CONTRIBUTING FACTORS OF THE 2020 U.S. CENSUS CHILDREN AND YOUNG CHILDREN UNDERCOUNT IN TEXAS

In partnership with Dr. Bill O’Hare from Count All Kids, this research report identifies the factors contributing to the 2020 Census undercount of Texas children and young children.

By: Dr. Francisco A. Castellanos-Sosa, Texas Census Institute, Senior Research Associate
Dr. William P. O’Hare, Count All Kids Campaign, Consultant

Recognition of undercounting in social science surveys is not new (Clogg et al., 1989; de la Puente, 1995; King & Magnuson, 1995; Martin & de la Puente, 1993; O’Hare, 2019; Tourangeau & Plewes, 2013; West & Fein, 1990); however, scholarly work has not yet achieved a consensus framework for studying its determinants.
It is important to note that the census coverage of children differs from that of adults or the whole population. Children cannot answer census questionnaires and have no direct participation in U.S. Census activities. Thus, the coverage error in children depends on whether an adult is available to provide information for the entire household and on the quality of the data provided by the adult.
Studies on sub-national accuracy assessment of the U.S. Census are limited. Moreover, exploring overall undercounting determinants has provided limited information on the causes of undercounting children (Cohn, 2011; Mayol-Garcia & Robinson, 2011; Robinson et al., 1993; Siegel et al., 1977).
Recent studies have provided insights into potential determinants of variation in young child undercounting in the 2010 Census (O’Hare, 2014, 2017; O’Hare et al., 2019). O’Hare (2014) suggests that at the state level, young child undercount rates are correlated most highly with the relative size of the Black population—Alone or in Combination—and the Hispanic population. He also found that characteristics such as linguistic isolation, the percentage of adults without a high school degree, and the unemployment rate were more highly related to variations in the young child undercount among states than were housing unit features.
O’Hare (2017) examined variations in young child undercount rates in the 2010 Census across counties. He found that the combined Black and Hispanic population was a key determinant of young child undercounts. King et al. (2018) also found that counties’ racial/Hispanic composition was closely related to county coverage rates for young children.
One of the most prominent research works on this issue is that of O’Hare et al. (2019). They used 2010 young child undercount rates for large counties and several potential causal variables to obtain county-level regression estimates showing the relationship between the potential causal variables and the young child undercount rate. They then combined the coefficients from that model with data from the 2013-2017 American Community Survey to predict the tract-level risk of young children being uncounted in the 2020 Census in 689 counties with at least 5,000 children under age 5. The authors explored 40 potential explanatory variables and clustered them into six domains:

1. Race and Hispanic Origin;
2. Socioeconomic Status;
3. Family Structure and Living Arrangements;
4. Other Demographic Measures;
5. Housing; and
6. Census Response/Return Rates.

O’Hare et al. (2019) examined the potential determinants of young child undercounts in the 2010 Census at the census tract level for large counties (those with 250,000 people or more in 2010). O’Hare et al. (2019) found a close association between young child undercount rates and:

The present analysis builds upon the above studies. We extend the previous work of O’Hare et al. (2019) by exploring multiple variables, organized in 6 domains, using all counties in Texas, studying not only young children (ages 0 to 4) but also children (ages 0 to 17). Moreover, our approach’s wide set of variables allows a clearer understanding of the mechanisms of the potential determinants of undercounting.
Note that all the studies reviewed above were focused on young children (ages 0 to 4). This study compares results for young children to those for all children. This analysis will help researchers understand if the same factors drive the coverage of all children as the coverage of young children. At present, it is unclear whether the determinants of variations in the net undercount rates for all children and young children are the same. This study will shed light on that question.

The Census Bureau’s Demographic Analysis (DA) assesses the quality of the 2020 decennial census for children by age groups (see Table 1). Data in Table 1 indicate that the youngest children (ages 0 to 4) had the highest net undercount at 5.4 percent, and the oldest children (ages 10 to 17) had the smallest net undercount at 0.6 percent.^[3]
One of the advantages of using to study the undercounting or overcounting of children in the Census is that it allows exploration by a single year of age. Figure 1 shows how the youngest children are highly undercounted. Children under 1 year have a 7.0% net undercount rate as measured by the Middle Series. As children grow, the undercount shrinks with each year of age until the 8-year-old age group, for which the undercount is 0.6 %.

O’Hare (2023) presents a nationwide analysis of the country’s county-level estimates of young child undercounts using Census Counts and the April 1, 2020 version of the Vintage 2020 Estimates. The coverage estimates are derived by calculating the difference between the county-level Census Counts and 2020 Vintage Population Estimates such that when the census is lower than the Vintage Estimates, there is an undercount, and when the census is higher than the Vintage Estimates, there is an overcount. We focus on children (ages 0 to 17) and young children (ages 0 to 4) because these age groups experienced a statistical undercount, according to the U.S. Census Bureau Demographic Analysis (Jensen, 2022; Jensen & Kennel, 2022). This method for calculating net coverage has been widely used by others, and it is used in this study.
Net coverage refers to net undercounts and net overcounts. Net undercounts and overcounts are based on a balance between people missed in the census and people counted more than once or included in the census inappropriately. Net undercounts are not the same as the number or rate of people missed in the census.

If we do not do something different in 2030 than we have done over the past 40 years, there is no reason to expect the count of young children to improve in the next census. If the net undercount of young children increases between 2020 and 2030 as it did between 2010 and 2020, it will be over 6 percent in the 2030 Census. No other demographic group has experienced the deterioration of census quality that young children have experienced since 1980. That makes young children a special population in planning for the 2030 Census. The key point of Figure 2 is that the net coverage of young children will likely worsen in 2030 unless we take bold action to change the trajectory of the past 40 years.
Second, it is also important to note that while the coverage of young children has deteriorated since 1980, the census coverage of older children and adults has not. The coverage of adults continued to increase to the point that they had a small overcount in the 2020 Census. The coverage of children ages 5 to 17 remained relatively stable after 1980.

For example, the total population in Hawaii had a net overcount of 6.8 percent in the 2020 Census, while young children had a net undercount rate of 8.6 percent. Delaware had a net overcount of 5.5 percent for the total population but a net undercount of 6.1 percent for young children.
The trends over the past 40 years and the lack of a relationship between total population coverage and young child coverage in states show that improvements in the quality of census counts for the total population are unlikely to lead to improvements in the count of young children. There has been a significant improvement in the total population coverage over the past 40 years, but there has been no improvement in counting young children.
Given this disconnect between the quality of census data for the total population and the quality of census data for young children, it is critical to focus on young children separately from the total population and older children.

Of the 7,432,438 children in Texas reflected in the Vintage 2020 Population Estimates, the census undercounted 153,633 or 2.1 percent (Castellanos-Sosa & O’Hare, 2023a). Of 254 counties in Texas, 190 experienced an undercount (74.8%), and 64 showed an overcount (25.2%).
On the other hand, of the 1,975,115 young children in Texas reflected in the Vintage 2020 Population Estimates, the 2020 Census undercounted 155,855 young children (7.9%) (Castellanos-Sosa & O’Hare, 2023b). Many counties in Texas not only had an undercount of children and young children, but their undercount was quite high. In total, 184 Texas counties (72.4%) had a high net young child undercount (rate, number, or both).[5] These results suggest that 85.8% of Texas young children live in a county with a high undercount (rate, number, or both) and that of the 184 Texas counties with a high net young child undercount, 31 Texas counties had a high net young child undercount rate and number, with a net 134,421 young child undercount (being this the 9.5% of their children).

The U.S. Census Bureau does not plan to produce sub-state measures of accuracy for all children in the 2020 census using either of the two main methods the Census Bureau uses to assess census accuracy (Post-Enumeration Survey () and Demographic Analysis . The Census Bureau will provide coverage rates for the population ages 0 to 4 for counties with at least 1,000 children in this age range.
It is unfortunate there are not more sub-national assessments of census quality because many stakeholders are seeking sub-state measures of census quality (Adlakha et al., 2003; American Statistical Association, 2021; Borusyak et al., 2021; Hill et al., 2022; National Academies of Sciences, Engineering, 2022; National Association of Latino Elected Officials Educational Fund, 2022; U.S. Census Bureau, 2014). This study responds to those requests for more sub-state quality measures for the 2020 Census.
It is important to study counties because state-wide numbers can mask big differences within a state. One set of counties may have high undercounts, which may be counterbalanced by another set of counties with high overcounts, leading to an average low overall coverage error for the state. For example, in the 2010 Census, nearly all the undercounts of young children (ages 0 to 4) in New York and Illinois were accounted for by large urban counties in those two states (U.S. Census Bureau, 2014).
This study employs a commonly used demographic benchmark to identify counties with net undercounts and overcounts of children in the 2020 Census. Specifically, the 2020 Decennial Census county-level counts of children (ages 0 to 17) and young children (ages 0 to 4) are compared to corresponding figures from the Census Bureau’s Vintage 2020 Population Estimates to ascertain census coverage for children and young children. The Vintage 2020 Population Estimates are based on the 2010 Census results with births, deaths, and net migration between 2010 and 2020 incorporated. When the 2020 Census is smaller than the Vintage 2020 Estimates, we observe an undercount (expressed with a negative undercount estimate), and when the 2020 Census is higher than the Vintage 2020 Estimates, we observe an overcount (denoting a positive undercount estimate). Then, net undercounts are reflected with negative values, and net overcounts are reflected with positive values.
According to the U.S. Census Bureau, “Both the 2020 DA Estimates and the Vintage 2020 Population Estimates can be used as demographic benchmarks for evaluating certain aspects of the 2020 Census results” (Jensen & Johnson, 2021). This study is closely linked to Jensen and Johnson’s (2021) work since we compare 2020 Census results and Vintage 2020 Population Estimates to assess the 2020 Census data quality for young children at the county level. According to Jensen and Johnson (2021), “Increasingly, data users are comparing the Population Estimates to the results of the 2020 Census to try and understand the quality of the census results.” This study adds to that research stream and uses the methodology developed by O’Hare (2023) and used by several others.^[9]
There are four main reasons why the Vintage Population Estimates are thought to be more accurate than the census counts for the population ages 0 to 4.

1. 1. There is a high net undercount of young children in the 2010 and 2020 censuses (Jensen, 2022).
   2. The Vintage Population Estimates for ages 0 to 4 are largely based on birth certificate data, which is widely considered reliable. Ninety-nine percent of the population ages 0 to 4 in the 2020 Population Estimates are based on birth certificates (U.S. Census Bureau, 2022).
   3. The data sources and methodology for producing Vintage Population Estimates are nearly identical to the Census Bureau’s Demographic Analysis method, which is the preferred method for estimating young child undercounts at the national level (U.S. Census Bureau, 2021b).
   4. The results of the Vintage Population Estimates for young children are nearly identical to the Census Bureau’s Demographic Analysis estimates at the national level, which underscores the suitability of using the Vintage Population Estimates to examine the subnational geographic distribution of the net undercount rates of young children (O’Hare, 2023c).

It is worth noting that some of the differences found here may be important even if differences do not reflect true undercounts. According to the U.S. Census Bureau, “…significant or unexpected differences can be useful for identifying areas for further investigation” (Hartley et al., 2021, p. 2). In other words, a difference between the census count and the Estimates may signal a problem with the underlying data.

As per O’Hare et al. (2019), our set of potential explanatory variables is sorted into six domains to reflect broad dimensions of the environment experienced by persons responding to the U.S. Census Bureau. These domains consider aspects that are inherent to the individual (Race and Origin and Age Characteristics), related to their lifetime achievements (Socioeconomic Status), linked to their interpersonal connections (Family Structure and Living Arrangements), descriptive of their housing units (Housing), and reflecting the geography and the implementation of the census itself (Census-and Geography-Related Features).^[10]
For Race and Origin, many racial and ethnic minorities have long had high net undercounts in the census. This analysis aims to identify which groups have the highest disparities. On Socioeconomic Status, our research would underscore the intersectionality of undercounting with poverty, education, and income-related features—characteristics that usually raise systemic challenges. In Family Structure and Living Arrangements, household complexity influences census accuracy. For instance, O’Hare et al. (2019) found that the share of children under age 18 living in a female-headed household with no spouse present or the share of children under age 6 living with a grandparent householder was related to the variation in net undercount rates for young children. Similarly, Housing characteristics might shed light on factors influencing census accuracy, such as overcrowding and tenure.
It is not the children or young children that respond to the census survey, but an adult person in the housing unit. Therefore, the Age Characteristics of people living or caring for children are explored. Lastly, in terms of Census-and Geography-Related Features, the response rate of people and geographic accessibility of the housing units are incorporated to cover the data collection stage of the census implementation.
Figure 4 shows the number of variables and their proportion from each domain’s number of variables. For instance, Socioeconomic Status, the domain with the largest number of variables, considers 95 potential explanatory variables, representing slightly more than half of the variables (52%).

Rather than splitting the Socioeconomic Status 95 variables, we preserved them within the same domain since we acknowledge that income, poverty status, access to technology, social security programs, and the value of the houses and monthly rent payments result from the lifetime achievements of families.
The second largest domain in terms of the number of variables is Race and Origin, with 39 (21.4%) of the variables). It includes variables related to race, nativity, migration, and language. The other domains, Family Structure and Living Arrangement, Age Characteristics, Housing, and Census-and Geography-Related Features, reflect 48 factors (26.4 %).

We use a two-stage approach for selected potential explanatory variables. The first stage helps us find variables statistically associated with undercounts but not highly correlated with other measures in the domain and across domains. The second stage consists of a stepwise regression analysis with those selected variables.
In the first stage, we refine the number of variables in three steps, as in O’Hare et al. (2019). We identify the variables within each dimension that have a statistically significant correlation with the undercount rates for children and young children. Among these, we preserve the variables with a low correlation with other variables within each dimension. Furthermore, we keep only those variables with a low correlation with variables in other dimensions. The last two steps help us avoid multicollinearity problems.
The second stage consists of a stepwise forward regression analysis.^[12] It consists of identifying the best one-variable model and extending it by comparing it against the best two-variable model, and so forth. We set the stepwise regression analysis to add more variables only if they are statistically significant, at least at the 10% level, and keep being significant at least at the 10% level when other variables are included (StataCorp, 2021). These are the steps used for the inclusion and exclusion of variables:

1. 1. Fit “empty” model.
   2. If the most-significant excluded term is “significant” at least at the 10% level, add it and reestimate; otherwise, stop.
   3. Do that again: if the most-significant excluded term is “significant” at least at the 10% level, add it and reestimate; otherwise, stop.
   4. Repeatedly, until neither of the next steps is possible,
      • if the least-significant included term is “insignificant” at the 10% level, remove it and reestimate; and
      • if the most-significant excluded term is “significant” at least at the 10% level, add it and reestimate.

Our undercount rate measures can be negative or positive. As described in the Methodology section, negative rates exist when the 2020 Census is smaller than the Vintage 2020 Estimates as a signal of undercount. Furthermore, positive rates exist when the 2020 Census exceeds the Vintage 2020 estimates, indicating overcount.
All of the factors were constructed so that a higher value of the factor is expected to be associated with a lower value (worse) of the net undercount rate. Recall that net undercount rates are reflected as negative numbers. A worse net undercount rate is a lower number.
The first stage of our statistical approach is finding variables statistically associated with undercount but not highly correlated with other measures in the domain and across domains. In this first step, it is essential to identify the variables that have a significant correlation with child and young child undercount rates. Correlation coefficients can go from -1 to 1, where a negative correlation coefficient suggests that an increase in the variables would associate with lower values of the undercount rate (this means more undercounting). A positive correlation would, therefore, suggest that an increase in the variables would relate to higher values of the undercount rate (this means more overcounting). Of the 182 variables, 99 had a negative correlation with the undercount rate of children, and 102 had a negative correlation with the undercount rate of young children. To identify whether the correlation coefficient is significant (i.e., the correlation coefficient is not the result of pure chance), we test the null hypothesis of zero correlation against the alternative hypothesis that the correlation coefficient is different from zero.
For the counties’ child undercount rate, we find 92 variables with a statistically significant correlation. We also find 90 statistically significant variables correlated to the young child undercount rate. After preserving the variables with a correlation below 0.7 with other variables within each domain and variables in other domains, these numbers go down to 54 and 51 for the study of child and young child undercount, respectively.
The second stage of our statistical approach consists of a stepwise forward regression analysis with the aforementioned selected variables. Table 2 shows the results from applying the stepwise forward regression analyses using the 54 and 51 potential contributing factors of the child and young child undercount. The dependent variable is the undercount rate for children in Model 1 and young children in Model 2. Of the 54 potential contributing factors used for children, only 14 are part of the final model. Of the 51 potential contributing factors for young children, only 18 are in the final model.
Our results suggest four factors contribute to both undercount and overcount rates of children and young children; 10 unique factors relate to the rates of children only; and 14 unique factors relate to the rates of young children only. The four variables contributing to the undercount and overcount of both children and young children are:

The share of addresses unable to be geocoded by the U.S. Census Bureau are those addresses in the Census Bureau’s Master Address File that were used to take the 2020 Census that were not possible to have a tract and block geocode at the beginning of the Local Update of Census Addresses (LUCA) operation (by January 2018). In other words, this variable reflects problems with the Census Bureau’s Master Address File at the county level.

Therefore, since the U.S. Census Bureau needs to have information on every household address, discrepancies in the Master Address File could easily lead to misleading outcomes during the imputation and filling process of those addresses unable to be geocoded (Cantwell, 2021).
Another important factor in the domain of Census-and Geography-Related Features, one that is particularly significant for the study of children (ages 0 to 17), is the nonresponse rate for reasons other than refusal. This is the proportion of housing units that did not participate in the 2020 ACS for any other reason different from refusal.^[13]

Although participation in the ACS is required by law, some housing units – nontraditional housing in sheds and outbuildings, trailer parks and even multi-family dwellings or multiple families in one dwelling – went un-located, or there was no one at home, the interviewed people faced language problems, people were temporarily absent, the maximum contact attempts were reached, or any other reason that impeded the housing unit from participating successfully in the ACS. In any case, a higher nonresponse rate for reasons other than refusal implies that more housing units did not complete the ACS interview and as a result, the Census Bureau did not get accurate data from the housing unit. People’s attitudes and motives for not responding to the ACS interview, even though a response is mandated by law, might be the same as those related to not responding to the 2020 Census. Non-response forces the U.S. Census Bureau to impute characteristics on those addresses and potentially endanger the accurate counting of certain population segments. Self-response is widely seen as the most accurate data in the census. Counties where self-response rates are low are likely to have less accurate data.
Another predictor of an inaccurate count is the share of children ages 1 to 17 that moved into the county, i.e., the proportion of children ages 1 to 17 (from those in the same age range) that did not live in the same county one year ago and that moved to the county from another county within the same state, from another state, or that moved from abroad. This is a reflection of mobility and might relate to an overcount (due to double counting) of children since neighbors might provide information for the former occupants of a housing unit.
The share of SNAP-eligible individuals that are young children (ages 0 to 4) is the proportion of eligible individuals in the counties that are young children ages 0 to 4, according to the Texas Health & Human Services Commission. This variable mostly reflects the share of counties’ young children in low-income households, regardless of receipt of SNAP benefits. It is widely believed that adults in low-income households are not present at home or do not have the time to respond to the census questionnaire due to their workload, which would lead to the undercount of the children and young children living in the housing unit.
Occupied units paying rent between $1,500 and $1,999 could be considered middle income, with a potentially higher ability and motivation to participate in civic activities and to attend to the census questionnaire (Chetty et al., 2022a, 2022b). A higher share of people in higher income groups could result in lower undercount rates due to civic engagement levels.
While these four potential explanatory variables are closely related to variables in the coverage of children and young children, it is unclear if they are driving the coverage of children or the coverage of the total population, including children. In other words, they are not child-specific factors. Table 3 summarizes the variables that appear to contribute to the undercount of children compared to young children, that is those contributing only to the undercount of children and those contributing only to the undercount of young children. These factors can appear through diverse mechanisms and deserve further research to identify how they relate to census coverage.

A standard way to know how well the data explain the variability of our dependent variables (undercount rates) is through the coefficient. The coefficient goes from 0 to 1, where 0 means the variables do not explain any of the variability of the undercount rates, and 1 means the variables perfectly explain the variability. The higher the , the better the model is at explaining the data. The coefficients in this study—0.63 and 0.66—are higher than those of the closest related work (0.38) in the best model of O’Hare et al. (2019)).

Our values are moderately high by social science standards. Another way of thinking about the is that these models explain about sixty six percent of the variation in coverage rates of children and young children across Texas counties.

Interpreting the results in Table 2 may be complicated because net undercounts are reflected as negative numbers. A negative sign in the coefficients from the regression suggests that the net undercount is worse when a given factor increases. A positive sign in the coefficients suggests that the net undercount is lower when a factor increases. The coefficients themselves express the magnitude of the changes.
This study finds that the set of explanatory variables for all children is different from the set of explanatory variables for young children. This underscores the extent to which distinct factors drive the coverage of older and younger children; therefore, this subsection contains two parts that explain the results of each Table 2 model separately.

Our findings suggest that two out of the eight variables in the Census-and Geography-Related Features domain worsen the undercount rate of children (ages 0 to 17) when they increase. In other words, the higher the nonresponse for reasons other than refusal, the worse the undercount rate of children. Similarly, the higher the share of addresses that the U.S. Census Bureau cannot geocode, the worse the undercount rate of children.
Two of the nineteen variables in the Family Structure and Living Arrangements domain seem to worsen the undercount rate of children (ages 0 to 17) when they increase. The higher the share of children who are not immediate family members, the worse the undercount rate of children.[14] Similarly, the higher the share of children that are not biological children, adopted children, or stepchildren, and that are other relatives, the worse the undercount rate of children.
In the Housing domain, our results suggest one factor contributes to a worse undercount rate of children (ages 0 to 17) when it increases. The higher the share of vacant households, the worse the undercount rate of children is.
Regarding the Socioeconomic Status domain, we find that three factors worsen the undercount rate of children (ages 0 to 17) when they increase. The higher the average share of SNAP-eligible individuals who are young children (ages 0 to 4), the worse the undercount rate of children. The higher the average share of families whose income was below the poverty level during the last 12 months with related children of the householder under 5 years, the worse the undercount rate of children. The higher the average share of married-couple families whose income was below the poverty level during the last 12 months with related children of the householder under 18 years, the worse the undercount rate of children. These could both be seen as different ways to measure poverty.

Our study extends the exploration of determinants of geographic variation in the coverage of children and young children in the decennial census. For instance, we found two operational measures (nonresponse rate for reasons other than refusal and share of addresses unable to be geocoded) were closely associated with the coverage rate of children and young children. Previous studies did not include operational measures.
One of the key findings of this study is that operational measures help explain variation in the coverage rates of children and young children. This has not been examined in previous studies. We also find that many of the factors that appear to drive variation in the coverage of young children are not the same as those that drive variation in all children’s coverage. This is consistent with other analyses showing that coverage of young children differs from that of older children or adults. By incorporating a more extensive set of variables, studying all counties in Texas, and using the latest available data, we provide more Texas-specific results and updated insights into the factors potentially contributing to the undercounting of American children.
In particular, we identify that the nonresponse rate for reasons other than refusal, the share of addresses unable to be geocoded by the U.S. Census Bureau, the share of SNAP-eligible individuals that are young children (ages 0 to 4), and the share of occupied units paying rent between $1,500 and $1,999, are factors that contributed to the variation of the net undercount rate of both children and young children, both positively and negatively.
The relevance of our findings extends beyond academic discourse. The identified contributing factors can guide targeted outreach and resource allocation for the 2030 Census, improving the accuracy of child and young child counts. Moreover, local data analysts and researchers can leverage this comprehensive set of determinants to gain deeper insights into the persistent issue of undercounting in the census. Hopefully, this study will be replicated for other states to discover whether the determinants found here are generalizable or are more Texas specific.
As we conclude, this research contributes to the ongoing dialogue on census accuracy assessment, emphasizing the need for continuous refinement of methodologies and a thorough understanding of factors influencing undercounting. The roadmap presented in this study lays the groundwork for future research, facilitating a more accurate and inclusive representation of children in the U.S. Census.

The undercount of children and young children are complex issues and, therefore, are hard to solve. For instance, the undercount of young children has been worse decade by decade since 1980. This study presents a set of variables explaining the variation in children and young children undercounting across Texas counties. Our results could help identify counties more likely to experience undercounts in the 2030 Census if nothing is done to prevent it. On the other hand, our results are helpful to target outreach efforts and awareness campaigns. On-the-ground research should be conducted to expand the understanding of the presented findings.

Acknowledgements: The authors appreciate the insightful support provided by Angela Broyles, Helen You, and Lloyd Potter in the development of this study.