Switchback

API reference

Randomized switchback schedules of two settings, and their randomization analysis.

tsdive.switchback.design draws a balanced schedule of K blocks and the reference assignments; tsdive.switchback.inference computes the B - A difference on the kept samples, its randomization p-value and the interval that inverts the test; tsdive.switchback.plan holds the schedule as a plan with a digest, checks it, and reads its power off a history window; tsdive.switchback.archive reads the target and covariates through the store and runs the analysis, and tsdive.switchback.render writes both as text and JSON. The package imports numpy and pandas and nothing heavier.

Name	Summary
tsdive.switchback_plan	Plan a balanced random schedule of settings A and B over one window.
tsdive.SwitchbackPlan	A balanced random schedule of settings A and B over one window.
tsdive.switchback_analyze	The difference between settings A and B on target under a verified plan.
tsdive.SwitchbackAnalysis	The difference between settings A and B on one target, under a verified plan.
tsdive.SwitchbackEstimate	One analysis of the target: direct (no covariates) or adjusted.
tsdive.switchback.analyze	The B - A coefficient of y on [1, z] or [1, z, X], its p-value and interval.
tsdive.switchback.cut_blocks	Blocks of length from time 0, K = schedule_blocks(span, length).
tsdive.switchback.design_size	(balanced assignments, enumerated, smallest attainable two-sided p) of K blocks.
tsdive.switchback.lag_response	First-order lag response to the per-sample setting u, starting from 0.
tsdive.switchback.make_design	Observed assignment and reference set, deterministic in seed.
tsdive.switchback.make_plan	The balanced schedule seed draws over start to end.
tsdive.switchback.plan_digest	SHA-256 hex digest of the plan's canonical schedule.
tsdive.switchback.schedule_blocks	K = floor(span / length), less one when it is odd.
tsdive.switchback.setting	Per-sample setting: 1 in B blocks, 0 in A blocks and outside the schedule.
tsdive.switchback.verify_plan	The design the plan's seed draws, after checking the plan against it.
tsdive.switchback.Analysis	One randomization analysis of a target, or the reason it has none.
tsdive.switchback.Blocks	Block of each sample (-1 outside the schedule) and its offset from the block start.
tsdive.switchback.Design	The observed assignment of K blocks and the assignments its p-value reads.
tsdive.switchback.PlannedBlock	One block of the schedule: when it runs, its setting, and when its washout ends.
tsdive.switchback.PowerReadout	Detection rates of the plan's design over a history window where nothing changed.
tsdive.switchback.SwitchbackPlan	A balanced random schedule of settings A and B over one window.
tsdive.switchback.ALPHA	Level of the two-sided randomization test.
tsdive.switchback.COLLINEAR	Refusal reason: the observed assignment lies in the span of the covariates.
tsdive.switchback.EMPTY_BLOCK	Refusal reason: a block with no kept sample.
tsdive.switchback.MIN_ASSIGNMENTS	Fewest balanced assignments a design may have (DesignTooSmall below it).
tsdive.switchback.NO_SPREAD	Refusal reason: the kept target has MAD 0, or the covariates reproduce it.
tsdive.switchback.PERMUTATIONS	Largest reference set enumerated in full.
tsdive.switchback.POWER_DELTAS	Shifts the power readout adds to the B blocks, in sigma (1.4826 MAD of the history).
tsdive.switchback.POWER_DRAWS	Seeded schedules the power readout draws over the history window.
tsdive.switchback.REASONS	Every refusal reason analyze can put in Analysis.reason.
tsdive.switchback.TOO_FEW	Refusal reason: fewer than 30 kept samples.
tsdive.switchback.TOO_MANY_COVARIATES	Refusal reason: more than one covariate per 10 kept samples.