Reaching for the Stars:
Who Pays for Talent in Innovative Industries?
FREDRIK ANDERSSON, CORNELL UNIVERSITY,
MATTHEW FREEDMAN, UNIVERSITY OF
MARYLAND,
JOHN HALTIWANGER, UNIVERSITY OF
MARYLAND AND NBER,
JULIA LANE, NSF
KATHRYN SHAW, STANFORD UNIVERSITY AND
NBER
Focus of paper
2
 What is the link between product market and labor market?
 Theory of innovation-based theory of production creates
implications for the structure of earnings.
 This paper examines how firms recruit, motivate and retain
talented workers in a particularly innovative industry –
software.
 We examine the relationship between the variation in the
returns to innovation and the variation in compensation.



Product innovation in the software industry is very closely tied to the
talents of the workforce.
Software industry is characterized by skewed returns: successful
innovations can produce an enormous payoff to the firm, while failed
products can lead to large losses. Also skewed compensation structure.
Variance of product payoffs is very different in different segments of the
industry.
Background Point 1: Firms pay a lot for star
software workers
3
Table 1
Summary Earnings Statistics, Workers 21-44
Mean
Median*
90th*
SD
(a) 2000 Decennial Census (PUMS) Data – Total Earnings 35+ Hours/Week & 35+ Weeks/Year
All Industries
40,918
31,891
70,160
183,134
Software Industry (SIC 7372)
80,787
63,782
127,563
334,906
Computer Software Engineers (Census Occupation Code 102) in the Software Industry
90,668
70,691
138,193
369,374
69,353
59,665
108,692
82,432
344,268
95,508
310,644
2,051,985
107,660
80,899
184,951
142,526
Worker Earnings at end of job spell (Censored and Uncensored)
2,532,500
670,993
6,688,470
* Average within a 10% band around the true percentile. ** Annualized earnings three quarters prior to last observed full quarter.
*** Includes only individuals for whom we observe a prior spell in the data.
6,064,204
(b) LEHD Data for Ten States - Earning $50,000+ Annualized
All workers in Software Industry
Starting Earnings (Excludes Left-Censored)
Worker Earnings at end of job spell (Censored and Uncensored)
Top Decile of Workers in Software Industry
Starting Earnings (Excludes Left-Censored)
Background Point 2: Stars earn more with experience
(or the variance of pay rises, comparing the distribution of
Beginning-of-Spell & End-of-Spell
Earnings)
4
Background Point 3: There is a high variance to the gains to
innovation in the software industry
5
(Table 2: Top Video Games, Ranked
by 2007 Sales Revenues)
Game
Producer
Units Sold
(millions)
Price/unit
Sales
($million)
Halo 3
Xbox 360 Microsoft
4.82
$59.99
$289.15
Wii Play w/remote
Wii, Nintendo
4.12
$49.99
$205.96
Call of Duty 4
Xbox 360, Activision
3.04
$59.99
$182.37
Guitar Hero III: Legends Of Rock w/guitar
PlayStation
Neversoft/Budcat/Activision)
2.72
89.99
244.77
Super Mario Galaxy
Wii, Nintendo
2.52
$49.99
$125.97
Pokemon Diamond
DS, Nintendo
2.48
$34.99
$86.78
Madden NFL 08
PS2 Electronic Arts
1.9
$29.99
$56.98
Guitar Hero II w/guitar
PS2 Activision
1.89
$89.99
$170.08
Assassin's Creed
Xbox 360 Ubisoft
1.87
$59.99
$112.18
Mario Party 8
Wii Nintendo
1.82
$49.99
$90.98
Paper Objectives: Linking the Background Points
6
 Hypothesis: firms operating in product markets with high payoff
dispersion will hire and reward stars more.
 We test this hypothesis using employer-employee matched data from
the software industry
Paper Objectives
7
 Bigger picture:
 Labor economics: The hypothesis is that the demand for
innovation has pushed up the demand for and the wages of the
most skilled knowledge workers, where skill is defined as the
ability to innovate and solve problems.
 Personnel economics: We are connecting the firm’s product
market strategy to its human resource management practices
– explaining why some firms choose practices of careful
selection and high incentive pay and some firms do not. Few
empirical studies have identified which firms gain from
specific HR practices.
Model of Innovation: Projects have a Payoff Distribution
(some projects have huge payoffs; some have losses)
8
What do ‘star’ workers do in firms?
9
 Stars are the innovative spark or creative talent that results in the success
of innovative projects: stars create or pick projects better than non-stars.
 Stars reduce the false positives and false negatives in project outcomes:
they shift out the payoff distribution – they accept fewer bad projects
(false positives) and reject fewer good projects (false negatives).
 The payoff gains for stars is lower in lower-risk payoff markets: (PA2PA1)>(PB2-PB1)
 firms in high variance product lines should hire more stars
Shifts in the Payoff Distribution Due to Reductions in
False Positive or False Negative Errors
(a) More Risky Payoff Distribution
10
(b) Less Risky Payoff Distribution
Advantages of Focusing on Software
11
 High variance product payoffs
 Knowledge workers are key inputs
 Production function: output is a function of personal
innovation
 Variance of product payoffs is very different in
different segments of the industry
 Industry contributes to economic growth
The Data Set
12
 Employer-employee matched data, and product-matched data for software
 Employee Data: individual income from Unemployment Insurance quarterly data
for all employees, for all employers, for the full universe of employers and employees
for 10 states
 Employer Data: Firm-specific product revenue information for every software firm,
from the Services Industry Economic Census of Software Publishing conducted every
five years
 N=83,497 employees with 143,485 job spells
 Advantages of the data: income is all income from salary, bonuses, and
exercised stock options, for all workers in all software companies
 Disadvantages of the data: no information on occupation or hours of work
 Three data sets created:
 Employees who earn more than $50,000 a year and ages 21-44  N=51,859
 Those who also have complete firm information  N=26,726
 A subset of these who are workers in software occupations in 2000 Decennial Census
of population  N=2,638
The Measures
13
 Focus on the software industry
 Calculate product line payoff dispersion for each firm using the Census of
Software Publishing for 1997

Steps to calculate the product-specific dispersion of sales per worker for all firms:




For each of the 30 product classes in software, using firms’ product line data, calculate the
90/50 ratio of log of sales per worker
Examples of Product classes: game and entertainment; business graphics design; layout
software; etc.
Given 90/50 for the 30 product classes, we create 90/50 for each firm using the firm’s actual
product sales mix weights
This “product payoff dispersion” measure reflects the firms actual product mix, but
not its actual revenue. A firm with a high Product Payoff Dispersion measure is not
necessarily a high or low performing firm, but rather has a product mix with a right
skewed distribution of payoffs
 Match all workers’ wages within all software firms to the Census data using the
UI wage data for all individuals in the software industry (for ten U.S. states).

Wages are measured for all full quarters of earnings, including wages, bonuses, and
exercised stock options.
Empirical Hypotheses
14
 Primary hypothesis: ‘star’ talent is sorted into firms
with a high payoff dispersion because these firms
value the star skills of project innovation the most.
 Testable hypothesis: given that talent is unobserved,
but wages are observed, we should find that the firms
with the highest payoff dispersion should pay the
highest wages, as talent sorts to those firms. [These
firms may also be offering incentive pay that raises
effort the most.]
Empirical Hypotheses
15
 Auxiliary hypotheses:
 Wages should be more sensitive to the firm’s payoff dispersion
for more highly skilled workers. In software companies, it is the
top talent (or the brilliant programmers) who should be paid the
most for their skills in the firms operating in product markets
with high payoff dispersion.
 The wage regression hypotheses should apply for workers at
different experience levels, and for wage growth rates.
Empirical Specification
16
ln( Wij )  X   Z      ij
'
i
'
j
P
j
i indexes workers and j indexes firms – dependent variable is log quarterly earnings for
a worker observed at some point in employment spell that is ongoing in 1997
(beginning, end, one year prior to end, etc.)
X is a vector of worker controls including quadratics of tenure at job (depending
on when in spell earnings are measured), tenure in industry, and age,
fully interacted with each other and with left and right censoring dummies
Z is a vector of firm controls include a quadratic (log) firm employment, dummies for firm age,
firm growth, and a dummy for whether firm is in a high density, high education and industrially
diversified county. Z also includes log revenue per worker and firm worker turnover. All firm variables
measured in 1997.
Main variable of interest:  j
Product line dispersion measure reflecting actual product mix but not actual revenue
P
Remarks: (i) Focus only on high skilled software workers ($50K+) from age 21-44;
(ii) Revenue per worker control to abstract from rent sharing explanations of findings;
(iii) Worker churning to abstract from risk (of worker turnover)
Tests
17
Do firms operating in software sectors that have high
variance payoffs pay higher wages? Is α>0?
2. How does this affect earnings distribution? Is effect at
90th percentile > 10th percentile?
3. What is contribution of skill vs. effort (i.e. importance
of screening)? Examine starting salaries
4. How do firms structure compensation? Examine
1.
1.
2.
5.
Experienced earnings including stock options (end of spell)
Experienced salaries excluding options and bonuses
Do firms reward loyalty? Examine within firm earnings
growth vs. between firm growth.
1 and 2: α significant at 90th percentile; not at mean
The Relationship between Product Payoff Dispersion and Different Earnings Measures
OLS
Quantile Regressions
10th Percentile
90th Percentile
0.0526
-0.1848
0.2129
(0.0331)
0.1840
(0.0460)***
0.1145
(0.0557)***
0.1406
(0.0563)***
(0.1149)
(0.0722)*
-0.0460
-0.1082
-0.0647
(0.0289)
(0.0390)***
(0.0580)
0.3868
0.0537
0.8279
(0.0629)***
(0.0340)
(0.0990)***
0.1312
-0.1674
0.4983
(0.0518)**
(0.0445)***
(0.1001)***
0.0551
-0.1251
0.3731
(0.0340)
(0.0343)***
(0.0697)***
0.1519
-0.0287
0.5709
(0.0552)***
(0.0346)
(0.1172)***
0.0706
-0.0060
0.1837
(0.0120)***
(0.0073)
(0.0191)***
-0.0025
-0.0136
0.0499
(0.0131)
(0.0092)
(0.0146)***
-0.2169
-0.2352
-0.2597
(0.0476)***
(0.0653)***
(0.0921)***
Dependent Variable
Starting Earnings
Earnings with Previous
Employer
Starting Earnings (with
control for previous
employer)
Experienced Earnings
Lagged Experienced
Earnings
Experienced Salary
Experienced Earnings
(with control for previous
employer)
Within-Job Earnings
Growth
Within Job Salary Growth
Between-Job Earnings
Growth
18
3: Screening important: α significant at 90th percentile;
not at mean; disappears with controls
The Relationship between Product Payoff Dispersion and Different Earnings Measures
OLS
10 Percentile
90th Percentile
0.0526
-0.1848
0.2129
(0.0331)
0.1840
(0.0460)***
0.1145
(0.0557)***
0.1406
(0.0563)***
(0.1149)
(0.0722)*
-0.0460
-0.1082
-0.0647
(0.0289)
(0.0390)***
(0.0580)
0.3868
0.0537
0.8279
(0.0629)***
(0.0340)
(0.0990)***
0.1312
-0.1674
0.4983
(0.0518)**
(0.0445)***
(0.1001)***
0.0551
-0.1251
0.3731
(0.0340)
(0.0343)***
(0.0697)***
0.1519
-0.0287
0.5709
(0.0552)***
(0.0346)
(0.1172)***
0.0706
-0.0060
0.1837
(0.0120)***
(0.0073)
(0.0191)***
-0.0025
-0.0136
0.0499
(0.0131)
(0.0092)
(0.0146)***
-0.2169
-0.2352
-0.2597
(0.0476)***
(0.0653)***
(0.0921)***
Dependent Variable
Starting Earnings
Earnings with Previous
Employer
Starting Earnings (with
control for previous
employer)
Experienced Earnings
Lagged Experienced
Earnings
Experienced Salary
Experienced Earnings
(with control for previous
employer)
Within-Job Earnings
Growth
Within Job Salary Growth
Between-Job Earnings
Growth
Quantile Regressions
th
19
4: How do firms structure compensation? Pay talent
more; both options and salary
The Relationship between Product Payoff Dispersion and Different Earnings Measures
OLS
10 Percentile
90th Percentile
0.0526
-0.1848
0.2129
(0.0331)
0.1840
(0.0460)***
0.1145
(0.0557)***
0.1406
(0.0563)***
(0.1149)
(0.0722)*
-0.0460
-0.1082
-0.0647
(0.0289)
(0.0390)***
(0.0580)
0.3868
0.0537
0.8279
(0.0629)***
(0.0340)
(0.0990)***
0.1312
-0.1674
0.4983
(0.0518)**
(0.0445)***
(0.1001)***
0.0551
-0.1251
0.3731
(0.0340)
(0.0343)***
(0.0697)***
0.1519
-0.0287
0.5709
(0.0552)***
(0.0346)
(0.1172)***
0.0706
-0.0060
0.1837
(0.0120)***
(0.0073)
(0.0191)***
-0.0025
-0.0136
0.0499
(0.0131)
(0.0092)
(0.0146)***
-0.2169
-0.2352
-0.2597
(0.0476)***
(0.0653)***
(0.0921)***
Dependent Variable
Starting Earnings
Earnings with Previous
Employer
Starting Earnings (with
control for previous
employer)
Experienced Earnings
Lagged Experienced
Earnings
Experienced Salary
Experienced Earnings
(with control for previous
employer)
Within-Job Earnings
Growth
Within Job Salary Growth
Between-Job Earnings
Growth
Quantile Regressions
th
20
5: How do firms retain workers? Reward loyalty.
The Relationship between Product Payoff Dispersion and Different Earnings Measures
OLS
10 Percentile
90th Percentile
0.0526
-0.1848
0.2129
(0.0331)
0.1840
(0.0460)***
0.1145
(0.0557)***
0.1406
(0.0563)***
(0.1149)
(0.0722)*
-0.0460
-0.1082
-0.0647
(0.0289)
(0.0390)***
(0.0580)
0.3868
0.0537
0.8279
(0.0629)***
(0.0340)
(0.0990)***
0.1312
-0.1674
0.4983
(0.0518)**
(0.0445)***
(0.1001)***
0.0551
-0.1251
0.3731
(0.0340)
(0.0343)***
(0.0697)***
0.1519
-0.0287
0.5709
(0.0552)***
(0.0346)
(0.1172)***
0.0706
-0.0060
0.1837
(0.0120)***
(0.0073)
(0.0191)***
-0.0025
-0.0136
0.0499
(0.0131)
(0.0092)
(0.0146)***
-0.2169
-0.2352
-0.2597
(0.0476)***
(0.0653)***
(0.0921)***
Dependent Variable
Starting Earnings
Earnings with Previous
Employer
Starting Earnings (with
control for previous
employer)
Experienced Earnings
Lagged Experienced
Earnings
Experienced Salary
Experienced Earnings
(with control for previous
employer)
Within-Job Earnings
Growth
Within Job Salary Growth
Between-Job Earnings
Growth
Quantile Regressions
th
21
Summarizing Results
22
 When firms operate in product markets that have high
payoff dispersion rates, these firms pay higher wages than
do other software firms




Starting salaries are higher in high payoff dispersion
Experienced-worker compensation is much higher in high payoff
dispersion firms
Wage growth is higher
These conclusions are not sensitive to different measures of income,
or to different subsamples of the data – we also test occupation
subsamples
 More highly skilled workers, in higher wage quantiles, earn
the most at high payoff firms)
 Workers are also paid more when the firm succeeds: firms
with high sales per worker pay more
Interpreting the Wage Regression Results
23
Workers are paid for performance:
 highly skilled (or high effort) sort to high payoff dispersion
firms – highly skilled are paid more ex ante at these firms
 when software firms perform well, workers are paid more
ex post
 The link between the firm’s payoff dispersion and wages
suggests that workers are being paid for innovation.
“Loyalty” Pays
24
 Wages rise with tenure, but more important:
 Workers in high payoff-dispersion firms have the highest returns to
tenure within the firm. High payoff-dispersion firms do not pay ‘job
hoppers’ higher wages.
 The most skilled workers have the highest returns to tenure in high
payoff-dispersion firms.
 The raw data shows that the vast majority of wage growth arises from
within the firm, not hopping across firms (Figure 4)
Since most wage growth for workers comes from ‘within’
job wage growth, we conclude that “loyalty pays,” and it
pays the most in high payoff firms.
Figure 3A
Predicted Starting Earnings at Minimum and Maximum
Product Market Payoff Dispersion
25
Figure 3B
Predicted Experienced Earnings at Minimum and Maximum
Product Market Payoff Dispersion
26
Figure 3C
Predicted Experienced Salary at Minimum and Maximum Product Market Payoff
Dispersion
27
Summary
28
 Our general hypothesis is that product market strategy
determines the use of a ‘star’ strategy in pay and
selection:


Firms in higher-risk payoff product lines are more likely to
pay higher wages – hiring or building stars
Workers are paid for loyalty and performance:


Workers achieve much higher wages (and wage increases) by
staying with a firm rather than hopping between firms  loyalty
pays
Workers are paid for performance over time within firms that
might succeed and within firms that do succeed
 Our results suggest that labor demand has risen for
workers who are skilled at innovating: wages are high
for workers in firms that value innovations the most.