Results (`increment.results`)

Sitewide behavior is exposed through Analysis.sitewide, with output values available from the receive-only results namespace.

Estimate

Estimate

A number with quantified uncertainty.

lb/ub (when set) represent an interval at level confidence. open_side names an unbounded endpoint of a one-sided interval.

Parameters

value : float Point estimate. lb, ub : float | None Interval bounds. Closed intervals set both; open intervals leave the endpoint named by open_side unset. open_side : {“lower”, “upper”} | None Unbounded endpoint for a one-sided interval, or None for a closed or unavailable interval. level : float | None Confidence/credible level, e.g. 0.95. Required iff lb/ub are set. At extreme alpha values this may round to 1.0; alpha retains the exact allocated noncoverage budget. alpha : float | None Allocated noncoverage budget; level is its nominal complement. Fixed open endpoints use this full quantile tail. Sequential open endpoints retain the symmetric parent budget, without a fixed-time tail interpretation. Attained coverage may be higher than nominal. Unlike level, it remains recoverable when the confidence level rounds to 1.0. log_mean, log_se : float | None Point estimate and standard error on the natural-log scale. Set by infer_lift to the raw pre-prior statistics, not the posterior mean/sd - the two differ under an informative prior.

LiftEstimate

LiftEstimate

A lift estimate, carrying the metadata to identify which (metric, method, group) and its persisted inference reference.

value is exp(mu_n) - 1 for scale="log" estimates (under a near-flat prior this equals treatment_mean / control_mean - 1), or mu_n directly for scale="linear" estimates, which already carry the posterior on the relative scale. The CI is the same posterior’s quantile, back-transformed to this scale.

The Normal posterior is never persisted: with a stored alpha the mu/sigma recovery from value/lb/alpha/scale is accurate up to rounding away from a zero critical value; open fixed rows use raw statistics to recover uncertainty at alpha=0.5. A closed alpha-less interval falls back to the level tail, which is exact at ordinary alphas but loses precision as alpha approaches the representable floor. A mixture posterior (prior_spec set) is not recoverable, so it persists the prior instead.

reference_kind persists the reference: “normal”, “t”, “sequential”, “binomial”, or “confidence_set”. Winsor confidence sets persist raw construction state and endpoint statuses, expose no posterior, and reinvert through reintervalize(alpha). Their public confidence_set.qualification is either pointwise_asymptotic_model_conditioned_v1 (bootstrap candidate) or uniform_support_conditioned_v1 (rank method); neither is an unqualified finite-sample promise. Bootstrap p-values invert stored roots; rank p-values are alpha when the set excludes the null, else one. dof retains cluster degrees of freedom, so dof=None does not imply Normal inference: a Welch reference has its own reference_df. Relative t p-values use this reference; posterior-derived stats refuse it.

inference labels the interval semantics: “fixed” is a single-look interval; “always_valid”/“asymptotic_mean” are valid at every look. value carries no stopping adjustment (winner’s curse); the interval endpoints are the safe summary. Decision-stat methods refuse on any non-”fixed” inference.

value_scale="absolute" marks a row whose value is not a relative lift at all - an encouragement design’s additive LATE, or an observational metric reported additively.

Clustered unadjusted and prior-free adjusted rows preserve additive uncertainty independently of relative_confidence_set. The joint set can be disconnected, unbounded, or present without a finite ratio point. Its reference is a qualified working approximation; relative_unavailable_reason identifies an indefinite or unrepresentable covariance without discarding additive output.

JointContrastReference

JointContrastReference

Working joint reference for the additive numerator and denominator.

RelativeConfidenceSet

RelativeConfidenceSet

Outward-enclosed Fieller set under an explicitly approximate joint reference.

LiftEstimates

LiftEstimates

list[LiftEstimate] with .to_frame() - :meth:~increment. analysis.Analysis.run’s return type.

BreakoutEstimate

BreakoutEstimate

A lift estimate for ONE segment (dimension value) of a breakout.

Carries the same identity fields as LiftEstimate (via _RowIdentity) plus dimension/dimension_value to identify the segment and source to disambiguate the rare same-name, different-source case.

reference_kind/reference_df preserve the source row’s sampling reference through serialization. A Welch t reference can have dof=None; it must not be reconstructed as a Normal interval.

run_breakout returns one row per (segment, metric, method, non-control arm) cell, dense - an unestimable cell still gets a row, with lift=None and excluded naming why (see :data:ExclusionReason).

low_reliability flags a real estimate whose control or treatment arm has fewer than reliability_floor units: estimable, but the Wald interval’s nominal coverage is not trustworthy there.

Call .to_frame() on the :class:BreakoutEstimates this returns rather than constructing a plain list[BreakoutEstimate].

BreakoutEstimates

BreakoutEstimates

list[BreakoutEstimate] with .to_frame() - :func:run_breakout’s (and :meth:~increment.analysis.Analysis.run_breakout’s) return type.

DailyMetricValue

DailyMetricValue

One arm’s absolute metric mean (with CI) for ONE day.

The time-series counterpart to a plain arm-level mean - what plots on a per-day chart of the raw value, as opposed to the relative lift between arms (see :class:DailyLiftEstimate). Produced by :func:run_daily directly from one daily_group_summary row.

dimension/dimension_value/source are populated together when :func:run_daily is given a dimension, and None otherwise.

ds_basis says what ds indexes: "calendar" (default) is an observation date; "cohort" means ds is the unit’s own exposure date, how a retention series is conventionally indexed.

Call .to_frame() on the :class:DailyMetricValues this function returns rather than constructing a plain list.

DailyMetricValues

DailyMetricValues

list[DailyMetricValue] with .to_frame() - :func:run_daily’s (and :meth:~increment.analysis.Analysis.run_daily/ run_asof’s) return type.

DailyLiftEstimate

DailyLiftEstimate

A relative lift estimate for ONE day of a daily time series.

Carries the same identity fields as :class:BreakoutEstimate and :class:~increment.estimation.results.LiftEstimate (via _RowIdentity) plus ds to identify which day’s slice of moments produced it.

estimand/value_scale/note mirror :class:BreakoutEstimate’s fields of the same name. ds retains calendar dates, numeric day indices, or structured string labels for as-of frame readouts. low_reliability marks a real estimate whose control or treatment arm has fewer than reliability_floor units that day.

dimension/dimension_value/source are populated together when :func:run_daily_lift is given a dimension, and None otherwise. ds_basis - see :class:DailyMetricValue.

DailyLiftEstimates

DailyLiftEstimates

list[DailyLiftEstimate] with .to_frame() - :func:run_daily_lift’s (and :meth:~increment.analysis.Analysis.run_daily_lift/ run_asof_lift’s) return type.

HeterogeneitySummary

HeterogeneitySummary

One row per (metric, method, group_id, dimension, source, estimand, value_scale) grouping key: Cochran’s Q, DerSimonian-Laird tau^2, Higgins-Thompson I^2, and an HKSJ pooled effect, over the segments declared for that key.

scale is "relative" (log-RR) or "absolute" (risk difference) - the two frequently disagree, so both ship. value_scale is the upstream rows’ own value scale; scale is which of the two heterogeneity passes this row belongs to. tau2/i2/i2_lb/i2_ub are None whenever n_excluded_outcome > 0 for this (key, scale) row: an outcome-based exclusion biases tau^2, so it is suppressed rather than reported as trustworthy. q/p_value/pooled are not.

HeterogeneitySummaries

HeterogeneitySummaries

list[HeterogeneitySummary] with .to_frame() - see :class:~increment.breakout.estimates.EstimateList.

SegmentEstimate

SegmentEstimate

One row per (segment, scale, estimator) for a HeterogeneitySummary key: two rows per estimable (segment, scale) (estimator="raw" and "shrunken"), plus an unavailable pair per excluded segment.

excluded is set from an upstream BreakoutEstimate.excluded, a live segment unusable on this scale only (excluded="zero_variance"), or a shrunken row withheld because that scale’s tau posterior could not be integrated within the numerical budget (excluded="estimation_failed").

raw reuses the segment’s own estimate (for scale="relative", BreakoutEstimate.lift verbatim). shrunken is the tau-marginalised posterior estimate, with its own shrink_k. baseline is the segment’s control-arm absolute mean, recovered from abs_diff and the raw log-scale lift.log_mean.

value_scale is the upstream rows’ own value scale; scale is which of the two heterogeneity passes this row belongs to.

role, discovery, family_axes, family_q, and family_threshold mirror the source BreakoutEstimate’s fields of the same name verbatim (see there) — a reader of this frame alone can otherwise not tell a discovery from a non-discovery, or a multiplicity-corrected interval from an uncorrected one. Both estimator rows for a segment carry the same source values.

SegmentEstimates

SegmentEstimates

list[SegmentEstimate] with .to_frame() - see :class:~increment.breakout.estimates.EstimateList.

SegmentRolloutResult

SegmentRolloutResult

:func:segment_rollout_recommendation’s return value - two independent, row-aligned-by-grouping-key result sets.

RolloutRecommendation

RolloutRecommendation

One row per (metric, method, group_id, dimension, source, estimand, value_scale) grouping key: which of that key’s segments to roll out, and an honest price on doing so.

UNWEIGHTED, and deliberately so: every value field treats each usable segment as an equal contributor regardless of exposure, so policy_value is a plain sum over the rolled-out segments, not traffic-weighted. An exposure-weighted variant was measured and dropped - its selection-bias correction did not meet the accuracy bar the unweighted correction is held to.

recommendation carries the estimator’s verdict verbatim: "rollout" (corrected value is positive), "no_net_benefit" (evidence cannot demonstrate positive value), or "refuse" (the offset guard fired - every value field is withheld).

rollout_cost is the relative-lift break-even actually used. k/n_excluded_design/n_excluded_outcome count how many of the declared segments the decision could and could not use, so a reported value can never silently understate its coverage.

RolloutRecommendations

RolloutRecommendations

list[RolloutRecommendation] with .to_frame() - see :class:~increment.breakout.estimates.EstimateList.

RolloutSegment

RolloutSegment

One row per segment declared for a :class:RolloutRecommendation key - dense: every segment of a key that ran appears, usable or not.

selected is the estimator’s subset membership on a usable segment, None on one this call could not use - populated even behind a "refuse" recommendation as evidence about the selection rule, not a deployable decision.

excluded is an upstream BreakoutEstimate.excluded, or "zero_variance" for a live segment whose log-scale statistic or variance is unusable here - the same tag segment_heterogeneity uses for its own per-scale drop.

estimand/value_scale identify the upstream rows’ own estimand and value scale; rows are grouped by both, so a family is priced on its own or not at all. A real LATE family (additive lift, no log-scale moments) never clears the pricing gate below.

RolloutSegments

RolloutSegments

list[RolloutSegment] with .to_frame() - see :class:~increment.breakout.estimates.EstimateList.

PowerResult

PowerResult

Result of a power-analysis solver.

Parameters

n_per_arm : int Number of units in the treatment arm. n_total : int Total N across both arms. power : float Planned power at the computed / given sample size, under the model named by power_basis. For required_sample_size and achieved_power it describes the SUPPLIED effect; for minimum_detectable_effect it describes the returned effect’s implied absolute alternative, which can exceed the target when the answer is a domain endpoint. power_basis : {“asymptotic”, “exact”, “approximate”} How power was computed. "asymptotic": the log-ratio Normal/noncentral-t planning model (every plan the runtime does not decide with the exact binomial risk-ratio test). "exact": the probability that the runtime’s unchanged exact binomial decision rejects, at the analyzed integer counts (up to at most about 1e-12 of omitted outer count mass). "approximate": the same decision replayed with Normal conditional tails, for binomial plans whose exact geometry exceeds the planning cell budget. mde_relative : float | None Minimum detectable relative effect on the complier scale, expressed RELATIVE TO the declared null: (exp(distance) - 1) / baseline.compliance, where distance is the search’s own log-scale gap from theta0 = log1p(decision.null_lift). At compliance=1.0 this is exactly exp(distance) - 1; a lower compliance rescales it up. To recover the implied ABSOLUTE alternative, compose against the null rather than adding lifts: expm1(log1p(decision.null_lift) + log1p(mde_relative * baseline.compliance)) / baseline.compliance. Positive for two-sided and one-sided “greater”; negative for one-sided “less”. None when no minimum detectable effect exists at the design’s target power — a valid supplied-effect answer can still be reported then — with the cause in mde_unavailable_reason. mde_unavailable_reason : {“unattainable”, “unrepresentable”, “numerical_resolution”} | None Why mde_relative is None: the target power exceeds every admissible alternative’s power (unattainable), the admissible answer has no float64 representation (unrepresentable), or the search could not resolve it within its numerical limits (numerical_resolution). None whenever mde_relative is available; a missing effect always carries a reason. effective_var : float Per-unit variance for asymptotic planning, after the effective decision method’s CUPED / factor-absorption reduction and cluster design effect. Sensitivity-only CUPED receives no reduction, so this may differ from the caller’s Baseline.effective_var. Exact and approximate binomial plans report it as metadata; their power uses event rates and counts. For a QuantileBaseline, the per-unit variance its pilot’s standard error implies (see QuantileBaseline.from_control_values). n_clusters_per_arm : int | None Randomization clusters needed in the treatment arm, ceil(n_per_arm / baseline.avg_cluster_size). None under unit randomization. n_clusters_total : int | None Randomization clusters across both arms, ceiled per arm and summed (not ceiled once on n_total, since a part-cluster in each arm costs two whole clusters). None under unit randomization. n_triggered_per_arm, n_triggered_total : int | None Units actually entering a triggered analysis (n_per_arm/n_total times trigger_rate). None when no trigger_rate was declared; n_per_arm/n_total always count assigned units. expected_n_total : int | None Expected total N a sequential design stops at under the solved-for effect: E[T] * n_total, where E[T] charges each look’s first-boundary-exit mass its own information fraction and the never-crossing remainder the final planned look (see increment.power.sequential.sequential_expected_information_fraction). This is the honest economic counterpart to n_total (the worst-case, never-stops-early size) — early stopping is the entire argument for monitoring sequentially. None when no inference spec was given (a fixed-horizon design always runs to n_total). inference_to_declare : InferenceSpec | None The exact runtime InferenceSpec this plan assumed — pass it to InferenceSpec (or a YAML inference: block) at runtime declaration so the runtime’s boundary is tuned from the same N planning assumed. None for a fixed-horizon result. planned_metric_name, planned_quantile : str | None, float | None The metric name and quantile level a QuantileBaseline was built for (Analysis.planning_baseline(metric)), echoed here so a caller who reuses one metric’s baseline to plan a different metric sees the mismatch stated in the answer instead of discovering it, if at all, from a silently mis-sized design. None for every non-quantile metric.

PowerCurvePoint

PowerCurvePoint

One evaluated point in a power or MDE curve.

mde_relative is None with mde_unavailable_reason set when no minimum detectable effect exists at that row’s size and target (see PowerResult); frames keep the column numeric with a null. power_basis names the planning model behind power and mde_relative, as on PowerResult.

PowerCurve

PowerCurve

List-like power-curve result with dict and dataframe conversion.

SRMResult

SRMResult

Result of an allocation sample-ratio-mismatch check.

fixed_p_value is the ordinary Pearson fixed-look p-value. log_e_value is the current uniform-Dirichlet mixture evidence for a predeclared allocation (unavailable when fixed inference infers equal shares from observed arms). inference states which quantity controls is_srm.

unassigned_units and mixed_assignment_units are accounting entries, not arms: they contribute no chi-square degree of freedom and appear in neither observed nor expected. mixed_assignment_units counts units observed in more than one arm, which shrinks every arm’s count symmetrically and so must be surfaced separately to be seen.

grain is "cluster" when randomization happened over clusters, so the chi-square belongs there; unit_counts then carries per-arm unit counts as descriptive context only (cluster size imbalance earns no degree of freedom). At grain="unit", unit_counts is empty.

min_expected_count is the smallest per-arm expected count (the minimum over arms of expected[k] * total), or None when unavailable. low_expected_count is True when that minimum falls below 5, the standard Cochran rule of thumb, flagging that the fixed asymptotic chi-square p-value may be unreliable there.

AllocationBand

AllocationBand

A Beta posterior credible band on one arm’s allocation share at one ds.

Purely descriptive: unlike SRMResult, this carries no pass/fail flag - sample_ratio_mismatch is the actual gate. posterior_a/ posterior_b are the Beta parameters the interval was derived from, carried so a caller can reconstruct the full posterior (they cannot be recovered from n/n_total alone without the prior, a call-site argument).

NotApplicable

NotApplicable

A diagnostic check that does not apply to the current design.

Returned in place of a check’s normal result (e.g. SRMResult) when the design makes the check meaningless - a sample-ratio test presumes a target randomized allocation an observational design does not have.

AbsorptionResult

AbsorptionResult

The absorbed average treatment effect and its diagnostics.

effect is on the absolute scale (treated mean minus control mean, factor absorbed). When the effect is homogeneous across levels this is the ATE; when it varies, effect converges to the precision-weighted average of per-level effects (weights n_t * n_c / n_g, Angrist 1998), down-weighting skewed-allocation levels relative to the unit-weighted ATE.

icc is always the estimated variance-component ratio; mean_shrinkage is the pooling weight actually applied, pinned to 0.0 or 1.0 when pooling is forced rather than estimated - the two can disagree when pooling is not "partial".

ClusterScore

ClusterScore

A pooled score keyed by canonical cluster identity, in canonical ID order.

CateScoreState

CateScoreState

Immutable portable scoring basis, fitted centering and effect coefficients.

CateResult

CateResult

A fitted Lin (2013) interacted regression.

ate is the treatment coefficient at the design centre (the weighted mean of the fitted treatment effects); se is its sandwich SE. Without clusters, se_unadjusted is Welch’s SE; with clusters it uses the same weights and cluster sandwich fitted on [1, d]. se_reduction reads the width the adjustment bought.

dimension, n_clusters (all observed IDs), immutable vcov and reference_df are the stored inference contract used by projections. Cluster intervals use t(K-1); the interaction Wald quadratic uses F(q, K-1) after division by q. The reference is cluster-asymptotic. unadjusted_vcov stores the separate two-column fit’s covariance. Unavailable required uncertainty refuses rather than returning a partial fit.

The state behind cate/contrast/score lives on excluded fields: a dumped result is report-only, so those need the live object. ate/se/lb/ub/heterogeneity are always the unpenalized fit’s; fit_cate(ard=True) only moves the interaction estimates, in beta_ard (None if ARD did not run).

Pass include_evaluation_population=True to CATE validation or targeting APIs to retain the immutable evaluation roster and its actual base weights. The default retains no identifiers. Read the snapshot from validation.evaluation_population, or from rule.validation.evaluation_population for fixed and selected rules. Selection captures only the outer evaluation split; its overlap provenance is independent of the inner selection population. Equal-cluster weights use retained cluster sizes after overlap trimming, before policy selection. Snapshot rows, cluster identities, and weights stay aligned through JSON serialization.

For synthetic validation, increment.simulate.cluster_dgp.evaluation_policy_truth(rule, population) consumes a saved rule and its ClusteredCATEResult. It returns exact policy truth and the retained-population ATE, using the stored IDs and actual base weights. Missing or mismatched evaluation rosters are refused, not reconstructed. This is truth for the same retained population, not an independent evaluation batch.

CateEvaluationPopulation

CateEvaluationPopulation

Immutable roster and weighting provenance for a reported evaluation.

CateValidation

CateValidation

Everything the held-out half says about a fitted CATE model.

passed is autoc.p_value < alpha; while false, the only defensible number is the average effect, holdout_ate - the effect on the same rows the groups and rank tests use. Randomized sources use a difference in means with a Welch SE; observational sources use the mean cross-fitted doubly robust score with its SE.

groups/clan intervals are Bonferroni-corrected across their own family (n_groups group intervals, len(clan) CLAN rows): a reader scanning every row for the one excluding zero pays the true familywise error, not the per-row nominal alpha (CDDF 2018). split_caveat names the single-split limitation this correction does not address. For declared clusters, top-level uncertainty_method, reference_df, and unavailable_reason describe the holdout ATE; each group/rank/CLAN row records its own bootstrap uncertainty. A missing AUTOC p-value closes the gate. Nuisances are frozen using training rows.

TargetingRule

TargetingRule

A frozen deployment candidate and its honest-split evidence gate.

fraction is the requested budget; achieved_fraction is its realized holdout share under cluster_weight. Cluster policies pool scores, break ties by canonical ID and take the longest feasible whole-cluster prefix. Their threshold is descriptive, never a substitute for the prefix rule. Unit policies retain their frozen score cutoff.

predict applies this candidate to new columns without fitting again. recommendation is "target" only when the evidence gate passes; otherwise it recommends treating everyone alike based on the average effect. A passing, nonempty candidate reports its conditional policy effect and uplift as points only: this same holdout gates and reports them, so nominal intervals would be invalid after selection. Empty candidates retain null effects and their exact unavailable reason, rather than fabricated zeros.

TargetingSelection

TargetingSelection

A fraction chosen honestly, and the locked rule’s untouched-test verdict.

inner is the selection table: each grid fraction’s net benefit E[1{targeted}(tau - cost)] on inner units scored by models that never saw them, using IPW contributions for randomized sources and cross-fitted doubly robust scores for observational sources. selected_fraction is its argmax (ties to the smaller fraction), locked before the outer test is read. rule is the ordinary :class:TargetingRule evaluated once on the outer half. population records when overlap trimming restricts the inner selection and training population; it is independent of rule.population, which records trimming on the outer test. seed and the grid are pre-commitments: rerunning with a new seed until the answer improves is the failure mode this workflow prevents. Clustered bootstrap availability and valid-repetition metadata mirror the selected inner row; outer-test uncertainty remains on rule.

MetricTrend

MetricTrend(table, metric, grain, window, week_start, denominator, start, end, by, alpha)

One metric’s calendar trend: a lazy ibis query plus metadata.

Columns (order not guaranteed, only names): metric, grain, window, period, [dims...], period_complete, n, value, ci_lb, ci_ub. n is the per-period unit denominator for entity-scoped metrics and NULL for total/active (an active metric’s count is its value). ci_lb/ci_ub are NULL for ratio (point estimate only in v1) and for total/active (no variance concept). grain keeps materialized rows self-describing across grains. window is the rolling trailing-day size (total/active only), NULL for every calendar-bucket trend.

period_complete is true once the metric’s own fact has any event at or past the period’s end (period_end <= max(fact ts)), so a period can be marked complete up to one partial day early. Fine for date-grain horizons; a consumer gating on the last complete period of a still-loading warehouse should ignore it or wait for the next load.

SitewideImpact

SitewideImpact

Whole-site impact of shipping a lift to every enrolled unit.

Parameters

delta, delta_se : float Per-unit absolute lift of the target arm and its standard error. treatment_group : str group_id of the target arm. n_control, n_treatment : float Control and target arm population sizes (float since a cluster grain population is K * mean cluster size, not an integer). n_enrolled : float Every enrolled unit across all arms (N_exp); equals n_control + n_treatment only when no other arm is enrolled. other_arm_ids : tuple[str, …] group_id of every other enrolled non-control arm, whose lift was netted out of baseline_volume. site_total_volume : float The observed site-window total this result was computed against. baseline_volume : float Counterfactual site-window total with nobody exposed to treatment. absolute_impact, absolute_impact_se, absolute_impact_lb, absolute_impact_ub : float Ship-to-all absolute impact, its SE, and its alpha-level interval. relative_impact, relative_impact_se, relative_impact_lb, relative_impact_ub : float Ship-to-all impact as a fraction of baseline_volume, its SE, and its alpha-level interval. alpha : float Two-sided significance level the intervals were built at. n_clusters : int | None None at iid grain, else the contrast’s total cluster count. absolute_dof, relative_dof : float | None None at iid grain (the Normal reference applies). Otherwise the dof absolute_impact’s and relative_impact’s own critical values were cut at - see “Degrees of freedom” in the module docstring for which reduction each is. They differ from each other, and both differ from the contrast’s pooled n_clusters - 2 once another arm is enrolled.

SitewideRatioImpact

SitewideRatioImpact

Whole-site impact of shipping a ratio metric’s lift to every unit.

Ratio metrics need two per-unit lifts (numerator, denominator) instead of :class:SitewideImpact’s single delta, hence a separate model. See “The ratio math” in the module docstring for the derivation.

Parameters

delta_num, delta_num_se : float Per-unit absolute lift of the target arm’s numerator and its SE. delta_den, delta_den_se : float Per-unit absolute lift of the target arm’s denominator and its SE. delta_cov : float Covariance of delta_num and delta_den. treatment_group : str group_id of the target arm. n_control, n_treatment : float Control and target arm population sizes (see :class:SitewideImpact for why this is a float). n_enrolled : float Every enrolled unit across all arms (N_exp). other_arm_ids : tuple[str, …] group_id of every other enrolled non-control arm, whose numerator and denominator lifts were netted out of the baselines. site_total_numerator, site_total_denominator : float The observed site-window totals this result was computed against. baseline_numerator, baseline_denominator, baseline_ratio : float Counterfactual site-window totals with nobody exposed to treatment (N0, D0), and their ratio. shipped_numerator, shipped_denominator, shipped_ratio : float Ship-to-all site-window totals (N1, D1), and their ratio. absolute_impact, absolute_impact_se, absolute_impact_lb, absolute_impact_ub : float Ship-to-all absolute impact (shipped_ratio - baseline_ratio), its delta-method SE, and its alpha-level Wald interval. relative_impact, relative_impact_se, relative_impact_lb, relative_impact_ub : float Ship-to-all impact as a fraction of baseline_ratio, its delta-method SE, and its alpha-level interval. alpha : float Two-sided significance level the interval was built at. n_clusters, absolute_dof, relative_dof : int | None, float | None, float | None Same contract as :class:SitewideImpact’s fields of the same name - unlike the sum-metric’s absolute_dof, this class’s absolute_dof is always the Satterthwaite reduction, never the plain pairwise dof (see “Degrees of freedom” in the module docstring).