U.S. Bureau of the Census
Washington, DC 20233-8800
Population Division Working Paper No. 19
The views expressed are attributable to the author and do not necessarily reflect the views of the U.S. Bureau of the Census.
If you have any questions concerning this report, please e-mail a message to firstname.lastname@example.org. Include the name of this report and author in the body of the message.
By David L. Word
INTRODUCTION AND SUMMARY
This study analyzes the geographic variation in mail response rates for tracts and block groups (local areas) that occurred in the 1990 Census. We will show that variation in local area response for 1990 is largely attributable to the "demographics" of the local area. Because local area demographics are not prone to large changes over a ten year interval, we believe that response rates for most tracts and block groups in 2000 will approximate 1990 levels.
We explore the degree to which individual household response in the 1990 Census is dependent on the interaction of four demographic variables: (1) race-ethnic, (2) tenure, (3) family composition, and (4) type of structure. We use 1990 national response rates for these four variables to generate synthetic 1990 response rates for combinations of individual tracts and block groups. These synthetic response rates are then compared with actual 1990 response rates. The degree to which the expected rates mirror the actual rates demonstrates the predictive power of the model.
Our research shows that response to the mailed 1990 Census questionnaires is highly correlated with race and tenure. This finding is consistent with 1990 Post Enumeration Survey (PES) undercount research. In addition to the race and tenure variables, we demonstrate that mail response is also strongly related to family composition--a finding not considered in the 1990 PES research. Since response rates and net undercount rates may be causally linked, we will use this study as a sounding board for commenting on the Integrated Coverage Measurement (ICM) opertion.
The final section of this paper includes a bibliography of other studies that deal with nonresponse to the decennial census.
Before presenting results, we need to differentiate two similar concepts, "Mail Return Rate" and "Mail Response Rate." These two terms have had different definitions in the past. Here, Mail Return Rate relates to housing unit addresses and is defined as the proportion of addresses (including vacant) which return census forms. Mail Response Rate relates to households (occupied housing units) and is defined as the proportion of households in the 1990 Mail universe which completed their 1990 census form without the aid of an enumerator. In this paper we are interested only in the effect of demographic factors on mail response rates and are not concerned with the problem of vacant unit enumeration.
In the remainder of this paper we invert Mail Response to Mail Nonresponse for the same reason that Bureau of Labor Statistics concentrates on unemployment rates rather than employment rates. Mail Nonresponse rate is the complement of mail response rate and is defined as the proportion of households (not persons) where the 1990 Census questionnaire was completed by an enumerator rather than the addressee. With that definition nonresponse is independent of vacancy status.
The analysis presented here could not have been done without George McLaughlin's proficiency in manipulating the 100% Detail files from the 1990 Census. Gregg Robinson's parallel research on a "targeting data base for hard to count areas" provides an independent check on any hypotheses generated from these data. Furthermore, I thank the readers, especially Cynthia Taeuber, for taking the time and effort to review, refine and rewrite earlier drafts of this document.
We hope that the five separate steps covered under STATISTICAL HIGHLIGHTS lead the reader to the conclusion that individual household demographics are highly correlated with the likelihood of responding/nonresponding to a mailed census form. Since block groups and tracts are composed of individual households (approximately 1,500 per tract and 400 per block group) we can accurately predict nonresponse rates for the geographical area provided that demographics of the neighborhood are known or can be estimated.
1.0 Overall nonresponse in 1990 was 25.3 percent but area geographic variability among tracts and block groups is high.
|Table 1||Tracts||Block Groups|
|Number of areas||58,858||205,881|
|Mean number of units||1,492||426|
2.0 The nonresponse for individual households is correlated with four demographic factors available on every shortform questionnaire. Race-ethnicity, tenure, family composition, and type of structure.
|1 White NonHispanic||22.0||(NA)|
|1 Spousal household||19.4||(NA)|
|2 Nonspousal household||38.1||30.1|
|1 Single Family Detached||19.1||(NA)|
|3 Mobile Homes||32.6||30.9|
Within each of the four separate demographic categories, the first entry's nonresponse rate is close to 20 percent while the nonresponse rates approach 40 percent for the remaining entries. The large differences in raw nonresponse rates are muted when we normalize the rates for compositional effect. For example, the raw nonresponse rate for Blacks (43.4 percent) is double the rate for White nonHispanics (22.0 percent). But Blacks are more likely to be renters, live in apartments, or to have Nonspousal living arrangements than are White nonHispanics.
The normalized nonresponse rate for Blacks (37.9 percent) is the aggregated nonresponse rate that Blacks would have had in 1990 if the distribution of tenure, spousal relationships, and structure for Blacks matched that of White nonHispanics. Even with the normalizing process, nonresponse rates for minorities continue to be substantially greater than the nonresponse rates for White nonHispanics.
Persons living in apartments have a normalized nonresponse rate (21.6 percent) that approaches the rate for persons living in single family detached units. This seeming anomaly results from the strong correlation between owner status and single family detached units and renter status with apartments. It is probably unnecessary for ICM purposes to include an apartment stratum if tenure already forms a stratum. On the other hand, the normalized nonresponse rates for persons living in mobile homes (30.9 percent) does not differ greatly from the raw rates (32.6 percent). It is evident that nonresponse for persons living in mobile homes is different from that of persons living in single family detached units or apartments. Mobile homes should be a separate stratum.
3.0 Nonresponse rates for individual demographic categories within tracts and block groups are essentially independent of the demographic composition within the local area. Because block groups are slightly more homogeneous than tracts, sections 3.1 through 3.4 are analyzed from a block group perspective, although an analysis based on tracts would be similar. It is obvious that each of the four tables (3.1, 3.2, 3.3, and 3.4) omit a middle grouping. For example, table 3.2 could have included a column "Block groups with more than 50 percent but less then 75 percent of housing units owned." For space reasons, we have omitted that data. But the nonresponse rates for this omitted group were midway between the rates that are displayed.
|Table 3.1||Block groups that are at
least 75 percent
|Block Groups that are less
than 50 percent
|Households in millions||66.1||12.0|
|Percent White nonHispanic||93.4||20.2|
Although the nonresponse rate for White nonHispanics increases when white nonHispanics move from being a strong numerical majority to a numerical minority (20.8 percent to 28.5 percent), there is little change in the nonresponse rates for racial and ethnic populations (38.3 percent vs 41.4 percent). This "independence" concept is useful in the development of section 5.0.
|Table 3.2||Block groups with
at least 75 percent of
housing units owned
|Block groups with
less than 50 percent of
housing units owned
|Households in millions||38.6||24.0|
In table 3.2, the differences in household nonresponse rates for owners and renters is minimally affected by the degree of home ownership within the block group. Virtually all the difference in total nonresponse rate--19.6 percent vs 34.8 percent--is related to the owner/renter composition within the block group.
3.3 Family composition
|Table 3.3||Block groups that are
at least 75 percent
|Block Groups that are
less than 50 percent
|Households in millions||37.0||12.0|
In table 3.3, a household is considered "Spousal" when one of two conditions is met: (1) the second person listed on the census form is the spouse of the householder or (2) the unit is occupied by exactly one person and that one person is over the age of 50.
For both Spousal and nonspousal units, nonresponse rates increase by nearly 15 percentage points when the proportion of spousal units shifts from being a strong majority "spousal" to a numerical minority. The nonresponse rates for spousal units increases from 16.2 to 29.6 percent and the nonresponse rate for the "nonspousals" move from 31.4 to 44.8 percent. It is difficult to understand why the degree of spousal units affects the nonresponse rates for this variable but has little effect on the remaining three variables discussed here.
3.4 Structure type
|Table 3.4||Block groups where at least
75 percent of households are
Single Family Detached
|Block groups where less than
50 percent of households are
Single Family Detached
|Households in millions||34.9||29.2|
|Apts or Mobile Homes||31.7||35.4|
Like the race and tenure variables (sections 3.1 and 3.2), nonresponse rates within structure type are not greatly affected by the proportion of units classified as single family detached (S/F/D) within the block group.
4.0 Characteristics affecting nonresponse are additive. In the previous section we presented nonresponse rates for a single demographic variable. Here we combine the four variables into individual strata. The number of households for specific racial and ethnic groups within "rented mobile homes" is so small that we have omitted some of the nonresponse rates for those strata.
From section 2 we found that the differences in raw and normalized nonresponse rates for single family detached units and apartments is small. For that reason we will combine single family detached units (S/F/D) and apartments (Apts) into a category called "Buildings." From this point forward, the separation among structures is "Building" vs "Mobile Home."
Tables 4.1 through 4.4 provide statistics on household nonresponse rates by various combinations of tenure, family composition and building type for four of the five race/ethnicity groups. Because the individual strata for the American Indian, Eskimo, and Aleut (AIEA) populations are so sparse, we do not show them here. However, rates for AIEA's were used in the development of geographic nonresponse rates for local areas discussed in section 5.
|Table 4.1||Households in Millions||Nonresponse Rate|
Table 4.1 demonstrates that variables on the high side of the nonresponse spectrum add about 10 percentage points to overall nonresponse. Among White nonHispanics the nonresponse rate for the Owner/Spousal/Building stratum is 13.2 percent. The second set of observations consist of three strata which successively eliminate spousal, building, and owner from the Owner/ Spousal/ Building stratum. The nonresponse rates in this second block range from 22.7 to 26.4 percent. The third set of observations (with two substitutions) have nonresponse rates ranging from 37.2 to 41.1 percent. In the final stratum, consisting of White nonHispanics who rent mobile homes and live in a nonspousal arrangement, the nonresponse rate balloons to 54.1 percent.
4.2 Black nonHispanic
|Table 4.2||Households in Millions||Nonresponse Rate|
The nonresponse rates for the Black population follow the same pattern as for Whites, except that the starting nonresponse rate for (Owner/Spousal/Building) is 15 percentage points higher.
Table 4.2.1 illustrates the strong degree of consistency in White/Black nonresponse rates after allowing for compositional differences in the two population groups:
4.2.1 Comparison of Nonresponse Rates: White nonHispanics and Black nonHispanics
|Table 4.3||Households in Millions||Nonresponse Rate|
Among the three Hispanic substrata with the greatest number of households, (Owner/Spousal/Building, Renter/Spousal/Building, and Renter/Nonspousal/Building), nonresponse rates are approximately 10 percentage points greater than the equivalent rates for White nonHispanics; but approximately 5 percentage points lower than nonresponse rates for the Black nonHispanic population.
4.4 Asian and Pacific Islander (API)
|Table 4.4||Households in Millions||Nonresponse Rate|
The nonresponse rates by stratum for the API population are almost identical to the nonresponse rates for the Hispanic population.
5.0 Applying national rates of nonresponse for the 40 strata described in section 4.0 provide useful estimates of local area nonresponse. Tables 5.1 and 5.2 provide a statistical analysis of the difference between actual and estimated nonresponse rates for all tracts and block groups when national stratum nonresponse rates are applied to the occupied household universe within the individual tract and block group. It is useful to remember that the lower limit of nonresponse is 13.2 percent (White nonHispanic Spousal Owners residing in a building) which is about half the overall nonresponse rate of the nation. The theoretical ceiling for a local area's expected nonresponse rate is 64.3 percent (the nonresponse rate for Hispanic nonspousal householders who are renting a mobile homes)
5.1 The unit of analysis in table 5.1 is, the tract; table 5.2 provides identical information for block groups.
|Table 5.1||No. of
|Tracts with expected
nonresponse rates of
|Less than 20 percent||13,888||21.2||16.7||18.2||6.2|
|20 to 30 percent||27,913||47.2||24.3||24.0||6.9|
|30 to 40 percent||9,799||15.2||35.0||34.0||7.8|
|More than 40 percent||3,539||4.3||45.1||44.1||8.8|
Applying the national household stratum nonresponse rates from Section 4 to each household within the 55,139 tracts provide an unbiased estimate of nonresponse rates with a standard error of only 7.0 percent. In terms of staffing needs for the nonresponse operation in Census 2000, it is more crucial to examine modeling errors for various levels of expected nonresponse.
As we expect, the standard error of the predicted level of nonresponse for individual tracts is correlated with increases in the expected nonresponse rate. But the predictive errors do not increase precipitously. Approximately 6 percent of the tracts in the country had estimated 1990 nonresponse rates exceeding 40 percent. For those 3,539 tracts the expected nonresponse rate was 44.1 percent--extremely close to the actual nonresponse rate of 45.1 percent. This research shows that we have the ability to identify and isolate local areas where nonresponse is expected to be high provided that there has not been a significant transformation in the demographic characteristics of the tract or block group.
The unit of analysis in section 5.2 is the block group
|Table 5.2||No. of
|All Block Groups||203,034||87.8||25.3||25.3||8.1|
|Block Groups with
expected nonresponse of
|Less than 20 percent||64,650||26.0||16.2||17.7||7.3|
|Between 20 and 30 percent||90,150||40.3||24.6||24.2||8.2|
|Between 30 and 40 percent||35,267||16.3||35.2||34.3||9.0|
|More than 40 percent||12,967||5.1||45.7||44.3||10.1|
The "average" block group contains 426 households and the "average" tract has 1,492 households. Therefore it is not surprising that the standard error of the estimation model is somewhat higher for block groups than tracts. Because block groups are more "demographically" homogeneous than tracts, we find that considerably more households are located in block groups (as opposed to tracts) with expected nonresponse rates below 20 percent or over 40 percent.
A very careful reader may note that the total number of households is 87.9 million in table 4.1 but only 87.8 million in table 5.2. That is not a typographic error. In this paper, our data set consisted not of all tracts but only tracts with 100 or more households. Likewise, our data set was limited to block groups containing at least 50 households. The reader may also note that the number of tracts and block groups in table 1 is somewhat different than what is shown here. In table 1, and only in table 1, tracts and block group lines crossing political boundaries created an artificial tract or block group. This anomaly issue was rectified in the later tables.
This paper introduced a systematic process for isolating small geographic areas (tracts and even block groups) where the Census Bureau might expect difficulty in conducting the 2000 Census. The premise for this model is that area specific nonresponse rates in one Census (e.g., 1990) will carry forward to the next unless there is a significant transformation in attitudes toward the census. Rather than using actual 1990 nonresponse rates, we have built an estimation model based on the relationship of four demographic variables that are highly correlated with nonresponse. The four characteristics which are available from every Census form are:
|Race/Ethnicity||White Non-Hispanics||All Remaining Racial/Ethnic Groups|
|Family Composition||Living with a Spouse||No Spouse Present|
|Type of Structure||Reside in Building||Live in Mobile Home|
The national nonresponse rate for White Non-Hispanics who own (as opposed to rent), live with a spouse, and do not reside in a mobile home in the 1990 Census was 13.2 percent--about one-half the all household rate of 25.3 percent. At the other end of the spectrum, there is one stratum--Hispanics without spouses who are renting a mobile home--where the national nonresponse rate is 64.3 percent. The first stratum consists of households whose characteristics appeared only on the left side of the preceding table. The second stratum contained persons whose characteristics matched the attributes on the right side of the table. In between these two extremes of nonresponse lie 38 other stratum combinations consisting of various racial/ethnic groups, tenure, family formation and building structure.
This analysis uncovered a "stairstep" pattern in nonresponse rates. That is, each shift from left (low nonresponse risk) to right (high nonrespones risk) increases the nonresponse rate by 10 to 15 percentage points. The illustration below begins with a stratum (White non-Hispanic owners who live with a spouse in a building) who are low in each of the four demographic attributes. In that stratum the nonresponse rate attains a minimum value of 13.2 percent, as shown on row 1 of the table on the next page.
Switching one of the attributes from low risk to high risk (owner to renter in row 2) raises the rate by 13 percentage points to 26.4 percent. Changing a second variable to high risk (Spousal to Nonspousal in row 3) increases the nonresponse rate another 14 percentage points to 40.8 percent. Including a third high risk variable (mobile home) raises the rate for this White Non-Hispanic group to 54.1 percent (row 4). Finally, by deleting the race/ethnic category (Non-Hispanic) and replacing it with Hispanic we create a "maximal" nonrespone rate of 64.3 percent.
|Type of Structure||Nonresponse
|White Non-Hispanic||Renter||Non-Spousal||Mobile Home||54.1||+13.3|
In order to estimate nonresponse rates for any specific geographic area, we assumed that we could apply national rates of nonresponse for each of the 40 stratum and accumulate to the boundaries of the desired area. For the 55,000 tracts--(average size of 1,500 units)--the standard error of the estimate was 7.0 percent; for the 203,000 block groups the standard error of the estimating procedure was slightly greater--8.1 percent. In summation, it should be possible to estimate relatives levels of response to the 2000 Census from knowledge of demographics in 1990. We may not be able to pinpoint the rates exactly but we can identify the local areas which are likely to present problems in 2000.
1. Studies that analyze mail return rates by demographic characteristics
Krenzke, Thomas (1997), "Profile of the Last 10% of Returns in the 1990 Census," DSSD Briefs, Information, and Topics Memorandum Series #E-1, Decennial Statistical Studies Division, January 15.
Robinson, J. Gregory (1996), "Demographic Analysis of Mail Return Rates," memorandum to Arthur J. Norton, Population Division, May 30.
Griffin, Deborah H. and Susan P. Love (no date) "1990 Decennial Census Results - Additional Analysis of Response Data," memorandum to Ruth Ann Killion.
Cecco, Kevin (1994) "Characteristics of Responding and Nonresponding Households From the 1990 Decennial Census," DSSD REX Memorandum Series #YY-3, Decennial Statistical Studies Division, March 29.
Fay, Robert E., Nancy Bates, and Jeffrey Moore (1991) "Lower Mail Response in the 1990 Census: A Preliminary Interpretation," Proceedings of the 1991 Annual Research Conference, pp. 3-32.
Kulka, Richard A., Nicholas A. Holt, Woody Carter, and Kathryn L. Dowd (1991) "Self- Reports of Time Pressures, Concerns for Privacy, and Participation in the 1990 Mail Census," Proceedings of the 1991 Annual Research Conference, pp. 33-54.
2. Studies that analyze mail response rates for geographic areas (e.g., States, test sites), but not by demographic characteristics
Treat James B. and Cresce, Art (1995) "Response Rate Analysis: Respondent-Friendly Implementation Evaluation and Multiple Sample Forms Evaluation," 1995 Census Test Results Memorandum No 8.
Robinson, J. Gregory (1995) "Analysis of 1995 Mail Response Rates," memorandum for the record, April 17.
Bolton, Debbie (1993) "Comparison of Mail Return Rates for the 1990 Test Census Sites," memorandum to Susan M. Miskura and Jim Dinwiddie, July 15.
Barrett, Diane F. (1992) "1990 Census Mail Return Rates," DSSD 1990 REX Memorandum Series #Q-13, U.S. Bureau of the Census.
Wilson, Jerusa C. (1991) "Factors Related to Census Mail-Response Rate and Census Undercount in the 1990 Decennial Census, Part One: Analysis of Mail-Response Rate," Center for Survey Methods Research, U.S. Bureau of the Census.
Keane, John G. (1988) "Detailed Report on the 1988 Dress Rehearsal Mail-Response Rates," memorandum to Robert Ortner, June 23.
Harner, Deborah A. (1986) "Mail Response History--1970, 1980 and 1990 Decennial Testing Cycles," memorandum to Susan M. Miskura, May 15.
3. Studies that use multivariate analysis to "predict" mail response patterns for geographic areas
Gbur, Phillip M. (1996) "Database for Targeting and Other Uses," Evaluation Project #8, 1995 Census Test Results.
Mulry, Mary and Richard Griffiths (1996) "Comparison of CensusPlus and Dual System Estimates," 1995 Test Census Results Memorandum No. 43.
Keeler, Jay (1992) "Using SAS Econometric Modeling to Predict Year 2000 Census Mail Response," memorandum to Susan Miskura, March 31.