Expert Meeting on Income, Poverty, and Health Insurance 3/20/15 (Part 1 of 3)

>>Okay. Good morning. Welcome to the Census
Bureau. I think we had a couple of options on days to have this meeting. One of them
was going to be Friday the 13th and the other one was the first day of spring. We chose
the first day of spring and we got snow. So, you can’t get away from Friday the 13th one
way or the other, so. Good morning. I’m Enrique Lamas. I’m the Associate Director for Demographic
Programs at the Census Bureau and I want to welcome you to — I was hoping to say sunny
Suitland — but to Suitland. What we really want to do today is get your thoughts and
you input on a couple of things that we are thinking about and a couple of options that
we have going forward with the ASEC, with the Current Population Annual Social Economic
Supplement. So ASEC, as you all well know we collect every spring of the year and we
use the ASEC to measure the economic well-being of the population, and that’s the primary
purpose of it. In doing that, we measure income, poverty, labor force, attachment during the
previous calendar year. We measure health insurance coverage, all of the things that
directly measure your economic well-being or related to it. And it is the official source
of the poverty estimates for the Federal Government. So, coming into last year, we had seen several
trends, in particular with income. We knew, obviously, that interest and dividend income
were not particularly well-reported in the CPS. Respondents may not exactly know how
much they have earned in interest in a particular type of income. So there’s some kind of recall
issues related to that. We also knew that the defined contribution programs or 401K
programs were growing, and whether people think of those as assets or withdrawals from
those as income was another big issue for us. So, on the income side we wanted to think
about how we might address some fairly well-known issues related to the measurement of income.
And on the health insurance side we knew that we had some measurement issues there where
people forgot about short spells of health insurance coverage. And this was all well-documented
in a lot of research that we had been doing over the previous decade, 15 years, both inside
the Bureau and outside the Bureau. So we wanted to make some changes, in particular prior
to the Affordable Care Act implementation going into effect which was starting this
year. So we wanted to have a good baseline to measure the impact of the Affordable Care
Act. So last year we did changes to our income and to our health insurance questions. In
the spring of ’14 we collected information about calendar year ’13, and in particular
what we did about income was that for 5/8 of the sample we collected the old income
questions so we could go back through time and measure it consistently until 2013. And
then for 3/8 of the sample we asked respondents new questions, totally redesigned income questions,
and that will be going forward. And you’ll hear all about this today. I’m going to go
over the agenda. This is just kind of the introductory piece. But what we really want
to do is think about how we look at last year’s data to compare it to the data that we’re
collecting this year. You’re going to hear a presentation from John Rothbaum that will
look at how we might model from the 3/8 of the sample to the other 5/8 to get a full
sample for the income side that we could then use to compare to the data that we’re collecting
this year. And we’re looking at multiple imputations, all kinds of technical things. So it’s an
interesting innovative work that John has done. The purpose of this meeting really is
to get your input. We didn’t want to make the decision totally on an insular basis within
the Census Bureau. We wanted to have people around the table that knew a lot about income,
poverty; knew a lot about the health insurance coverage. And then get your input as we make
a decision going forward. We will probably make a decision, and we’ll talk a little bit
more of this at the end of the day, in the May timeframe. And the reason for that is
twofold. One is we have a lot of work that we do once we get the March supplement or
the ASEC data, and between there and when we release the data in September of each year.
So we have to get going with all of our processing and doing all of our tabulations and the report-writing
and all of that stuff. The second main reason is that we do not want to get perceived as
having the data for 2015 and then making a methodological decision on what to compare
it to for last year based on the results. We wanted to be transparent, make the decision
beforehand without having the knowledge of what the data that we’re collecting this year
are all about. So that’s kind of the approach. We wanted to have this meeting to get your
input. I hope that we can have a very good discussion as we go through. What we’ll do
today is — Tory Velkoff, who’s to my right, will make a presentation and will tell you
all that we did, that 3/8, 5/8, all of that comparison, kind of give you a good walkthrough
through it. And then Marina is going to talk a little bit about the health insurance results
that we had last year. Edward Welniak will talk about the income results. Trudi Renwick
will talk about the poverty and the work that we’re going. And then Jonathan Rothbaum will
talk about that modeling that we’re thinking of doing. So the idea that — we’ll lay it
all out for you. And then Chuck and Jennifer and Jennifer. Chuck Nelson and then Jennifer
Day and Jennifer Madans will talk about kind of how we plan to go forward from here until
September when we release all of our data. So, with that kind of as an overview of what
we would like to do today, if that — any questions? Did I confuse all of you? Are we
good? You’ll note that we have cameras around the room. The purpose for that is that this
meeting is going to be webcast. It’s webcast to other people that are interested in this
topic. And we will also put a copy of it up on our Website so that people can refer to
this meeting later on, so, without any questions or anything like that, what I thought we would
do is go around the room so that you get a chance to introduce yourself. You see there’s
a big red button or actually it’s a white button that will turn red once it gets around
to you all. And maybe we can introduce ourselves, those — name and rank and serial number.
>>My name is Tori Velkoff . I’m the Chief of the Social, Economic, and Housing Statistics
Division here at Census.>>I am Jennifer Day. I’m the Assistant Division
Chief for Employment Characteristics in the Social, Economic, and Housing Statistics,
the Census Bureau.>>I’m Jennifer Madans. I’m the Associate
Director for Science at the National Center for Health Statistics.
>>I’m Thesia Garner, Senior Researcher at the Bureau of Labor Statistics, and I work
with Consumer Expenditure Survey and Supplemental Poverty Measurement.
>>I’m Joe Schafer. I’m from the Center for Statistical Research and Methodology here
at the Census Bureau.>>I’m Howard Iams. I’m a Senior Researcher
at the Social Security Administration.>>I’m Robin Cohen. I’m a Statistician at
NCHS, and I primarily work on health insurance.>>I’m Stephen Blumberg. I’m the Associate
Director for Science for the National Health Interview Survey at the National Center for
Health Statistics.>>Good morning. I’m Jessica Batham, Deputy
Assistant Director in the Health Division at the Congressional Budget Office.
>>Good morning. I’m Don Olarch [assumed spelling]. I’m the Deputy Chief Economist in the Office
of the Assistant Secretary for Planning and Evaluation at HHS.
>>Hello. I’m Laurel Wheaton from the Urban Institute and Senior Fellow.
>>Hi. I’m Mike Boudreaux, Assistant Professor at the University of Maryland.
>>John Czajka, Senior Fellow at Mathematica Policy Research.
^M00:10:02>>I’m Gary Claxton with the Kaiser Family
Foundation.>>I’m Rachel Garfield. I’m a Senior Researcher
at the Kaiser Family Foundation.>>Stephen Tordella at Decision Demographics.
>>Connie Citro, Director of the Committee on National Statistics at the National Academy
of Sciences.>>Jim Ziliak, Professor of Economics at University
of Kentucky.>>Jevalo Riunaton [assumed spelling], but
everybody calls me Ragu. I’m the Director of [inaudible] at the University of Michigan.
>>Kathleen Call at the University of Minnesota State Health Access Data Assistance Center,
SHADAC.>>Joanna Turner, Senior Research Fellow,
also at the University of Minnesota, SHADAC.>>Tom Lewis, Associate Director for Research
and Methodology and FROG. [ Inaudible Speaker ]
>>In our city first.>>Charles Nelson, Assistant Division Chief
for Economic Characteristics, Social Housing, and Economics Statistics Division of the Census
Bureau.>>John Rothbaum, Income Statistics Branch
in the Social and Economic Housing Statistics Division, Census.
>>Trudi Renwick, Chief of the Poverty Statistics Branch.
>>Ed Welnick [assumed spelling], Chief [inaudible] Branch.
>>Marina Vornovitsky, Chief of the Health and Disability Statistics Branch.
>>Okay. Thank you. So for the folks at the table that don’t know everybody around the
outside of the room, I’m going to ask the folks on the outside — you have a mic somewhere?
>>Yeah, right there.>>They’re all going to have to walk up to
the mic?>>They’ll see in webcast.
>>If you could just stand up and give your name and where you work, that would be useful.
Jessica [inaudible] from the Income Statistics Branch.
>>Carla Mediya, Health and Disability Statistics Branch.
>>Thank you.>>Sharon Stern, Special Assistant to Tori.
>>Emily Soscopoletti [assumed spelling], and I’m a demographer at the Census Bureau,
and I work for Tori as well.>>Sharon Long, Senior Fellow at the Urban
Institute.>>Greg O’Hara, Census Bureau.
>>Amy Steinway, Health and Disability Statistics Branch.
>>[Inaudible] Malloy, the Health Division of the Congressional Budget Office.
>>Robert Aashmeen, Center for Statistical Research and Methodology.
>>Rick Ganby, Demographic Surveys Division in the Census Bureau.
>>Timothy Kennel, Demographics Statistical Methods Division, Census Bureau.
>>Carman DeNavas-Walt , Income Statistics Branch, Census.
>>Melissa Kohler, Income Statistics Branch.>>Doug Allegate [assumed spelling], Center
for Statistical Research and Methodology.>>Jennifer Ertman, Assistant Division Chief
for Social Characteristics in the Social, Economic, and Housing Statistics Division.
>>Art Prechey [assumed spelling], Assistant Division Chief for Housing Characteristics,
Social, Economic, and Housing Statistics Division.>>Aaron Dixon, Poverty Statistics Branch.
>>Laryssa MyKyta, Poverty Statistics Branch.>>Joelle Abramowitz, Health and Stability
Statistics.>>Danielle Taylor, Health and Disability
Statistics.>>Kathleen Short, Census Bureau.
>>Al Gottschalk, Assistant Division Chief, Small Area and Longitudinal Estimates.
>>Marissa Gadrace [assumed spelling], Bureau of Labor Statistics.
>>Wes Basil, Small Area Estimates Branch.>>Shannon Locke with the CBO.
>>Lauren Bowers, Small Area Estimates Branch.>>Lucy Dalzell, Chief of the Small Area Estimates
Branch. We manage SAHIE.>>David Powers, Small Area Estimates Branch.
^E00:14:08 ^B00:14:17
>>Okay.>>Thank you. Good morning and thank you all
for coming. As Enrique said, we’re very interested in getting your feedback on our plans for
CPS ASEC. I’m going to briefly review the changes we made to the 2014 ASEC and then
talk about our 2015 data collection plans. As some of you are aware, we made changes
to the 2014 CPS ASEC. We made changes to some of the income questions and to many of the
health insurance questions. I will review these changes with you and explain how we
implemented the changes last year. I’ll also talk about what we released last year. And
then I’ll talk about what we’re doing this year in terms of data collection. And, finally,
I want to talk a little bit about our Small Area Health Insurance Estimates, or SAHIE
as we call them. In 2014, we introduced both redesigned income and redesigned health insurance
questions into the ASEC. For income, we wanted to reduce non-response. We also wanted to
improve some questions, particularly on pension withdrawals and asset income. For health insurance,
we redesigned questions to improve the accuracy of the ASEC and to ensure that we could properly
measure coverage under the Affordable Care Act. So why did we need to change the income
questions? Well, we hadn’t changed them, basically, since 1980. We found that the pension questions
were not adequately collecting to find contribution plan withdrawals. We saw that non-response
was growing and asset income was not well-reported. So we changed the questions in a few ways.
First, we asked about recipiency before collecting amounts. We customized the order of questions
based on the household characteristics. We collected amount ranges for people who were
unable to give us an exact amount. We added questions on retirement account withdrawals
and we added more detailed questions on asset types and amounts. And Ed Welniak will talk
more about these changes later this morning. The changes that we made to the health insurance
question were based on years of research. We worked with other Federal agencies, academics,
private researchers, and state agencies on this research. We conducted two field tests
on the health insurance questions — one in 2010 and one in 2013. And we put out several
^IT Federal Register ^NO notices announcing both the tests and the changes to the questions.
The key finding from the research was that there was an under-reporting of health insurance
coverage in the ASEC. The old health insurance questions were at the household level for
the previous calendar year. Research using administrative data showed that short spells
of coverage were missed. And Marina Vornovitsky will go into more details about this research
and the changes in the next presentation. So, why did we think we needed to change the
health insurance questions? As I said, research found that the estimates of the uninsured
population were too high. There was confusion about current-year coverage and previous calendar-year
coverage. We also found that the estimates were biased due to respondent recall issues.
Again, we saw undercounts of some types of health insurance and we were missing some
people’s health insurance. And then there was a law change. In 2010 the Affordable Care
Act was passed and this meant that we needed to ask additional questions about the marketplace
and changes to employer-sponsored insurance. As I said, the Census Bureau did a lot of
research and testing on the health insurance questions over the last several years. We
were also testing some redesigned income questions at the same time. In 2011 we did a cognitive
test of these income questions and then we tested both the redesign income and the redesigned
health insurance questions in the 2013 CPS ASEC Content Test. That test had a reference
year of 2012. After that test we incorporated both the new income questions and the new
health insurance questions into the 2014 CPS ASEC which has a reference year of 2013. So
how we implemented the questions for income and health insurance differed. For income,
we used a split panel approach. In 2014 we had about 98,000 addresses in the CPS ASEC
full sample. We asked the new questions of about 30,000 of those addresses. And the remaining
68,000 addresses received the traditional income questions. We needed the split panel
design for income because it preserves the time series and provides a bridge between
the old and the new series. The CPS ASEC is the source of the official U.S. Poverty Estimates
so a consistent time series was a necessity, and we thought that the best way to make improvements
and create a bridge was to use the split panel approach. For health insurance we implemented
the questions differently. The time series for health insurance is also important but
knowing that we had other data sources out there, specifically the American Community
Survey and the National Health Interview Survey, and knowing that we needed a very strong baseline
in 2013 with the new health insurance questions, we decided to ask the redesigned health insurance
questions of the entire sample. We needed to establish this baseline in 2013 with the
new health insurance questions and we wanted this baseline in place before the major effects
of the Affordable Care Act took effect. And we can now use this 2013 baseline as a comparison
for future years. And only a full ASEC sample provides reliable estimates for small groups,
some of which are most impacted by the Affordable Care Act.
^M00:20:00 This slide just gives you a visual of how
this was implemented. Again, 68,000 addresses received the traditional or old income questions,
and about 30,000 received the redesigned income questions. The redesigned health insurance
questions were asked of the full sample of around 98,000 addresses. Again, the split
panel design for income and poverty preserves the time series and provides a bridge between
the old and the new series, and the full ASEC sample for health insurance provides reliable
estimates for small groups and establishes a baseline before the effects of the Affordable
Care Act take effect. In September of 2014 we released two reports, an Income and Poverty
Report and a Report on Health Insurance. These two reports, the source of which was the 2014
ASEC referred to calendar year 2013. Both of these reports were based on the 68,000
addresses that received the traditional income questions. Again, just giving you a visual.
The reports on Income and Poverty and the Health Insurance that we released last September
were based on the sample of 68,000 addresses that received the traditional income questions.
We did this for a couple of reasons. Income and health insurance are very closely related,
so we wanted to have a consistent set of income questions for the health insurance report.
Note that the sample is nationally representative of the total U.S. population and also note
that last year in our health insurance report we only focused on calendar year 2013. We
did not look at year-to-year change with the CPS data. We released many other products
based on the 2014 ASEC. We released a public use file based on the 68,000 addresses when
we released the reports. We released the supplemental poverty measure and the report that goes along
with it based on the 68,000 addresses. That was released in October. In January, we released
a public use file that was based on the 30,000 address that received the redesigned income
questions. We also released an extract that contained full file weights, and this extract
gives external users the ability to combine the 68,000 sample with the 30,000 sample and
put it into one full file. And we, the Census Bureau, also released many table packages
on migration and family status, and these were based on a full file, but none of these
variables were cost by income. Since last September we’ve done a lot of research on
the redesigned income questions, examining how they compare with the traditional income
questions. We presented several papers at the Allied Social Science Association meetings
in January, and Ed and Trudy will talk about these results later this morning. And as Enrique
said, we’re also exploring if we can use the data from the redesigned income questions
to model select income variables for the sample that receive the traditional income questions.
And John will talk about that later today. We’re very interested in hearing your feedback
on all of this work. As I said, we changed the health insurance questions in the 2014
ASEC based on years of research and so we would have that strong baseline for 2013.
However, there were some concerns raised about changing the health insurance questions when
we did. And these concerns prompted a request that we ask the old questions as well as the
redesigned questions in 2015. In fact, in our budget language it was fairly specific
that we should collect both the redesigned and the old health insurance questions. Note
the text in purple that says, “The Bureau of the Census shall collect data for the annual
social and economic supplement to the current population survey using the same health insurance
questions included in previous years in addition to the revised questions implemented in the
current population survey beginning in February of 2014.” So, what does this mean for us?
This means in 2015 we’re doing two separate data collections. The first is the production’s
ASEC. This has both the redesigned income and the redesigned health insurance questions,
and they are asked of the full sample. These are the data we will release in September.
These are the data that the official poverty estimates will be based on as well as the
income and health insurance estimates that we release in September. These are also the
data that the supplemental poverty measure will be based on. The second data collection
we’re calling the Parallel ASEC. This parallel sample is going out to about 28,000 addresses.
We designed the sample to be nationally representative, and assuming an 80% response rate, we expect
to be able to measure a 0.3 percentage point change in the uninsured rate. In this parallel
sample we are asking both the old, or traditional, income and the old health insurance questions.
The survey went into the field on March 1st. The data collection was suspended for CPS
Week which is actually this week right now. But we will go back into the field after the
production CPS interviewing is finished for March. For this parallel sample we plan to
release a research file with a working paper. The date of this release is yet to be determined.
Note that while we were asked to ask this additional sample we didn’t receive any additional
resources. So the same people who are processing and analyzing the production ASEC are also
going to be processing and analyzing this parallel sample. Let me talk a little bit
about the goals for our parallel sample. First, note the graph on the slide of this graph.
The blue solid symbols represent the calendar years for which we will have data collected
with the old health insurance questions which are calendar year 2012 and calendar year 2014.
The red line with the red symbols indicate the calendar years for which we will have
data from the redesigned health insurance questions which are calendar year 2013 and
calendar year 2014. We plan to use the parallel sample to compare health insurance over time.
That is, looking at calendar year 2012 and calendar year 2014. We also think that we
can use data from both the old and the redesigned questions to model what coverage might have
looked like in 2013 if we had asked the old health insurance questions. Finally, we can
compare health insurance coverage between the old and the redesigned questions for 2014.
We can look at measures such as percent with health insurance coverage, percent not covered
at any time during the year, and so on. So, in some ways, this parallel sample gives us
a unique opportunity to compare the old and the new questions one more time. Now I’m going
to switch gears a bit and talk about the Small Area Health Insurance Estimates or SAHIE.
The Census Bureau releases model-based health insurance estimates for all U.S. counties
through the SAHIE Program. We also release health insurance estimates at the state level
as well through this program that has race and Hispanic origin detail. In fact, we released
our SAHIE estimates on Tuesday of this week. We produce SAHIE estimates because they provide
the only single year estimates of health insurance for every county in the United States, and
we used a model to produce estimates that typically have lower variances than the survey
estimates. So why am I talking to you about SAHIE at a CPS conference? Well, SAHIE is
one of the health insurance estimates that the Census Bureau puts out. And for the 2014
SAHIE estimates we have some decisions to make and we wanted to talk to you about those
decisions. As I said, the SAHIE estimates are model-based. The inputs for SAHIE are
data from the American Communities Survey, both one- and five-year estimates, data from
Census 2010, data from the Census Bureau’s population estimates, and county business
patterns. We also use other data in the models such as information from tax returns, information
from the Supplemental Nutritional Assistance Program, and from Medicaid and CHP. For the
2014 SAHIE estimates we need to rethink the inputs. Typically, the Medicaid data that
we use in the model are lagged by a few years. For example, the SAHIE estimates that we release
this week, the 2013 SAHIE estimates, used Medicaid data from 2011. We use lag data because
it’s the latest dataset available from CMS that has all of the information that we need.
For the 2014 estimates, using lagged Medicaid data will present an issue. Obviously, using
2012 Medicaid data to model health insurance for 2014 can be a problem. We expect that
health insurance coverage will change in 2014 due to the Affordable Care Act, so we really
need to ensure that the data we’re using on our model can capture this change. So we need
to rethink where we get the Medicaid data or what goes into our model. And, of course,
we would appreciate any advice you have on this. So, thank you. That was my presentation.
I’ll now turn the mic over to Marina Vornovitsky.>>I know I’m breaking what staff were thinking
of doing, but if you have any clarification questions at this point that you want to ask
of Tori, why don’t we do that. And then if not, then we’ll go to Marina’s presentation.
>>Will the main ASEC sample size be the same as usual?
>>Yes.>>The idea was that we wanted to maintain
the same level of reliability for the ASEC year-to-year, especially since it’s the official
source of poverty. ^M00:30:00
And so we had to find additional budget resources to cover the parallel sample. And that’s why
it was 28,000 and all that. Connie.>>In the SAHIE estimates, will there at some
point be an effort to kind of calibrate the CPS with the new questions, with the ACS,
which is the basis of the SAHIE estimates?>>I’m looking at Lucy. Did you hear that
question? Want to repeat the question, Connie?>>It’s about getting the SAHIE estimates
which were based on the ACS [inaudible] a course represent a moving time period and
all that with the new health insurance questions in the CPS and if there’s going to be any
effort at some point to try to, you know, get those in sync.
[ Inaudible Speaker ] ^E00:30:54
^B00:31:03>>Yes. We have tried in the past to use the
CPS along with the ACS. We’ve done that with a sister program to SAHIE, which is the Small
Area Income and Poverty estimates, and we found that there’s really no impact with the
size of the CPS samples so small relative to the ACS that it really doesn’t make a difference.
So, that was the last time we considered that option. We really haven’t talked about it
specific to SAHIE, but it’s something to think about.
>>But, Connie, the CPS and the ACS do track closely together for health insurance.
>>And they’re parallel.>>What’s the sample frame for the parallel?
>>I’m looking at Tim. It’s the retired 2000 CPS. Is that correct? Retired and reserved
sample from the 2000. But Tim Kennel has more information on that.
>>It’s a combination. Most of the sample will come from cases that have finished all
of their CPS interviews. So it will be cases on their ninth, tenth, or in some cases, 11th
interview from CPS that have already rotated out. Then we’re adding some sample that we
selected in 2000 called our Reserve Sample from our unit frame that we’ve selected just
extra in case we needed a sample, Bruce. But it’s all part of our 2000 sample design.
>>And the old questions were all 2000 Census-based.>>Okay.
^E00:32:29 ^B00:32:42
>>Good morning. Hey. So today I’m going to talk about traditional health insurance questionnaire
design in the current population survey. I’m also going to talk about some of the improvements
that we recently made to it and discuss some of the results from this redesign. That would
have helped, right? [ Inaudible Speaker ]
Okay. Some health insurance data collected by multiple surveys. These surveys differ
in the timing of data collection, the reference period, the timeframe of the resulting health
insurance coverage estimates, and uses of the data. This infographic illustrates some
of the differences in measurement and uses. So the CPS provides estimates of the population
result insurance for the entire previous calendar year. This means that if someone had insurance,
even for one day, they would not be included in this uninsured count. The benefit of CPS
is the combination of detailed employment and detailed income. There’s a time series
that stretches back decades at the national level. In September, 2014, we released estimates
of how many people did not have health insurance for all of 2013. Similarly, the American Communities
Survey produces annual estimates of the uninsured. However, they’re based upon the average of
responses collected during the whole year. These respondents providing their health insurance
coverage status as of the time of their interview. The strength of the ACS is it visits large
sample size which enables us to drill down to smaller geographies and provide health
insurance estimates for most communities. Last fall, we published estimates reflecting
2013. The survey of income of program participation collects monthly coverage data for every month
of the prior year. Its strength lies in measuring transitions on and off health insurance as
well as changes in types of coverage. Finally, the National Health Interview Survey collects
current coverage at the time of the interview. The benefit of this survey is that it combines
health insurance measures with detailed health conditions, forming a picture of health well-being.
Last year, we implemented a redesigned set of questions about health insurance coverage
in CPS ASEC. These survey improvements will better measure health insurance coverage for
calendar year 2013 and beyond. And while the 2013 estimates from the CPS are based on a
different set of questions and, thus, not comparable to previous CPS estimates, the
new questionnaire provides a strong baseline to measure changes in health insurance coverage
due to the Affordable Care Act since the questionnaire change was implemented prior to 2014. So a
questionnaire change that occurred last year was hardly the first change to the way we
collect and process health insurance data in the history of the CPS. This infographic
provides a brief history of the many efforts that the Census Bureau has undertaken in order
to improve our data collection methods and produce the most accurate data on health insurance
coverage in the United States. Detailed here are all the question changes, sample changes,
and processing changes that the Census Bureau has implemented since 1987 in an effort to
provide the most up-to-date information for all estimates. All of these changes are noted
in our annual published reports. I know you can’t see this very well, but while changes
to the CPS are not uncommon, they require thorough and extensive research and testing
to ensure that they will perform in a production environment. The Census Bureau has to follow
a strict set of internal and external guidelines and regulations that I’m sure that any changes
that are implemented are implement thoughtfully and carefully in a transparent fashion and
in full public view. This most recent change to the health insurance questionnaire was
also subject to these guidelines and regulations. This infographic details just some of the
research and testing going back to 1998 that went into making this questionnaire change.
In addition to a large number of research projects the Census Bureau has conducted multiple
cognitive tests, a content test in 2010, and following the passage of the Affordable Care
Act another content test in 2013. The Census Bureau has also collaborated with researchers
in other Federal agencies and in academia, and sought feedback from subject matter experts.
It is worth mentioning that over this time period the research community outside of the
Census Bureau has also developed an extensive body of literature that indicated the need
to update the health insurance questionnaire. So, what pointed to the need to reassess how
the CPS collects data on health insurance coverage? This chart illustrates some of the
research. This particular research is a multi-phase research project. It was the Center for Medicare
and Medicaid Services, agencies within the HHS including NCHS, SHADEC, and the Census
Bureau to explain why discrepancies exist between Census Bureau survey estimates of
enrollment in Medicaid and the number of — and release reported in state and national administrative
data. As you can see, the blue bars here are the counts from Medicaid, and the green are
counts of people reporting Medicaid in the CPS. So for 2008 there were 26% fewer people
in the CPS saying that they had Medicaid compared to how many people had benefits in the Medicaid
statistical information system, the administrative data for Medicaid. You can see that over the
years the difference is rather consistent with CPS underreporting the population with
Medicaid health insurance benefits. On a more general level, the research conducted both
within and outside of the Census Bureau pointed to the old CPS health insurance estimates
having some reporting problems. On the left-hand side you can see that the introductory question
asked. “At any time in 2012, was anyone in the household covered by a health insurance
plan provided through their current or former employer or union?” Thus the questions were
retrospective in nature about health insurance during the previous calendar year.
^M00:40:01 This is problematic because, as we know, health
insurance coverage status can change over the course of a year. So answering questions
about this long reference period may lead to response errors. For example, some people
may report their insurance coverage status at the time of the interview rather than their
coverage status during the previous calendar year. Research also pointed to other problems
such as respondents struggling with the meaning of the questions and getting confused among
plan types. A verification question on the left-hand side illustrates some of these issues.
When being read a long list of plan types respondents could pre-report their coverage
at the first plan type that seems somewhat appropriate resulting either in under- or
over-reporting of coverage. In addition, the household level design of health insurance
questions increased the likelihood of some household numbers being forgotten since they
were not mentioned by name. So how are the new questions different? The redesigned health
insurance questionnaire begins the conversation about health insurance by asking respondents
about their current coverage situation, and then uses the answers provided to obtain information
on health insurance coverage during the previous calendar year. In terms of plan types, the
instrument starts with general coverage questions first, and then drills down to specific types
of coverage, their different paths. These paths depend on respondents’ early answers.
This should make it cognitively easier for respondents resulting in more correct answers.
The instrument also changed from a household level design to one that helps us capture
health insurance coverage for all members of the household. They ask, “Who else in the
household had that plan type?” and ask about household members by name, to address gaps
in household coverage. Further, the CPS now includes questions to measure marketplace
participation as well as additional questions on employer-sponsored insurance offers and
take-up. We also revised questions on medical out-of-pocket expenses. It is worth emphasizing
that the redesigned CPS instrument is also very dynamic. That is, each respondents pass
through the instrument depends on answers to prior questions, resulting in lower respondent
burden. So, how did we do? One of the issues with the old CPS estimate was that it was
not in line with other survey estimates of health insurance coverage. As you can see
here, the red line at the top are the CPS estimates which are much higher than the NHIS
last 12-month estimate. In 2013 the CPS estimate for 2013 was 13.4, indicating that the new
questionnaire brought the estimates more in line with estimates from other surveys. As
a matter of fact, you can see here that prior to 2013 the CPS estimates were much closer
to the estimates of current coverage in other surveys than to the estimates of previous
year uninsured rate from NHIS. What else can we tell about how well the new questions worked?
I mentioned before that as part of the new health insurance question sequence we begin
by anchoring respondents with a question asking if they had health insurance coverage at the
time of the interview. This current coverage question provides us with a new measure in
the CPS. We expected this measure of coverage now to differ from the coverage all of last
year measure, though not dramatically, as most people who had coverage last year will
also have it at the time of the interview. We see here that the CPS and the NHIS estimates
of current coverage are very close. This is another indicator that the new question sequence
may yield improved health insurance measures from the CPS. We did not include this new
measure in our regular reports last year. Here you can also see some comparisons of
current coverage estimates from the CPS and the NHIS select demographic groups. These
were produced in collaboration with the National Center for Health Statistics. For the most
part, differences between the two are not statistically significant. At the same time,
it is important to keep in mind that these two surveys differ in the timing of data collection,
context of questions, sampling procedures, and data processing. So some difference in
estimates ought to be expected. Other findings the Census Bureau released last fall included
estimates of the number and percentage of people by health insurance status. Most people,
8- to 8.6%, had health insurance coverage at some point during the calendar year. Overall,
64.2% of the population had private health insurance coverage with more than one-half,
53.9%, covered by employment-based coverage, and 11% covered by direct-purchase insurance.
Government health programs provided coverage for about 1/3 of the population with 34.3%.
Between the two largest government health programs, Medicaid covered more people than
did Medicare — 17.3% compared to 15.6%. As I mentioned before, the 2013 estimates from
CPS are based on a different set of questions and, thus, not comparable to previous CPS
estimates. For that reason, last year we turned to the ACS for analysis of year-to-year changes
in health insurance coverage. Here you can see how the estimates from the two surveys
compare to each other between 2008 and 2012. Due to differences in methodologies, the levels
of the uninsured rate between CPS and ACS differ. However, both surveys produce uninsured
rates that track closely over time. On our website we have a technical paper that formally
examines the differences between the estimates produced by the two surveys. So last year
we turned to the ACS to answer the question of whether the uninsured rate changed and
by how much. Here we see the uninsured rate from the ACS from 2008 to 2013. Of course,
the ACS began collecting health insurance information in 2008. As measured by the ACS,
the uninsured rate from 2008 to 2010 increased from 14.6% to 15.5% in 2010, and then fell
between 2010 and 2013. Between 2012 and 2013 the percentage of people who did not have
health insurance decreased from 14.7 to 14.5%. The ACS, which has a much larger sample than
the CPS, is also a useful source for estimating and identifying changes in the uninsured population
at the state level. During 2013, the state with the lowest percentage of people without
health insurance was Massachusetts at 3.7%, while the highest uninsured rate was for Texas
at 22.1%. You can see the darker colors along the southern states indicating a greater percentage
of people with no health insurance coverage, and the lighter colors concentrated along
the northeastern and north central states. So in summary, today I talked about the facts
that multiple surveys over health insurance coverage measures. These surveys have different
methodologies and different uses. Federal statistical agencies work continuously to
improve data collections, and our understanding of those data. And any changes to a survey,
to any survey, require years of research and testing. CPS ASEC improvements for 2014 result
in a better measure of health insurance coverage for calendar year 2013 and provide a strong
baseline to measure future changes in health insurance coverage due to the Affordable Care
Act. Thank you.>>Thank you, Marina. Are there any questions
for Marina, quickly, before Ed gives his talk? Okay.
>>Thank you. [ Inaudible Speaker ]
^E00:49:42 ^B00:50:00
>>I don’t think, do you [inaudible]. ^E00:50:02
^B00:50:13>>Good morning. I’m going to expand on background
about the current population survey’s annual, social, and economic supplement, and the changes
we made to the survey instrument in 2014, and what effect those changes had on some
of the income estimates. All of the information I’m presenting here today was originally presented
by my co-author, Jessica Semega, at the January Allied Social Science Association Meetings
in Boston. It’s also available on our Website and the link is shown on this page here. These
are the topics I’m going to cover briefly this morning. First, why we redesigned the
questionnaire. Go over the list of the things, of the changes, we made; the impact those
changes had on the income estimates. And I’ll finish with a summary of our findings. While
the ASEC is a very important survey and the source of the official poverty estimates,
it’s long been known to clearly have and clearly documented that the ASEC suffers from misclassification
and underreporting of income. The goal of the redesign was to address these shortcomings
and to develop a redesigned questionnaire to improve the collection of ASEC income data.
What we wanted to address was underreporting, item non-response, and errors resulting from
respondent fatigue, all while taking better advantage of the functionality and flexibility
afforded by a computer-assisted questionnaire. The redesign also offered the Census an opportunity
to address the changing retirement income environment. Retirement pensions and annuities
are historically all underreported in the ASEC. While retirement income is still dominated
by Social Security and traditional pensions, the aggregate holdings in newer types of retirement
vehicles, such as tax-advantaged IRAs and 401K plans, already exceeds those of traditional
pension plans by a substantial margin. Also, one of the largest aggregate shortfalls in
measured ASEC income is asset income. We wanted to redesign ASEC to better capture income
from assets such as the interest and dividends earned in both retirement and non-retirement
asset accounts. Will that be a problem? The Census Bureau began the redesign effort using
a limited telephone interview test in March, 2013, using a retired ASEC sample of approximately
23,000 households. Based on the encouraging results we saw from this test which showed
likely increases in income recipiency and higher amounts, the Census Bureau conducted
a second split panel test. This test was administered to a sub-sample of households in the 2014
ASEC. About 30,000 addresses were randomly assigned to be eligible to receive a set of
redesigned income questions. The remaining sample of 68,000 households were eligible
to receive the traditional questions. The remainder of this presentation shows the results
of this split panel test. I will compare the estimates from the traditional set of questions
and the redesigned questions using the full production sample conducted in 2014. Let me
list the changes in the redesign. I’ll go over each of these in more detail after I
go through this list. We converted to a dual pass approach for identifying income recipients
and income amounts. That is, we asked all recipiency first followed by amounts. We incorporated
three tailored income question skip patterns. We removed the income screener. The traditional
ASEC only asks households with incomes less than $75,000 about means tested transfer programs
such as Food Stamps and TANF. There is evidence that the traditional questionnaire was inappropriately
screening out some households that would be eligible for one or more of these transfer
programs. This could happen if there was a non-family member that may be receiving such
benefits or if the current economic situation of the household was different from the previous
calendar year where they may have also received low income program benefits in the previous
year. We added income range follow-ups for income don’t knows and refusals. We added
questions to differentiate property income from retirement accounts and other interest-earning
assets. We asked about retirement account withdrawals and distributions, and we added
more detailed disability survivor and retirement questions. The following slides give details
about these changes. The traditional questionnaire uses an interleaf approach where income receipt
is immediately followed by questions on amounts. The redesign uses a dual-pass approach. The
first pass identifies all income sources received in the household. The second pass collects
income amounts for each of these sources. Here you can see how this would work using
Social Security as an example. A similar approach of income-reporting is used in the Survey
of Income and Program Participation or what’s commonly known as SIPP. Tailored skip patterns
use known characteristics of the household composition to ask more relevant questions
earlier in the interview to help reduce respondent fatigue and keep the questions relevant. The
three skip patterns are lower income households which prioritizes questions on means tested
programs such as public assistance and Food Stamps. Here the income screener is used only
to identify lower income households and not to skip income questions. The households with
a member aged 62 and older, which prioritizes disability, retirement, and pension questions,
and the default which is used if the household does not qualify as either lower income or
having senior present. It closely reflects the traditional instruments questionnaire
order. All questions, regardless of household composition are asked, just in a different
order. The redesign also uses unfolding range follow-up questions anytime a respondent doesn’t
know or refuses to provide an amount for an income source. The unfolding aspect is for
respondents that initially select the lowest range. A follow-up question is then asked.
The income amounts presented — excuse me. The income amounts presented in the range
questions depend on the source of income. The redesign uses high, middle, and low ranges
based on the type of income. This slide shows the sequence of follow-up middle-range questions
used for Social Security income. The objective of these income range questions are to reduce
amount non-response by allowing respondents to provide a less precise response. Again,
this approach is not new. It’s used in other surveys such as the Consumer Expenditure Survey
and the SIPP. To better capture retirement income, the redesigned ASEC specifically asks
about both traditional pensions, that is, defined benefits, and retirement accounts,
defined contributions such as IRAs, 401, and other accounts designed specifically for retirement
savings, and also about annuities. The traditional ASEC asked one broad question that combined
pensions, retirement, and annuity income. If the respondent has a retirement account,
the redesigned ASEC instrument asks the respondent to identify the specific types of accounts
and whether there were withdrawals or distributions from those retirement accounts. Questions
on withdrawals and distributions from retirement accounts are new in the redesign. There are
no directly comparable questions in the traditional ASEC. It only asks, “How much did you receive
in pension or retirement income?” The new questions use an account-type fill that can
use the exact wording given by the respondent for easier identification. For respondents
over 70 years old, the redesigned question text changes slightly to reflect the fact
that they may be required to take a distribution. All withdrawals are followed up with questions
asking whether any of the money was rolled over or reinvested into another account which
would then not be counted as income for that year. To better capture asset income, interest
and dividend income received on retirement accounts is asked separately from non-retirement
accounts in the redesign. The traditional questions make no distinction between investment
income received in a retirement account or investment income received outside of retirement
accounts. Also new in the redesign for people who refuse or don’t know asset income amounts
are questions that ask the total value of the account at the end of the year that allow
the Census Bureau to estimate what the income might have been from those assets. ^M01:00:00
Interest-earning checking accounts, savings accounts, money market funds, CDs, savings
bonds, and shares of stock in corporations or mutual funds are all non-retirement accounts.
The redesign asks a series of questions specifically about each of these possible sources of income.
Respondents with shares of stocks in corporations or mutual funds are asked follow-up questions
on dividends received along with questions on the receipt of capital gains. No questions
on capital gains are in the traditional ASEC. Capital gains are not included in our current
definition of income, nor income from the redesign.
^E01:00:42 ^B01:00:51
This table shows median household income and the percentage difference between the redesign
and the traditional ASEC estimates by selected household characteristics. All results shown
are statistically significant at the 90% confidence level. Other characteristics we typically
examine from year to year showed that none had statistically significant lower median
incomes as a result of using the redesigned income questions. This graph shows the percentage
change in recipiency by sources of income for people age 15 and older based on the redesign,
which is in purple, and the traditional, which is in green. The stated goal of the redesign
was to increase income source reporting using the dual-pass approach of identifying all
sources of income before asking amounts and using the tailored skip pattern to ask more
pertinent questions earlier in the interview based on household composition. With the exception
of Worker’s Compensation, all income recipiency was higher or not statistically different
in the redesign. The increased recipiency of these types of income sources resulted
in some lower mean amounts by source but increased overall aggregate income. As you recall, another
goal of the redesign was to improve reporting of retirement income. Here we see more retirement
income from both pensions and, though relatively small compared to some of the other income
sources, retirement accounts. Targeting retirement account income with expanded questions resulted
in a 419-1/2% increase in people that received income from IRA, Keough, or 401-type plans.
Here we see the changes in aggregate income by source. Aggregate income was higher in
the redesign for all the same sources of income that had higher recipiency except dividends.
Total aggregate income was up 4.2% in the redesign. One of the focuses of the redesign
was to improve the reporting of means tested income by removing the income screener, tailoring
the income questions for low-income households, and using the dual path approach. Public assistance
aggregate income was up 36.7% higher in the redesign compared to the traditional. Aggregate
income and aggregate interest income was up over 111% though aggregate dividend income
was 20% lower. This is probably a result of the redesign questionnaires better classifying
what constitutes dividend income, separate from interest or capital gains. Collectively,
aggregate interest and dividend income was up 52.8% in the redesign. Let me also add
some perspective here with regard to these income sources. While every income source
is important, we need to keep in mind that most income comes from earnings. That’s wages
and salary and self-employment. Earned income accounts for about 75% of all income collected
in the CPS ASEC. There was no change in the earnings questions in the redesign and, not
surprising, there was no statistically significant difference between the traditional and redesigned
recipiency or aggregate earned income. This table shows the shares of aggregate income
by percentile. Even with the overall increase in income recipiency, an aggregate income
including means-tested income sources, we found that the share of income going to the
first and second quintile were lower in the redesign. An examination of the lowest quintile
of households showed that aggregate income of some income sources were, in fact, lower.
Trudy Renwick looks into this in more detail in her presentation on the effect the redesign
had on poverty estimates. So, to summarize, the redesigned ASEC showed increases in household
medians, income recipiency, and aggregates. The redesign questionnaire also seemed to
improve the reporting for targeted income sources such as public assistance, retirement,
and asset income. These improvements did not seem to be equally distributed, however, with
the low end of the income distribution showing less aggregate income for some income sources.
Based on these results, the Census Bureau went forward with the redesigned ASEC as the
full production instrument this year. Okay, thank you. And I guess if we’re going to take
questions?>>Yeah. Are there any questions?
>>Anybody, please.>>At Connie and then Jim. Or Jim and then
Connie.>>I had a quick question. How were wave kind
of second round supplement recipient interviewees treated, right, in the split panel? So, if
this is my second March interview, as it were, right, were they notified that, you know,
this is a different structure, or were they actually just given the original questions?
>>No, nobody knew what they were getting until they pulled up to — the FRIs had no
idea. It was just a random number that was generated and, depending on whatever random
number was generated, that would depend on what questionnaire came up for the FR in the
field. So the FR out in the field had no idea as well until they opened the case.
>>Do we know anything about differences in response rates between the — from the first
wave to the — or the second interview for these individuals?
[ Inaudible Speaker ]>>No. I mean, we probably could look at that
— yeah, I think we have the information available, but we haven’t looked at that.
>>Connie?>>Will you be planning comparisons with benchmarks
for various income sources for the new and the old questions?
>>Yes. John’s going to give the presentation earlier. He’s done updates of the benchmarks
that were done earlier before the implementation of the redesign. So, yes, we would certainly
like to do some comparison to benchmarks in the future. I think one thing that we would
really like to test before moving too far forward is we have the ability to look at
how these unfolding income ranges work, or how good they are, by comparing the earnings
data that we collect that also use these unfolding income ranges, and compare them to the DER
records we have access to as well. I think the most recent data we may have access to
would be for the 2013 test, the smaller field test that we conducted two years ago, to take
a look to see how well people are able to report these ranges compared to administrative
data. But, yes, we hope to do that in the near future.
>>Okay.>>Mike and then —
>>You need to turn [inaudible].>>With the dual pass approach it strikes
me that ^M01:08:04
[ Inaudible Speaker ] ^M01:08:11
>>Yes, we did. As a matter of fact, that was in one of the slides here. So you can
take a look at the income recipiency by source. So the green shows the traditional; the blue,
the redesign. So these are all up and the recipiency by all of our income sources, none
of them showed a decline or statistically significant difference. And these were all
showing increases except for Worker’s Compensation. We have no idea what went on there.
>>Thanks, I’ll just read the slide.>>Right. Howard?
>>I guess I don’t need to [inaudible].>>On the Social Security — oh. On the Social
Security question, the Social Security received, how do you establish the amount that Social
Security is sending to pay for Medicare premiums which now amount to about $1400 a year per
person?>>Okay. So the questions that we ask in the
CPS ASEC are — or I’ll just say twofold. We ask, you know, “Did you receive Social
Security income.” “Yes, we’ve received it.” “Who received it?” And then we ask, “What’s
the easiest way for you to report that income to us? Annually, monthly, or whatever?” We
allow the respondent to choose the easiest way. After they give us an amount, we ask
them whether or not that was [throat clearing], excuse me, before or after any deductions,
and then we’re able to gauge from, you know, how much. If they say, “After those deductions,”
then we ask how much those deductions were. And so we want to collect gross Social Security
before any deductions. So we’re collecting those two components.
[ Inaudible Speaker ]>>Yeah. Did you make these comparisons with
and without imputed data? ^M01:10:00
>>I’m trying to think if we did or not. No. I’m going to say no. I have my co-author shaking
her head and I’m going to go by her call on that. So, no, we haven’t.
>>Well, I think we did find that the rates were — that the imputation rates were pretty
similar between the –>>Oh, yeah.
>>– between the two. I mean the rates of imputation were pretty similar between the
two treatments, so –>>And there was very little difference in
the aggregate income imputed as well.>>Right.
>>So were there big differences in the imputation routines with the new data added?
>>We ran them through the same — we ran them through the same imputation routine.
So we had to do some collapsing of more detailed asset types and pension types to run them
through the current processing system, yeah.>>So we’re behind on two things. One is,
well, one more thing, and that is implementing a new processing system to take advantage
of all of the new detail that we are collecting. So, again, we’d like to do an evaluation of
some of the information by comparing to administrative sources and then try to update our processing
system to take advantage of all these extra income.
>>Yeah. But I have a couple of questions. One is, how does this compare with the questions
that are asked in the Consumer Expenditure Survey? So do you make any comparisons how
it — your [inaudible] patterns are closer to the Consumer Expenditure Survey or [inaudible]?
>>I’m going to have to say right now I don’t know. We haven’t made any comparisons.
>>The second is, have you made any comparisons with any administrative data sources? Which
one is giving you a better — closer answer?>>From the redesign questions, we have not
had an opportunity to do that yet. We just recently within the last year finished up
an update to comparisons to administrative sources just before implementing the redesign,
so after the redesign we’ll have the data from 2015 I believe, or 2014. So start taking
a look at how it compares to administrative sources as well. Typically our first pass
would look at the comparisons to our national income and product accounts. John did some
work in that area as well, so I’ll probably be calling on him to update that using the
new questions and the new questionnaire.>>But the fact that we’re under on, you know,
surveys always get [inaudible] where the admin records data show by type. In fact, we’re
getting more higher [inaudible] for virtually every kind of income shows that we’re closer,
closer to benchmarks. Just a question of how much closer we are.
>>Laura.>>So you’re not going to say maybe use of
the unfolding ranges [inaudible].>>We have not made any use of the unfolding
ranges yet. Our first test again would hopefully compare to the DER just to see how well people
are reporting compared to administrative data. And then very likely what we would do then,
second, would be to make some changes to our existing hot deck system that we’ll use that
as one of the match variables for each of the people to get them a little closer within
the range that they think that they belong.>>I had a second question, too, which was
if somebody is elderly and they’re of low income, which [inaudible].
>>Oh, good question.>>Low income. The difference is so close.
Lisa.>>Okay. I have a comment and then a question.
Maybe it’s a clarification issue. Regarding the Consumer Expenditure Survey, we at BLS
have actually done a comparison using the older question data compared to the Consumer
Expenditure Survey, and they’re very similar. The Consumer Expenditure Survey income section
was a major redesign in 2013. I’ve calculated poverty rates using the official poverty measure
and for text money income and not including Food Stamps, which is the definition that
the Census Bureau uses, and what happened for us is that our poverty rate went up, and
the Census Bureau’s goes down for 2013. So we need to look into a little bit more what’s
going on with our changing and our questionnaire in terms of income. Second, this might — I’m
a little confused about the withdrawals from pensions and what you’re getting out of that,
and you’re counting it as income. I recall that —
>>Double counting.>>Yeah. I’m a little concerned about that
because if you’re not — because the question originally the way it was designed, you know.
1980s. At that time most retirement plans were defined benefit plans and people didn’t
contribute to them. Okay. But now, so most people are contributing, are you taking as
a subtraction what people contributing to come up with your net income? Because otherwise
you’re counting what people are contributing in their income and you’re counting what they’re
receiving, because in some cases people are going to be receiving and they’re still going
to be contributing, especially if you’re in the 70 range where you’re getting money out
from your, you know, 401K but you still may be contributing to an annuity or something
like that. And if — I know these are person, but if you’re talking about household it’s
not unusual that you’re going to have some people still contributing while someone else
is receiving. And so you’re counting it as income but you’re not subtracting when there’s
a contribution in terms of what I heard. Maybe I’m wrong.
>>We shouldn’t do that, right?>>Right.
>>Okay [laughter]. And that’s good.>>Maybe subtraction if you’re going to make
an addition.>>If we meet again next year I will be able
to answer that question by saying we also are asking questions on how much you’re contributing
to the plan. So, this year –>>So you should include it in your definition.
Maybe use an alternative definition.>>And, yes, I think we’ll need to explore
that. But right now we don’t collect enough information to allow us to do that, mainly
how much are you contributing to these plans? That is going to be a change for 2016, so
we will have all the components we can use to maybe look at another alternative income
definition, and I think that would play very nicely to the supplemental poverty measure
in being able to tease that out.>>Right. No, it would be very useful as we
look at assets and how it –>>Apparently we don’t.
>>– plays in the [inaudible] amount of poverty measure. Well, I encourage you at that point,
if you don’t do it before, to look at the Consumer Expenditure Survey because we do
look at allocations to retirement. Okay, so.>>Thank you. Howard.
>>I think it’s very important to count the distributions coming out of retirement accounts
and have maintained that for at least a decade. I believe, if you’re worried about contributions,
you can use the DER and you can establish whether, when people are reporting their earnings,
they’re reporting earnings plus the Box 12 withheld. My expectation is they’re giving
you the close salary without counting the money that is being contributed to their 401K
in Box 12. But you have the data in the DER. It’s a simple thing to compare Box 1 plus
Box 12 to Box 1 and which one they are reporting.>>That’s a very good point. Thank you, Howard.
>>Right.>>Okay. I think we need to move on to Trudy’s
presentation. Thank you, Ed.>>Thank you.
^E01:17:29 ^B01:17:46
>>Because I need to move it to [inaudible]. Do-do-do. Ah. Good morning. My name’s Trudi
Renwick. I am Chief of the Poverty Statistics Branch and I’m going to talk a little bit
about the comparison of poverty rates between the addresses that got the traditional questions
and the poverty rate from the addresses that got the redesigned questions. This is a summary
of work that I’ve done with my co-author Josh Mitchell that was presented at the Allied
Social Sciences meetings in Boston in January, and the entire paper’s on our Website with
a lot more tables and a lot more estimates if you want to dig into this a little bit
further. So what I’m going to do today is do this comparison of poverty estimates from
the two samples. I’m going to look at the official poverty estimates. I’m going to just
spend a moment summarizing some work by my colleague, Kathleen Short, that looked at
the comparison of the supplemental poverty rates between the two samples, and look in
more detail at some of the more detailed demographic groups. Finally, you’ll see that we found
differences in poverty rates for children and the older population, and I’m going to
spend some time digging into that a little bit further. Summary of my findings is basically
we didn’t find much, okay? That most of the differences in poverty rates across the two
samples were not statistically significant at the 90% confidence interval. There were
a few exceptions. There were lower poverty rates for people with a disability and that
was consistent with what Ed just reported with increases in disability income from the
new questions. The poverty rate for people who worked at least one week during the past
year was also lower in the sample with the redesigned questions. But given the increases
in aggregate income that Ed reported, increases in median income for all households, increases
in income for family households, increases in income for households with householder
age 65 years or older, were somewhat surprising is that poverty rates went up for children
and for people age 65 and older. And so we’ll spend a little bit more time looking at that.
Just basic review. I’m looking around the room, I don’t think many people need this
review of how we measure poverty at the Census Bureau.
^M01:20:00 But we take thresholds that — the official
poverty measure uses a set of thresholds that vary by the size of the family, the number
of children, and the age of the householder. We then compare this threshold to pre-tax
total money income for each family, and if total money income is less than the thresholds,
all members of the family are categorized as in poverty. Last September we released
our 2013 Official Poverty Estimates — 14.5% of the population were in poverty, plus or
minus 0.3 percentage points. And that Official Poverty Estimate was based on the sample that
received the traditional income questions. The overall poverty rate for the sample with
the redesigned income questions was 14.7% plus or minus 0.5 percentage points. So, the
difference was not statistically significant. Note that the confidence interval for the
estimates from the smaller sample was larger than the confidence interval from the 60,000
addresses sent and my colleague, John, will talk about that a little bit in his presentation.
Let me just spend a minute talking about the supplemental poverty measure. Since October,
2011, the Census Bureau has been releasing an alternative National poverty Estimate,
the Supplemental Poverty Measure. Based on a series of suggestions from an inter-agency
technical working group, the new measure creates a more complex statistical picture incorporating
additional items such as tax payments, work expenses, medical out-of-pocket expenses,
and the value of non-cash governmental assistance such as SNAP benefits and housing assistance
in the resource measure. The thresholds are estimated by the Bureau of Labor Statistics
and are derived from consumer expenditure survey data and adjusted for geographic differences
in the cost of housing. Last October, we released SPM estimates using the data from the sample
drawn from the 68,000 addresses who were eligible to receive the traditional income questions.
The overall SPM poverty rate was 15.5%. Using the sample with the redesigned income questions,
the overall SPM rate was 15.2%. Not statistically different from the rate from the traditional
sample. There were only a few groups with statistically significant differences in SPM
rates across the two samples, and no major group in the redesigned sample had a higher
poverty rate than from the traditional sample. For more detail on this, my colleague, Kathleen
Short, presented a paper at the Allied Social Science Association meetings in Boston this
year, and her paper is on our Website. In addition, as was mentioned earlier, in addition
to the changes in the health insurance questions, there were some small changes in the questions
about medical out-of-pocket expenses. And those are also reflected in the supplemental
poverty rate estimates from the redesigned questions. Okay. Let’s go back to the estimates
using the official poverty measure. In the next sequence of slides I will show you the
differences in poverty rates for specific demographic subgroups. We will focus on differences
in poverty rates across the two samples, emphasizing the few that had statistically significant
differences. Let’s start with the poverty rates for major age groups. The green bars
show the poverty rates from the sample that received the traditional income questions
and are the same rates as the rates published last September in our official income and
poverty report. The purple bars show the rates for the sample that received the redesigned
income questions. The poverty rate for children under the age of 18 was 21.3% in the sample
with the redesigned questions, 1.4 percentage points higher than the rate for children in
the traditional sample. The difference for those age 18/64 was not statistically significant.
The poverty rate for the population age 65 and older was 0.8 percentage points higher
in the redesigned sample than the traditional sample. Now let’s look at poverty rates by
family type. For all people in family, the differences were not statistically significant.
For families with children the poverty rate for the sample with the redesigned questions
was 20.7%, 1.2 percentage points higher than the rate from the traditional sample, very
much consistent with the increased poverty rate for children. The other groups did not
have significant differences. Looking at race and ethnicity, the only group with a statistically
significant change in the poverty rate was the Asian alone group whose poverty rate was
2.5 percentage points higher in the redesigned sample. As those who use the CPS ASEC know
that our sample for Asians in a single year of the CPS ASEC is quite small and does tend
to have a large variance. But this was statistically significant. Looking at poverty rates by sex,
while women had a higher poverty rate than men in both samples, the differences across
the two samples were not statistically significant. Here we see poverty rates by nativity and
citizenship. Across the samples those differences in poverty rates for these groups were not
statistically significant. Looking at poverty rates by region. Again, while the South has
a higher poverty rate than the other regions, for estimates from both samples, none of the
differences across the two samples were statistically significant. Examining poverty rates by place
of residence, the story’s the same. While there are differences by place of residence,
for example, higher poverty rates inside principal cities than outside principal cities, the
differences across the samples were not statistically significant. This slide looks at poverty rates
for the working-age population, those aged 18 to 64, by work experience in the previous
year. The poverty rate for anyone who worked at least one week during 2013 was 0.4 percentage
points lower in the sample with the redesigned questions than the sample with the traditional
questions. Looking at the subgroups of workers we can see that this difference in the poverty
rate was driven by the 1.8 percentage point lower poverty rate for those who worked less
than fulltime year-round in the redesigned sample. The poverty rates for those who worked
fulltime year-round were not — the differences were not statistically significant, nor for
those who did not work. And, finally, we’ll look at poverty by disability status. The
poverty rate for people with a disability was 28.8% in the traditional sample and 27.6%
in the sample with the redesigned questions, a 1.2 percentage point difference consistent
with what Ed found with an increase in disability income in the redesigned questions. So, let’s
go back to this question, why did poverty rates go up for children and for the older
population when we switched from the traditional to the redesigned sample? For children, we
started by looking at child poverty rates broken down by many of the same characteristics
as we explored for the overall poverty rate. And all of that detail is in our paper on
the Website. We found that differences in child poverty rates across family types were
not statistically significant across the two samples. But what we did find was differences
in the demographic composition of the two samples that can help explain the differences
in overall child poverty rates. Both samples were weighted to be nationally representative.
But our weights control for race, age, ethnicity, and state of residence. The weighting will
not guarantee that other characteristics will be the same across the two samples. This chart
shows the distribution of children across family types. In the sample with the traditional
questions, 67.2% of children lived in married-couple families. For the sample with the redesigned
questions, 65.7% of children lived in these families. In numerical terms, there were approximately
1 million fewer children in married-couple families in the redesigned sample. These differences
in sample composition explain about half the differences in child poverty rates across
the two samples. We did an exercise in our paper. You can see the details. But, if the
distribution of the children across family types had been the same in the two samples,
the differences in the child poverty rates would not have been statistically significant.
So, next we looked at the same set of characteristics for the older population and we found that
the higher poverty rates for the redesigned sample were driven by higher poverty rates
for those who lived outside of families. Those are either living alone or living with non-relatives.
But unlike children, we did not find a difference in the sample composition for the older population.
So we kind of [inaudible] with this finding that these were the people who driving the
higher poverty rate from the redesigned questions, but we didn’t — it wasn’t sample composition
that was driving it. So we then looked at looking at the income distribution and looking
at how uniform were these increases in income across the income distribution. We knew from
Ed’s work that aggregate income for all households was up by 4.2%, and that median household
income, which would measure the middle of the income distribution, was up 3%.
^M01:30:00 But we found when we looked at that bottom
quintile, so the lowest fifth of households in the income distribution, the cut-off was
about $20,900. Actually from both samples the cutoff was fairly similar. And looking
at those we found that overall their income did not change. The differences were not statistically
significant and that there were only really two sources of income that had increases.
Those were public assistance and disability benefits. And thanks to some work from Laura
we know that the increase in public assistance was not TANF. It was people reporting other
public assistance. But probably most importantly for looking at the population 65 years and
older is that the changes in interest, dividends, and retirement income, what we might have
expected to drive down the poverty rate for people age 65 or older did not increase for
the people in the bottom quintile. So while those were up in the aggregate for the whole
distribution, they were not up for households in that bottom quintile. And in our paper
we go into much more detail looking at the bottom quintile of the elderly population
— the older population. So, in conclusion, I mean the story is for the most part there
were really very few statistically significant differences in poverty between the two samples.
We did find higher poverty rate for children in those age 65 or older. The child poverty
rate we feel can be pretty much explained by the difference in the composition of the
two samples. The increase in poverty for the older population is being driven by those
unrelated individuals and when we look at the — we break down aggregate income increases
by quintile, we see that that bottom quintile did not have an increase in the income sources
that were increased in the overall population. So, if you have any more questions, here’s
my contact information. Be happy to answer questions.
>>Trudi, I wonder if you would split the sample a little bit further. Of course it
always gets a little bit more noisy when we do that, but with families with children under
age 18, if you would also split it by fulltime work status, right? So, qualitatively there
was in increase in poverty for that group, and so —
>>We spent a lot of time looking at families with children, and we did not find — we looked
at earnings, we looked at hours worked, and we looked at a lot of labor force variables,
and we did not find that — you can see that in our paper. We did not find that those helped
us to explain it.>>Interesting. And then, with respect to
seniors, did you focus at all on seniors raising children, that sub-sample to see —
>>No.>>– whether or not that’s helped in accounting
for the rise in elderly poverty at all?>>I would say since they were unrelated individuals
— so, that the group for which the poverty rate went up were unrelated individuals. So
they, by definition, are not — they’re not raising a related child. They may be raising
–>>But the families could just be husband/wife,
>>And not young children. So I guess it’s kind of taking that families group and focusing
in on whether or not there’s a child present within senior families.
>>That’s right.>>But we did not — but for senior families
we did not find an increase in poverty. All the increase in poverty for people 65 and
older were unrelated individuals who were living outside of families. So they were living
alone.>>I guess I’ll —
>>Because I’m saying is that those families are going to be an aggregation of those with
and without children.>>Uh-huh.
>>Right? And so if you just zero in on those senior families with the child present. Okay.
>>Okay. We could. We didn’t, but we [inaudible].>>In the — and I noticed there that you
had tailored skip paterns –>>Mm-hmm.
>>– whether or not those order affect might also have an impact on this by — did you
check whether or not there is a order effect in any other studies?
>>Well, what I did look at was the elimination of that screener, okay? So I looked at the
people whose — this is for the insiders — whose control card income was less than or greater
than $75,000. So, people that wouldn’t have gotten those questions before. And I found
that that did not make any difference. There were I think 13 cases in the redesigned questions
of people whose control card income was above 75,000 who reported receipt of public assistance.
And there are people in the traditional questions that also had control card income above 75,000.
So that was not — I ran it that way and didn’t matter.
>>The skip patterns between the households with seniors and the low income households,
it doesn’t vary all that much. For the seniors we’re asking about the retirement and disability
questions first. Then we ask about the low income questions. So the real switching that
occurred between low income and seniors was the seniors get asked the retirement questions
first and then the low-income questions, and then for the low income they get asked the
low-income questions and then lastly the asset questions.
[ Inaudible Speaker ]>>Was there any experiment done in terms
of looking whether or not there will be an order effect? Because if you are asking the
first questions, retirement questions first, then that — as opposed to asking it in a
regular fashion, there could be a difference in terms of recall and also the respondent
fatigue.>>I don’t recall. We’ll have to look at some
of the experiments that were done back prior to 2013 when we first went out with our first
field test and tried to order the questions. But I’m not sure if the Mathematica or Urban
had checked any of that in their early work. So, I don’t know.
>>Yeah. So, other — Trudi, I’ll make a statement and just throw it out there, and you can react
and tell me if I’m right or wrong. It seems to me that the way you went about it might
be a recipe for finding nothing because if you look at the difference between the two
sample parts and look for things that are significant, and then look for compositional
differences in the two samples to try to explain away the results that were significant, you’ll
find that things that had asterisks, but before you adjusted for the compositional differences,
those asterisks, a lot of them would go away.>>Mm-hmm.
>>But what if you were to equalize the characteristics of the two sample parts and then wouldn’t
there be some things that were not significant before that suddenly became significant?
>>Mm-hmm. Good point, good point. ^E01:37:22
^B01:37:30>>Yes. I’m wondering generally about the
extent to which there may just be some differences between that 3/8 and that 5/8 sample. And
I did a little work looking at the child poverty rates and so I thought, well, earnings is,
you know, what’s most important for bulk of income and also the earnings questions weren’t
changed. So I looked at what’s child poverty just looking at earnings. I was looking at
the 5/8 sample and comparing to the full sample because I wasn’t doing 5/8 versus 3/8 comparison.
But I found when it went up from the 5/8 sample, just looking at poverty rate based on earnings
was 24.9. But then going up to the full sample it went up to 25.8. And it made me think,
okay, it looks like this 3/8 portion, this 3/8 sample, is just a lower earnings group.
And with earnings being such a important part of child poverty, they’re starting out poorer.
So even with improvements in income and some areas of public assistance getting picked
up, that’s not enough to show the kind of result we’d be expecting to see. And it is
because of this just starting with a poorer sample. So I was wondering if you’d been able
to look at that any.>>I mean, we looked at that a little bit
in our paper and we showed that there were increases in means tested benefits. Really,
higher incidents in the 3/8 sample than the 5/8 sample. But whether that’s because of
changing the skip pattern or — it’s hard to dissect what’s driving that. But there’s
certainly lots of evidence like that.>>It’s not how we selected the sample, though.
Looking for the poor households and put them in.
>>I know. It wasn’t we intend [inaudible].>>Right. But we had them looking at the samples
and we’re seeing some differences along that in a couple of other ways, so we’re wondering
the same thing.>>And wasn’t there some issue that some groups
of — one of the — among the sample groups or something, was automatically put in one
sample by accident, so — I don’t know if you found that made a difference.
>>Yeah. There was a Hispanic over-sample group that got put into I forget which one.
>>The 5/8.>>Five eights. But the weighting should control
for that in theory. So, but, those kinds of differences are what we’re — in the back
of our mind thinking about. ^M01:40:01
>>I had just one more question or comment really on the household screen, the family
income screen. I think, Trudi, you’re comment was looking at — that didn’t have any effect
on public assistance income, but I would expect there would be bigger effect on SNAP and WIC
and school lunches. And I was curious if you had looked at that. That’s where I’d really
expect to see more of an effect from that change.
>>I think the only thing we looked at in our paper was SNAP, and there was a higher
incidence of receipt of SNAP from the 3/8 sample. Yeah. But I can’t really — I don’t
know if it’s because of the screener or the order.
>>Thanks.>>All of this shows why, when we do randomized
studies, we don’t just stop with the randomization. I mean, those, you know, studies that are
looking for effects include regression adjustments because you can’t end up with identical samples.
The work that we did as part of the contract for ASPI and for you was looking at the matched
samples, because we have people who were asked the same questions the prior year as this
year, and people who were asked the old questions the prior year, and the new questions this
year. And it basically shows the same results. That, you know, for the pension interest and
so on items, there are big changes and really not much elsewhere. One of the other things
that came out of this is that when you look at people who were asked the same questions
in two years, there are extraordinary differences between the two years. It just shows how really
noisy income data are. I mean, we see differences, you know, 30% in recipiency. The same respondent
in the two years. And you got to factor all of that into the kinds of, you know, potentially
small changes that we’re looking for here.>>So, John, you did not just look at the
addresses. You looked at the people themselves?>>Yeah, we went at this a few ways. We wanted
the largest group, but we limited it to ones that had the same householder in the two,
and then beyond that we looked at a group that had minimal changes in household composition.
And then, lastly, we looked at the groups that had the same respondent —
>>Yeah.>>– between the two years to try to, you
know, address exactly that.>>Yeah, so, we don’t follow movers CPS —
>>Right, exactly.>>– so that’s a good thing. And the other
thing is, could there be some proxy versus primary respondent differences you might want
to think about, but –>>Yeah. When you have the same respondent,
it’s –>>Yeah.
>>Let’s take a 15 minute break and come back at five after 11:00. Thank you.

Tags:, ,

Add a Comment

Your email address will not be published. Required fields are marked *