1. Updated Unified Category System for 1960-2000 Census Occupations
Peter B. Meyer
US Bureau of Labor Statistics
(but none of this represents official measurement or policy)
SSHA 2006, Minneapolis; Nov 4, 2006
Outline
Tentative standard categories
Users and bug fixes
How Census assigns occupation codes
Imputation practice
Social Science History Association conference
It's in Minneapolis, Nov 2-5, 2006
Minneapolis, Nov 2-5, 2006

2. Census Occupational Classifications
U.S. Bureau of Census determines a list of 3 digit occupation codes each ten years
Then puts one for employed respondents to the decennial Census and some other surveys
Vast data is available in these categories:
CPS, ATUS, SIPP, NLS, ACS, decennial Census
But not always consistently over long time spans
Research efforts may require some standard

3. Tradeoffs in Classification Systems
Precise job distinctions vs. Consistency, duration, and sample size
High tech occupations vs. other technical occupations
blacksmith, database admin (shorter, more precise series)
electrical engineer (longer evolving series)
“Superstars” jobs like athletes and musicians (need precision)
Licensed jobs (need long comparable occupations)
Conformity with other data
Avoid “sparseness” – many missing year-occ cells
Meaning of occupation: function, tasks, skills, background, social class
There is no perfect classification but there are tools & criteria for better ones.
Tradeoffs when mapping occupations together into unified categories
Many choices designed to achieve consistency must introduce error
No set of choices is ideal for all purposes– tradeoffs between precision and accuracy for example.
Different research questions will require different category groups, and links to other data sets
Samples size of each occupation are needed to calculate standard deviation
Number of comparable occupations are neded to determine that an particular occupation is significantly different from others
Exact earlier slide.
Duration of category vs. accuracy and precision of category
E.g. blacksmith, electrical engineer
Sample sizes of each occupation vs. number of comparable occupations
standard deviation of incomes within occ requires more cases
But narrow distinctions may be the ones of interest
English professors vs. other humanities professors
High tech occupations vs. other technical/mathematical occupations
Conformity with many other sources of data
Past Censuses, CPS, NLSY, DOT, O*NET, ISCO, HISCO
So there is no one best classification
Often (I think) poor fitting ones are used.
Researchers may need tools to make them from existing data

4. Baselines to improve on
IPUMS defined occ1950 for US workers recorded in ANY Census
Working paper (Meyer and Osborne, 2005) defined classification of digit occupations codes from 1960 to present
It was adapted from the 500+ categories in 1990 Census:
379 categories have same name or almost same as 1990
125 were eliminated to help harmonize with other years
(Example to follow)
19 categories have expanded (changed name or a not-elsewhere-classified category was given more scope)
3 categories added for 1960 data which doesn’t fit in

5. Some distinctions are lost in standardization
Census reports and IPUMS data show how many respondents would be coded
in each of two classification systems.
1970
code
occupation
title
1980/90 component category titles
Civilian
Labor Force
% of
category
284
Sales workers,
except
clerks,
retail trade
263
Sales workers, motor vehicles and boats
185,160
37.06%
266
Sales workers, furniture and home furnishings
98,941
19.80%
267
Sales workers; radio, television, hi fi, and appliances
76,674
15.35%
268
Sales workers, hardware and building supplies
81,668
16.35%
269
Sales workers, parts
39,120
7.83%
274
Sales workers, other commodities
16,008
3.20%
277
Street and door to door sales workers
2,082
0.42%
However, the title of 1980 occupation 263 is specifically restricted to motor vehicles and boats, while the 1970 title is not. If we were to use the 1980 category name and apply it to 1970 data, we would have had a category that explicitly mislabeled most of its members. Instead, we combined the workers in category 284 in 1970 into the category called “Salespersons not elsewhere classified”. Because occ1950 uses the predefined 1950 categories, no categories were renamed, or “n.e.c.” categories created or expanded, to extend consistency in definition across years.

6. User input and new data since 2005
Sent these programs to 19 people who expressed interest
Open-source code idea (helps find errors; also is public property)
Corrections from users did come in
Philip Cohen, UNC Sociology, identified some problems/mistakes.
Sarah Porter, research assistant at U of Iowa working with Jennifer Glass, wrote a program to do some similar mappings. Comparing to that program I found mistakes in mine.
Dual-coded 1990/2000 data sets highlighted some surprises
Experimented with imputations (example to follow)
Visited the Census office where they assign these codes.

7. Census Bureau's National Processing Center in Jeffersonville, IN
All occupation and industry codes are assigned there Security was pretty high.
I was not a sworn Census agent and this got in the way a bit.
 Louisville, KY, is just south of it
I interviewed four specialists who assign occupation & industry codes.

8. Information used when coding
“what kind of work"
“most important activities or duties"
employer name
“what kind of industry”
city and state ("PSU") of respondent's home
industry type (manufacturing, service, other)
years of education, age, sex
not income, although it was available before Jan '94 software.
Tens of thousands of job titles are mapped to a code in a reference book they have, if industry also matches.
Some cases may be "autocoded" by software and coder checks
After coding, public use samples have 3-digit occupation code and 3-digit industry code
Quality of assignments from public use samples are limited
coder cannot do this; referralist can

9. Imputation: Statisticians and Actuaries
These were separate categories in and after 1970
But in 1960 they were all in “statisticians and actuaries”
When standardizing (2005) they were put in “statisticians”
Counts of Actuaries and Statisticians in Census Sample
1960
1970
1980
1990
Actuaries . 45
129
182
Statisticians
199
237
352
338
I’m aware that no one has expressed interest in the question of how many actuaries there were in 1960.
However if you wanted time series on occupations in general going back to 1960, for comparative reasons, it helps to extend any of them back is of interest because there were so few semiconductors and computers then data is my control for what computers have done.
Also I thought if you were going to probabilistically assign anybody to an occupation, statisticians were the right group to start with. We can’t offend them; they do this kind of thing to other people all the time.
-- 65% of actuaries worked in industry 717 (the insurance industry), whereas only 10% of statisticians did.
-- 88% of actuaries worked in the private sector, whereas only 60% of statisticians did
-- 10% of statisticians were foreign-born; only 4% of actuaries were
-- About half of statisticians were female. Only a third of actuaries were.
-- The mean salary of actuaries was 50% higher than the mean salary of statisticians
-- Actuaries had much higher mean business income.
Will try to infer which of the 1960 people were actuaries.

10. Statisticians and Actuaries
Pooled all statisticians and actuaries
Good predictors of whether respondent is an actuary:
Recorded in a later year
Employed in insurance, accounting/auditing, or professional services industries
Employed in private sector
High salary income
High business income, or to earn mostly business income
Is employed
Lives in Connecticut, Minnesota, Nebraska, or Wisconsin
Ran many logistic regressions predicting the actuaries

11. Statisticians and Actuaries
For 1970 data that logistic regression predicts occupation right 88% of the time
Revised counts of actuaries and statisticians after imputation
1960
1970
1980
1990
Actuaries 29 45
129
182
Statisticians
170
237
352
338

12. Statisticians and Actuaries
Why work this arcane problem?
More accurate standardized “statistician” category
Longer actuary time series
Reduces sparseness – empty cells
Builds a technique for this data mining
Benefits scale up through IPUMS

13. Imputing judges In 1960 Census, lawyers and judges were one category
Later, they’re separate, and separate in “standard” system
Without more info, we categorize all in 1960 as “lawyers”.
We wish to impute which ones are judges
Useful fact: private sector ones were all called lawyers
Predictors for the public sector ones, of who’s a judge:
Older
Employed in state government
High salary income
Low business income
Educated less than 16 years
Employed at time of survey

14. Logit regression predicting judges in 1970-90 Census
Dependent variable: maximum likelihood probability this individual is a judge.
Coefficient
Std error
p-value
Year
-0.005
0.011
0.633
Age
0.155
0.033
0.000
Age-squared
-0.001
0.040
Federal government employee
-1.440
0.137
State government employee
0.499
0.263
0.058
Ln(salary)
-1.795
3.094
0.562
Ln(salary) squared
0.052
0.333
0.877
Ln(salary) cubed
0.003
0.012
0.798
Ln(business income)
-0.041
0.036
0.261
Fraction of earned income that is business income
-0.714
1.053
0.498
Education less than 16 years
2.235
0.320
Years of formal education
-0.044
0.046
0.336
Is employed at time of survey
0.224
0.241
0.352
Constant
13.017
23.428
0.578
Judges are likely to be OLDER,

15. Thus we assign judge occupation code
gen logitindex = * year * age
* age * age * indfed
* indstate * lnwage
* lnwage * lnwage * lnwage * lnwage * lnwage
* lnbus * busfrac * (educyrs<16)
* educyrs * employed
/* constant */ ; …
gen logitval=exp(logitindex)/(1.0+exp(logitindex))
replace logitval=.0001 if !govtemployee /* this is a perfect predictor */
replace logitval=.0001 if !indfed & !indstate & !indlocal /* this too */
gen assigned = logitval>.46 /* Now ‘assigned’ has a 1 for imputed judges */
Threshhold probability is chosen to match the number of judges expected to be there, based on annual trend.
Can get 83% accurate predictions from such a rule on 1970 data.
This mis-assigns a few who should have stayed lawyers.
Having data-mined the coefficients, we forge this tool to cut between the lawyers and the judges.

16. Newly Imputed Judges 1960 1970 1980 1990 Lawyers 1971 2570 5082 7603
Respondents in Census samples after imputation
1960
1970
1980
1990
Lawyers
1971
2570
5082
7603
Judges 82
123
298
331

17. What's next?
Use dual-coded CPS datasets with 1990 and 2000 codes to make a few more imputations
Keep listening, seek more help, make it better.
Publish variable at IPUMS.org
Keep going? & 1980 dual coded data sets exist.

18. Industry and occupation coding
Industry codes and occupations codes are assigned by the same group of people, at the same time for each respondent.
Industry is almost always decided first.
The people who do that are “coders”
Procedures are carefully documented
I wasn’t a “sworn” Census agent and couldn’t see it done, live

19. Desirable Attributes of a Classification
For each occupation, well-behaved time-series of:
mean wage
wage variance
fraction of the population
New criterion: SPARSENESS
One prefers a classification not be sparse, meaning it does not have many empty occ-year cells
A good table is the one on CV from 1990 to Technicians n.e.c. shows a change that we could maybe fix.
My input:
For each occupation which exists in two consecutive Census samples we calculate these summary statistics in both years, and compare them.
IF something did jump around there is not necessarily much we can do.
We might change the definition, or think through whether the phenomenon of change is a real-world change, or a redefinition of the category. Possibly we can document this
Where possible we want to the occupations to be a repeated cross section or panel.
THINK ABOUT THE EXAMPLE.
PROPOSED_STANDARD wfreqnum_obs1960 wfreqnum_obs1970 wfreqnum_obs1980 wfreqnum_obs1990 wfreqnum_obs2000

20. What new information would help referralists?
Information about a job title
Information about employer's city and state
not respondent’s
But asking more questions would extend the CPS interview

21. Problems faced by referralists
Too little information from respondent
“Computer work" for “kind of work”
Exaggeration (example: dot com businesses)
Ambiguities:
"water company" for industry or employer
"surveyor" occupation
"boot" vs "boat" in handwriting
Having to hurry
Referralists confer with each other routinely, but sometimes make different choices from one another
Does technological change go along with occupational ambiguity? YES. Problems with computer work, biotech. Still no nanotech in classification.

22. The information coders have

23. Who's Doing the Coding
There were about 12 coders and 14 referralists in October 2006
Referralists have been coders before and usually have 9+ years of experience
I interviewed three referralists, and a supervisor
The ones I met handled referrals from several surveys:
CPS, ATUS, SIPP, NLS, ACS
others on contract
All these use decennial Census occupation codes
They DON’T handle the decennial Censuses

24. Information available to referralist
Can match Employer name to a known employer from their Employer Name List (ENL), same as SSEL or Business Registry.
Can look on the web for that employer
Can study “little red book” - SOC manual
or (less often) giant Dict Occ Titles 1991
or, I’m told, look up employer in Dun and Broadstreet data
They try to make a coherent choice for industry and occupation together.

25. “Coders” and “Referralists”
Coders follow carefully documented procedures from the Census headquarters in Suitland, MD
Coders with two years of experience are expected to assign 94 codes an hour, with 95% accuracy (which is checked)
If there is not enough information to assign industry and occupation codes by procedure, the case is forwarded electronically ("referred") to a “Referralist" (aka statistical assistant)