Program Recipiency

Program Recipiency

Karima Nagi

Transfer Program in the 1997 Cohort of the NLSY

Dan BlackUniversity of Chicago & NORC

Four types of transfer payments

• TANF (AFDC), Temporary Assistance to Needy Families, what is generally referred to as “welfare”

• Food Stamps • WIC (Women, Infants, and Children)• Other transfer payments – primarily SSI,

or Supplemental Security Income, which may include some “insurance” programs such as Disability Insurance

Four types of variables

• Status: an indicator for each month whether or not the household is participating in the program since age 14

• Amount received

• Household member(s) receiving the aid

• Deny: Indicates that the respondent denies previously reported receipt

Two types of social insurance

• Workers compensation (only available in first two rounds)

• Unemployment Insurance

• Social insurance programs are different than transfer payments because your “contributions” or taxes fund the program

• I will focus on Unemployment Insurance for this talk

Strengths of data

• While this data clearly allows you to track the incidence of, say, Food Stamp use, other data sets such as the CPS will give you much larger and much broader samples

• Our data covers adolescent “participation” in programs as well as young adult with reasonably large samples

• Our data allow you to see what fraction of time they have been on a program

Strengths of data

• Our data allow you to estimate the duration of the spell of recipiency

• Such duration models are also often referred to as “hazard models”

• Policy relevance: for a program of a given size you might want to know is this a large group of individuals with relatively short stays or a small group with long stays

A gentle but incomplete introduction to hazard models

• These are discrete-time hazard models because the data are monthly

• Most standard software is not appropriate because it assumes underlying data are continuous (no observations with the same length of spell)

• With monthly data, you will get a very large number of observations with the same spell length

• Stata has ado file that will do estimation, pgmhaz.ado


• Basic idea: Given that the spell of transfer recipiency lasted until time t0, what is the probability that it ends at time (t0+1)?

• Mathematically, this is just

• If you have a complete set of conditional probabilities, you can recover probability function

00 0

0

( 1| )Pr( 1| , )

1 ( | )

f t Xt t X

F t X


• A couple of complications• First, you may not observed the end of the spell

because of data limitations or other censoring mechanisms, right-hand censoring

• Second, you may not observe the beginning of the spell because of data limitations (left-hand censoring)

• The first problem is easy to handle (and pgmhaz.ado does it for you), the third is a lot harder


• To estimate these models, you would arrange the data where each month is an observation. For persons’ whose spells last 3 months, they will have 3 observations. For persons’ whose spells last 5 months, they will have 5 observations

• Each dependent variable will be a 0 until the month they leave, which will be a 1

• Recipiency may not be monthly – may be weekly


• Consider the probability we are estimating

• In months 1, the probability of “surviving” to month 0 has no effect so you use the whole sample

• In month 2, you use only those who survived month 1 so your sample is conditional

00 0

0

( 1| )Pr( 1| , )

1 ( | )

f t Xt t X

F t X


• Again, we wish to estimate

• This begins to look like a logit or probit problem, and the pmghaz.ado uses a logit-type model, but constrains to coefficients to be the same on the X’s across months

• Because the sample is limited to survivors for months 2 and beyond we may have a problem

00 0

0

( 1| )Pr( 1| , )

1 ( | )

f t Xt t X

F t X


• The surviving sample is selected on survivorship so any unobserved differences (or heterogeneity) results in biased estimates

• One approach is to assume that the heterogeneity results in a gamma mixture model as proposed by Meyer (and you guessed it, pgmhaz.ado does this for you)

• More complicated adjustments (Heckman-Singer’s nonparametric adjustment) are not yet implemented


• Further reading– Basic approach

• Meyer, B. D. “Unemployment insurance and unemployment spells,” Econometrica August 1990 58(4) 757-782.

– More complicated heterogeneity adjustment• Heckman, J.J., and B. Singer. “Econometric duration

analysis” Journal of Econometrics January/February 1984 24(1-2) 63-132

– Left-hand censoring• Berger, M.C. and D. A. Black. “The Duration of Medicaid

Spells: An Analysis Using Flow and Stock Samples,” Review of Economics and Statistics November 1998 80(4) 667‑674.

A very short bit on UI

• UI is an insurance program that only covers workers with an adequate work history and who are not dismissed for cause or who do not quit

• A UI claim covers one year and allows you to take full benefits for 6 months (generally) within the year

• May be periodically extended

How do you go from the data to estimation?

• Start with the User Manual!!!• Begin to pull the data you need, including

covariates• This is a slow process, but do not rush it.

You will just have to come back• Can pull some variables from Harris

School’s (Bob Michael’s) “flat files” at http://harrisschool.uchicago.edu/Research/faculty_projects/NLSY97_flat_files/

http://harrisschool.uchicago.edu/Research/faculty_projects/NLSY97_flat_files/

http://harrisschool.uchicago.edu/Research/faculty_projects/NLSY97_flat_files/

Data to estimation

Name Tag Question Variable Title Year

1 R9147700 UNEMP_STATUS_2002.01 2002 UNEMPLOYMENT INSURANCE: R RECEIVED IN MONTH 01 2004






Data to estimation

• Web investigator will dutifully give you 8984 observations, one for each public id

• Most of these are not going to be used. For instance, in 2002, only 208 report any spell of unemployment

• You need to remove data with no UI spells (which is easy in Stata using the egen anycount option– egen x = anycount(ui*), v(1)

Data to estimation

• Let’s look at some data

• I took the data from 2002 and dropped all observations that had no unemployment spells

• I am going to name variable mxx, where xx is the number of the month from 01 to 12. This a terrible convention (I actually used ui2002xx, but it was hard to display)

Data to estimation

+-----------------------------------------------------------------------+m01 m02 m03 m04 m05 m06 m07 m08 m09 m10 m11 m12 -----------------------------------------------------------------------

1. No No No No No No Yes Yes Yes No No Yes 2. No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes 3. No No No No No No No No No No Yes Yes 4. No No No No No No No No Yes Yes Yes Yes 5. No No No No No No No Yes Yes No No No

-----------------------------------------------------------------------6. No No No No No No No Yes No No No No 7. No No No No No No No Yes Yes Yes No No 8. No No No No No No Yes Yes No No No No 9. Yes No No No No No No No No No No No 10. No No Yes Yes Yes Yes Yes No No No No No

Data to estimation

+-----------------------------------------------------------------------+m01 m02 m03 m04 m05 m06 m07 m08 m09 m10 m11 m12 -----------------------------------------------------------------------

52. No No No No No No No No No Yes Yes . 120. No No Yes Yes . . . . . . . . 142. Yes . . . . . . . . . . . 153. Yes No No . . . . . . . . .

Data to estimation

• Need to then keep relevant months and reshape the data to meet your estimation needs (see Stata’s reshape command)

• Will need to collect and append any time invariant variables – race, sex, ethnicity, ASVAB score – to each month of the data

• If you believe that calendar month affects transitions, you need to keep track of time as well

Further considerations

• You may want some covariates to vary with time (age, education, number of children). How will you do this?

• Age is relatively easy as you can augment it by a month

• Education: What if attending school?

• Children: need to be very precise about timing

Program Recipiency

Documents

data sets

underlying data

talkstrengths of data

large samplesour data

incomplete introduction

broader samplesour data

discretetime hazard

spell of transfer recipiency