Technology, job characteristics, and retirement of aged workers: evidence from automation and IT adoption of firms in Korea

Abstract

We investigated how the adoption of a new production technology differently affects the risk of job separation of young and old employees in South Korea by analyzing establishment-level panel data linked with administrative employment insurance records on individual workers. To address potential endogeneity associated with a firm’s technology adoption, we conducted instrumental variable estimations with a two-stage residual inclusion (2SRI) approach. The results suggest that technology (indicated by newly adopted automation and increased purchase of Information Technology equipment) positively affects the overall employment of incumbent workers. However, the employment of aged workers is less favorably affected by newly adopted technologies compared to that of younger workers. In some conditions, technology adoptions increase the retirement risk of older workers absolutely as well as relative to that of younger workers. Newly adopted automation negatively affects the employment of aged workers who are engaged in clerical and support occupations or employed in the wholesale and retail industry. Estimation results according to the reason for retirement suggest that the negative effect of technology adoption on employment may be related to both labor demand and supply.

1. Introduction

How will new advances in technology, often symbolized by artificial intelligence (AI) and automation in production, change the labor market in the future? A few studies have investigated and predicted the labor market consequences of the technological changes of the Fourth Industrial Revolution. Debates are ongoing on whether newly invented machines, such as robot adoption, will radically replace human labor (Graetz and Michaels, 2018; Acemoglu and Restrepo, 2020) and on what kinds of jobs will be vulnerable to the effects of technological change (Autor et al., 2003; Goos et al., 2014; Autor, 2015; Fonseca et al., 2018).

Growing evidence suggests that the labor market consequences of technological changes, if any, will probably be heterogeneous across jobs with disparate human capital requirements and workplace characteristics. Given that aged workers tend to have more obsolete skills, be less efficient in learning, and be less mobile across jobs compared to younger workers, their labor market status could be affected differently by technological advances in production and managerial practices. As for the mechanism, which will be covered later in detail, technological changes can affect the employment of older workers by changing their relative productivity and by altering the quality of the matching between the workers and their jobs. However, only a few studies have empirically investigated how technological changes affect young and old workers differently (Bartel and Sicherman, 1993; Friedberg, 2003; Aubert et al., 2006; Lee, 2015), compared with the attention given to the differences across workers with disparate human capital or skills. No evidence regarding the issue has been suggested for Korea.

The primary purpose of this study is to investigate how the adoption of new production technology¹ affects the employment of older South Korean workers by using establishment-level panel data that were newly linked with administrative records. The workplace panel surveys (WPSs), conducted by the Korea Labor Institute since 2005, provide detailed information on each establishment, including variables pertaining to workplace innovations in production, organization, and human resource management. The Korean Employment Insurance records, matched with WPS data, offer information on labor market transitions and personal characteristics of the individuals employed in the workplaces included in WPSs.

Using the data, we investigated how indices of new technology adoption affected the risk of employees leaving their jobs and how the effects differed between older and younger workers. For constructing variables on technology adoption, we utilized responses to the following two questions: (i) whether new automation was adopted by the firm and (ii) how much IT equipment purchases increased. These changes may substitute jobs or take over tasks that workers used to perform.

We conducted survival analyses (Cox proportional hazard model) that included each of the indices of technology adoption and its interaction with a variable indicating if the worker is an aged person, along with variables pertaining to personal and job characteristics. The decisions to adopt new technology, however, could be endogenously influenced by unobservable factors related to the retirement risk of aged and younger workers. We addressed this problem by relying on an instrumental variable estimation using a set of instruments: (i) small group activities to discuss quality improvement, productivity enhancement, cost savings, and the resolution of customer complaints and (ii) a code of conduct regarding internet use. We employed the two-stage residual inclusion (2SRI) approach, a preferred method to obtain a consistent estimator when using instruments in nonlinear models (Terza et al., 2008).

The paper is organized as follows: in Section 2, we discuss the conceptual framework of this paper and review previous studies on the subject; in Section 3, we introduce the data and empirical strategy used in the study; Section 4 presents the results of baseline and additional regression analyses; and the final section concludes the study.

2. Conceptual framework and related literature

As life expectancy rapidly increases before the pension system fully matures, the labor force participation of the elderly in Korea is relatively high among the Organization for Economic Cooperation and Development countries.² Traditionally, Korean firms tended to implement seniority-based wage schemes where wage grows along with job tenure. Since firms have to be overburdened when employees’ job tenure increases, they also implemented the early retirement system with higher allowance or set the retirement age lower than the pensionable ages.³ Reflecting the elderly’s hopes for working longer, however, the National Assembly of Korea passed a bill on the minimum retirement age in 2013. According to the new bill, workplaces were required to set the retirement age at least 60 from the beginning of 2016. Hence, firms in Korea have more incentive to reduce elderly labor than before: a wage-peak system (a salary cap not to increase along with job tenure) for older workers near the retirement age has become spread recently, but the goal also can be realized by technological progress.

The adoption of new production technology can affect the overall employment of workers through several different pathways. Previous studies investigated the relationship by employing a lens of innovation (Harrison et al., 2014; Calvino, 2019; Crespi et al., 2019). Specifically, they suggested that a classification between product and process innovation is helpful to unravel the impact of innovation: the former means new or significantly improved goods or services released and the latter includes installing new equipment/software or a change in the method of production to save cost or increase quality. Since measures of firms’ technology adoption can be indicators of process innovation (Hall et al., 2008), the purpose of our study is close to finding the employment effect of process innovation.

Researchers have identified two distinct labor market effects of process innovation, namely, displacement effect and compensation effect (Harrison et al., 2014). First, process innovation is directly connected with negative employment effects because it typically aims to labor-saving (Calvino, 2019; Crespi et al., 2019).⁴ Hence, technological progress can diminish labor demand and salary equilibrium by taking over tasks previously conducted by human labor (displacement effect). Second, process innovations vary in their own indirect effects depending on the extent to which reduced labor can be compensated by improved productivity or demand (Vivarelli, 2014). The adoption of a new technology may increase the demand for labor by improving labor productivity. Also, technological progress from the adoption can create new tasks and, if labor has comparative advantages in the newly created tasks, increase demand for labor (compensation effect).

The overall employment effect of technology adoption would be determined by the balance between these two effects. Harrison et al. (2014), analyzing 20,000 European firms, observed reduced employment due to the displacement effect of innovation, but they showed that the overall employment increases by the compensation effect from enhanced productivity offsetting the negative effect. Zimmermann (2009), examining the relationship between the employment growth rate and innovation by using small and middle firms in Germany, also found a positive employment impact of process innovation.

On the contrary, empirical evidence suggesting that the displacement effect dominates the compensation effect has also been provided. Prettner (2019) found that increased automation explains 14% of labor share reduction in the United States during the 1970–2016 period. He also reported that automation might accelerate the inequality between capital owners and job-losing workers over the last decades. Using cross-country data covering four decades (from 1970 to 2007), Autor and Salomons (2018) examined the impact of automation (industry level) on employment and labor’s share of value added in total productivity. They found that automation reallocates employees across industries and has a negative impact on labor’s share of productivity value for decades. Using variations in robot exposure across US industries between 1990 and 2007, Acemoglu and Restrepo (2020) examined the impact of industrial robots on local labor markets. They found that robot adoption lowered both the employment-to-population ratio and average wages across the commuting zones. Graetz and Michaels (2018) examined the economic contributions of robot adoption using industry- and country-level panel data. They found that robot adoption did not significantly affect labor hours and total employment, whereas it reduced the employment of low-skilled workers.

As the labor force is aging over time, the relationship between aging and new technology has received increasing attention. Using the US commuting zone data, Acemoglu et al. (2022) examined how robot adoption affected the employment of middle-aged and older workers. They found that robot adoption reduced the employment and earnings of middle-aged workers but had no effect on those of older workers. The negative employment effect found among middle-aged workers was attributed to their greater concentration in blue-collar jobs that can be more easily automated by robots. The study also found that countries experiencing rapid aging are more likely to invest in robots. Their estimates suggested that aging explains 40% to 65% of the cross-country variations in robot adoption.

Even within the same industry, the employment effect of technological progress may differ by the age of workers. Any technology adoption with disparate effects on productivity depending on the age of workers would affect young and aged employees differently. If the tasks of old workers can be automated more easily, for example, a newly adopted technology would be associated with a relative decrease in demand for the elderly. As marginal workers in the labor market, aged people may likely be vulnerable to radical economic changes, such as the emergence of new technology. As noted by Lankisch et al. (2019), the potential disadvantages associated with aging may stem from the lower level of skills possessed by the elderly compared with the young.

Even with the same quality of human capital, the labor market effects of technological changes could be more strongly felt among the elderly than the young. For example, new technology could affect the employment of aged workers by deteriorating the quality of matching between workers and their jobs. Technological progress is often associated with radical changes in job requirements and working conditions. Adoption of new technology can make it increasingly difficult for aged workers to continue working as the speed and intensity of work, as well as the requirements for skills, increase, possibly beyond their physical and mental capacities. Given their deteriorated physical strength and health, obsolete skills and knowledge, and lack of education compared with young cohorts, aged workers have lower capabilities (or incentives) to learn to meet new work requirements. A return to training generally decreases with age; thus, employers would be unwilling to invest in the training of aged workers, thereby increasing the severity of their disadvantages.

Only a few studies have empirically investigated how technological changes affect young and old workers differently, compared with the attention given to the differences across workers with disparate human capital or skills. Hurd (1996) demonstrated that labor rigidities, such as the inability to change working hours, could become a greater burden to old workers as their preference shifts from work to leisure. According to Lee (2015), technological changes in the course of the “Second Industrial Revolution” during the early 20th century increased the intensity of work and diminished job flexibility. Lee suggested that these changes negatively affected the employment of aged US manufacturing workers. Lee and Lee (2013) found that older workers engaged in a more flexible job show a lower probability of retirement than others when other conditions are equal.

Acquiring a new skill to adjust to a technological change could be more difficult for older workers. Bartel and Sicherman (1993) demonstrated that unexpected changes in the rate of technological changes provoked disutility for working and induced older workers to retire earlier. To examine these effects empirically, several studies have examined the impact of IT innovations on the employment of older workers. Friedberg (2003), using the US data, found that the spread of computers in a workplace had a negative impact on the employment of old workers, but only for workers close to the timing of their mandatory retirement. Aubert et al. (2006) offered firm-level evidence that the use of computers and the internet tended to reduce the wage bill share of older workers. Peng et al. (2017), using a dataset of European countries from the 1970s to the 2000s, suggested that the impact of IT on labor demand is skill-biased; it reduces wage shares of low-skilled workers. Jerbashian (2019) showed that a reduction in the price of IT induced an increase in the share of employment in high-wage occupations but a decrease in the share of middle-wage occupations.

In this study, we attempted to investigate how the adoption of new technology affects the overall employment and the relative employment of the young and the old. More specifically, we analyzed the technology adoption variables and their interaction with the variable for older workers. If the conceptual framework discussed earlier is applied, the estimated coefficient for the interaction term captures how a newly adopted technology alters the productive efficiency and the quality of job matching of older workers relative to those of younger ones, whereas the coefficient for the technology variable shows how it changes the overall productivity of all employees in the establishment.

3. Data and empirical strategy

3.1 Data

It is difficult to obtain data containing information on both technological changes adopted by firms and the employment of individual workers in the firms. Thanks to the cooperation and support of the Korea Labor Institute, we obtained and used a unique dataset that was produced by linking firm-level panel data with administrative employment records for the individuals employed in the firms. More specifically, the following micro datasets were linked and used: the WPSs and the Korean Employment Insurance records.

The WPSs, conducted by the Korea Labor Institute since 2005, provide detailed information on each establishment included in the survey, such as variables pertaining to workplace innovations in production, organization, and human resource management. As WPSs started to ask questions regarding new technologies adopted by each firm since 2015, we largely use the 2015 survey, which includes 3431 firms.

The Korean Employment Insurance records matched with WPSs offer information on labor market transitions and personal characteristics of the individuals who were employed in the workplaces included in the WPSs. The significant advantage of the linked data is that it includes all workers employed at least once in these firms, which allows us to follow up on labor market changes of the individuals, including the exact timing and reason for job separations. A drawback of the linked data is that it provides only a limited set of variables regarding the information on each individual, especially his or her job characteristics (e.g., full/part-time job).

We restricted our sample to individuals who met the following conditions: (i) employed in the WPS firms at the beginning of 2015, (ii) aged 25–69 in 2015, and (iii) working in a firm with more than 10 employees.⁵ As a consequence of the selection, we ended up with a sample of 962,404 persons employed in 3033 firms, and the size of each birth cohort ranged from 13,734 (those aged 25) to 1320 (those aged 69). We classified individuals aged 50 and older in 2015 as aged workers in the baseline analysis. In the additional analysis, we use various age cutoffs for defining older workers as a robustness check.

3.2 Empirical strategy

The 2015 WPS includes questions regarding newly adopted technologies. The following indices of technology adoption were considered in the present study: (i) newly adopted automation and (ii) the increased purchase of IT-related equipment. New automation refers to the circumstance that any process or part of work for a product or service is newly automated and is considered along with the extent of automation already completed. The measure of IT-related equipment is obtained from responses using a five-point Likert scale (“not at all,” “not much,” “neutral,” “somewhat,” and “very much do”). We constructed a dummy variable that has a value of 1 if a firm selected “somewhat” or “very much do.”

We investigated how the adoption of new technologies in a firm affects the probability that an employee will leave the firm. The 2015 WPS was linked to the employment insurance records from 2015 to 2017, which allowed us to follow up on individuals employed in the firm in 2015 for 2 years. Taking advantage of the longitudinal features of the data, we examined the effect of technology adoption on the retirement hazard by estimating a Cox proportional hazard model. Motivated by the fact that certain types of retirement, such as disciplinary dismissal, could be irrelevant for this study, we limited the definition of retirement based on the reason for retirement. The Korean Employment Insurance provides reasons for job separation as follows: (i) voluntary retirement due to a personal reason; (ii) voluntary retirement due to a relocation of the workplace, a change in working condition, or wage arrears; (iii) going out of business; (iv) inevitable managerial issues; (v) disciplinary dismissal; (vi) mandatory retirement; and (vii) termination of a contract. We assumed that technology adoption might induce workers to retire through the following two channels. First, firms would lay off workers as a result of adopting new technologies. Second, workers would quit their jobs because of a deteriorated quality of matching with the newly assigned task. Accordingly, we chose to consider two types of retirement in the analysis, namely, inevitable managerial issues (i.e., a forced layoff) and voluntary retirement due to a personal reason or a change in working conditions.⁶ The retirement risk for workers is specified as follows:

In equation (1), |${\lambda _0}\left( t \right)$| denotes the baseline retirement risk of workers at time |$t$|⁠. Subscripts i and j denote individual and firm, respectively. T stands for the dummy variable for technology adoption. A indicates a dummy variable for aged workers (aged 50 and older). X denotes a set of a matrix that includes each worker’s personal characteristics, such as gender, occupation, job tenure, and being subjected to extending mandatory retirement.⁷  F indicates variables pertaining to firm characteristics, including size (the number of employees) and industry. These variables are expected to be associated with labor market conditions, institutional features, and work environments that could affect decisions on job separation.

The variable of primary interest to us is the interaction between variables for technology adoption and aged workers. The estimated coefficient for the interaction term (⁠|${\beta _1})$|⁠) is expected to capture how a newly adopted technology changes the productive efficiency and the quality of job matching of older workers relative to those of younger ones. If the sign of the parameter is estimated to be positive (negative), it indicates that the employment effect of the technology adoption is more unfavorable (favorable) for older workers as compared to the effect for younger employees. The coefficient for the technology variable (⁠|${\beta _2})$|⁠) shows how the adoption changes the overall productivity of the employees in the establishment. If the estimated parameter is positive (negative), it tells us that the overall effect of the technology adoption is unfavorable (favorable) for the employment of incumbent workers. The sum of the two coefficients (⁠|${\beta _1} + {\beta _2}$|⁠) shows the overall employment effect for older workers.⁸ As explained in Section 2, they indicate the net effect of both displacement and compensation effects of technology adoption (i.e., process innovation) combined.

Assessing the direction of causality between technology adoption and labor has been a challenge in previous studies on the topic, as new technology adoption is potentially endogenous. Suppose our model includes unobserved confounders, |${U_j}$|⁠:

If |${T_j}$| is correlated with |${U_j}$|⁠, then the estimation from equation (1) would result in a biased inference.

To address this issue, we employed the 2SRI approach, which was suggested first by Hausman (1978). The basic idea of the 2SRI approach is similar to that of the two-stage predictor substitution (2SPS), as an endogenous variable is regressed on exogenous variables and instruments in the first stage:

where |${Z_j}$| stands for a set of instrumental variables. Since technology adoption is determined by a firm, the first-stage estimation is conducted at the firm level (⁠|${{\rm A}_j}$| is the proportion of older workers in |$j$|⁠). Unlike 2SPS, where the endogenous variable is replaced by predictors, the 2SRI approach includes residuals |$u_j^1$| and |$u_j^2$| as additional regressors in the second stage.

2SRI gives equivalent results to the 2SPS in a linear setting. In a nonlinear setting, however, Terza et al. (2008) showed that only the parameters estimated from 2SRI are consistent and that those from 2SPS are not. Accordingly, nonlinear estimation with the 2SRI approach has been widely employed (Guo et al., 2015; Yeung, 2017; Cheng, 2018; Lazuka, 2018).

For each of the endogenous variables, we used a different instrument. First, we used the presence of small group activities for product/service innovation to assess the adoption of new automation. The activity discusses quality improvement, productivity enhancement, cost savings, or the resolution of customer complaints to assess. In WPS 2015, about 43% of firms responded that the activities were put into practice, and 80% of them consisted of employees. It is likely that the proportion of employees participating in such activities is positively related to the probability of adopting new technology. However, such a bottom-up organization would be unlikely to have a direct effect on employment. Second, we used a code of conduct regarding internet use to measure investment in IT-related equipment. The dependence on IT devices in a firm’s working environment would raise security or communication issues and increase the need to make rules to prevent potential problems. Such rules, however, are not likely to be strongly correlated with unobservable firm characteristics influencing the employment of workers.

Besides the separation from employment we mainly focus on, technology adoption may affect the influx of employment. Since survival analysis cannot measure the impact of technology on employment in relation to new hiring, we employed another empirical strategy examining the employment and hiring effect of technology at the firm level. Detailed descriptions and results are reported in  Appendix.

4. Technology adoption and risk of job separation by age: results

4.1 Summary statistics

Table 1 presents the sample means of the variables used in this study, including demographic characteristics, labor market behaviors, workers’ occupations, technology adoptions, and features of establishments. Column 1 provides the statistics for the full sample, and columns 2 and 3 compare those for older and young workers. Columns 4 and 5 report sample means for males and females, and the rest of the columns show how those variables differ between industries.

The first column shows that 63.6% of the entire sample consists of male workers, and aged workers (aged 50 and older) account for 21.4% of the full sample. The average length of tenure is 7.1 years, and 14.6% of workers employed in WPS firms at the beginning of 2015 (13.3% of males and 16.7% of females) retired from the job by the end of 2017. The proportion of workers who left the firms by the end of 2017 is slightly lower for aged workers (13.3%) than for younger workers (14.9%). Among the retired, 23% of them retired due to inevitable managerial issues such as forced layoff and 77% voluntarily retired for personal reasons. For convenience, we abbreviate the term “involuntary retirement due to inevitable managerial issues” as “involuntary retirement” and “voluntary retirement due to a personal reason or a change in working conditions” as “voluntary retirement.”

The linked dataset provides information on the occupations of individual workers on the basis of the International Standard Classification of Occupations 2008 (ISCO-08). The nine classifications represent high-skilled (senior officials and managers, professionals, and technicians and associate professionals), medium-skilled (clerical support workers, service and sales workers, skilled agricultural, forestry, and fishery workers, craft and related trades workers, and plant and machine operators and assemblers), and low-skilled (elementary occupations such as cleaners and helpers, food preparation assistants, and agricultural laborers) jobs. Excluding agricultural, forestry, and fishery workers, workers are relatively evenly distributed across occupations.

Among the sample, 21.3% of individuals are in firms where any process or part of the main product/service was newly automated. In small group activities for innovation, the corresponding instrument consists of 43.3% of the sample. The proportion of workers whose firms expanded investments in IT-related equipment is 23.5%. Also, 33.2% of workers are engaged in firms implementing the code of conduct regarding internet use.

Differences in statistics are observed depending on age and gender. Older workers tended to be engaged in low-skilled jobs compared to others. They were also more likely to retire due to involuntary reasons. Workers under 50 years old had a high proportion of working as professionals and clerical workers. The statistics also suggest that retirement among young workers was attributed to the supply (voluntary) side more than the demand (involuntary) side of the labor market.

While male workers were relatively evenly distributed across the occupations, more than half of females were engaged in service and sales occupations. Also, nearly half of female workers worked in the wholesale and retail trade industry, which is known to have vulnerable jobs in Korea. In the same vein, a higher proportion of female workers were employed in small-sized firms, indicating their job security may be lower than male workers.

In terms of industry, the differences between the manufacturing and service sectors are apparent in many aspects. As expected, workers in the manufacturing industry were more likely to experience newly adopted automation. The production process would be closely related to robots and other automation systems Although the firm size was larger, employees were younger, and the retirement rate was lower, the proportion of low-skilled workers in the manufacturing industry was higher compared to the service industries.

Service industries account for 32% of the full sample.⁹ Although the use of IT is widespread across the industries, workers in the service sector were probably more accustomed to the use of the internet and IT devices. The mean value of “expanded investment in IT-related equipment” is similar to the value of the full sample, but firms belonging to the “information and communication,” “financial and insurance activities,” and “professional, scientific, and technical activities” areas of the service sector show very high rates of increasing expenditure on IT-related equipment.

Although the wholesale and retail industry is the sub-industry of service sectors, we report them separately because it is mainly representative of employing low-skilled female workers and we also observed large differences in the estimation results. Due to the nature of the industry, the concentration of workers in service and sales occupations is severe. They also show the highest ratio of voluntary retirement as well as the lowest job tenure among the three groups, suggesting that their job conditions were not satisfactory.

The WPS surveyed human resources officers regarding how firms adjusted employment and tasks in response to technology adoption. Figure 1 shows how firms adopting new automation changed the number and allocation of their employees. The result shows that nearly 60% of the firms maintained the majority of existing workers in the same tasks. Only about 10% responded that they reduced the number of employees. Similarly, Figure 2 shows how increased investment in IT affected employment in the firms. Approximately 10% of firms responded that they reduced the number of employees as a result of making additional investments in IT over the past 2 years. These results suggest that the overall displacement effect of technology adoption, if any, was probably small in magnitude.

Newly introduced technology often requires the reallocation of workers across different tasks. Figure 1 shows that more than 30% of firms adopting automation reallocated their employees to new tasks. As noted in Section 2, such changes in tasks could deteriorate the quality of matching between the job and the worker, especially for aged workers. Moreover, it would be costlier for workers with a low learning capability (such as older workers) to be reassigned to a new task. Consistent with these conjectures, Figure 3 shows a negative relationship between the proportion of older workers and the two indices of technology adoption, namely, automation and IT device use. The result may indicate the greater difficulty of innovation among firms with a higher proportion of older workers or/and higher rates of retirement of older workers in firms adopting new technology.

4.2 Baseline results

We estimated a Cox proportional hazard model to determine the factors of retirement risk to examine how the adoption of new technology differently affected the employment of younger and older workers. Table 2 presents the baseline results based on using two different measures of technology adoption (newly adopted automation and increased purchase of IT-related equipment). The main independent variables include the adoption of new technology, aged workers, and the interaction between the two. In addition, variables on the employee’s gender, occupation, job tenure, and being subjected to the mandatory retirement reform are controlled. To take into account the unobservable heterogeneity across firms, we also included variables on firm size (categorical), industry fixed effect, and occupation.

The results shown in Table 2 suggest that technology adoptions, measured by newly introduced automation and the increased purchase of IT-related equipment, reduce the retirement hazard of incumbent workers. The estimated hazard ratios of retirement for the two technology variables are significantly less than 1, indicating that technology adoption tends to lower the risk of retirement for workers who had been employed at the firm prior to the change.

More significantly, the estimated hazard ratios for the interaction term (of which the magnitude is greater than 1) suggest that the employment effect of technology adoption differs between aged and younger workers, with the former less favorably affected by newly adopted technologies than the latter. In columns 1 and 2, conditional on adopting new production technology, aged workers were more likely to retire than younger workers by 20.3% (new automation) or 24.4% (investment in IT equipment). However, the overall effect of technological change on aged workers’ employment is actually positive (reducing the hazard of job separation absolutely), indicated by the multiplication result of the two hazard ratios (those for technology adoption and its interaction with that for elderly workers) being less than 1.

The results for the other independent variables are expected ones. As descriptive statistics revealed in Table 1, older workers are positively associated with the hazard of involuntarily leaving the firm due to the inevitable necessity of management. The length of job tenure and being subject to extending mandatory retirement are negatively related to the risk of retirement.

In addition to the Cox proportional hazard model, we estimated 2SRI models with instrumental variables. Since there are two endogenous variables of interest (each index of technology adoption and its interaction with a variable for the aged worker), two Ordinary Least Squares regressions were conducted separately in the first stage. All firm-specific variables considered in the baseline analysis, as well as the instrumental variables, were included in the model. The results of the first-stage regressions, shown in Table 3, reveal a positive relationship between the two endogenous variables and corresponding instruments across columns. The presence of small group activities for innovation increases the probability that the firm adopts a new automation system by 16.5 percentage points (column 1), and its interaction with older workers also increases the corresponded dependent variable by 8.9 percentage points (column 2). Similarly, having a code of conduct for internet use increases the probability of increasing the purchase of IT-related equipment by 14.5 percentage points (column 3). The interaction term with aged workers also increases the corresponded dependent variable by 5.6 percentage points (column 4). The F statistics for weak instrument tests are all greater than 50, confirming that instrumental variables are not weak by the standard of the Stock–Yogo criteria (Stock and Yogo, 2002).

Table 4 reports the estimation results of 2SRI, where the predicted residuals are included in the proportional hazard model. The results are considerably different in terms of technology adoption effect on all incumbent workers. The impact of automation on retirement slightly diminished (column 1). Instead, the positive effect of investment in IT equipment becomes slightly stronger: it changes from 0.790 to 0.285 (column 2), suggesting the unobserved confounder absorbs the hazard of retirement.

The estimated hazards of its interaction term with aged workers are similar to those obtained from the baseline model, with a disadvantage for aged workers compared to the young. However, the overall effect of technology adoption on older workers’ risk of retirement significantly changes due to the technology adoption effect itself. In column 2, for example, the overall effect of IT equipment purchase on aged workers’ retirement decreases from 0.98 (1.244 × 0.790) to 0.31 (1.076 × 0.285) when conducting the 2SRI model, which indicates that the proportional hazard model overestimates the negative effect.

Given that the predicted residuals are significant in all specifications, it is likely that endogeneity biases exist. The direction and magnitudes of such biases depend on the technology adoption and the type of retirement. Regardless of the bias, however, the difference in the technology adoption effect on old and young workers may also suggest that the displacement effect of the new technology would be stronger, or the compensation effect is weaker for elderly workers. Either way, the results show that technological progress may bring unfavorable labor market consequences for older people compared with the young.

4.3 Sensitivity to choice of age cutoff

Another possible concern arises concerning whether or not the results are sensitive to the cutoff age for defining “aged worker,” which was set to 50 years old in the baseline specification. We conducted a sensitivity test that examines how the regression results change with alternative cutoff ages for aged workers ranging from 45 to 55. Figures 4 and 5 plot the estimated hazard ratios for the interaction between older workers and each of the technology adoption indices. In general, the results suggest that the interaction between aging and the influences of technology adoption could differ by the type (or feature) of the technology that is put into practice.

Regarding the adopted automation (Figure 4), the estimated hazard ratio for the interaction continues to increase with age from the mid-40s, and the trend declines at the age of 53. The result could be explained by a particularly strong displacement effect among older workers, which is slightly different from Acemoglu et al. (2022), who found that automation technologies tended to substitute for tasks performed largely by middle-aged workers. A possible explanation for this finding is that automation could create a need for task reallocation (see Figure 1). As noted in Section 2, transferring to a new task would be more costly for elderly workers. The reduction of the magnitude after 53 years old could also be attributed to the official retirement age that older workers reach simultaneously as the technology is introduced. While reduced, the coefficient remains greater than 1 until the cutoff age reaches 55.

In terms of expanded purchase of IT-related equipment (Figure 5), the estimated hazard ratio for the interaction term does not change much with the choice of cutoff age. It remains at around 1.0 and 1.1 across the different cutoff ages. The outcome confirms the baseline result that the employment effect of technological change is less favorable for older workers compared to the young. The result also suggests that the disadvantages associated with aging in terms of coping with IT perhaps start from middle age. Typical employees in the IT industry are relatively young, and the speed of IT development is faster there than in any other industry. For this reason, employees in their mid- to late-40s in the industry could face difficulties in catching up on the latest state-of-the-art technology, such as transitions to new programming languages.

4.4 Heterogeneity

Reflecting the previous studies that the impact of technology on the labor market is skill-biased, it is necessary to confirm the heterogeneity of our estimation results among occupations. Tables 5–7 report estimation results of three occupations of the high-, medium-, and (the other) medium- and low-skilled jobs, respectively. For high-skilled labor (Table 5), technologies generally lowered their retirement hazard. While the effect on retirement is relatively unfavorable to old workers, the overall effect regarding technology and its interaction term is less than 1.

Table 6 reports the retirement hazard of medium-skilled workers. Since the sample size of skilled agricultural, forestry, and fishery workers are not sufficient, we focused on the first two occupations: “clerical support workers” and “service and sales workers.” Table 6 shows that automation significantly increased the hazard of clerical workers’ retirement hazard. The impact is two times higher for old clerical workers compared to young workers. Investment in IT-related equipment also increases the retirement hazard of service and sales workers. The results are consistent with the previous studies showing that recent technology is more harmful to the survival of routinized jobs (Goos et al., 2014; Autor, 2015), suggesting the displacement effect overweighs the compensation (productivity and reinstatement) effect.

Table 7 reports the estimation results for the other two medium-skilled occupations and elementary occupations, the low-skilled jobs. Compared to the first two medium-skilled workers, craft and related workers, plant and machine operators, and assemblers faced more positive employment effects of technology adoption. Technology adoption also lowers the retirement hazard of low-skilled workers, but its negative effect on older workers remains.

We also examined whether the effect of technology adoption differs across industries by comparing the results of similar regressions conducted separately for workers employed in the manufacturing, service, and wholesale and retail industries. Table 8 reports the comparison results. In manufacturing, similar to the results for the entire sample, older workers were at a higher risk of leaving the firm than younger workers when new technology was adopted. Technologies (automation and purchase of IT equipment) reduce the overall retirement hazard. However, the overall employment effects of technological change among older manufacturing workers are more harmful than those observed for the entire sample: automation increases the probability of retirement for older manufacturing workers. In service industries, too, technology adoption reduced older workers’ retirement hazards less than the young workers’. Automation increases the retirement hazard of aged workers 1.095 times. While newly adopted IT-related equipment reduces the overall probability of job separation, the relative disadvantage associated with old age is significant.

Meanwhile, in the wholesale and retail industry, technology adoptions increase the retirement hazard of overall employees. Considering the results by occupations, the adverse effect of technologies can be attributed to workers’ tasks. More than half of the workers are service and sales workers (related to the impact of IT equipment), and clerical jobs (related to the impact of automation) comprise the second highest percentage. Hence, the distinction reflects the results from columns 1 to 4 in Table 6.

We conducted similar hazard analyses separately for males and females. The results presented in Table 9 reveal significant gender differences. The estimated hazard ratios for the interaction terms are substantially larger for males than those estimated from the entire sample as well as females. Nevertheless, old male workers’ retirement hazards are below 1. Conversely, females do not reveal clear disadvantages associated with aging. For the indices of technology adoption, the relative handicap of older females is insignificant.

Another gender difference is observed in the overall employment effect of technology adoption. Whereas automation positively affects male employment, it increases females’ risk of leaving the firm. Since clerical and support occupations account for the second highest percentage of female workers’ jobs, the negative effect of automation on female workers’ employment is attributed to their tasks. While many male workers are also engaged in clerical jobs, their retirement hazards (including the elderly) are below 1. The result suggests that the negative impact of automation on job security concentrates on female clerical workers.

Last, we investigated how the effect of technological change differed by the reason for retirement. We conducted 2SRI regressions using each of these two types of retirement—involuntary and voluntary retirement—as the dependent variables.

Table 10 shows the estimation results. If the separated definitions of retirement are applied, the negative effect of new technology on elderly workers’ employment still stands out. The effects on the overall risk of forced layoffs are statistically significant and less than 1, but the estimated hazard ratios for the interaction terms are all greater than 1 and also significant. The effect of technological change reduced the hazard of voluntary retirement for all employees, but the effect is relatively unfavorable for older people.

The results suggest that technological changes may affect elderly workers’ employment through various pathways. First, from the perspective of labor demand, adopting new technology played a role in replacing the tasks previously conducted by older workers. Automation and IT equipment appear to bring a particularly strong displacement effect for older workers, given that it raises their risk of forced layoff by more than 1.4 times. Second, it is notable that technologies also relatively increase the possibility of voluntary retirement for personal reasons among older workers. These reasons are more related to the labor supply. One possible explanation is that the quality of matching between the workers and their jobs deteriorates because of the workplace changes accompanied by the introduction of new devices.

5. Conclusion

This study has investigated how the adoption of new production technology affects the employment of older workers in Korea by using establishment-level panel data (WPS) that were newly linked with administrative records (Korean Employment Insurance data).

More specifically, we estimated proportional hazard models to examine how indices of new technology adoption affect the probability that employees leave their jobs and how the effects differ between old and young workers. For constructing variables on technology adoption, we utilized responses to the following two questions: (i) whether new automation was adopted and (ii) how much the purchase of IT equipment increased. To address the potential bias arising from endogenous technology adoption, we used the 2SRI method.

The baseline results suggest that the employment effect of technology adoption differs between aged and younger workers, with the former less favorably affected by newly adopted technologies than the latter. Technology adoptions, measured by newly introduced automation and investment in IT equipment, tend to lower the overall risk of retirement for workers who had been employed in the firm prior to the technological adoption. However, the retirement risk of older workers compared to that of younger workers is increased by the adoption of new technology. These types of technology significantly increase both the possibilities of involuntary and voluntary retirement of older workers. Possible explanations for this are (i) the demand side: firms’ managerial decisions are unfavorable for the employment of older workers and (ii) the supply side: a deterioration in the quality of matching between the workers and their jobs caused by the introduction of new devices.

In some conditions, technologies absolutely increase the retirement risk of the elderly in certain types of occupations. Newly adopted automation negatively affects the employment of aged workers who are in clerical and support occupations or in the wholesale and retail industry. These results indicate that automation could play a role in replacing the tasks previously conducted by them. We also found that the purchase of IT equipment absolutely increases the retirement hazard of older workers as well as young workers engaged in service and sales occupations. We believed these heterogeneous effects across occupations are associated with the difference in estimation results by industry and gender.

Our study has various limitations. First, we used two specific indices of production technology available from our data, which provide only a partial picture of the effect of technological change. Second, our data do not provide detailed information on the personal characteristics of individual workers. A more sophisticated examination would be possible given information on wages or the education level of workers. Third, we only provided some speculative explanations and highly circumstantial evidence regarding the result that technology adoptions are less favorable for older workers. Further studies are required to fully understand the mechanisms behind the relationships between technological advances in production and the employment of workers with heterogeneous characteristics. Last, since our study covers only 3 years of employment status, we were only able to examine the short-run effect of technology adoption. Given that the compensation effects of technological changes may not emerge instantly, an extended period needs to be analyzed to capture possible long-term effects fully.

In spite of these limitations, the evidence provided in the paper raises the possibility that technological changes could negatively affect the employment of older workers. Of course, the results based on several specific types of technology adoption in a single country can hardly be generalized. The features of ongoing technological progress may differ from one another, and certain types of new technologies may help people become active and productive until very old age by alleviating the disadvantages associated with aging. However, given that automation and IT are the two major types of the current wave of technology innovations, our study strongly suggests that radical changes in technology may bring unfavorable labor market consequences for older people, at least compared with the young. In this sense, technological changes would likely produce an additional challenge for a rapidly aging society.

Acknowledgments

We thank Dae Il Kim, Yoo Bin Kim, Joseph Han, Sok Chul Hong, Jisoo Hwang, Jungmin Lee, Sangyoon Song, Chunghyun Nam, and seminar participants at the 2019 WPS Conference, 2022 BOK Mid-term Seminar, and 2022 Korea Empirical Applied Microeconomics Conference for their helpful comments and discussions. C.L. gratefully acknowledges that the study was supported by the Center for Distributive Justice of the Seoul National University Institute of Economic Research. This paper is a revised version of the second chapter of Chung’s doctoral thesis.

Footnotes

References

Appendix

An assessment of the technology adoption effect on employment or hiring is required in the context of previous studies dealing with process innovation and labor. The impact of technology adoption on employment, however, cannot be measured by the survival analyses we employed. Hence, we employed a new empirical strategy that examines the employment and hiring effect of technology at the firm level. The new analyses provide the net impact of automation and IT on employment.

The outcome is the difference of logged levels (employment and hiring) between the beginning and the end of the timeline (⁠|$\Delta{{\rm{y}}_{\rm{j}}}$|⁠) in our dataset. We estimated the impact of technology adoption |$T$| by employing two-stage least-squares estimation. We used the same instruments and firm-specific variables (F) that we employed in the survival analysis.

Tables A1 and A2 report the estimation results for employment and hiring, respectively. The results suggest that firms adopting new technology significantly increase employment and hiring. The increases are attributed to young workers, not the elderly, suggesting that the results are consistent with the original estimation results that workers employed in the firm adopting technology experience a lower retirement risk but less favorable to old workers.