The words pyfixest uses for its econometric and estimation concepts in code,
docstrings, errors, and documentation. Terms are written out in prose; fixest
abbreviations such as vcov or fixef are API spellings and appear under
Avoid as "(in prose)".
Estimation function:
A public function that fits one or more models from a formula and data, such
as feols(), fepois(), feglm(), quantreg(), or event_study(). One
estimation function can fit several estimators.
Avoid: method, estimation method, estimator (for the function)
Estimator: The statistical procedure that defines a model's estimates, such as OLS, 2SLS, Poisson, probit, quantile regression, TWFE, or DID2S. Avoid: method, model type
Solver: A numerical procedure that computes an estimator's solution, such as a linear solve, the Frisch-Newton interior-point method, or a demeaner. Changing the solver moves the estimates only within its tolerance. Avoid: method, algorithm (as a separate concept)
IWLS: Iteratively reweighted least squares, the solver that fits a GLM through repeated weighted least-squares fits. Avoid: IRLS, iterated weighted least squares
Model: A model formula and an estimator applied to one estimation sample: the thing being estimated, whether or not it has been fitted. Its inference is not part of it. Avoid: specification, regression (as synonyms)
Fitted model: The object an estimation function returns for one model; "fit" is its short form. Avoid: model object, fitted result, estimation result, Fixest object, model (for the object)
Multiple estimation: One call to an estimation function that fits several models, through stepwise operators, several dependent variables, a sample split, or several quantiles. It is shorthand for the single fits: each model equals the one fitted on its own. Avoid: stepwise regression, multi-model fit
Estimation options: The settings a fitted model was estimated with, other than its formula and data. Avoid: settings, config, arguments
Retention:
What a fitted model keeps from its estimation, such as its data, design
matrix, or residuals, as chosen with store_data and lean. An operation
that needs removed state fails rather than recomputing it.
Avoid: storage, caching, lean mode (for the concept)
Formula: The model description passed to an estimation function. With stepwise operators or several dependent variables, it describes several models. Avoid: formula string, fml (in prose)
Model formula:
The formula of one model after stepwise operators and dependent variables are
expanded, written in a canonical spelling that can differ from the user's,
for example with an explicit 0 +.
Avoid: expanded formula, rendered formula, parsed formula, model name
Stepwise operator:
A formula function that expands into several models: sw() uses its
arguments one at a time, csw() adds them cumulatively, mvsw() uses every
combination of them, and sw0(), csw0(), and mvsw() also give a model
with none of them.
Avoid: multiple estimation syntax
Term:
One summand of a formula, such as X1, X1:X2, C(f), or f1[x]. a*b
expands to the terms a + b + a:b.
Avoid: factor
Dependent variable: The left-hand-side variable of a model. Avoid: depvar (in prose), response (for the variable), Y, target; outcome, except in causal and difference-in-differences prose
Regressor: A right-hand-side variable with an estimated coefficient. In an IV model it is an exogenous regressor or an endogenous regressor. Avoid: covariate, independent variable, explanatory variable, control, X
Intercept:
The constant regressor a formula includes unless it writes 0 or -1, or a
fixed effect with level intercepts absorbs it.
Avoid: constant, structural intercept
Fixed effect:
A term after | in a formula whose levels are absorbed instead of being
reported as coefficients. C() and i() terms are categorical regressors,
not fixed effects, even when they estimate the same intercepts.
Avoid: FE, fixef (in prose); fixed effect (for C() or i() terms)
Level: One category of a fixed effect or a categorical regressor. Avoid: group, cell
Varying slope:
A slope that differs by fixed-effect level, written f1[x]. A slope-only
effect, f1[[x]], has the slopes without the level intercepts.
Avoid: random slope, interacted slope
Categorical regressor:
A categorical variable entered as regressors, one per level except a reference
level where one is needed, written C(f) or i(f). i(f, x) interacts those
regressors with a second variable.
Avoid: factor, dummies, fixed effects, interacted fixed effects
Reference level:
The level of a categorical regressor that gets no regressor of its own,
because the intercept or another term already spans it; the other levels'
coefficients are contrasts with it. Chosen with ref= in i().
Avoid: base level, reference category, omitted category
Main equation: The equation of the dependent variable on the regressors that every model has. Avoid: second stage (outside IV)
Model matrix: The data a model is estimated from, by role: the dependent variable, the design matrix, the fixed-effect levels and varying-slope variables, the endogenous regressors and instruments, the weights, and the offset. Avoid: model frame, design matrix (for the whole)
Design matrix: The regressors of the main equation, one column per estimated coefficient. In an IV model it holds the endogenous regressors, not their first-stage fitted values. Avoid: X (in prose), regressor matrix
Offset: A variable added to the linear predictor with its coefficient fixed at one, such as log exposure in a Poisson model. Avoid: exposure (for the offset)
Collinear regressor: A regressor that is, within a tolerance, a linear combination of other regressors in the within data. It is dropped from the design matrix and gets no coefficient. Avoid: multicollinear variable
Instrument block:
The [endogenous ~ instruments] part of a formula.
Avoid: IV block, multistage block
Endogenous regressor: A regressor on the left of the instrument block; the first stage predicts it from the instrument matrix. Avoid: endogenous variable
Exogenous regressor: A regressor of the main equation outside the instrument block.
Instrument: A variable in the instrument block that is not a regressor of the main equation; it enters the model only through the first stage. An exogenous regressor repeated in the block is redundant, not an instrument. Avoid: excluded instrument
Instrument matrix: The instruments together with the exogenous regressors. Avoid: instruments (for the matrix), Z
First stage: In an IV model, the regression of an endogenous regressor on the exogenous regressors and the instruments. Avoid: first stage (for DID2S, unqualified)
Second stage: In an IV model, the main equation with the endogenous regressors replaced by their first-stage fitted values.
Demeaning: Projecting a variable on the fixed effects, including their varying slopes, and keeping the residual. Avoid: residualizing, partialling out, projecting out, within transformation
Absorb: What a fixed effect does to its levels, and to the intercept when it has level intercepts: they are accounted for by demeaning and not reported as coefficients.
Within data: The dependent variable and regressors after demeaning. Avoid: demeaned data, transformed data
Demeaner: The solver that performs demeaning, such as the method of alternating projections (MAP) or LSMR. Avoid: backend (alone), demeaning algorithm
Demeaner backend: The library or kernel that implements a demeaner, such as Rust or numba for MAP, and within or PyTorch for LSMR. Avoid: backend (alone), engine
Preconditioner: A transformation an iterative demeaner such as LSMR applies to the fixed-effect system so that it converges in fewer iterations, such as the diagonal or the additive Schwarz preconditioner. Like any solver choice, it moves the estimates only within the tolerance. Avoid: preconditioning method
Estimation sample: The observations a model is fitted on, after sample splitting and every dropping stage. Avoid: sample (unqualified), estimation data
Dropped row: An input row excluded from the estimation sample for a missing or infinite value, a zero weight, a singleton, or separation. Avoid: NA row, removed observation
Singleton: A level of a fixed effect with level intercepts that occurs in exactly one row. Its row is dropped by default. Avoid: singleton observation, singleton group, singleton fixed effect (for the row)
Separation: The condition of a GLM observation whose dependent variable sits at a bound that the mean reaches only as the linear predictor diverges, such as zero in a Poisson model. Separated rows are dropped. Avoid: perfect fit, perfect prediction
Separation check: A procedure that detects separated observations: the fixed-effect check or the iterative rectifier. Avoid: separation method
Sample split: Fitting the same formula separately on the subsets defined by the values of one variable; each subset gives its own model. It is a form of multiple estimation. Avoid: subsample, by-group estimation
Observation: What the number of observations counts. A row counts as one observation, except under frequency weights, where it counts as many as its weight. Avoid: N (unqualified), effective sample size
Observation weights: User-supplied per-row weights with one interpretation, analytic or frequency. Avoid: weights (unqualified), regression weights, sample weights, WLS weights
Analytic weights:
Observation weights that scale a row's contribution without changing the
number of observations. Written aweights.
Frequency weights:
Observation weights under which a row stands for as many identical
observations as its weight. Written fweights.
Working weights: The weights of the final IWLS iteration. They already contain the observation weights. Avoid: weights (unqualified)
Bootstrap weights: Random multipliers that the wild bootstrap draws per observation or per cluster, such as Rademacher or Webb weights. Avoid: weights type (in prose), weights (unqualified)
Inference: Standard errors, tests, p-values, and confidence intervals for a fitted model's coefficients. Avoid: standard errors (as the umbrella)
Variance-covariance matrix: The estimated sampling covariance of a fitted model's coefficients under one variance estimator. Avoid: vcov (in prose), covariance (alone)
Variance estimator: The procedure that estimates the variance-covariance matrix, such as iid, heteroskedasticity-robust (HC1 to HC3), cluster-robust (CRV1, CRV3), HAC, or nid. Changing it changes the inference, not the estimates. Avoid: vcov type, covariance estimator, standard-error type, inference type, estimator (alone)
Inference type: Whether intervals and tests hold for each coefficient on its own, for several coefficients simultaneously, or at any time under continuous monitoring (anytime-valid). Avoid: interval type
Significance level: The error probability α of a test. Confidence intervals at significance level α cover 1 − α. Avoid: alpha (in prose), level, confidence level (for α)
Small-sample correction: The finite-sample scaling of the variance-covariance matrix and the degrees of freedom of its tests. Avoid: ssc (in prose), dof adjustment, small sample adjustment
Adjustment factor: The number by which the small-sample correction multiplies the variance-covariance matrix. Avoid: ssc (for the number)
Cluster variable: A data column whose values define the clusters of cluster-robust inference; several cluster variables give multiway clustering. Avoid: clustervar (in prose), cluster (for the column), group
Number of clusters: The count of distinct values of a cluster variable. For non-clustered variance estimators, the small-sample correction counts each observation (heteroskedasticity-robust) or each period (HAC) as a cluster. Avoid: G (in prose)
Coefficient table: A fitted model's estimates with their standard errors, test statistics, p-values, and confidence intervals under its current variance-covariance matrix. Avoid: coeftable (in prose), results table, summary table
Wald test: A test of linear restrictions Rβ = q on the coefficients, built from the variance-covariance matrix, whether its statistic is reported on the χ² or the F scale. Avoid: F test (as a synonym)
Wild bootstrap: Resampling inference that perturbs residuals with bootstrap weights drawn per observation or per cluster. Avoid: bootstrap (unqualified)
Randomization inference: Inference that compares a coefficient's test statistic with its distribution over random reassignments of a treatment variable. Avoid: permutation test, ritest (in prose)
Multiple testing: Testing several hypotheses at once, with p-values adjusted to control the family-wise error rate, such as by Bonferroni, Romano-Wolf, or Westfall-Young. Avoid: multiple hypothesis testing, multiple hypothesis corrections, multiple comparisons
Simultaneous confidence intervals: Confidence intervals for several coefficients that cover all of them jointly with probability 1 − α, unlike pointwise intervals, which each cover one coefficient. Avoid: joint confidence intervals, confidence bands, joint confidence bands
Causal cluster variance: A design-based variance estimator for treatment effects that accounts for the share of clusters sampled and how treatment varies within clusters. Avoid: CCV (in prose)
Capability: An inference or post-estimation operation an estimator declares it supports, such as CRV3 inference, prediction, or randomization inference. A fitted model rejects an operation its estimator does not declare. Avoid: feature flag, support check
Link scale: The scale of the linear predictor.
Response scale: The scale of the dependent variable's mean. It coincides with the link scale for linear models.
Fitted values: A fitted model's in-sample predictions, on the link or the response scale. Avoid: yhat, predictions (for in-sample values)
Unseen level: A level in prediction data that the estimation sample does not contain. Its row's prediction is missing. Avoid: new level
Residual: The dependent variable minus the fitted mean, in the dependent variable's units. Unqualified, it never means the working residual. Avoid: response residual, raw residual, u_hat (in prose)
Working residual: The residual of the weighted regression in the final IWLS iteration.
Fit statistics: Goodness-of-fit measures of a fitted model, such as R², deviance, or log-likelihood. Avoid: fitstat (in prose), GOF, model stats
Fixed-effect estimates: The estimated level intercepts and varying slopes of a model's fixed effects, recovered after estimation. Avoid: fixed effects (for the estimates), alpha, fixef (in prose)
Refit: Rerunning a fitted model's estimator with its estimation options on other data, or on its own estimation sample with changed columns. Avoid: update, re-estimate
Sherman-Morrison update: Updating an OLS fit's coefficients for appended observations without refitting. Avoid: update (in prose), online update
Unit: The panel entity observed across periods. Avoid: id, individual, idname (in prose)
Period: A calendar time point of the panel. Avoid: time (alone), tname (in prose)
Cohort: The units whose treatment starts in the same period. Avoid: group, treatment group, gname (in prose)
Event time: The number of periods between a period and a unit's first treated period. Avoid: relative time, relative period
ATT: The average treatment effect on the treated.