GEOGRAPHY and WEIGHTS IN THE NLS

1
GEOGRAPHY and WEIGHTS IN THE NLS

• By
• Randall Olsen

2
The Plan for this Module

• Basics of geographic data
• Geography and sampling
• Level of detail available in the NLS and GIS data
• How you get geographic data
• Weighting
• Correct standard errors and geo-variables
• Using GIS and family-level data to enhance your
  analysis plan

3
Basics of Geographic Data

• Four Census Regions this is the finest level of
  geography available for original cohorts without
  going to a Census Research Data Center.
• State and Counties
• Census Tracts size varies, thousands of persons
• Block Groups neighborhoods (almost)

4
Counties

• Finest level routinely available in NLSY79 and
  NLSY97
• 3100 in U.S. Texas (254), Delaware (3), Georgia
  (159) FIPS codes to designate
• Extensive socio-economic demographic data
  available at county level (nee City-County Data
  Book) you merge them in using FIPS codes

5
Census Tracts

• In 2000 all U.S. partitioned into tracts
• Size and population varies, but can contain
  several thousand people in urban areas or a few
  hundred in rural areas
• Using these data requires clearance from BLS

6
Block Group

• In Urban areas, consist of groups of blocks (did
  you guess?)
• This is the finest level of aggregation for
  Census geography (building blocks for
  reapportionment)
• What non-rural folk think of as a neighborhood,
  except for those near boundary of the block group

7
Census Regions
8
Census Tracts NYC Metro Area
9
Census Tracts in Wyoming
10
Sampling - Original Cohorts

• Original Cohorts drawn from experimental CPS
  sample frame in 1960s Title 13 confidentiality
  restrictions prohibit release of geographic data
  below Census Region
• We recently geocoded these data (latitude and
  longitude) and one may use them at a Census
  Research Data Center
• The exact sampling structure was kept secret a la
  Raiders of the Lost Ark details may exist in a
  musty file in Suitland, MD although the latitude
  and longitude data allows one to reverse engineer
  the sampling

11
Sampling for the NLSYsMultiple stages

• U.S. divided into Primary Sampling Units (PSUs)
  Major Met areas, counties or groups of counties
  (rural areas)
• Selection probability proportional to population
  of interest large cities always chosen
• Dividing PSUs into groups can insure correct
  fraction of rural suburban areas chosen this
  can reduce the sampling variance relative to a
  simple random sample (SRS)

12
Next Stages

• Select Tracts or Block groups list in order by
  income or ethnic composition to pick every nth
  one insures even distribution over the ordered
  characteristics. Again, this can reduce sampling
  variance relative to a SRS. Segments of streets
  selected within block groups.
• List all addresses in selected segments randomly
  select units to do a screening interview to
  identify eligible persons
• This process generates area clusters of nearly
  contiguous respondents

13
Examples of PSU Clustering

• NLSY97 100 PSUs in cross sectional sample. 100
  PSUs in minority oversample.
• NLSY79 102 PSUs in cross sectional sample. 100
  PSUs in oversample. 38 PSUs in Military
  oversample.
• We average about 50 respondents per PSU the
  effect of clustering on statistical properties
  increases with the size of the cluster and degree
  to which variables are correlated within cluster.
• Correlations within clusters increase the
  sampling variance and usually overcome the
  advantages of stratification.

14

• PSUs in NLSY79 initial screening (done in 1978)
  over yield

• PSUs in NLSY97 initial screening (done in 1997
  screen and go) under yield

15
Geographic Detail in NLS

• States and counties available in Geocode release
• Zipcode data kept at BLS and CHRR
• Census tracts and block group identifiers at CHRR
  and BLS
• Latitude and longitude (accurate to about 50
  feet) at CHRR

16
Geocode How you get it

• You need to apply to BLS (see Web site)
• Describe how you plan to use the data
• If BLS approves you, CHRR sends you a CD
• You need to return the CD when finished and you
  are subject to audit and legal liabilities if you
  violate terms of agreement with BLS. BLS
  performs many audits keep yourself in
  compliance.

17
Geocode How you use it

• You use the state and county codes to merge in
  the data you need
• Use standard FIPS codes
• There is a variable indicating when R is in a
  central city (this was done using zipcodes -
  before 1998 missing values show zips that are not
  unambiguously central/non-central)
• Data merge is a do-it-yourself project

18
Zipcode Data

• CD is at BLS or CHRR and is not released
• The CD has Zipcodes, but matching and merging in
  the data you need is a do-it-yourself project
• You can have CHRR create a variable you need,
  with BLS approval
• Zipcode centroid can be used as rough location of
  respondent for simple distance calculations

19
Fine Level Location

• Modern Geographic Information Systems data use
  latitude and longitude as the basis for linking
  data
• We geocode respondent addresses with latitude and
  longitude sometimes with GPS units (all years
  except 1980)
• We place R within about 50 feet
• Opportunities to extend analysis abound

20
Distance from R to

• Fast food restaurants
• Employers
• Doctors offices
• Hospitals
• Freeways
• Schools, public private
• Post offices

• Banks
• Bus stops
• Train stations
• State licensed day care centers
• Drug seizures prices
• Air quality measures
• Toxic waste sites

21
Data at Tract and Block Group Level

• Based on Decennial Census Long Form or American
  Community Survey (recent years)
• Ethnicity and Color of people in area
• Average income, poverty rate, dispersion in
  income, housing attributes
• Population density, education, employment rates

22
Other Sensitive Data for Analysis

• CHRR maintains the names of employers for each
  respondent in each round
• With BLS approval we can identify persons working
  for a particular sort of employer or match in
  employer characteristics
• The guiding principal is that these specialized
  extracts must not give you the ability to
  re-identify the respondent

23
Ideas using detailed geography

• Does proximity to fast-food restaurants now and
  in the past correlate with BMI?
• Does current and past air quality have a
  relationship to the incidence of asthma?
• Does proximity to health care correlate with
  health outcomes?
• Is local income inequality related to health?

24

• Respondent location is generally chosen by
  respondent this problem of endogenous location
  may be attenuated or solved using locational
  attributes at either screening or age 15
  locations reflecting primarily parental choice,
  not respondent choice.
• These past locational attributes can be used
  as either regressors or instrumental variables
  (IV). IV creates a variable that stands in for
  a regressor that is correlated with the error
  term.

25

• Some respondent choices may be endogenous to an
  outcome, such as smoking and birth weight of
  ones infant. One could use the incidence of
  smoking by ones peers in the original PSU (or by
  ones siblings) as an instrumental variable.
• Peer smoking reflects shared socio-economic
  forces, but weight of Rs baby unlikely to have
  an effect on smoking behavior of Rs peers.
• Need to avoid weak instruments, that is
  instruments that do not explain much of the
  variation in the variable they stand in for.

26
Using Fine-level Geography

• Make application to BLS
• CHRR can often create the variable for you if it
  does not threaten re-identification
• Rounding data reduces precision and reduces
  threat of re-identification of tract, block group
  or zipcode
• Do the analysis at BLS or CHRR

27
DIFFUSION OF THE SAMPLES PSU Clusters in
original NLSY97 Sample
But this clustering has broken down over time.
Here is where people live as of Round 6 in NLSY97
28
PSU Clusters in original NLSY79 Sample 12,000
By Round 20 in NLSY79 Sample there is even more
geographic dispersion. 9,000
29
Example of Segment Clustering

• In the NLSY97 a cluster of respondents were
  picked from the Lower East side of Manhattan and
  a cluster from around Yankee Stadium.

30
Implications of Sample Design for Routine Data Use

• All NLS samples contain oversamples of Blacks and
  NLSYs oversample Hispanics. Poor whites and
  military members have discontinued oversamples in
  NLSY79.
• NLSY looks different from a Simple Random Sample
• Clusters of Rs may share unobservable
  characteristics

31
Weighting

• Weight summary statistics to describe population
• For regressions Gauss Markov rules
• OLS is BLUE under standard conditions, including
  correct specification of the model
• Model heterogeneity does not call for weighted
  regression but rather weighting the various
  regression coefficients

32
Weighting Horn of the Dilemma
33
Using Weights

• Weights for the NLSY97 Round 7 range from a high
  of 1,785,202 to a low of 90,060 two implied
  decimal places
• One respondent represents from 900 to 17,852
  people, average is about 2,500
• Zero weights indicate person not interviewed
• NLSY97 and NLSY79 have single round weights
  representing population in 1997 and 1978 not
  immigrants since screening
• NLSY97 has weights for cross section (no
  oversamples) as well as panel weights

34
A NLSY79 Example From 1994

• Blacks and Hispanics on average have lower wages
  than whites (see WeightingWageData.Sas).
• Unweighted
• Mean Wage 12.50 per hour
• Median Wage 10.15 per hour
• Weighted with 1994 sample weight (R50804.00) to
  correct for oversampling
• Mean Wage 13.60 per hour
• Median Wage 11.10 per hour
• Weighting increases average wage by roughly 1.00
  per hour

35
How Do I Weight Multiple Years?

• NLS has a custom weighting program that provides
  users with the ability to go beyond weighting
  just a single round
• Web Version http//www.nlsinfo.org/web-investiga
  tor. Allows you to weight a set of survey
  rounds.
• PC-SAS Version Allows you to use the code that
  runs the web version on your own PC. Enables you
  to weight any set of respondent ids. This allows
  you to take into account event history data and
  item non-response. This is a powerful tool.

36
Web Version
37
PC-SAS Custom Weight Program

• Contact NLS User Services. They will send you a
  pair of PC-SAS programs, a set of data files and
  an input file. Jay Zagorsky at CHRR will help
  you.
• You must be comfortable making minor
  modifications to SAS programs and must have SAS
  installed on your computer.
• Program takes as input a sorted list of ids, one
  id per line. Program produces same output as web
  version
• This program allows you to weight data from an
  event history or other complex designs

38
Clustering Standard Errors

• NLS has numerous clusters of respondents who are
  alike same person in different rounds, siblings,
  people in same neighborhood
• Clustering means all observations are not
  independent (not i.i.d.) heterogeneity across
  persons and families plus spatial correlation
• PSU clustering more a problem than family
  clustering for variances d.e. 1p(k-1)
  (adjust s.e. by sqrt). Large k produces problems
  clusters larger than families. But same person
  in different rounds means large p.

39
Clustering Standard Errors(cont.)

• If intra-cluster correlations are high, number of
  effective observations number of clusters, not
  number of observations
• OLS is still consistent and unbiased must use
  GLS for correct standard errors
• Design effects in regressions are perhaps better
  described as misspecification effects as the
  intracluster correlation is due to unobserved
  variables affecting the cluster

40
Household Clustering

• NLSY97 4,027 respondents came from homes that
  had multiple respondents.
• There were six homes that each provided five
  respondents.
• NLSY79 5,914 respondents came from homes that
  had multiple respondents.
• There were four homes that each provided six
  respondents.
• Data on siblings allows us to separate effects of
  household versus individual characteristics
• For original cohorts, refer to multiple
  respondent file to detect parents and children
  across cohorts and siblings both within and
  across cohorts

41
Effect of Clustering on Std Errors

• NLSY79 to explain log of male hourly wage
• Regress hourly wages on race, age, education,
  AFQT score and marital status. Details are in
  WageData.sas
• Y XB ui vij wijk zijkt
• ui is error for PSU i, vij is component for PSU
  i and family j, wijk is component for PSU i,
  family j and person k, zijkt is idiosyncratic

42
OLS Results From SAS

• Results using OLS with SAS. Note high T-values.

Variable Coefficient T Value
Constant 5.21 133
Black -0.16 18
Hispanic -0.04 4.4
Age 0.02 18
High Grade 0.06 38
AFQT 0.001 16
Married 0.19 26
43
How To Fix Problem

• There are at least two statistical packages
  designed to fix the clustering problem.
• Sudaan (www.rti.org/sudaan) is a special purpose
  package designed to fix clustering issues.
  Integrates with SAS.
• Stata (www.stata.com) is a general purpose
  statistical program. To adjust for clustering
  for means use the Svyset command for
  regression use robust cluster (Huber-White).
• No clustering data available for Original Cohorts

44
OLS Results From Sudaan

• Here we correct for the surveys clustering on
  PSU (not on person or family)

Variable Coefficient SASs T Value Sudaans T Value
Constant 5.21 133 83
Black -0.16 18 8.8
Hispanic -0.04 4.4 1.5
Age 0.02 18 13.4
High Grade 0.06 38 18.7
AFQT 0.001 16 8.3
Married 0.19 26 15.8
45
What Happened?

• Adjusting for clustering using Sudaan resulted in
  most of the T-values falling by half. Most are
  still highly significant.
• The Hispanic variable, which was considered
  highly significant with the SAS results (Pr lt
  0.0001) is now no longer statistically
  significant (Pr lt 0.15) by most commonly used
  levels. (Problem more severe with clustered
  characteristics)

46
What Steps Are Needed To Adjust?

• First, get geocode clearance. You need this
  clearance to access replicate and PSU data.
• Second, extract all variables for your research
  plus the replicate and PSU values.
• NLSY79 The PSU variable is R02191.45, titled
  Stratum Number For Primary Sampling Units and
  the replicate variable is R02191.46, titled
  Within Stratum Replicate Of Primary Sampling
  Unit. PSU10R02191.45R02191.46
• NLSY97 The PSU variables is R13082.00, titled
  PRIMARY SAMPLING UNIT (CODED). The replicate
  variable is not released. Set replicate1 in
  your work.

47
What Steps Are Needed To Adjust?

• Third, sort your data set by replicate and PSU.
• Fourth run Sudaan. We used the following
  command.
• Proc Regress
• data"C\Documents and Settings\All
  Users\Desktop\ClusteringandWeighting\WageData.dbs"
  
• filetypeascii designwr DEFT1 est_no24000
• weight _ONE_
• nest REPLICAT PSU / MISSUNIT
• Model Ln_Pay Black Hispanic Age HGC AFQT Marry

48
Small Extension

• The SAS file we used to create the previous
  example is called WageData.sas.
• What happens when we add one more explanatory
  variable, height in inches?
• Adding this variable investigates if taller
  people earn higher wages.
• The created variable height is already part of
  the SAS data set.

49
Extension Results

• Using SAS the OLS regression results show
  heights coefficient is 0.004 and the t-value is
  3.34.
• In simple language this means each extra inch of
  height is associated with a 0.4 increase in
  hourly wages. The 3.34 T-value shows the
  coefficient is robust at the 99.9 level of
  significance, suggesting height and wages are
  definitely related.
• Using Sudaan to take into account clustering
  lowers the T-value to 2.0. Sudaan computes the
  heights coefficient significance level at 95.
  Hence, adjusting for clustering means we no
  longer have almost complete statistical certainty
  in the relationship.

50
What If You Do Not Have Sudaan (or Stata)?

• One method of getting roughly similar results is
  to add extra geographic variables which track
  each PSUs characteristics to the regression.
• Using just SAS we reran the wage function and
  included for each respondents 1979 location
  percent black, percent Hispanic, median income,
  did the respondent reside in a SMSA of 2 million
  people and dummies for USA regions (see the file
  named WageDataPlusGeoVariables.sas).
• Note we get results much like Sudaan just using
  location characteristics that are 20 years old.

51
Result of Adding Geographic Variables

• The left 3 columns are the original wage
  equation. The right 2 columns are the results
  after adding the geographic variables. Like
  Sudaan regressions, adding the extra geographic
  indicators dramatically lowers the T-statistics.

Variable Original Coefficient Orig. T Value New Coefficient New T Value
Constant 5.21 133 4.42 39.7
Black -0.16 18 -0.22 10.4
Hispanic -0.04 4.4 -0.065 3.1
Age 0.02 18 0.02 9.5
High Grade 0.06 38 0.06 18.2
I.Q. 0.001 16 0.0015 9.2
Married 0.19 26 0.22 15.1
52
Children of NLSY YAG

• Children had already diffused by their teen years
  relative to their mothers
• Clustering is more subtle, includes kin networks
• Appropriateness of Sudaan other routines more
  problematic
• The two alternatives are a complex random effects
  model or using geographic descriptors to explain
  the error components responsible for the design
  effect problem

53
Bottom Line

• GIS systems have become essential to social
  scientists. The NLSYs have a lot of data on
  location, but use is restricted.
• The oversamples and clustering of the NLSYs
  require you to think carefully about the impact
  of heterogeneity, weighting and clustering on
  your analysis. Weighting is not usually correct
  except when estimating univariate population
  moments.
• Using geographic descriptors as regressors
  attenuates design effects.
• For original cohorts, geography is very limited.