Appendix E

Evidence on Strengths and Limitations of IRS and Survey Data to Measure EITC and CTC Take-Up

Survey and administrative data have strengths and limitations. Nationally representative large-scale surveys, hosted by federal agencies, also have strengths and limitations relative to more targeted surveys (e.g., they extend over a certain time period or for a defined population that might be too small or too difficult to find in large population-level surveys). The Current Population Survey (CPS), for example, represents the civilian noninstitutionalized population, thus excluding people in the armed forces, or individuals in detention centers, prisons, jails, and nursing and related mental health facilities, who are in fact eligible for tax credits. It is also more difficult for ongoing national surveys, including the CPS, to capture certain eligible populations because of their changing circumstances (e.g., being unhoused). Survey data also rely on self-reports of receipt, which can suffer from recall, social desirability, and related biases. Many surveys fail to ask about tax filing status or receipt of tax credits, further complicating matters. Data from the Internal Revenue Service (IRS) and comparable agencies overcome many of these challenges by more completely representing the population receiving benefits and, importantly, accounting for benefit receipt more accurately. However, identifying information for certain demographic groups is less reliable and no information is available for people who are not connected to data collection (e.g., detailed information is often lacking for individuals who do not file taxes). In this appendix, these and related issues are discussed in the context of considering trade-offs and their implications for estimates of Earned Income Tax Credit (EITC) and Child Tax Credit (CTC) take-up or receipt.

FINDINGS ON MEASUREMENT OF INCOME AND PROGRAM RECEIPT AND BENEFITS IN LINKED CPS ASEC IRS DATA

Previous work has highlighted underreporting of income and program receipt in survey data, which may contribute to underestimates of program take-up and of earned income (Meyer et al., 2009, 2022). Further, survey datasets may fail to measure tax filing or receipt of tax credits. The CPS Annual Social and Economic Supplement (ASEC) deals with this by using tax models to impute EITC, CTC, and other tax credit accrual (see Chapter 3). Indeed, several studies that used survey data matched to administrative data documented discrepancies between the two measures of earned income and adjusted gross income (AGI; e.g., Bee et al., 2023a; Jones & Ziliak, 2022; Meyer et al., 2022; Unrath, 2022). Jones and Ziliak (2022), for instance, matched the CPS ASEC to IRS 1040 information to estimate the impact of the EITC on reducing poverty between 2005 and 2016. They found that the CPS ASEC overstated the antipoverty impact of the EITC by as much as 35%, and up to 45% for children.

According to Jones and Ziliak (2022), a very important reason why the CPS ASEC overstates the impact of the EITC on poverty is discrepancies in earnings between survey and administrative data. They showed that the CPS ASEC produced very similar numbers as IRS data in terms of the number of people lifted out of poverty after accounting for imputed earnings and individuals who cannot be matched to IRS data. They recommended that researchers using CPS ASEC data remove individuals with imputed earnings, as well as those who have the entire CPS ASEC imputed, and reweight the remaining data to account for this adjustment (see also Bollinger & Hirsch, 2013). They further recommended removing noncitizen Hispanic individuals from the data and reweighting accordingly to account for individuals who cannot be matched to IRS data. This approximated individuals who could not be matched to IRS data using Protected Identification Keys (PIKs)¹ and often encompassed individuals who were not eligible for the EITC.

After making these adjustments, Jones and Ziliak (2022) found that the CPS ASEC approximated 94% of the IRS outlays. The remaining reasons for the discrepancies between the two data sources were largely driven by differences in the number of children claimed (i.e., there were more individuals in the IRS 1040 data who appeared to have qualifying children compared to the CPS ASEC) and differences driven by those with self-employment income. Inconsistencies in tax filing status (e.g., a married couple files as two separate head-of-household (HoH) tax units when they should file jointly) represented a smaller issue contributing to the mismatch.

___________________

¹ PIKs are a unique personal identifier assigned when linking records.

While Jones and Ziliak (2022) argued that making these adjustments brought the CPS ASEC estimates closer to the IRS outlays, they cautioned that there were two main sources of mismeasurement between the data sources underlying this match: the CPS ASEC assumed 100% take-up in their tax calculator/modeling, when in reality, take-up of the EITC was closer to 80% (Jones, 2014); and IRS 1040 data potentially contained erroneous payments. These two issues, they argued, offset each other and thus the aggregate CPS ASEC estimates were quite close to the IRS estimates after making the adjustments discussed above.

In interpreting these results, it is important to keep in mind that it is impossible to say with certainty which of the two data sources is the “correct” source. The definition of a qualifying child is complex and relies on residency information that is not available in survey data or IRS 1040 data. For instance, for a tax filer to claim a dependent as a qualifying child for the purposes of the EITC, the dependent must reside with the filer for at least six months of the year. It is impossible to ascertain this information from either the CPS ASEC or IRS 1040 data. In instances where a dependent is present in IRS 1040 data but not present in the CPS ASEC household roster, it is unclear whether this is a fraudulent claim, or whether the CPS ASEC household roster does not accurately reflect the residential situation of the tax filing unit (TFU). This is further complicated by the fact that the CPS ASEC is administered in March of each year and asks respondents about their income from the past tax year, but all household roster information reflects the housing arrangement at the time of the survey. This presents an issue, particularly in more complex family arrangements in which children might not always reside in the same household. For instance, a child of divorced parents who splits their time equally between both parents may show up on one parent’s household roster but be claimed as a dependent by the other parent. These might be bona fide TFUs that are not accurately represented in CPS ASEC households.

Meyer et al. (2022) conducted a similar exercise as Jones and Ziliak (2022) but used a wider array of administrative data sources such as the Social Security Administration’s Detailed Earnings Records, information from 1099-Rs, which provide information on retirement distributions, and 1099-Gs, which provide information on Unemployment Insurance income. Their goal was to simulate tax liabilities and credits more broadly, while Jones and Ziliak (2022) focused specifically on matching EITC totals. Meyer et al. (2022) showed that when earnings were imputed into the CPS ASEC using findings from their administrative data sources, the resulting estimates of tax liabilities and credits more closely matched IRS Statistics of Income (SOI) data than when using CPS ASEC data alone or a simpler IRS-based imputation (e.g., relying only on 1040 tax return information). After imputing extensive information from administrative data, the authors

found that tax liabilities and credits fell within 10% to 13% of published SOI numbers, while a more limited imputation of information only from IRS 1040 tax return data fell short of SOI numbers by more than 20% (Meyer et al., 2022).

FORMATION OF TAX FILING UNITS AND DETERMINING DEPENDENTS: FINDINGS FROM LINKED CPS ASEC AND IRS DATA

Unrath (2022, 2024) focused discrepancies in definitions of TFUs between the CPS ASEC and IRS 1040 information for the 2019 tax year, to quantify the magnitude of this issue. He relied on CPS ASEC-constructed TFUs, which are imputed by the Census Bureau. Using a similar approach as Jones and Ziliak (2022) in matching the two data sources, Unrath (2022) documented that less than 60% of modeled TFUs in the CPS ASEC directly mapped onto identical TFUs in IRS 1040 data. Discrepancies were higher for HoH and single filers, and lower for married couples, particularly those with no dependents.

Like Jones and Ziliak (2022), Unrath (2022, 2024) found that there were generally more children claimed in the IRS 1040 data than appeared in the CPS ASEC; among children who did appear in the CPS ASEC, the vast majority were claimed by the individual predicted by the CPS ASEC tax calculator. Among children in the CPS ASEC who could be matched to IRS 1040 data, 87% were claimed by the tax filing head or spouse predicted from the CPS ASEC tax model; the vast majority of the remaining children (more than 80%) were claimed by someone in the CPS ASEC household—in most cases by the birth parent, but otherwise another related individual in the household. Distributionally, these discrepancies were more common among lower-income households, as well as for Black and Hispanic children. In particular, Unrath showed that less than 50% of children residing in CPS ASEC households where the TFU had zero W-2 income were claimed as dependents by that TFU. Above $20,000 in W-2 income, he found fairly uniform claiming rates of children across the income distribution.

Despite these discrepancies in TFUs between the two data sources, Unrath (2022) illustrated that reassigning TFUs based on IRS 1040 data did little to impact post-tax and transfer income across the income distribution. Unfortunately, due to the sensitivity of the data, Unrath (2022) was unable to provide concrete guidance to users of the public-use CPS ASEC to address these issues, but future work by those at the Census Bureau will hopefully provide these recommendations.

FINDINGS FROM LINKED CPS ASEC AND IRS DATA FOR TAX YEAR 2021

To assess the accuracy of the model, Bee et al. (2023b) analyzed the modeling of the eligibility and credit amounts for the 2021 provisions of the CTC using the CPS ASEC tax model and compared them against information in linked IRS 1040 tax forms and aggregate payment results compiled by the IRS’s SOI. This analysis is particularly relevant for the analyses conducted for the 2021 CTC changes produced using the Urban Institute’s Transfer Income Model version 3 (TRIM3) that are discussed in Chapter 8. As discussed in Chapter 8 and in Appendix H, TRIM3 generates imputations of income sources, notably various safety net program benefits, and makes somewhat different assumptions in constructing TFUs than is done by the CPS ASEC tax model. Findings from Bee et al. (2023b) provided some quantitative evidence regarding how model-based estimates of CTC eligibility and amounts, based solely on 2022 CPS ASEC data, captured the control totals for CTC receipt and its distribution in 2021 reported by the IRS’s SOI for tax year (TY) 2021.

To validate the CPS ASEC tax model’s predictions of CTC credit receipt/eligibility and amounts, Bee et al. (2023b) compared model-based estimates of each based solely on CPS ASEC data with estimates that made use of the information on earnings and dependents in linked IRS 1040 records. They found that 83.1% of individuals in the CPS ASEC were in TFUs that were predicted to be eligible for the CTC by both the CPS ASEC tax model and based on IRS 1040 earnings and dependents information. At the same time, for those cases in which the CTC amounts predicted by the CPS ASEC tax model differed from estimates based on IRS 1040 information, 17.6% underestimated the CTC amount, while 22.8% overestimated it. In short, they found evidence that CTC amounts based on CPS ASEC survey information did not always match those modeled from linked IRS 1040 information. While 48.4% of TFUs have the same number of children based on the CPS ASEC tax model and linked IRS 1040 records, 16.2% of units with a child in the linked 1040 data did not have a child in the CPS ASEC household roster, while 35.9% of the CPS ASEC TFUs had a child but did not have one in the linked IRS 1040 data. Bee et al. noted that this latter discrepancy in qualifying children was not due to differences in the total number of children in these two sources of information, but rather to the misallocation of children to TFUs, which may result from children who were not in the CPS ASEC household roster recorded in spring 2022 being claimed by these sample households on their 2021 tax returns.

Finally, Bee et al. (2023b) examined how the (population-weighted) estimates of CTC credit eligibility/receipt amounts received and the number of qualifying children in TFUs estimated to receive these credits based

on the CPS ASEC tax model compared in aggregate with control totals reported in IRS’s SOI for TY 2021. In aggregate, they found that estimates from the CPS ASEC tax model were fairly close to the IRS control totals reported in the SOI. However, they also found substantial differences between the model-based estimates and the control totals in the IRS SOI across the distribution of AGI. In particular, the CPS ASEC tax model overpredicted the number of TFUs receiving the CTC, the total amount of credits received, and the number of qualifying children for AGI less than $10,000. It underpredicts these values for AGI between $10,000 and $50,000, with substantial underprediction of TFUs near Supplemental Poverty Measure (SPM) poverty thresholds (i.e., AGI between $20,000 and $30,000). It also overpredicts these same values for high levels of AGI ($75,000 and above), especially for those TFUs with very high incomes ($200,000 or more). As the authors noted, the finding that the CPS ASEC tax model substantially underpredicted IRS control totals for the CTC aggregates in TY 2021 for those with incomes near SPM poverty thresholds raised concern that previous Census Bureau estimates of the impact of the American Rescue Plan Act of 2021 CTC Expansion on child poverty were substantially underestimated (Burns & Fox, 2022).

FINDINGS FROM LINKED CPS ASEC AND IRS DATA FOR TAX YEAR 2018

This section describes the findings from an analysis of tabulations from CPS ASEC data linked to TY 2018 IRS tax return data,² assembled by Bee and Unrath (2025).³ CPS ASEC data for this year as well as for 2022 (reporting on 2021 outcomes) are the data source for TRIM3 in the analyses reported elsewhere in this report, although TRIM3 modeling for the formation of TFUs and for predictions of EITC and CTC eligibility and amounts are not the same as used by Bee and Unrath (2025).

The analysis presented here compared how the Census Bureau’s tax model, referred to below as the CPS ASEC tax model, constructs TFUs, (referred to as CPS ASEC TFUs) and their tax filing statuses (TFSs). These comparisons were generated to determine how well the CPS ASEC tax model predicts the IRS filing statuses. In addition, the analysis reported on how well the CPS ASEC tax model predicts EITC and CTC eligibility and credit amounts to be received by CPS ASEC TFUs whose TFSs match the IRS TFS and those that do not. This latter comparison, while indirect,

___________________

² The 2019 CPS ASEC asked households to report on their income, work experience, and other economic characteristics for the previous calendar year (i.e., calendar and tax year 2018).

³ This technical document and three accompanying tables (in Excel format) are found at https://www.census.gov/library/working-papers/2024/demo/sehsd-wp2024-32.html

provides some information on how well the CPS ASEC tax model can predict EITC and CTC eligibility and credit amounts.

Finally, findings in Bee and Unrath (2025) are examined with respect to differences between the number of dependents CPS ASEC TFUs are estimated to claim and the number actually claimed on their tax returns. These data provide indirect evidence on the likelihood and extent to which CPS ASEC TFUs may be claiming dependents who are not in the CPS ASEC households at the time the survey is conducted. Bee and Unrath (2025) did not attempt to explain the findings presented in the tabulations. Such analyses will presumably be provided in future reports. Rather, these tabulations are intended to provide “aid [to] technically advanced users of CPS ASEC tax variables to adjust corresponding variables in the public-use CPS ASEC” (Bee & Unrath, 2025, p. 2).

Before summarizing the outcomes from the linked data described above, the CPS ASEC tax model is briefly described, along with key features of the process used to link heads of CPS ASEC TFUs to IRS tax returns, with a focus on the steps undertaken in constructing the tabulations.

The CPS ASEC Tax Model

The 2019 CPS ASEC did not include questions about whether anyone in the surveyed households filed a tax return. Using the CPS ASEC tax model, TFUs are determined based on data from household rosters in the CPS ASEC sample. The model selects a TFU head and assigns children and other adults within the household to a TFU. These units are referred to as CPS ASEC TFUs. The CPS ASEC tax model determines these TFUs using only CPS ASEC data, without reference to IRS linked data.⁴ Note that there could be more than one TFU per CPS ASEC household.

The CPS ASEC tax model then predicts the TFS for each CPS ASEC TFU. Again, this determination is based only on CPS ASEC data. The model allows for the following TFSs: married filing jointly (MFJ), HoH, and single (with no children).⁵

The CPS ASEC tax model also determines whether a CPS ASEC TFU is a nonfiler (i.e., that TFU is predicted to have not filed a tax return). This is based on a set of filing requirements, including income above the IRS filing threshold, positive self-employment income, and others (see Lin, 2022, for details).

___________________

⁴ See Lin (2022) for a detailed description of how the CPS ASEC tax model constructs TFUs and determines their tax filing statuses. These procedures are summarized in Bee and Unrath (2025).

⁵ The single with no dependents category also includes those TFUs filing as widows and other groups, although they are a relatively small share of this filing status. See Bee and Unrath (2025) for the rationale behind combining these TFSs into a single group.

Note that the CPS ASEC tax model cannot identify qualifying children for a tax return if those children do not reside within the CPS ASEC household. For example, a parent in a CPS ASEC household cannot claim children that reside elsewhere at the time of the survey. This implies that a CPS ASEC TFU may not include all the qualifying children claimed on the tax return in that TFU and thus the model may misclassify the TFS of a TFU. For example, it may classify the TFU as a MFJ with no children filer or a single filer, rather than a MFJ with children or a HoH with children filer. (See Lin, 2022, for other limitations of the tax model.)

The CPS ASEC tax model was used to determine the CPS ASEC TFUs and their TFSs for the 2019 CPS ASEC data for TY 2018.

Matching CPS ASEC TFUs to IRS Data

CPS ASEC TFUs for TY 2018 were matched to TY 2018 IRS tax return data for the Census Bureau’s National Experimental Well-Being Statistics (NEWS) project team. The NEWS team first obtained PIKs for each CPS ASEC TFU head. The PIK allows linkage to IRS tax return data and other databases, such as Social Security Administration records.

Predicted Tax Credit Eligibility and Amounts for CPS ASEC TFUs

As noted above, the 2019 CPS ASEC did not ask whether anyone in its households filed a tax return. Furthermore, the 1040 tax forms linked to CPS ASEC TFUs did not contain information on whether particular tax credits were claimed or the amounts TFUs received. Bee and Unrath (2025) used TAXSIM⁶ to model EITC and CTC eligibility and the credit amounts those predicted to be eligible would receive. Eligibility and credit amounts were predicted for each CPS ASEC TFU based on estimates of income and earnings and assignment of dependents. The working paper describes recent efforts to improve households’ underreporting of various income sources in the CPS ASEC that were identified in previous work using versions of the CPS ASEC linked to tax returns and other data sources.⁷

Structure of Tabulations

Findings described in Bee and Unrath (2025) are based on tabulations in Excel files that accompany the paper. These tabulations are organized

___________________

⁶ See Feenberg and Coutts (1993) for details on TAXSIM.

⁷ For previous work on this issue, including its consequences for measuring poverty, see Bee and Mitchell (2017), Bee et al. (2023a, 2025), Bollinger et al. (2019), Jones and Ziliak (2022), Larrimore et al. (2020), Meyer and Mittag (2019), Meyer et al. (2020, 2022).

into race/ethnic groups by levels of earnings from the CPS ASEC data, for four race/ethnic groups (non-Hispanic White, Hispanic, non-Hispanic Black, and non-Hispanic Asian individuals and other races) and four earnings (E) groups (E < $25,000, $25,000 ≤ E < $40,000, $40,000 ≤ E < $60,000, E ≥ $60,000). Tabulations in Bee and Unrath (2025) are based on applying the CPS ASEC tax model described above. Within each race/ethnicity and earnings group, they included detailed tabulations for specific TFSs (i.e., MFJ, HoH, etc.) that also included numbers of dependents and the genders of the heads of HoH TFUs. The analysis aggregated over the latter details and reported only on TFSs—MFJ, HoH, and single—as well as for nonfiling status.

Tabulations in the files accompanying Bee and Unrath (2025) were subject to suppression of cells with TFUs that had small sample sizes, to reduce disclosure risks per Census Bureau rules. In general, cell suppression was greater for minority groups, especially the non-Hispanic Asian individuals and other races group, and for some cells representing higher earnings groups.⁸ The approach used to handle suppressed cells in the analyses presented here is described below. Suppression potentially complicates the interpretation of certain results.

Tables E-1 through E-4 present results from comparisons between the CPS ASEC tax model and actual outcomes found in linked IRS 1040 forms. In all cases, tables were created by aggregating information from the tabulations in Bee and Unrath (2025). Weighted averages of particularly disaggregated cells were calculated, using the weighted shares of the CPS ASEC TFU samples associated with those cells. Suppressed cells were treated as missing and excluded from the weighted averages calculations. Additional details of table construction are provided in the notes accompanying each table.

Match (or PIK) Rates for CPS ASEC TFUs

The shares (in percentages) of adults in CPS ASEC households who have PIKs and can thus be linked to IRS tax returns are displayed in Table E-1. Panel A presents PIK rates by CPS ASEC tax model TFSs, while Panel B presents PIK rates for race/ethnic and earnings groups. Among all CPS ASEC household adults, 85.9% had PIKs, which is similar to PIK rates found in other studies of linked CPS ASEC–IRS data. As shown in Panel A, PIK rates are higher for the MFJ and single (no dependents) statuses and lower for TFUs that were predicted to be nonfilers or HoH filers by the

___________________

⁸ The cells used to compute match rates of CPS ASEC and 1040 filing statuses for nonfilers in earnings groups $25,000 ≤ E < $40,000, $40,000 ≤ E < $60,000, and E ≥ $60,000 were all suppressed. Similarly, tabulations for EITC and CTC eligibility and average credit amounts that accompany Bee and Unrath (2025) were also subjected to extensive suppression.

TABLE E-1 Rates of Adults in CPS ASEC Households with Protected Identification Keys (PIKs)

NOTES: Entries are aggregated from detailed tabulations in taxcreditimputationv5at1.xlsx distributed with Bee and Unrath (2025). PIK rates used are from column B of this file. These rates were the share of all CPS ASEC TFUs formed by the CPS ASEC tax model, including TFUs in which all adults in a TFU had a PIK and those that did not meet this condition. The total numbers of TFUs formed by the CPS ASEC tax model for each filing status were determined by dividing the number of PIKed TFUs in column Q by the PIK rates in column B. Cell-specific numbers of PIKed TFUs in column B were rounded as per Census Bureau disclosure rules. Some cells in column B were suppressed due to small sample sizes, again per Census Bureau disclosure rules. Suppressed cells were omitted from the calculations of weighted averages used to form the aggregated values included in this table.

^a Sample sizes (N) in this table are the sum of PIKed and nonPIKed CPS ASEC TFUs. The number of nonPIKed TFUs is estimated by dividing the number of PIKed TFUs by the corresponding PIK rate.

CPS ASEC tax model. These PIK rates are similar to those found in Unrath (2022, 2024) using data from the 2019 and 2021 CPS ASEC linked to IRS tax records for 2018 and 2020, respectively; although the PIK rate in Table E-1 for nonfilers was higher than in Unrath (2022, 2024; 80.4% vs. 68%) and slightly lower for HoH filers (88.0% vs. 93%) for TY 2020.

As shown in Panel B, PIK rates were lower for minority groups, especially Hispanic individuals (77.4%) and non-Hispanic Asian individuals and other races (79.1%). PIK rates also tended to be lower for adults in TFUs with lower earnings—E < $25,000 (83.6%) and $25,000 ≤ E < $40,000 (83.7%)—compared to adults in TFUs with higher earnings.

As noted in the Table E-1 notes, PIK rates were affected by suppression of PIK rates in the tabulations accompanying Bee and Unrath (2025), which was more frequent for both non-Hispanic Black individuals and non-Hispanic Asian individuals and other races and for the two middle earnings groups.

CPS ASEC Tax Model Predictions of Incidence of Tax Filing and Filing Statuses Versus Incidence in Matched IRS Tax Returns

Table E-2 compares TFSs predicted by the CPS ASEC tax model and the TFSs in the linked IRS 1040 data for CPS ASEC TFUs. Panel A displays results across all earnings and race/ethnic groups, Panel B displays the same tabulations broken out for each earnings group, and Panel C for each race/ethnic group. In each panel, the incidence of each TFS is provided based on the CPS ASEC tax model and the TFSs found in 1040 tax returns, as well as the differences between them. Finally, each panel provides the percentages of rates for which the TFSs from the CPS ASEC tax model match the statuses on 1040 linked tax data.

Across all race/ethnic and earnings groups (Panel A), the CPS ASEC tax model overpredicted nonfiling status by 11.2 percentage points; the MFJ and HoH statuses were slightly underpredicted by the CPS ASEC tax model (–3.1 and –4.3 percentage points, respectively) while the single status was slightly overpredicted by the CPS ASEC tax model (4.0 percentage points). These overall tabulations mask important differences found for the various earnings and race/ethnic groups in Panels B and C. For the lowest income group (i.e., E < $25,000) the CPS ASEC tax model overpredicted TFUs as being nonfilers compared to the actual incidence of nonfiling based on the IRS 1040 data (Panel B). This was also true for each race/ethnic group, with overprediction being higher for Hispanic, non-Hispanic Black, and Asian individuals and other races (Panel C). This overprediction of nonfilers by the CPS ASEC tax model for the lowest earnings group accounts, in part, for the shortfall of tax filers claiming the EITC and CTC in the CPS ASEC relative to control totals for tax filing reported by the IRS. In contrast, for

TABLE E-2 Tax Filing Status Based on CPS ASEC Tax Model and IRS 1040s and Match Rates Between Them

NOTES: Entries are aggregated from detailed tabulations in taxcreditimputationv5at1.xlsx, distributed with Bee and Unrath (2025). The units in this file consist of all CPS ASEC TFUs in which all adults in a TFU had a PIK. The number of TFUs displayed in this table, labeled N, are the sums of entries from column Q of the Excel file for the various TFSs. Entries in this table for shares of the various TFSs, reported as percentages, are formed from entries in columns C–F of the Excel file, weighting them by the sample sizes of PIKed CPS ASEC TFUs from column Q in the Excel file. The sample sizes of the rows in the Excel file were rounded as per Census Bureau disclosure rules. Some cells in the Excel file were suppressed due to Census Bureau disclosure rules. Suppressed cells were omitted from the calculations of the weighted averages used to form the aggregated values in this table.

^a Match rates (as percentages) for the CPS ASEC filing statuses with those on 1040s are found in columns C–F of taxcreditimputationv5at1.xlsx. Entries in the rows of the Excel file are aggregated, weighting them by the numbers of PIKed CPS ASEC TFUs in column Q of the file. See the table note for how suppressed cells were treated in these calculations.

^b Sample sizes (N) in this table are the number of PIKed CPS ASEC TFUs.

higher-income TFUs (Panel B), the CPS ASEC tax model tended to underpredict tax filing relative to nonfilers based on CPS ASEC data linked to IRS 1040 data. (While not shown in Table E-2, this underprediction was even higher for Hispanic and non-Hispanic Black individuals in the lowest earnings group.)

With respect to prediction of TFSs, the CPS ASEC tax model tended to overpredict the incidence of MFJ for higher-income CPS ASEC TFUs (i.e., E > $25,000) compared to IRS 1040 data, but it underpredicted this filing status for those TFUs with low incomes (i.e., those with E < $25,000) and underpredicted TFUs filing as HoH at higher income levels compared to linked IRS 1040 data (see Panel B). Patterns for under- and overprediction by the CPS ASEC tax model relative to statuses in linked IRS 1040 data across race/ethnic groups in Panel C for those who filed tax returns as MFJ, HoH, or single statues were more varied.

Across the three panels, match rates between TFSs from the CPS ASEC tax models and those in the IRS 1040 data, including nonfilers, vary by CPS ASEC TFS. For CPS ASEC–designated nonfilers, the match rate was 47.5% across all earnings and race/ethnic groups, with higher rates observed among non-Hispanic Black individuals. However, for TFUs with earnings of $40,000 or more, match rates for the nonfiler status were significantly lower. In these cases, match rates could not be determined due to the suppression of nonfiler cells in the tabulations distributed with Bee and Unrath (2025). At the same time, match rates for MFJ status were much higher (86.2% for all earnings and race/ethnic groups combined), with the match rates for HoH and single TFSs falling between the nonfiler and MFJ statuses (71.9% and 70.5%, respectively). Similar patterns held across earnings groups (Panel B) and race/ethnic groups (Panel C), although match rates for the HoH status tended to be lower for higher-earnings groups, and rates for the MFJ status were lower for non-Hispanic Black and Hispanic individuals, higher for HoH status, and lower for single status.

Differences in Number of Dependents in CPS ASEC TFUs Versus IRS Tax Returns

Table E-3 presents data on the differences between the number of dependents that the CPS ASEC tax model determined TFUs had versus the number found on IRS 1040 forms for those filing tax returns. Details of estimate generation are provided in table notes. Overall, the CPS ASEC tax model underpredicted the number of dependents found in IRS 1040 data by −0.05. However, tax model predictions differed by TFU earnings and race/ethnicity. In particular, the difference between the number of dependents predicted by the model and the number of TFUs actually claimed on tax returns declined across earnings groups, with those in the highest group

TABLE E-3 Mean Differences in Number of Dependents in CPS ASEC TFUs Versus Those Found in IRS 1040s by CPS ASEC Tax Model Tax Filing Statuses

NOTES: Entries are aggregated from detailed tabulations contained columns B–D in taxcreditimputationv5at2.xlsx, distributed with Bee and Unrath (2025). To calculate the average differences in this table, weighted averages of the differences in number of dependents by the 1040 TFSs (MFJ, HoH, single) are calculated in columns B–D in each row in the Excel file, using the shares of tax filers with these statuses from columns D–F in the Excel file as weights. Aggregated values in this table were formed by taking weighted averages of the cells in the Excel file, using the sample sizes in column Q. Suppressed cells were omitted from the calculations of the weighted averages used to form the aggregated values in this table.

(E > $60,000) claiming slightly more dependents than predicted by the CPS ASEC model (0.06). There are also important differences by race and ethnicity. The CPS ASEC tax model, on average, slightly overpredicted the number of dependents that non-Hispanic White TFUs have relative to the numbers claimed on IRS 1040s (0.02), with this average driven by those with higher earnings. In contrast, for the remaining race/ethnic groups, the tax model underpredicted numbers of dependents in TFUs, on average, and within each earnings group.

The overall finding that the CPS ASEC tax model underpredicted the number of dependents in CPS ASEC TFUs that file tax returns relative to the number of dependents they actually claim is also documented in Unrath (2022, 2024) for various tax years of linked CPS ASEC and IRS 1040 data. This underprediction, especially for low-earning and minority groups, has important consequences for the tax model’s predictions of CTC eligibility and the amounts received for both the EITC and CTC. The reasons for this underprediction remain unclear in the existing literature. For example, while the procedures used by the tax model to assign minor dependents to heads of TFUs seem appropriate—such as assigning dependents to adults in CPS ASEC households based on the survey’s relationship coding of household members—the CPS ASEC does not collect information on who

resided in households during the preceding calendar year, which is the basis for determining which minor dependents can be claimed on that year’s tax returns. Determining whether this issue or others may account for the CPS ASEC tax model’s underprediction of dependents would appear to be of high priority for understanding and improving the model.

EITC and CTC Eligibility and Estimated Credit Amounts by Whether CPS ASEC Filing Status Matched or Did Not Match IRS Tax Data

Table E-4 compares EITC and CTC eligibility and mean benefits received for CPS ASEC TFUs based on linked IRS 1040 data. As discussed in Bee and Unrath (2025), eligibility and credit amounts for both data sources were estimated using TAXSIM.⁹ The table presents averages for credit eligibility and credit amounts for those whose CPS ASEC–modeled filing status has the same filing status category as the linked 1040 records, and for those whose CPS ASEC TFS differs from the linked 1040 records. Mean credit amounts are conditional on having a positive credit amount. As indicated in the notes to Table E-4, the averages contained in more detailed tabulations from Bee and Unrath (2025) were aggregated using sample sizes of TFUs from their tables to form the weighted averages in Table E-4.

For all earnings and race/ethnic groups (Panel A) and the separate tabulations for different earnings groups (Panel B), the predicted CTC eligibility of CPS ASEC TFUs that match the filing status in their tax records was almost always lower than the predicted CTC eligibility for CPS ASEC TFUs that do not match the filing status in their linked 1040 tax records. The only exceptions are for the two highest-earning groups. The differences in the incidence of EITC and CTC eligibility between matched and unmatched TFUs are not particularly large, except for CPS ASEC TFUs with very low earnings (E < $25,000). For this earnings group, the estimated CTC eligibility rate for TFUs whose TFSs match those in the tax records are substantially lower than for TFUs whose CPS ASEC-predicted TFS does not match what was found in linked 1040s. Table E-2 is consistent with this finding, as it shows that the CPS ASEC tax model overpredicts that CPS ASEC TFUs in this earnings group would not file tax returns (and, correspondingly, underpredicts filing) relative to their actual filing status based on linked IRS data. The reason for this discrepancy in CTC eligibility is not clear from the tabulations provided with Bee and Unrath (2025). It is not clear whether the CPS ASEC tax model is misclassifying the TFS, including nonfiling, of CPS ASEC TFUs and/or predicting fewer TFUs with qualifying children (i.e., dependents) for either of these credits.

___________________

⁹ The linked 1040 forms do not contain information on whether a TFU was eligible for either the EITC or CTC, or the amounts received for either credit.

TABLE E-4 EITC and CTC Eligibility and Mean Estimated Credit Amounts by Whether TFUs CPS ASEC Tax Model Filing Status Does or Does Not Match 1040 IRS Status and for All TFUs

NOTES: Entries are aggregated from detailed tabulations contained in taxcreditimputationv5at1.xlsx distributed with Bee and Unrath (2025). The rows for matched filing status in this table use entries for the proportions of TFUs determined by TAXSIM to be eligible for the EITC and CTC, respectively, in columns H and I of the Excel file and mean credit amounts in columns J and K of the Excel file. The rows in this table for nonmatches use entries for the shares of TFUs eligible for the EITC and CTC from columns M and N of the Excel file, and for mean amounts of these credits from columns O and P. To form the aggregated entries for these two rows in this table, averages of the disaggregated data in the Excel file were formed by weighting by corresponding sample sizes of PIKed CPS ASEC TFUs found in column Q of the Excel file. As noted in the previous tables in this appendix, suppressed cells were omitted from calculations of weighted averages used to form the aggregated values in this table. The incidence of suppressed cells for EITC and CTC credit eligibility and amounts in the Excel file was relatively high and, as a result, had a nonnegligible effect on the tabulations in this table. TAXSIM was used to predict credit eligibility and credit amounts for each CPS ASEC TFU. With respect to estimated credit amount per TFU member, each individual (adults and dependents) was assigned an equal share of the estimated credit amounts for each credit.

^a Entries are estimates of credit eligibility incidence and mean amounts based on 1040 data for CPS ASEC TFUs. Estimates are weighted averages of matched and unmatched entries. Note that because of the differences in cell suppressions in the matched and unmatched rows across CPS ASEC TFSs in the Excel file, resulting estimates in the “TFUs based on 1040 data” rows are not weighted averages of the corresponding matched and unmatched entries displayed in this table.

With respect to the estimated amount of CTC credits received by eligible filers, all three panels of Table E-4 show that TFUs for which the TFS predicted by the CPS ASEC tax model matched the status in their linked tax returns received credits that were consistently greater than did those TFUs whose TFS did not match the status in their tax returns. Again, the reasons for this underprediction by the CPS ASEC tax model are not clear, nor, as noted above, are they considered in Bee and Unrath (2025).

With respect to estimated incidence of EITC eligibility, data again indicate that EITC eligibility for TFUs for which the TFS predicted by the CPS ASEC tax model matched the status in 1040 records is lower than that for TFUs whose TFS predicted by the CPS ASEC tax model did not match the status in tax return records (see all three panels of Table E-4). The magnitude of the discrepancy between estimated EITC eligibility for these two sets of TFUs is much smaller than is the case for CTC eligibility.

With respect to the discrepancies between the estimated amounts of EITC credits received on average between TFUs whose TFS was predicted by the CPS ASEC tax model and for those TFUs for which the prediction did not match, averages in all three panels of Table E-4 indicate that, on average, the amounts for TFUs whose TFS matched those in their tax returns are lower than for those TFUs whose filing statuses did not match. While the magnitudes of the differences in average EITC across race/ethnic

groups are very similar (Panel C), these magnitudes increase for higher-earning TFUs. Again, the reasons for these patterns are not clear and will require further analyses.

Finally, Table E-4 presents estimates of EITC and CTC eligibility and estimated credit amounts, conditional on receipt, that are based on information in linked IRS 1040 records for all CPS ASEC TFUs. These estimates were formed by taking weighted averages of the matched and unmatched values of each in the disaggregated tabulations that accompany Bee and Unrath (2025) using the corresponding match rates and then aggregating them. For reference, Table E-4 re-displays the aggregated match rates (as percents). These estimates are provided in the rows “TFUs based on 1040 data.” In general, these estimates fall between the corresponding values of the matched and unmatched estimates and suggest that a greater incidence of eligibility and estimated amounts for both credits is found for TFUs whose filing status predicted by the CPS ASEC tax model is the same as that found in linked 1040 data. Some of these weighted average estimates do not fall between those for matched and unmatched TFUs. This is due to the pattern of cell suppressions in Bee and Unrath’s (2025) tabulations, which mitigated the usual property of weighted averages falling between the values of the two values being averaged.¹⁰ (See the Table E-4 notes for details of estimates construction.)

The incidence of CTC eligibility generated by TAXSIM using linked 1040 tax return data on filing status and dependents is fairly high, with an average of 38.8% for all CPS ASEC TFUs (Panel A). Furthermore, CTC eligibility increases across earnings groups (Panel B). The pattern of CTC eligibility across income groups is consistent with the 2018 CTC schedule, which provides no benefits to TFUs with no or low earnings, reaches its maximum credit amount between $15,000 and $45,000 of earnings, and remains at its maximum through earnings above $200,000. The incidence of EITC eligibility based on 1040 tax return information tends to be lower than that for CTC eligibility, and it rises and then declines across earnings groups. Again, this is consistent with the 2018 EITC schedule, which provides no benefits to TFUs with no or low incomes (earnings), provides increasing benefits until earnings reach $15,000 to $18,000, begins phasing out around $27,000, and provides no credit between $45,000 and $60,000. The estimated incidence of CTC eligibility is almost identical for non-Hispanic White, Black, and Asian individuals and other races, but considerably higher for Hispanic individuals (Panel C). With respect to incidence of EITC eligibility by race and ethnicity, a similar pattern to CTC eligibility holds, with EITC eligibility being lower for non-Hispanic White individuals and

___________________

¹⁰ In particular, the cells that are suppressed for matched TFUs by the CPS ASEC tax filing statuses do not always correspond to those suppressed for unmatched ones.

Asian individuals and other races compared to Hispanic and non-Hispanic Black individuals.

Estimated mean credit amounts for those eligible to receive the EITC and CTC are $1,610 and $1,205, respectively (Panel A). The mean CTC credit amount rises across the earnings groups, while the mean EITC credit amount is not monotonic across earnings groups (Panel B). Estimated CTC credit amounts are similar by race and ethnicity, with non-Hispanic Black individuals having slightly lower credit amounts (Panel C). EITC credit amounts also are similar across race and ethnic groups, albeit slightly higher for Hispanic individuals.

Tabulations accompanying Bee and Unrath (2025) lack information on EITC and CTC eligibility and credit amounts for tax year 2018 based solely on the CPS ASEC tax model using data from only the 2019 CPS ASEC. Direct comparisons of the model that would predict credit eligibility and amounts are not possible without the use of linked tax data. However, findings with respect to the tax model’s determination of TFSs, especially nonfiling status, and its configuration of dependents in the CPS ASEC TFUs that it forms suggest that this model likely underpredicts eligibility and credit amounts for both credits, especially for low-earning TFUs and for minorities. It is not clear that evidence from the tabulations provided by Bee and Unrath (2025) with linked 1040 tax data reconciles the known differences between IRS control totals for EITC and CTC take-up and credits paid and the CPS ASEC tax model estimates using only CPS ASEC data. Such a reconciliation represents an important consideration for the study of the EITC and CTC, whether using the CPS ASEC tax model or any other existing model.

Finally, to reiterate, the results in Table E-4 and the corresponding tabulations accompanying Bee and Unrath (2025) are estimates of EITC and CTC eligibility and credit amounts derived from a model (TAXSIM) applied to both CPS ASEC data derived from the CPS ASEC tax model and to linked 1040 tax return data. Importantly, neither of these sets of estimates are based on credits actually received by tax filers, as 1040 tax forms do not contain that information. Determination of which TFUs actually receive either credit occur through routine checks of tax returns, including dependents claimed on multiple tax returns and an audit (referred to as a “correspondence audit”) of a random subset of tax returns for low-income TFUs, before including such credits in filers’ tax refunds. For one such set of correspondence audits for the EITC, Guyton et al. (2024) documented a widespread disallowance of the EITC among those audited, due to failure of tax filers to either respond to the audit at all or to provide a sufficient response.

While this type of audit is only performed on a relatively small subset of low-income tax filers (e.g., in tax year 2017, only approximately 350,000

filers were subjected to such an audit), this example clearly suggests that estimates of EITC and CTC eligibility, even when using tax return information, are likely to overestimate actual eligibility—even if these credits are actually included in the tax returns of apparently eligible tax filers. Put differently, the findings from the tabulations generated by Bee and Unrath (2025) and many of the other studies using linked data discussed earlier in this appendix cannot be considered “the truth” about EITC and CTC eligibility and credit receipt.