TY2022 state one-time rebates: TAXSIM books them in 2022 siitax; PE uses the liability-year convention (GA/VA/NM/ME/CO/MA/ID…)

[PavelMakarchuk]

Extending the eCPS analysis (#1059/#1062) to TY2022 — the big state-rebate year — shows the rebate-timing convention difference (#716 class, third+ instance) at scale. 2022 genuine state diffs (income-scaled tolerance): 20.7% of records, and the flat-amount signature is unmistakable:

Flat clusters in the 2022 state-tax diff (PE − TAXSIM): $500 × 114 records, $250 × 73, $1,000 × 46, $375 × 29, $850 × 37, $300 × 37 — i.e. TAXSIM subtracts the 2022 one-time rebates from TY2022 siitax; PE books rebates to the liability year:

State	2022 diff pattern	Rebate
GA	+$250/$375/$500, 100% PE-higher (n=103)	2022 surplus refund (HB 1302: $250/$375/$500)
VA	+$250/$500 (n=89, 98%)	2022 rebate
NM	+~$500/$1,000 (n=121)	2022 rebates
ME	+$850 × 37	$850 relief payment
HI	±$300	Act 115
MA	proportional +$1,890 median (n=143, 92%)	62F refund = 14.0312% of 2021 liability
CO	TAXSIM siitax deeply negative (flat −$1,703 single incl. zero-income records); +$3,331 median normal-income	Colorado Cash Back $750/$1,500 + TABOR
ID	+$977 median (n=136, 100%)	2022 rebates (12%/special session)
MN	+$2,565 median (n=85)	separate — the known 2021-22 under-computation (#1015)

The 2024 mirror confirms the convention difference: TY2024 shows flat −$200/−$400 clusters (38+29 records) — PE booking the 2025 Virginia rebate to its 2024 liability year, while TAXSIM books it in 2025 (where we saw the mirror +$200/+$400, #1062). Same rebate, opposite sides of the year boundary. GA's recurring surplus refund also appears in 2024 (+$375/$250/$500, n=106).

The question / decision

This isn't a computation bug on either side — it's the rebate-timing convention (#716): PE assigns one-time rebates to the tax year whose liability determines them; TAXSIM subtracts them in the payout/trigger year. With 2022 in scope it's now the largest single driver of state-side differences (hundreds of records/year). Worth settling the convention with @feenberg once (per #716) rather than state-by-state — happy to produce the full state×amount×year mapping if useful.

[PavelMakarchuk]

The TY2021 mirror confirms the convention difference end-to-end. PE books each rebate to the year whose liability determines it; TAXSIM subtracts it in the payout year. The same rebates therefore appear as PE-lower flat clusters in TY2021 and PE-higher in TY2022:

State	TY2021 (assessment yr, PE applies)	TY2022 (payout yr, TAXSIM applies)	Rebate
VA	−$486 med, 98% PE-lower (n=80)	+$250/$500	capped at 2021 liability
MA	−$718, 97% (n=116)	+$1,890	62F = 14.0312% of 2021 liability
ME	−$850 exactly, 100% (n=51)	+$850	$850 relief
GA	−$375, 99% (n=102)	+$375	surplus refund
NM	−$1,293, 100% (n=122)	+~$500/$1,000	2021-liability rebates
HI	−$410, 100% (n=126)	+$300	Act 115
CO	−$758, 82% (n=94)	Cash Back	2021 TABOR
IL	−$100, 99% (n=117)	—	2021-liability rebate

TY2021 flat clusters: −$500×106, −$250×81, −$800×54, −$1,300×37, −$850×24, −$750×24.

Because the difference is a timing one, it hits the match rate twice — once in each year, opposite signs — so settling the convention (#716) recovers alignment in both 2021 and 2022 at once. Exception: ID has no 2021 mirror (its 2022 rebates keyed to 2020/2019 liability, so PE books them before our window). Alongside the VA-2025↔2024 mirror already documented, every payout-year cluster now has its assessment-year counterpart located.

[feenberg]

On Mon, 6 Jul 2026, Pavel Makarchuk wrote: PavelMakarchuk created an issue (PolicyEngine/policyengine-taxsim#1068) Extending the eCPS analysis (#1059/#1062) to **TY2022** ? the big state-rebate year ? shows the rebate-timing convention difference (#716 class, third+ instance) at scale. 2022 genuine state diffs (income-scaled tolerance): 20.7% of records, and the flat-amount signature is unmistakable: **Flat clusters in the 2022 state-tax diff (PE ? TAXSIM):** $500 ? 114 records, $250 ? 73, $1,000 ? 46, $375 ? 29, $850 ? 37, $300 ? 37 ? i.e. TAXSIM subtracts the 2022 one-time rebates from TY2022 `siitax`; PE books rebates to the liability year: | State | 2022 diff pattern | Rebate | ---|---|---| GA | | +$250/$375/$500, 100% PE-higher (n=103) | 2022 surplus refund (HB | 1302: $250/$375/$500) | VA | +$250/$500 (n=89, 98%) | 2022 rebate | NM | | +~$500/$1,000 (n=121) | 2022 rebates | ME | +$850 ? 37 | $850 relief | payment | HI | ?$300 | Act 115 | MA | **proportional** +$1,890 median | (n=143, 92%) | 62F refund = 14.0312% of 2021 liability | CO | TAXSIM | siitax deeply negative (flat ?$1,703 single incl. zero-income | records); +$3,331 median normal-income | Colorado Cash Back | $750/$1,500 + TABOR | ID | +$977 median (n=136, 100%) | 2022 rebates | (12%/special session) | MN | +$2,565 median (n=85) | separate ? the | known 2021-22 under-computation (#1015) | **The 2024 mirror confirms the convention difference:** TY2024 shows flat **?$200/?$400** clusters (38+29 records) ? PE booking the **2025 Virginia rebate to its 2024 liability year**, while TAXSIM books it in 2025 (where we saw the mirror **+$200/+$400**, #1062). Same rebate, opposite sides of the year boundary. GA's recurring surplus refund also appears in 2024 (+$375/$250/$500, n=106). ## The question / decision This isn't a computation bug on either side ? it's the **rebate-timing convention** (#716): PE assigns one-time rebates to the tax year whose liability determines them; TAXSIM subtracts them in the payout/trigger year. With 2022 in scope it's now the **largest single driver of state-side differences** (hundreds of records/year). Worth settling the convention with @feenberg once (per #716) rather than state-by-state ? happy to produce the full state?amount?year mapping if useful.

Taxsim has option 30 to do a number of things to make taxsim more consistent with PE. IT sets options: 27 - rebates assigned to eligibility year 88 - turn off smoothing where possible 91 - stop iterating (no deduction of income taxes) The production default for 27 is zero, that selects the paid year. It looks like option 27 isn't being set for you. Is that possible? Generally for testing option 30 should be set. It would be like: mstat,year,pwages,opt1,opt1v 2,2025,10000,30,1 in the txpydata.raw file. I want to use the year paid so that users requiring after tax income get the most suitable result. The year eligible would be better for users that want the marginal tax rate, but even for those users the value isn't really right, since it wasn't known during that tax year. Dan

…

-- Reply to this email directly or view it on GitHub: #1068 You are receiving this because you were mentioned. Message ID: ***@***.***>

[PavelMakarchuk]

Implementation up: PR #1070. The emulator now emits its own srebate (computed as state_income_tax with one-time rebates zeroed minus actual — exactly what rebates moved siitax), and compare --net-of-rebates scores siitax + srebate on both sides, removing the payout-year vs liability-year convention from match scoring without changing either engine.

Validated on the 8k eCPS: 2021 state match 75.50% → 87.08% (+11.58), 2022 79.30% → 83.30% (+4.00), federal identical both years. ME 2021 srebate = $850.00 and VA = $250.00 to the penny. So @feenberg — no convention change needed on either side for comparison purposes; the residual (~13-17%) is now genuinely non-rebate material.

[feenberg]

The txpydata.csv file supplied in #1050 does not set option 30 or option 27, so it is no surprise that the rebates are attributed to the wrong year.

[PavelMakarchuk]

You're right, and this changes the picture substantially — thanks for the pointer. I mapped opt(27)=1 (rebates in eligible year — PolicyEngine's convention) across the rebate states on the 2026-06-22 binary, one $60k joint wage record per state × 2021/2022:

State	default → opt27, TY2021	default → opt27, TY2022	vs PE
VA	2,568.05 → 2,068.05 (srebate 500)	1,665.55 → 2,165.55	✅ matches PE exactly
GA	2,444.50 → 1,944.50 (500)	1,881.25 → 2,381.25	✅
CO	1,487.77 → −12.23 (1,500 Cash Back)	−1,834.09 → −334.09	✅
MA	2,460.00 → 2,114.84 (62F 345.16)	2,114.84 → 2,460.00	✅
SC	1,914.20 → 1,114.20 (800)	—	✅
DE, ID	unchanged in 2021 (rebates keyed to 2019/2020 liability)	2022 rebate removed	✅ consistent
ME	unchanged — the $850 relief is booked in neither year under opt27	611.60 → 1,461.60 (removed from 2022)	❌ PE books $850 in 2021
HI	unchanged both modes/years — Act 115 absent	unchanged	❌ PE books it in 2021
MT	2,518.22 → 18.22 ($2,500 income-tax rebate into 2021)	—	❌ reversed: PE computes mt_income_tax_rebate but doesn't net it into siitax — that's a PolicyEngine wiring gap we'll file
IL, IN	srebate reported but siitax unchanged in both modes		worth a look

Also verified opt(30)=1 resolves the MI Home Heating Credit (#1050, withdrawn — matches PE to the penny) and removes the 2021 OR kicker (#1071).

Protocol conclusion: our bulk comparisons ran the binary in default mode while your record testing runs opt(30)=1 — so the rebate clusters in this issue measured the mode difference, not disagreement. We'll switch the eCPS comparisons to opt(30)=1 (and keep the srebate/net-of-rebates scoring for default-mode runs). Remaining substantive items from the matrix: ME and HI missing from the opt27 set on your side, and the MT wiring gap on ours.

[feenberg]

Which states do I need to look at in Taxsim?

…

On Tue, Jul 7, 2026 at 4:05 PM Pavel Makarchuk ***@***.***> wrote: *PavelMakarchuk* left a comment (PolicyEngine/policyengine-taxsim#1068) <#1068 (comment)> You're right, and this changes the picture substantially — thanks for the pointer. I mapped *opt(27)=1* (rebates in eligible year — PolicyEngine's convention) across the rebate states on the 2026-06-22 binary, one $60k joint wage record per state × 2021/2022: State default → opt27, TY2021 default → opt27, TY2022 vs PE VA 2,568.05 → *2,068.05* (srebate 500) 1,665.55 → 2,165.55 ✅ matches PE exactly GA 2,444.50 → *1,944.50* (500) 1,881.25 → 2,381.25 ✅ CO 1,487.77 → *−12.23* (1,500 Cash Back) −1,834.09 → −334.09 ✅ MA 2,460.00 → *2,114.84* (62F 345.16) 2,114.84 → 2,460.00 ✅ SC 1,914.20 → *1,114.20* (800) — ✅ DE, ID unchanged in 2021 (rebates keyed to 2019/2020 liability) 2022 rebate removed ✅ consistent *ME* unchanged — the $850 relief is booked in *neither* year under opt27 611.60 → 1,461.60 (removed from 2022) ❌ PE books $850 in 2021 *HI* unchanged both modes/years — Act 115 absent unchanged ❌ PE books it in 2021 *MT* 2,518.22 → *18.22* ($2,500 income-tax rebate into 2021) — ❌ reversed: PE computes mt_income_tax_rebate but doesn't net it into siitax — that's a PolicyEngine wiring gap we'll file IL, IN srebate reported but siitax unchanged in both modes worth a look Also verified opt(30)=1 resolves the MI Home Heating Credit (#1050 <#1050>, withdrawn — matches PE to the penny) and removes the 2021 OR kicker (#1071 <#1071>). *Protocol conclusion:* our bulk comparisons ran the binary in default mode while your record testing runs opt(30)=1 — so the rebate clusters in this issue measured the mode difference, not disagreement. We'll switch the eCPS comparisons to opt(30)=1 (and keep the srebate/net-of-rebates scoring for default-mode runs). Remaining substantive items from the matrix: ME and HI missing from the opt27 set on your side, and the MT wiring gap on ours. — Reply to this email directly, view it on GitHub <#1068?email_source=notifications&email_token=AB55AVOWKUKTW7EBLE64MMT5DVJZPA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOJQHAYTOOBQHEZ2M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4908178093>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB55AVKKYOJEMZJYDI5HFPL5DVJZPAVCNFSNUABFKJSXA33TNF2G64TZHM4DAOBTGYZDOOBXHNEXG43VMU5TIOBSGIZTKNBVGY2KC5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AB55AVPIV4T6GXIGM3DSMY35DVJZPA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOJQHAYTOOBQHEZ2M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/AB55AVNWRYJ2VHOUXCEXQE35DVJZPA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOJQHAYTOOBQHEZ2M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>