Penelope Weinberger, AASHTO, Pweinberger@aashto.org
The CTPP Oversight Board met in October following the Census Data Conference held in Irvine, California. One big decision the board made was to extend the CTPP program for another year, through 2013. The funds are available and the need is certainly there, as the CTPP tabulation based on five-year ACS is not expected to be delivered until 2013. We are reviewing proposals for Commuting in America IV and expect a product which includes a web-based portal for information on commuting trends.
The CTPP training team continues to "spread the word" by visiting States and MPOs to provide day-and-a-half-long training, and attending conferences and meetings with workshops and presentations, covering what the CTPP is, how to best use it and caveats for use. Additional ways to learn about CTPP products and issues are through the e-Learning modules http://ctpp.training.transportation.org/ Topics include ACS, CTPP based on ACS, Geography, Margins of Error, and changes to the CTPP related to the discontinuation of the Census Long Form. More e-Learning modules are expected soon.
Jonette Kreideweis, TRB Conference Chair, Jonette.firstname.lastname@example.org
A TRB sponsored "Using Census Data for Transportation" conference held in October 2011 affirmed the critical importance of demographic, mode and travel flow data from the U.S. Census Bureau for transportation planning applications. Presentations from the conference are available at http://www.trb.org/conferences/Census2011.aspx. Katie Turnbull from TTI is preparing the conference summary.
Nearly 120 Federal, state, MPO, consultant, and academic participants shared their experiences in using census data and identified opportunities for overcoming issues and improving census data for decision-making. Presentations demonstrated the value of census data for planning, policy, travel demand modeling, environmental justice, and transit studies.
Participants heard how Census Bureau's (CB) American Community Survey (ACS) standard tables are being integrated and supplemented with data from the Census Transportation Planning Products (CTPP) program, the Longitudinal Household Dynamics (LED) program and an increasing number of private and publicly available data sources.
Data providers, researchers and technical experts shared how they are dealing with data issues and challenges and discussed new products and processes to improve the utility and availability of census data products. Breakout sessions identified themes to guide future census data efforts:
Research Areas (selected):
Demographic Forecasting at NCTCOG
Figure 1. Map of 12-county MPA.
The travel demand model at the North Central Texas Council of Government (NCTCOG) covers the metropolitan planning area (MPA) that includes 12 counties with a total area of about 10,000 square miles, as shown in Figure 1. The total population of the region has grown from 5,199,317 in year 2000 to 6,417,724 in year 2010, a total increase of 23.43 percent, as shown in Table 1. NCTCOG has historically used the DRAM/EMPAL model for the distribution of the population and employment control totals in the region. In the recent update to the demographic forecasts the G-LUM open-source application written in MATLAB script (www.mathworks.com) at The University of Texas at Austin (http://www.ce.utexas.edu) was used. The travel model at NCTCOG has 5,386 traffic survey zones (TSZ) and 242 demographic forecasting districts. Note that the 2010 Decennial Census and ACS (2005-2009) population data was not available until after the completion of the forecasting process. The population and employment control totals used in the forecasting process were purchased from the Perryman Group in year 2007 and verified against the other available data sources (State Data Center, Texas Water Development Board, NCTCOG historical datasets, and ACS data). The 2000 household dataset was created based on the Decennial Census data and modified based on local input. However, the 2000 employment and 2005 household and employment datasets were constructed based on an in-house development monitoring program.
Table 1. Total Population Growth by County
|County||Census 2000||Census 2010||Difference||Percent Total Growth||Percent Simple Annual Growth||Percent Share of Total Growth|
The household population in year 2005 and all forecast years were calculated based on the estimated number of households (HH) and the average HH size in each corresponding year. Therefore, it was necessary to first obtain the average HH size in year 2005 and to establish a method for calculating the future average HH sizes. The census data products indicate that the average HH size in the NCTCOG region increased from 2.70 in year 2000 (Decennial Census) to 2.81 in year 2005 (ACS 2005). This increase was considered significant and therefore triggered the need for reviewing the historical trend of the average HH size in the region based on the data available in the Census Fact Book and the Decennial Census, as shown in Figure 2. We also compared the average HH sizes based on the 1990 and 2000 Decennial Census against the values reported by ACS products, as shown in Figure 3.
These comparisons did not show a logical pattern that could be used for forecasting the future HH sizes in the region. However, it did show differences in the average HH size between urban and rural counties and that the average HH size becomes more stable as an area becomes more urbanized. The purchased Woods and Poole data indicates that the average HH size for the NCTCOG 16-county area will not exceed 2.70 in the next 20 years. The final HH sizes utilized in the forecasting process were calculated starting from the ACS 2005 values and based on the assumptions listed below:
The 2010 Census and ACS (2005-2009) became available after the forecasts were produced. Hence, they were only used to further verify the changes in the average HH size of the states against the Census 2000 data and to evaluate the difference between the ACS (2005-2009) and 2010 Decennial Census. This comparison indicated that based on the ACS (2005-2009), 19 states have shown an increase in their average HH size, as displayed in Table 2. However, the 2010 Decennial Census indicated that the average HH size only increased in five states: Nevada, California, Florida, Delaware, and Texas. The ACS (2005-2009), overestimated the average HH size with rather small margins of error (MOE) when compared to the 2010 Decennial Census in all the states where it represented an increase in average HH size compared to the 2000 Decennial Census. In four of the states (Nevada, California, Maryland, and Tennessee), the 2010 Decennial Census average HH size was at the lower boundary of the ACS (2005-2009) estimated range.
Population to Employment Ratio
The population to employment ratio (P/E) in the NCTCOG region has been historically around 1.60. It was more conservative for travel model forecasting to assume that it will remain at 1.60 in all forecast years. The control totals as provided by the Perryman Group also followed the same trend in all forecast years. The original NCTCOG calibration dataset for year 2005 was modified such that the total number of households and average HH size matched the ACS 2005 data in the 12-county MPA, and the P/E ratio was 1.60. These adjustments resulted in a reduction of about 200,000 in the household population and an increase of about 300,000 in the number of employees in the region for year 2005 compared to the original 2005 dataset.
|State||Census 2000||Census 2010||Percent Difference vs. Census 2000||ACS (2005-2009)||M
|Percent Difference vs. Census 2000||Percent Difference vs. Census 2010|
|District of Columbia||2.16||2.11||-2.31||2.21||0.02||2.31||4.74|
Table 2. Comparison of Average HH Size
The ACS (2005-2009) HH data was used during the calibration and validation of the forecasting model. For this purpose, the estimate of the 2010 HHs was calculated based on the ACS (2005-2009) values grown for six months based on the growth rates observed between ACS 2008 and ACS 2009, and then compared to the 2010 forecasts. The results of this comparison indicated that the total 2010 HHs resulting from the forecasting process was only 0.49 percent higher than the estimates we had calculated based on the ACS (2005-2009) data, with acceptable errors at the county levels, as shown in Table 3. The 2010 Decennial Census data became available at a later stage in the process (April 2011) and hence only could be used as another comparison point for validating the 2010 forecasts that were prepared based on the ACS (2005-2009) data, as shown in Table 4. This comparison indicated that the forecasted household population for year 2010 was only -0.28 percent different in total compared to the 2010 Decennial Census, with acceptable errors at county levels. We also validated the 2010 HH population forecasts in randomly selected TSZs in a 10-mile radius. All the comparisons indicated that the 2010 forecasts match the available reference data for year 2010 with an acceptable margin of error.
Table 3. Validation of 2010 HH Forecasts
2010 Decennial Census
Since the 2010 Decennial Census data was not available during the forecasting process, it was not until after the fact that we were able to compare the estimated 2010 HHs based on the ACS (2005-2009) with the actual 2010 Census numbers at the 242 demographics forecasting district level. This comparison showed that the two data sources provide rather comparable data at the region and county levels, as can be seen by reviewing the data in Tables 3 and 4. However, there are some major inconsistencies in the ACS (2005-2009) data in the areas that were of most concern for the region due to their rapid growth or decline in population in the recent years, as shown in Figure 4. The areas shaded in red indicate that the ACS (2005-2009) data underestimated the 2010 household population by more than 20 percent. The dark green areas indicate an overestimation of more than 20 percent. The areas of concern, which are circled in black, include northern/southern Tarrant County and most of Collin and Denton Counties which experienced a rapid growth in the recent years, and northern Dallas County that experienced some reduction in population and employment in the previous years, as shown in Figure 5. The comparison of Figures 4 and 5 also shows that the ACS (2005-2009) has the tendency to overestimate the HH population in areas with small growth and underestimate it in areas with rapid growth in the DFW area.
Table 4. Validation of 2010 HH Population Forecasts
|County||2010 Forecast||2010 Census||Difference||Percent Difference|
Aside from the numerous data-related technical issues that we had to resolve during this process, there were a couple of nontechnical issues, related to the perceived accuracy of the ACS and Decennial Census, which had to be addressed as well: 1) the overall disagreement and resistance of the demographers regarding the use of ACS products due to the sampling nature of it and the associated margins of error; and 2) the disagreement of some local governments with the data published by the Decennial Census products claiming that it underreports the minority population.
Figure 4. Comparison of HH Population based on ACS (2005-2009) estimates and Census 2010.
Figure 5. Comparison of HH Population based on 2000 and 2010 Decennial Census.
The Census data products make valuable and up-to-date data available to the transportation community in fairly small geographies. The uniqueness of this data source makes it unrealistic to assume that it will not be used due to the associated accuracy disclaimers. This data is often times used for short-term decisions that have real consequences. Therefore, the need for consistency is an actual issue with which the users are faced. It seems that improvement in the expansion of the data in the subcounty level can improve the usability and credibility of the ACS datasets to a large degree.
This report describes a preliminary look at ACS data on Means of Transportation to Work and Travel Time to Work at the block group level. The findings are a useful preview of what users can expect from ACS-based CTPP for very small areas.
Even after combining five years of data collection, the ACS sample used to provide small area data is considerably smaller than that of the Census Long Form - with 2005-2009 ACS data based on responses from only 7.6 percent of housing units, compared with 15.8 percent from the 2000 Census Long Form. As shown in Table 5, the ACS data for over 90 percent of block groups are based on fewer than 100 interviews.
Given the relatively small ACS sample, large margins of error (MOE) are expected, and Table 6 shows how often each cell in the Means of Transportation table has an MOE larger than the estimated value. The table also reveals that in many of these cases, the cell value was zero (as expected since many modes are rare in many areas). And as Table 7 illustrates, for a block group in Orange County, California, the ACS reports large MOEs for cells with a zero estimate. Travel Time to Work reflects a more even distribution, with fewer zero values, and as shown in Table 8, MOEs often exceed cell values even where cell values are greater than zero. In other words, the large MOEs cannot be blamed entirely on cells with zero or very small values.
Table 9 shows an unusually high number walking to work in another Orange County, California block group, leading one to wonder if it might be an ACS error. But given the location at the University of California, Irvine, and the fact that the 2000 census showed a similar distribution, the preponderance of walkers is plausible.
With a sample smaller than that of the Census Long Form, one would expect a higher frequency of zero values in the ACS block group data, and Table 10 confirms this tendency for Travel Time to Work. For example, for travel time "Less than 5 Minutes," the ACS reports 80,253 block groups with a value of zero compared with only 53,328 for the 2000 Census Long Form.
An examination of uncommon transportation modes confirmed that the ACS puts large numbers in expected areas. For example, block groups with the highest percent of workers commuting by "subway or elevated" were found in counties, including New York, Bronx, Queens, and Kings in New York. Workers commuting by "ferryboat" were most common in block groups in Hudson, New Jersey, Kitsap, Washington, and Richmond, New York. Just as important, the ACS seems not to show large numbers using uncommon modes in areas where they would not be expected. But there are occasional exceptions, as illustrated by Table 11 - which shows a conspicuously large number of workers (618) commuting by bicycle in a block group in Larimer County, Colorado.
Unlike the UC Irvine block group, where the large number of walkers could be expected (and was backed by the 2000 Census), there is no apparent explanation for the dominance of bicycles in the Larimer County block group (and 2000 Census provides no backup). The more likely explanation is that the ACS captured one or a few bicycle commuters, and weighted up to 618. But why would ACS weight to such an unrealistically large number? The probable explanation is that overestimation in this block group (and presumably others) compensates for a large number of block groups that actually have one or a few bicycle riders, but show zero because none were captured in the ACS sample. In other words, overestimation of this type is the flip side to the large number of zero cells reported by the ACS, and enhances the accuracy of ACS data for aggregate areas. And recall that the Census Bureau recommends that ACS block group data be used only for aggregate areas.
A broader assessment of the ACS block group data was provided by computing the index of dissimilarity (IOD) between ACS distributions and those from the 2000 Census. A measure of the difference between two percent distributions, the IOD ranges from 0 for distributions that are identical, to 100 for distributions that have nothing in common (where 100 percent of records in one distribution would have be in a different category to replicate the percent distribution of the other). Differences are expected between ACS and 2000 Census distributions, but there should be considerable similarity, and it can be useful to see where consistency is greatest.
As summarized in Table 12, the mean IOD is consistently lower for Means of Transportation than for Travel Time - probably because the dominance of "drive alone" is reflected in both ACS and the 2000 Census. For all block groups, the IOD was 15.1 for Means of Transportation and 30.7 for Travel Time. Limiting to block groups with stable population (change less than five percent from 2000 to 2010) reduces the IODs only slightly. It is when limiting to block groups with 100 or more ACS interviews that the IODs drop sharply to 8.2 and 18.3. Block groups in counties with populations of 500,000 or more actually have the highest IODs at 17.2 and 32.2. With large metropolitan counties typically of interest to transportation planners, one might wonder why consistency with 2000 Census is relatively low in these areas.
The relatively large IODs in large metropolitan counties might be a byproduct of a small town bias in the ACS sample. The ACS samples heavily in small towns to the detriment of the sample allocated to small statistical geographies in large metropolitan areas. Consequently, while 9.3 percent of households responded to ACS nationwide, the percent was only 7.7 for block groups in counties with populations of 500,000 or more.
Frequency of updates is a highly touted ACS benefit, and a review of ACS data confirms its ability to add value in areas with rapid population growth. For example, Table 13 shows the ACS Means of Transportation distribution for a block group that was part of Denver's former Stapleton Airport, and was developed following the 2000 Census. The census counted zero households in 2000, but 4,084 by 2010.
The preliminary analysis suggests a mix of strengths and limitations for small area ACS data. Users can expect large margins of error, but many are of little consequence, as they relate to reasonable estimates of zero or very small numbers. For small areas, the number of ACS interviews (unweighted sample housing units) might provide a better sense for data quality than the MOEs. The ACS certainly estimates "zero" in many cells that should have small numbers, and occasionally inflates cell values to unrealistically high levels. But some of the most questionable estimates for individual block groups contribute to enhanced accuracy for aggregate areas. Transportation planners might prefer that more ACS samples be allocated to small statistical geographies in metropolitan areas (as opposed to small towns), but the benefits of frequent updates are apparent in areas with rapidly changing populations. In short, both challenges and opportunities are apparent in the ACS block group journey to work data, and users are likely to encounter similar challenges and opportunities in the forthcoming ACS-based CTPP data.
|Missing (no ACS)||1,533||0.7|
|N suppressed (1 or 2)||801||0.4|
|500 or more||73||0.0|
|Means of Transportation||Pct MOE
GT Cell Value
Cell Value = 0
|Bus or trolley bus||90.1||63.7|
|Worked at home||76.5||34.6|
|Means of Transportation||Workers||MOE|
|Total Workers||271||+/- 114|
|Drive alone||207||+/- 91|
|Bus or trolley bus||0||+/- 132|
|Other means||15||+/- 25|
|Worked at home||30||+/- 34|
|Travel Time to Work||Pct MOE
GT Cell Value
Cell Value = 0
|LT 5 minutes||80.7||39.2|
|5 to 9 minutes||44.7||12.2|
|10 to 14 minutes||32.0||6.9|
|15 to 19 minutes||29.8||6.0|
|20 to 24 minutes||33.4||7.5|
|25 to 29 minutes||65.7||24.4|
|30 to 34 minutes||39.4||9.7|
|35 to 39 minutes||85.3||46.5|
|40 to 44 minutes||80.5||40.4|
|45 to 59 minutes||61.2||22.7|
|60 to 89 minutes||72.3||31.8|
|90 or more minutes||88.4||49.3|
|Means of Transportation||Workers||MOE||2000 Census|
|Total Workers||4,289||+/- 1,286||4,121|
|Drive alone||1,138||+/- 307||1,456|
|Bus or trolley bus||22||+/- 26||16|
|Other means||0||+/- 132||0|
|Worked at home||185||+/- 213||130|
|Travel Time to Work||2005-2009
|Census 2000 SF3||ACS Increase|
|LT 5 minutes||80,253||53,328||26,925|
|5 to 9 minutes||23,839||10,387||13,452|
|10 to 14 minutes||12,766||4,458||8,308|
|15 to 19 minutes||11,033||5,352||5,681|
|20 to 24 minutes||14,151||4,995||9,156|
|25 to 29 minutes||49,332||26,079||23,253|
|30 to 34 minutes||18,696||6,610||12,086|
|35 to 39 minutes||95,559||66,290||29,269|
|40 to 44 minutes||82,696||56,001||26,695|
|45 to 59 minutes||45,850||23,259||22,591|
|60 to 89 minutes||64,843||35,692||29,151|
|90 or more minutes||101,444||57,749||43,695|
|Means of Transportation||ACS||MOE||2000 Census|
|Bus or trolley bus||0||123||0|
|Worked at home||12||19||14|
|Block Group Type||Means of Transportation||Travel Time|
|All Block Groups||15.1||30.7|
|Pop change less than 5 pct||14.2||29.7|
|100+ ACS Interviews||8.2||18.3|
|In county with Pop 500,000+||17.2||32.2|
|Means of Transportation||2000 Census||ACS||MOE|
|Total Workers||0||4,177||+/- 237|
|Drive alone||0||3,221||+/- 221|
|Bus or trolley bus||0||94||+/- 68|
|Other means||0||43||+/- 34|
|Worked at home||0||476||+/- 124|
CTPP Hotline - 202/366-5000
CTPP Website: http://www.fhwa.dot.gov/planning/census_issues/ctpp/
FHWA Website for Census issues: http://www.fhwa.dot.gov/planning/census_issues/
2005-2007 ACS Profiles: http://download.ctpp.transportation.org/profiles_2005-2007/ctpp_profiles.html
AASHTO Website for CTPP: http://ctpp.transportation.org
1990 and 2000 CTPP downloadable via Transtats: http://transtats.bts.gov/
TRB Subcommittee on census data: http://www.trbcensus.com
Jennifer Toth, AZDOT
Susan Gorski, MI DOT
Census Bureau: Housing and Household Economic Statistics Division
The CTPP Listserv serves as a web-forum for posting questions, and sharing information on Census and ACS. Currently, over 700 users are subscribed to the listserv. To subscribe, please register by completing a form posted at: http://www.chrispy.net/mailman/listinfo/ctpp-news On the form, you can indicate if you want e-mails to be batched in a daily digest. The web site also includes an archive of past e-mails posted to the listserv.