Examining Return on Human Capital Investments in the Context of - - PDF document

1

Examining Return on Human Capital Investments in the Context of Offshore IT Workers1

Ravi Bapna*^, Ram Gopal~, Alok Gupta*, Nishtha Langer^, Amit Mehra^ (*University of Minnesota, ^SRITNE, Indian School of Business, ~ University of Connecticut) “What we measure affects what we do” – Joseph Stiglitz, Sep. 14, 2009

• 1. Motivation and Background

In today’s knowledge economy, firms need to continually nurture their human capital to gain lasting competitive advantage. This is especially true for the IT services industry, where the fast pace of technological and process changes necessitate continual rebuilding of technical and

• ther expertise (Lee et al. 1995), and where employee costs and associated productivity are the

major determinants of gross profits. Arguably, human capital is the key tangible and intangible resource for firms in the IT services industry. As of 2007, the Indian IT‐ITeS industry accounts for $71.7B of the global $967B, employing 2.23 million knowledge workers and growing by double digits for the last decade. This makes the study of performance impacts of human capital investments in such industries, the subject of this paper, particularly interesting from a theoretical and a managerial perspective. In this research, we examine whether human capital investments by Indian IT services firms in training their employees improve employee performance and productivity. Our rich employee level panel data permit us to ask whether general and specific training (Becker 1975) vary in influencing employee performance. We also build on the IS literature (Lee et al 1995, Joseph et al 2009) on required skills of IT workers by comparing the differential impact of domain and technical training. Controlling for unobservable employee characteristics and possible selection bias, we find significant positive impact of training on employee performance. However, we find that both general and specific training as well as domain and technical training are substitutes. This suggests that the value of training is conditional upon a focused curricular approach that emphasizes a structured competency development program, as opposed to the widely

• bserved ad hoc approach.

Our work is motivated by the observed trends of training investments by Indian IT firms; industry surveys show that IT firms are increasing investments in training at close to double digits in percentage terms.2 Our primary research motivation is to ask whether these training investments yield any measurable performance benefits to the workers. In the context of IT

• ffshoring, the answer to this question has productivity implications not just for the IT services

firms making these massive‐scale human capital investments, but also, in direct measure, to

1 We thank Anindya Ghose, Il‐Horn Hann, Anjana Susarla, and participants of 2009 SCECR conference for their

helpful comments on early versions of this work. We thank Tan Moorthy of Infosys for motivating us to undertake this study and for providing us with the panel data.

2 Price Waterhouse Coopers survey available at

http://www.pwc.com/extweb/pwcpublications.nsf/docid/2711a28073ec82238525706c001eaec4

2

more favorable outcomes (lower costs and/or improved quality) for IT services consuming firms across the globe. Broadly speaking, because training is costly, the question of measuring returns to investment in employee training has interested human capital researchers, but the lack of suitable data and methodological difficulties have resulted in a paucity of studies that have looked at the returns

• f human capital investments at the employee‐employer level. We bridge this gap in the

literature by using detailed archival training and performance data at the employee level from a leading Indian IT services firm that serves a global clientele. We have details about every firm provided training module taken by a random selection of close to 8000 employees over a three year period as well as detailed performance appraisal ratings for these employees. This panel structure allows us to identify the impact of training in the presence of omitted variable bias (say an employee’s unobserved motivation) as well as selection bias. We do this in two stages modeled along the lines of Verbeek and Nijman (1996)’s extension of the basic Heckman procedure.

• 2. Conceptual Framework and Hypotheses

The conceptual framework of our study is illustrated in Figure 1. Its main features are the linkage between employee training and performance; the breakdown of training into a) general versus specific, and b) technical versus domain, in an overlapping fashion; and the presence of significant identification challenges as exhibited by the bidirectional links between training and performance and observed and unobserved employee characteristics. Figure 1: Conceptual Framework

• 3. Data and Estimation

Research Setting: To empirically validate our hypotheses, we conducted an in‐depth study at a leading IT outsourcing vendor head quartered in India. The company was assessed at CMM level 5 for its stringent quality processes and at People CMM (PCMM) level 5 certification for its commendable Human Resources (HR) practices during the period of study.3 It offers technical, domain, process, project management, and behavioral courses. In addition to training, the firm also has an elaborate performance evaluation process, which includes feedback from team members, peers, and supervisors, etc., leading to unbiased evaluations. Employee ratings are between the scale of 1 to 4, with 1 indicating the highest performance level and 4 the lowest.

3 The certification aims for improving workforce capabilities and thus entails continuous workforce innovation through training,

appraisals, mentoring, and performance alignment with organizational goals (Curtis, Heffley, Miller 2001). Training

General

• Technical
• Domain

Specific

• Process
• Project Management
• Soft Skills

Controls

• Observed Employee Characteristics
• Fixed Effects

Employee Performance

3

We collect detailed training and performance data on 7399 employees between 2005 and 2007. Data and Measurement: In addition to the employee performance data, we have employee demographic data, such as age, gender, and total as well as firm level experience, and whether the employee is a direct or a lateral hire. We also have data on what training an employee took and when they took it. We use (‐1*PerformanceRating) as our dependent variable since a rating

• f 1 is better than a rating of 4 (Espinosa et al., 2007).

Training variables: Our training data included the number and types (domain, technical, etc.)

• f training courses an employee undertook in a year. We sum total number of courses taken in

a year to measure TotalTraining. Since we are interested in not only measuring the overall impact of training but also how different kinds of training affect performance, we needed to sum up the training along these different dimensions, such as domain (DomainTrng) and technical training (TechnicalTrng). We also distinguish between general and specific training (Becker, 1975): Domain and technical courses, such as expertise in technologies like Java or knowledge of say the Sarbanes‐Oxley Act, improved performance, these were the kinds of skills that an employee can use outside of the firm in question, and therefore constitute

• GeneralTrng. In contrast, SpecificTrng included process, project management, and behavioral

courses, because these were firm specific. The process or project management courses, for example, provided knowledge about internal processes or tools; the behavioral courses are also tailored to the firm’s context, and related to the notion of practical intelligence (Wagner and Sternberg 1985, Slaughter et al. 2007). Our other controls include the employee’s age in years (Age), is the employee’s work experience in years at the firm (FirmLevelExp), employee gender (dGender: male (1) or a female (0)), and whether the employee is a direct (1) or a lateral (0) hire (dDirectHire). We centered the relevant variables prior to analysis to alleviate collinearity issues in models using interaction effects, and make the results easier to interpret (Aiken and West 1991) .

• 4. Analysis and Results

We develop a series of models to test our hypotheses. As in all models where individual makes a choice decision – such as undertaking training – we worry about the potential selection bias4. At the same time, we are also concerned about the unobservable individual characteristics which can bias our main model. Attributes such as motivation, drive, and persona are intangible and not directly observable in our data, but likely to be correlated with the errors of any model

4 To test this we created a balanced panel of people who got training for two years and contrasted that with the

unbalanced panel where some individuals got training for none of the two years, some got training or were selected for one year and others got training for both years. Then, as per Verbeek (2001) we ran the identical fixed effects models on the balanced and unbalanced panel. If there was some systematic information in the selection of who got training, the vector of coefficients and the variance‐covariance matrix of the balanced panel and the unbalanced panel should not be significantly different. We calculated the test statistic for this to be 66.45 (> Chi Square = 4.5 with the appropriate degrees of freedom at p=0.05) and found there to be significant difference between the coefficients of the balanced and unbalanced panel. This confirms the existence of selection bias in the sample.

4

that links training with performance. We use a modified version of the Verbeek and Nijman (1996) extension of the basic Heckman procedure. To correct for the selection bias, we use a random effects model; and use a fixed effects model instead of a random effects in the second step of the Verbeek and Nijman model. The random effects model in the selection equation takes advantage of time‐invariant demographics, such as age, gender, whether the employees is a lateral hire or fresh out of college. The fixed effects model in the second step captures the additional unobserved effects which are directly related to the individual. To estimate the inverse mills ratio and correct for the selection bias, we use the following random effects probit model for employee ‘i’ for year ‘t’ :

, 1 ,( 1) 2 , 3 , 4 5

Pr( 1) ( . . . . . ),

i t i t i t i t i i it

gotTraining PerformanceRating Age FirmLevelExp dGender dDirectHire u α α α α α α

−

= = Φ + + + + + +

(1) where

1 if an employee ' ' takes training in year ' ', and , 0 otherwise. i t gotTrainingi t ⎧ = ⎨ ⎩

and where

2

; (0, ), (0,1)

it i it i it

u N N

α

α η α σ η = + ∼ ∼

In the above equation, PerformanceRatingi,(t‐1) is the (transposed) performance rating (‐ 1*PerformanceRating)of the employee for the previous year. Note that the variable dGender helps identify our model. Gender is likely to impact the propensity to take training, but unlikely to impact the performance of an individual. To test for the overall impact of training on employee performance, we test the following fixed effects model: (2) In this equation, TotalTrng i,(t‐1) is the total number of courses that the employee took in the previous year, and InvMills is the inverse mills ratio derived from the selection equation. We estimate this model for the overall sample. To check how the coefficients vary for different employee categories, we estimate the same model for a) direct hires with low experience, b) lateral hires with low experience, c) direct hires with high experience, d) lateral hires with high experience. To test how different types of training affect performance, we estimate the following series

• f fixed effects models. We analyzed our data for overall as well the entire sample as well as for
• ther employee specific characteristics such as experience (e.g., Joseph et al. 2009), and

whether the employee was a direct or a lateral hire (Slaughter et al. 2007).

, 1 ,( 1) 2 ,( 1) 3 , 4 ,( 1) ,( 1)

. . . . .

i t i t i t i t i t i t it

PerformanceRating DomainTrng TechnicalTrng InvMills DomainTrng XTechnicalTrng γ γ γ γ γ η

− − − −

= + + + + +

(3) The above model compares the effect of domain versus technical training. Again, this model is estimated for the entire sample as well as for the different employee categories already

• specified. The results are shown in table 3.

, 1 ,( 1) 2 ,( 1) 3 , 4 ,( 1) ,( 1)

. . . . .

i t i t i t i t i t i t it

PerformanceRating GeneralTrng SpecificTrng InvMills GeneralTrng XSpecificTrng λ λ λ λ λ η

− − − −

= + + + + +

(4) Finally, we compare the impact of general versus specific training on performance. The above model compares the effect of domain versus technical training. Again, this model is

, 1 ,( 1) 2 ,

. . .

i t i t i t it

PerformanceRating TotalTrng InvMills β β β η

−

= + + +

5

estimated for the entire sample as well as for the different employee categories already specified.

• 5. Current (Preliminary) Analysis

We now discuss our preliminary results. The results are shown in tables 1 through 4. Propensity of training (Table 1): We find that while the age of the employee and whether the employee was a direct hire did not exhibit a significant impact on the propensity to undertake training, 1) females employees are more likely to avail of opportunities for training than their male counterparts (α4 : ‐0.09**); 2) employees who have been with the firm longer are more likely to engage in training (α3 = 0.28**), and 3) surprisingly, we find that employees who perform better on the job are more likely to undertake training (α1 = 0.22**). Employees appear to perceive training as a ‘luxury good’ which they can less afford when the performance is sub‐par. The focus is more on immediate ‘on‐the‐job’ efforts to improve performance. Training, whose perceived impact may be not as immediate, is deferred until the performance is improved. Impact of total training (Table 2): We find that Inverse Mills Ratio is significant in all specifications suggesting the existence of selection. The results also consistently reveal a positive and significant impact of training on performance. In order to assess the relative performance impact of training we have estimated the model along the dimensions of work experience and on whether an employee is a direct or a lateral hire. A stark result from this analysis is that training has a much higher impact on performance for lateral hires than it does for direct hires. Lateral hires, who have the benefit of a wider work exposure, seem to be able imbibe training and translate the skill sets learnt from training towards their job responsibilities. This effect is particularly pronounced when examined along the dimension of work experience; the training effect on performance diminishes with experience for direct hires whereas it increases with experience for lateral hires. Impact of domain versus technical training (Table 3): We find that both types of training yield performance benefits, however, domain training is nearly twice as effect in boosting

• performance. The interaction effects between the two types of training are uniformly and

consistently negative suggesting the substitutive nature of these trainings. This is a clear indication that mixing technical and domain in a given year is counter‐productive. Focus appears to be the key to reaping the optimal returns from training. Impact of general versus specific training (Table 4): The results reveal that specific training has no significant impact on performance, and this finding remains consistent regardless of how the data was partitioned. A logical conclusion from the finding is that, while firm specific skills may indeed be an important driver of performance, formal training courses to impart these skills are not productive. Other avenues such as mentoring, networking and on‐the‐job learning may feed these skills and with a little further value addition coming from formal training

• programs. As general skills are the summation of technical and domain skills, the results

pertaining to general skills are consistent with the previous discussion.

6

In summary, our analysis indicates that an additional training module leads to a significant increase of 0.064 points on a 1‐4 employees appraisal scale on average. We believe this effect to be economically significant, given that the employees, on average, takes 1.1 course per annum, and given that their appraisal reflects a broad assessment of their activities throughout the year. The effect is almost double in magnitude, on average, for laterals suggesting significant differences in the ability of the lateral hires to translate their learning into firm valued performance. Our analysis takes care of possible selection bias as well as unobserved employee characteristics that could otherwise confound any linkage between training and

• performance. To the best of our knowledge this research is the first empirical examination of

the predictions of an IT services contextualized human capital theory. The primary managerial implication of this research is that provides guidance to the rapidly growing offshore IT services industry about the differential impact of the various types of training investments. Our analysis also indicates that while training overall has a significant positive influence, firms should go about offering these modules to employees in a focused and systematic manner. We find that both general and specific training, as well as domain and technical training, interact with each

• ther in a significantly negative way. This should serve as a warning to firms that without a

focused effort some of their investments can have a cancelling effect on other investments. Given that software development is a complex group task, we expect future work to examine the impact of training on team and project performance as well as client satisfaction. More detailed information on employee‐project characteristics will also facilitate the examination of the relative tradeoff between on the job learning and formal training. This falls into the unobserved time variant component of the error term in the present analysis. We expect that our full and final analysis will result in findings have both theoretical and practical significance, most important of which is that they justify increased human capital investments to fuel future growth of this important component of the global economy. We look forward to presenting the full set of results at WISE 2009! Table 1: Selection Equation5

GotTraining Coefficient (Std. Error) PerformanceRating (t‐ 1) 0.217 (0.017)** Age ‐0.002 (0.007) FirmLevelExp 0.284 (0.011)** dGender ‐0.086 (0.028)** dDirectHire ‐0.01 (0.027) Table 2: Impact of training on performance PerformanceRating Overall Low Exp High Exp Direct Hire Lateral Hire Direct Hire Lateral Hire TotalTrng (t‐1) 0.064 (0.008)** 0.044 (0.016)** 0.065 (0.023)** 0.027 (0.015)+ 0.098 (0.016)**

5 Note for all estimation tables: + Significant at 10%; * significant at 5%; ** significant at 1%.

7

PerformanceRating Overall Low Exp High Exp Direct Hire Lateral Hire Direct Hire Lateral Hire invmills 0.4 (0.051)** 0.913 (0.088)** 0.652 (0.116)** ‐0.291 (0.123)* 0.148 (0.09) Table 3: Impact of domain vs. technical training on performance PerformanceRating Overall + Interactions Low Exp High Exp Direct Hire Lateral Hire Direct Hire Lateral Hire DomainTrng (t‐1) 0.087 (0.014)** 0.045 (0.012)** 0.107 (0.028)** 0.03 (0.044) 0.06 (0.024)* 0.151 (0.029)** TechTrng (t‐1) 0.049 (0.021)* 0.036 (0.022)+ 0.048 (0.047) 0.227 (0.059)** 0.12 (0.06)* 0.12 (0.046)** InvMills 0.412 (0.051)** 0 (0.029) 0.974 (0.09)** 0.686 (0.119)** ‐0.183 (0.126) 0.197 (0.095)* DomainTrng (t‐1) X TechTrng (t‐1) ‐0.041 (0.026) ‐0.126 (0.061)* ‐0.148 (0.08)+ ‐0.147 (0.059)* ‐0.113 (0.056)* Table 4: Impact of general vs specific training on performance PerformanceRating Overall + Interactions Low Exp High Exp Direct Hire Lateral Hire Direct Hire Lateral Hire GeneralTrng (t‐1) 0.072 (0.009)** 0.031 (0.008)** 0.052 (0.017)** 0.069 (0.024)** 0.047 (0.017)** 0.099 (0.017)** SpecificTrng (t‐1) 0.011 (0.029) ‐0.059 (0.036)+ ‐0.089 (0.114) ‐0.208 (0.151) ‐0.012 (0.061) 0.063 (0.08) InvMills 0.413 (0.051)** ‐0.012 (0.029) 0.923 (0.089)** 0.64 (0.116)** ‐0.241 (0.125)+ 0.146 (0.092) GeneralTrng (t‐1) X SpecificTrng (t‐1) 0.024 (0.026) 0.04 (0.081) 0.199 (0.109)+ ‐0.044 (0.043) 0.025 (0.058)

References are available at http://bit.ly/UORic