controls` and `dummies` become `cols_control`, a list with the
+  controls of each table column, every term times `vj_vij_job`. A power of
+  `vj_vij_job` is `vj_vij_job` to one less times `vj_vij_job`, and a set of
+  dummies is `i.<column>`:
+
+  ```python
+  # 0.3.0
+  reg.regression_table(long, powers={'Squared': [2]},
+                       controls={'Age, gender': ['age', 'female']},
+                       dummies={'DISCED': 'disced_6d'})
+  # 0.4.0
+  sq, own = {'Squared': [c.vj_vij_job]}, {'Age, gender': ['age', 'female']}
+  reg.reg_control_table(long, cols_control=[{}, sq, {**sq, **own},
+                                            {**sq, **own, 'DISCED': ['i.disced_6d']}])
+  ```
+
+- `base` and `fe` become `common_control`, `{label: terms}`, where
+  `i.<column>` is a fixed effect. To report `vj_vij_job` rather than check it:
+
+  ```python
+  # 0.3.0
+  reg.regression_table(long, coefs={**COEFS, **BASE}, base={})
+  # 0.4.0
+  reg.reg_control_table(long, coefs={**COEFS, 'vj_vij_job': '$' + MIS[1:-1] + '$'},
+                        common_control={'Person FE': ['i.pnr']})
+  ```
+
+- In `labels`, `BASE`, `FE`, `CONTROLS` and `DUMMIES` are gone, and `POWERS`
+  holds the terms `[c(X), c(X) ** 2]` rather than the powers `[2, 3]`.
+  `COMMON_CONTROL`, `COLS_CONTROL` and the blocks `PRE_CONTROLS`,
+  `EDUC_FIELD` and `DISCED` are new.
+
+### Added
+
+- `cols_control`: one table column per entry, `{label: terms}`, each term
+  times `x`; `{}` is a column with the common controls only. A label's
+  check-mark row is checked in each column that has it.
+- `common_control`: the controls of every column, as they are. Several
+  fixed effects (`i.<column>`) are absorbed together, e.g. a job-type fixed
+  effect next to the person's; with none, the regression has a constant.
+  The check for separated rows takes them all.
+- A term is a column, `i.<column>` (as Stata's `i.`) or a polars expression,
+  named by its `.alias()` or else written out (`c.vj_vij_job ** 2` is
+  `vj_vij_job_pow_2`). In `cols_control`, `i.age` is Stata's
+  `i.age#c.vj_vij_job`; the column may be a string.
+- In `i.<column>`, as dummies or fixed effects, null and NaN are one value
+  of their own, `missing`, with its own dummy or group, so their rows stay.
+- `person`: the column counted in the Individuals row and held to `min_n`
+  (was the fixed effect's).
+- `verbose=True` in `reg_control_table` and `reg_control_fits` logs every
+  coefficient of every column with its standard error and p-value, and
+  names the terms dropped as collinear. These are not held to `min_n`.
+- The example uses `i.age` for active job seekers, `verbose=True` in the
+  quadratic table, and a job-type fixed effect next to the person's.
+- Data test: the pre-treatment controls have no missing values (null or
+  NaN), so the columns of a panel rest on the same rows but for the Poisson
+  drops.
+- Data test: pyfixest's estimation sample (`fit._data`), whose persons the
+  Individuals row and `min_n` count, has the fit's *N* rows; it replaces
+  the check that a fit counts no more rows or persons than the frame. The
+  Individuals row is also checked against a count made in polars.
+
+### Changed
+
+- A table with one outcome writes its title over all the columns, above
+  the column numbers, as Table A8 of the paper (it was a panel row below
+  them). Several outcomes keep their lettered panels.
+- A null in a column made into dummies keeps its row, with the dummy of
+  the value `missing`, which NaN joins (in 0.3.0 a null dropped the row
+  from that table column, and NaN was a value of its own). In the default
+  table this changes columns (4) and (5) where `educ_field` or `disced_6d`
+  is null; to drop those rows as before, filter them out before the call.
+- The dummies leave out the numerically lowest value of an integer column
+  (they left out the lowest as a string: 10 rather than 9). The other
+  coefficients are unchanged, as are the default tables, whose dummy
+  columns are strings.
+- `labels.PRETREATMENT` uses the server's names of the expected wages:
+  `Vij_1k_1`, `Vij_2k_2`, `Vij_3k_3`, `Vij5_1k_1`, `Vij5_2k_2` and
+  `Vij5_3k_3` (were `Vij_1` … `Vij5_3`).
+- The check for separated rows differences out the fixed effect with the
+  most groups, whatever the order of `common_control` (it took the first,
+  and made a person fixed effect listed second into a dummy per person).
+  The rows found are the same.
+- `regressions.METHODS` is `{method: pyfixest function}` (was a tuple of
+  the names).
+- The simulated `vj_vij_job` is NaN in half of its missing rows and null in
+  the other half (was null in all), so the tests meet both. The example
+  tables are unchanged.
+- The README leaves the arguments, what drops out of a column and the input
+  checks to the docstrings, which now list the checks.
+
+### Fixed
+
+- In `cols_control`, an `i.<column>` with two values whose dummies get one
+  name, such as `'a b'` and `'a-b'` (only letters, digits and _ are kept),
+  raises. Before, one of the two silently joined the reference.
+- The data test of `T_vj_vij_job` counts NaN as missing, as the regressions
+  do; it failed on a frame with NaN rather than null in `vj_vij_job`.
+- The data tests find the education columns of `labels.COLS_CONTROL` when
+  a block holds a single term rather than a list; before, they would have
+  dropped them from the one-value-per-person test without a failure.
+- The notes in `tex/examples.tex` say "parentheses".
+
+## 0.3.0 (2026-09-30)
+
+The job-level columns take their names in the server's long frame, where
+every variable of the job type has the suffix `_job`, and the last column
+has the DISCED dummies instead of both those and the field dummies.
+
+### Update a server script
+
+- `pip install occu-choice-pf==0.3.0` (or `pip install occu_choice_pf_1==0.3.0`).
+- Nothing else changes for a call with the defaults. Column (5) of the
+  default table has the DISCED dummies without the field dummies; since
+  DISCED codes lie within fields, its estimates are the same.
+- The defaults now read `vj_vij_job`, `T_vj_vij_job` and `type_sizerank_job`
+  (were `vj_vij`, `T_vj_vij` and `type_sizerank`, which the long frame does
+  not have). A call with the defaults needs no change. A call that names
+  them passes the new names, e.g. `coefs={'T_vj_vij_job': ...}`,
+  `x='vj_vij_job'`.
+
+### Changed
+
+- `labels.X`, `labels.COEFS`, `labels.BASE` and `labels.JOB` use the `_job`
+  names, and so do the simulated long frame, the example and the tests.
+- The fixed effect is labelled `Person FE` (was `Person fixed effects`), and
+  the notes in `tex/examples.tex` say FE.
+- Each set of `dummies` is an alternative rather than added up: column (5)
+  has the DISCED dummies and no longer the field dummies too (DISCED codes
+  lie within fields), and each dummies row is checked in its own column
+  only. Powers and controls still add up column by column.
+
+## 0.2.0 (2026-09-30)
+
+The tables follow the layout of Table 5 of the paper, with `vj_vij` checked
+under a `Controls:` row rather than reported. The default outcomes take the
+server's column names, and the functions take only a polars DataFrame.
+
+### Update a server script
+
+- `pip install occu-choice-pf==0.2.0` (or `pip install occu_choice_pf_1==0.2.0`).
+- The tables no longer report the coefficient on `vj_vij`: it is in every
+  column through the new argument `base` and checked under a `Controls:`
+  row, as in Table 5 of the paper. To report it again:
+
+  ```python
+  from occu_choice_pf.labels import BASE, COEFS
+  reg.regression_table(long, coefs={**COEFS, **BASE}, base={})
+  ```
+
+- A call without `outcomes` now reads `spell_1_new_job` and `appl_job` (was
+  the placeholders `job_j` and `appl_j`), so it can drop the argument:
+
+  ```python
+  # 0.1.0
+  reg.regression_table(long, {'spell_1_new_job': 'Found a job in the job type',
+                              'appl_job': 'Number of applications to the job type'})
+  # 0.2.0
+  reg.regression_table(long)
+  ```
+
+### Changed
+
+- The check-mark block follows Table 5 of the paper: a `Controls:` row, then
+  one indented row per control (`base`, the fixed effect, each step). The
+  default labels of `POWERS`, `CONTROLS` and `DUMMIES` fit on one line.
+- `labels.COEFS` is `T_vj_vij` only; `labels.BASE` is `vj_vij`.
+- `labels.OUTCOMES` is `{'spell_1_new_job': ..., 'appl_job': ...}`, and the
+  simulated long frame has these columns.
+- The active-job-seeker example controls for studying rather than for active
+  search, which is constant in that sample.
+- `estimate` and `regression_table` take only a polars DataFrame and raise a
+  TypeError on anything else. The data work, the check for separated rows
+  included, is done in polars; pyfixest gets a pandas copy of each
+  regression's columns. The estimates are unchanged.
+
+### Added
+
+- `tex/examples.tex`: the example tables with notes, in the paper's
+  `\showtable` layout (`cd tex && latexmk -pdf -outdir=out_dir examples.tex`).
+
+## 0.1.0 (2026-09-30)
+
+First version.
+
+### Update a server script
+
+- `pip install occu-choice-pf==0.1.0` (or `pip install occu_choice_pf_1==0.1.0`).
+  pyfixest is new on the server and installs with the package, with pandas,
+  scipy and formulaic.
+
+### Added
+
+- `regressions.regression_table(df, ...)`: Poisson (`method='poisson'`,
+  pyfixest.fepois) or OLS (`method='ols'`, pyfixest.feols) regressions of
+  each outcome on `T_vj_vij` and `vj_vij` with a person fixed effect, one
+  panel per outcome, with `vj_vij` squared and cubed, the pre-treatment
+  controls times `vj_vij`, and education field and DISCED dummies times
+  `vj_vij` added column by column. Standard errors clustered on the person.
+- `regressions.estimate(df, ...)`: the pyfixest fits behind a table.
+- Separated rows are found with an exact linear program and dropped before
+  the Poisson regression.
+- `labels`: `COEFS`, `OUTCOMES` (placeholder names), `PRETREATMENT`,
+  `POWERS`, `CONTROLS`, `DUMMIES`, `FE`.
+- `run_tests(frames)`, `setup_log`, `save_tex`, `example_root`, as in
+  ad-data-pf.

occu_choice_pf_2-0.6.1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Asker Nygaard David
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.