Understanding Sheepskin Effects in the Returns to Education: The - - PowerPoint PPT Presentation

SLIDE 1

1

Understanding ‘Sheepskin Effects’ in the Returns to Education: The Role of Cognitive Skills

• W. Craig Riddell

University of British Columbia CLSRN Workshop University of Toronto November 18-19, 2008

SLIDE 2

2

Objective

• To investigate the role of cognitive skills in “sheepskin effects” in

impacts of education

• Sheepskin effects refer to outcomes associated with the completion
• f a diploma or degree, controlling for years of schooling
• Credential effects can be interpreted as the value of program

completion -- the difference in outcomes between those with a diploma or degree and non-completers with the same years of schooling

• Study uses measures of prose literacy, document literacy, numeracy

and problem solving

SLIDE 3

3

Why sheepskin effects may matter I

• Completion of educational programs is associated with noteworthy

changes in outcomes

• Earnings: discrete jumps in earnings associated with credentials
• Crime: substantial drop in criminal activity associated with HS

graduation (Figure 1)

• Re-employment success: discrete jumps in re-employment

probabilities at 12 and 16 years of schooling (Figures 2 and 3)

SLIDE 4

4

Figure 1. Regression-Adjusted Probability of Incarceration, by Years of Schooling

SLIDE 5

5

Figure 2. Regression-Adjusted Probability of Re-employment Conditional on Being Unemployed for More than 12 Weeks in the Previous Year by Years of Schooling Data source: 1980 Census merged with compulsory schooling laws and child labor laws Number of observations: 204,853

Note: Regression-adjusted probability of re-employment is obtained by conditioning on state of birth, state of residence, gender, race, and cohort of birth (1916-1925, 1936-1945, etc.). The graphs display the coefficient estimates on the complete set of schooling dummies. The intercept applies to the base category – white males who were born in California between 1936 and 1945, had eight years of schooling or less, lived in California at the time of the survey.

SLIDE 6

6

Figure 3. Regression-Adjusted Probability of Re-employment Conditional on Being Unemployed in the Previous Year

Data source: Canadian Census (1981-2001) Number of observations: 458,641

Note: Regression-adjusted probabilities of re-employment were obtained by conditioning on survey year, province/territory, CMA (Toronto, Montreal, Vancouver, or other CMA), nine age groups (age 20- 24, 25-29, and so on), gender, marital status, census family size, and language. The graphs display the coefficient estimates on the complete set of schooling dummies. The intercept applies to the base category -- males surveyed in 1981 who were 35 to 39 years of age, had eight years of schooling or less, were married, lived in a CMA other than Toronto, Montreal, or Vancouver, lived in Ontario, and

• nly spoke English at the time of the survey.

SLIDE 7

7

Why sheepskin effects may matter II

• Help to understand economic returns to education
• Most common specification (human capital earnings function) is

linear in years of schooling

• According to this model each year of school raises earnings by the

same proportion

• Alternative popular specification uses highest level of education
• This amounts to a step function in which only credentials matter
• Sheepskin specification nests both years of schooling and program

completion within a more general model (see Figure 4)

SLIDE 8

8

Figure 4. Graphic representations of Five Models

• f the Economic Returns to Education

Source: Goodman, Jerry D. “The Economic Returns of Education: An assessment of Alternative Models” Social Science Quarterly 60 (September 1979) pp. 269-283

Ln(Earnings) Linear

0Elementary School 8High School 12College 16Grad18

Educational Attainment Ln(Earnings) Credentialist

0Elementary School 8High School 12College 16Grad18

Educational Attainment

18

Ln(Earnings) Linear/Credentialist

0Elementary School 8High School 12College 16Grad

Educational Attainment

SLIDE 9

9

Why sheepskin effects may matter III

• Relevant to how we measure human capital over time or across

jurisdictions/countries

• Most international comparisons are credential-based
• Years of schooling may be more comparable measure of

educational inputs

• Census and household surveys have been moving away from years
• f schooling

SLIDE 10

10

Why sheepskin effects may matter IV

• May help distinguishing among alternative theories of education
• Layard and Psacharopoulos (1974) dismiss signaling theory on

basis of absence of sheepskin effects in their data

• However, evidence not decisive
• Sheepskin effects consistent with human capital theory if package of

complementary educational inputs matter

• Also consistent with both signaling/screening and human capital

views is graduates are more skilled than non-graduates

SLIDE 11

11

Why sheepskin effects may matter V

• Education policy often focuses on program completion
• Many initiatives to encourage high school completion
• Recent concern that many university enrollees do not graduate
• If only years of schooling matter, program completion irrelevant

SLIDE 12

12

Previous literature

• Layard and Psacharopoulos (JPE 1974) crude comparison of dropouts and grads
• Hungerford and Solon (REStat 1987) linear spline with nodes at S = 8, 12, 16, 17, 18

Find SE at S = 16 (largest) and S = 12. Don’t observe program completion

• Jaeger and Page (REStat 1996) match respondents to old and new CPS

Find years of study an imperfect measure of program completion Estimated sheepskin effects much larger when program completion observed EG HS grad: 11% vs 3% with HS (1987) specification EG College grad: 31% vs 12% with HS (1987) specification

• Park (1999) reports similar results
• Ferrer and Riddell (CJE 2002) use Canadian census data

Estimated sheepskin effects for Canadian born: – HS grad 4%-6% – PS without HS 6% – PS with HS 3%-5% – Univ BA 21% – Univ MA 7%-10% – Univ Professional 30%

SLIDE 13

13

Data: IALSS 2003

• Several key advantages
• Direct measures of cognitive skills
• Information on years of completed schooling and highest level of

education

• Separate question on high school graduation
• Large sample

SLIDE 14

14

Cognitive skills measures

• Four skills assessed: prose literacy, document literacy, numeracy,

problem solving

• Tests assess ability to apply skills in everyday settings
• Results presented for average of four skills

SLIDE 15

15

Sample restrictions

• Full sample: focus on Canadian born, non-aboriginal population
• Drop those whose main activity is “student”
• Worker sample: drop self-employed, unemployed, non-participants,

wage outliers

• Sample sizes: Main sample 14,637

Worker sample 7,766

SLIDE 16

16

Educational attainment

Six main categories:

Less than HS HS graduate Non-university post-secondary without high school completion Non-university post-secondary with high school completion University bachelor’s degree University postgraduate and professional degree

SLIDE 17

17 Table 1 Summary Statistics for the Full Sample

Males Females Both Genders Age 44 45.8 44.9 Experience 25.1 26.9 26 Years of Schooling 12.9 13 12.9 Educational Attainment

% Less than High School 26.5 23.1 24.7 % High School 30.5 33.7 32.1 % Non Univ PS w/o HS 3.1 1.9 2.5 % Non Univ PS with HS 21.9 24 23 % University BA 12.7 13.6 13.2 % Univ Postgrad 5.3 3.7 4.5

Mother’s education

% Less than High School 41.2 48.3 44.8 % High School 29 23.4 26.1 % Some Post Secondary 12.7 15.2 14 % BA or higher 8.2 6.8 7.5 % None reported 8.9 6.3 7.6

Father’s education

% Less than High School 47 50.6 48.8 % High School 20.3 18.9 19.6 % Some Post Secondary 12 11.2 11.6 % BA or higher 11.7 10.5 11.1 % None reported 9 8.8 8.9

Immigrant Parents

% Immigrant mother 16.4 15.6 16 % Immigrant father 19.4 17.9 18.6

Math in high school

% Good math grades 70.8 66 68.3 % Teachers went too fast 21.4 30.8 26.2

Prose Literacy 278.1 282.9 280.6 Document Literacy 280.6 275.8 278.2 Numeracy 277.4 261.2 269.2 Problem Solving 273.8 273 273.4 Average Skill Score 277.5 273.2 275.3 Number of Observations 6693 7944 14637

SLIDE 18

18

Are graduates more skilled?

• Investigate whether sheepskin effects in generation of skills
• Figures 1 and 2 show plots of skills versus years of schooling
• No obvious discrete jumps at S = 12 or S = 16

SLIDE 19

19

Figure 1 Cognitive skills by years of schooling

SLIDE 20

20

Figure 2 Average skill score by average years of schooling

SLIDE 21

21

Skills generation (log skills regressions: Table 2)

• Small gender difference
• Essentially no relationship between skills and age in cross-section
• Strong relationship with education, but diminishing returns
• Impact of one extra year of S: @ S = 12 3.2% @ S = 16 2.1%
• Moderately large sheepskin effects (column 3)
• Column 4 adds controls for parental education and parental

immigrant status

SLIDE 22

22

Table 2 Sheepskin effects in the determinants of skills

OLS 1 OLS 2 OLS 3 OLS 4 OLS 5 Female

• 0.0137***

(0.0042)

• 0.0125***

(0.0043)

• 0.0148***

(0.0041)

• 0.0131***

(0.0041)

• 0.0097**

(0.0041) Years of schooling 0.0660*** (0.0045)

• 0.0556***

(0.0053) 0.0524*** (0.0050) 0.0425*** (0.0055) Schooling squared

• 0.0014***

(0.0002)

• 0.0013***

(0.0002)

• 0.0013***

(0.0002)

• 0.0010***

(0.0002) Age 0.0033*** (0.0007) 0.0025*** (0.0007) 0.0025*** (0.0007) 0.0052*** (0.0007) 0.0052*** (0.0007) Age squared (/100)

• 0.0001***

(0.0000)

• 0.0001***

(0.0000)

• 0.0001***

(0.0000)

• 0.0001***

(0.0000)

• 0.0001***

(0.0000) Educational Attainment High School

• 0.1545***

(0.0069)

• Non Univ PS w/o HS
• 0.1072***

(0.0138)

• Non Univ PS with HS

0.1998*** (0.0072) University BA

• 0.2682***

(0.0077)

• Univ Postgrad
• 0.3224***

(0.0102)

• Sheepskin effects

HS grad

• 0.0673***

(0.0082) 0.0602*** (0.0079) 0.0617*** (0.0077) Non Univ PS w/o HS

• 0.0525***

(0.0135) 0.0414*** (0.0135) 0.0412*** (0.0134) Non Univ PS with HS

• 0.0107**

(0.0054) 0.0079 (0.0053) 0.0099* (0.0053) Univ BA

• 0.0429***

(0.0076) 0.0346*** (0.0073) 0.0338*** (0.0072) Univ Postgrad

• 0.0394***

(0.0099) 0.0359*** (0.0097) 0.0306*** (0.0095)

SLIDE 23

23

Table 2 Sheepskin effects in the determinants of skills (continued)

Standard errors in parenthesis. *, **, *** statistically significant at 10%, 5%, 1% level.

OLS 1 OLS 2 OLS 3 OLS 4 OLS 5 Mother’s education Less than High School

• 0.0344***

(0.0052)

• 0.0358***

(0.0051) Some Post Secondary

• 0.0009

(0.0061)

• 0.0017

(0.0059) BA or higher

• 0.0158*

(0.0095) 0.0134 (0.0091) None reported

• 0.0556***

(0.0105)

• 0.0567***

(0.0102) Father’s education Less than High School

• 0.0244***

(0.0062)

• 0.0234***

(0.0062) Some Post Secondary

• 0.0017

(0.0068) 0.0026 (0.0066) BA or higher

• 0.0065

(0.0078) 0.0100 (0.0076) None reported

• 0.0465***

(0.0104)

• 0.0450***

(0.0104) Immigrant Parents Immigrant mother

• 0.0066

(0.0071) 0.0061 (0.0069) Immigrant father

• 0.0065

(0.0064) 0.0076 (0.0062) Math in high school Good math grades

• 0.0224***

(0.0051) Teachers went too fast

• 0.0264***

(0.0053) Constant 5.0103*** (0.0351) 5.5061*** (0.0165) 5.0952*** (0.0389) 5.0962*** (0.0380) 5.1633*** (0.0425) Observations 14,637 14,637 14,637 14,637 14,637 R-squared 0.5084 0.4767 0.5224 0.5403 0.5506

SLIDE 24

24

Skills generation: possible omitted variables bias

• Educational attainment and cognitive skills may be influenced by

innate ability

• Ideal control would be IQ type measure at young age
• IV strategy difficult in sheepskin specification
• Proxies for innate ability have significant and moderately large

effects (combined effect about 5%)

• Modest decline in coefficient on years of schooling (about 15%)
• Little impact on estimated sheepskin effects

SLIDE 25

25

Earnings, education and cognitive skills I

• Columns 1 to 3 in Table 4 report standard log earnings regressions
• Note importance of distinguishing post-secondary grads with and

without HS

• Pattern of estimated sheepskin effects in IALSS data similar to

census data but magnitudes larger: – HS graduate + 21% – Post-secondary without HS + 23% – Post-secondary with HS + 17% – University BA + 43%

SLIDE 26

26

OLS 1 OLS 2 OLS 3 OLS 4 OLS 5 OLS 6 Female

• 0.4106***

(0.0242)

• 0.4178***

(0.0234)

• 0.4201***

(0.0235)

• 0.4064***

(0.0236)

• 0.4124***

(0.0231)

• 0.4146***

(0.0232) Experience 0.0656*** (0.0036) 0.0635*** (0.0036) 0.0637*** (0.0037) 0.0659*** (0.0035) 0.0642*** (0.0036) 0.0643*** (0.0037) Experience squared (/100)

• 0.1127***

(0.0080)

• 0.1138***

(0.0084)

• 0.1110***

(0.0085)

• 0.1106***

(0.0079)

• 0.1117***

(0.0082)

• 0.1099***

(0.0083) Years of schooling 0.0890*** (0.0042)

• 0.0331***

(0.0066) 0.0706*** (0.0047)

• 0.0243***

(0.0065) Educational Attainment High School

• 0.2935***

(0.0384)

• 0.2178***

(0.0398)

• Non Univ PS w/o HS
• 0.2858***

(0.0962)

• 0.2429***

(0.0911)

• Non Univ PS with HS
• 0.5229***

(0.0428)

• 0.4201***

(0.0462)

• University BA
• 0.8585***

(0.0424)

• 0.7070***

(0.0469)

• Univ Postgrad
• 1.0124***

(0.0606)

• 0.8264***

(0.0677)

• Sheepskin effects

HS grad

• 0.2062***

(0.0409)

• 0.1600***

(0.0413) Non Univ PS w/o HS

• 0.2329**

(0.0961)

• 0.2075**

(0.0913) Non Univ PS with HS

• 0.1719***

(0.0336)

• 0.1624***

(0.0333) Univ BA

• 0.4253***

(0.0415)

• 0.3931***

(0.0409) Univ Postgrad

• 0.0738

(0.0634)

• 0.0635

(0.0628) Average Skill Score

• 0.0028***

(0.0003) 0.0025*** (0.0003) 0.0023*** (0.0003) Constant 4.7416*** (0.0679) 5.5715*** (0.0422) 5.2250*** (0.0809) 4.1507*** (0.0971) 4.8998*** (0.1005) 4.7036*** (0.1143) Observations 7,766 7,766 7,766 7,766 7,766 7,766 R-squared 0.3718 0.3949 0.4003 0.3887 0.4083 0.4111

Table 4 Earnings regressions, Worker sample

Standard errors in parenthesis. *, **, *** statistically significant at 10%, 5%, 1% level.

SLIDE 27

27

Earnings, education and cognitive skills II

• Columns 4 to 6 add controls for cognitive skills
• Coefficients on years of schooling and level of education drop by

about 20%

• Indicates that education has a large impact on earnings in addition

to its impact that operates through cognitive skills generation

• Estimated sheepskin effects also decline, especially that for HS grad

(about 20%)

• Decline in other sheepskin effects relatively modest
• EG effects for college diploma or university degree fall by < 10%

SLIDE 28

28

Earnings, education and cognitive skills III

• Table 5 adds proxies for innate ability
• These are insignificant once we control for cognitive skills
• Little change in estimated sheepskin effects

SLIDE 29

29

OLS 1 OLS 2 OLS 3 OLS 4 OLS 5 OLS 6 Female

• 0.3968***

(0.0245)

• 0.3961***

(0.0241)

• 0.4062***

(0.0238)

• 0.4040***

(0.0236)

• 0.4086***

(0.0239)

• 0.4060***

(0.0237) Experience 0.0661*** (0.0036) 0.0665*** (0.0035) 0.0636*** (0.0036) 0.0644*** (0.0036) 0.0640*** (0.0037) 0.0646*** (0.0037) Experience squared (/100)

• 0.1149***

(0.0081)

• 0.1131***

(0.0080)

• 0.1144***

(0.0084)

• 0.1129***

(0.0082)

• 0.1122***

(0.0085)

• 0.1114***

(0.0083) Years of schooling 0.0874*** (0.0043) 0.0709*** (0.0048)

• 0.0332***

(0.0067) 0.0255*** (0.0066) Educational Attainment High School

• 0.2892***

(0.0391) 0.2225*** (0.0403)

• Non Univ PS w/o HS
• 0.2939***

(0.0961) 0.2547*** (0.0911)

• Non Univ PS with HS
• 0.5173***

(0.0431) 0.4252*** (0.0465)

• University BA
• 0.8389***

(0.0432) 0.7045*** (0.0472)

• Univ Postgrad
• 0.9869***

(0.0615) 0.8211*** (0.0681)

• Sheepskin effects

HS grad

• 0.2033***

(0.0413) 0.1618*** (0.0417) Non Univ PS w/o HS

• 0.2423**

(0.0960) 0.2181** (0.0913) Non Univ PS with HS

• 0.1709***

(0.0335) 0.1611*** (0.0333) Univ BA

• 0.4107***

(0.0417) 0.3812*** (0.0413) Univ Postgrad

• 0.0679

(0.0634) 0.0578 (0.0631) Average Skill Score

• 0.0027***

(0.0004)

• 0.0024***

(0.0004)

• 0.0022***

(0.0004) Math in high school Good math grades 0.0576* (0.0300) 0.0377 (0.0291) 0.0488* (0.0290) 0.0330 (0.0283) 0.0448 (0.0290) 0.0312 (0.0284) Teachers went too fast

• 0.0777***

(0.0296)

• 0.0554*

(0.0299)

• 0.0727**

(0.0288)

• 0.0519*

(0.0292)

• 0.0699**

(0.0288)

• 0.0516*

(0.0292) Constant 4.7374*** (0.0746) 4.1665*** (0.1066) 5.5595*** (0.0470) 4.9230*** (0.1090) 5.2119*** (0.0870) 4.7119*** (0.1240) Observations 7,766 7,766 7,766 7,766 7,766 7,766 R-squared 0.3773 0.3922 0.3989 0.4105 0.4040 0.4134

Table 5 Earnings regressions with proxies for ability, Worker sample

Standard errors in parenthesis. *, **, *** statistically significant at 10%, 5%, 1% level.

SLIDE 30

30

Summary of main results

• Evidence of sheepskin effects in production of cognitive skills
• Graduates are more skilled than dropouts with same years of

schooling

• Previous estimated sheepskin effects do in part reflect (typically)

unobserved skills

• Addition of measures of skills results in a decline in estimated

sheepskin effects in the order of 20%

• Estimated sheepskin effects remain large even with controls for

cognitive skills

• Credential effects that remain may reflect non-cognitive skills – eg

perseverance – or cognitive skills not assessed in IALSS