Skip to main content

The struggle of small firms to retain high-skill workers: job duration and the importance of knowledge intensity

Abstract

In the knowledge economy, skilled workers play an important role in innovation and economic growth. However, small firms may not be able to keep these workers. We study how the knowledge-skill complementarity relates to job duration in small and large firms, using a Portuguese linked employer-employee data set. We select workers displaced by firm closure and estimate a discrete-time hazard model with unobserved heterogeneity on the subsequent job relationship. To account for the initial sorting of displaced workers to firms, we introduce weights in the model according to the individual propensity of employment in a small firm. Our results show a lower premium on skills in terms of job duration for small firms. Furthermore, we find evidence of a strong knowledge-skill complementarity in large firms, where the accumulation of firm-specific human capital also plays a more important role in determining the hazard of job separation. For small firms, the complementarity does not translate into longer job duration, even for those with pay policies above the market. Overall, small knowledge-intensive firms struggle to retain high skill workers and find it harder to leverage the knowledge-skill complementarity.

Plain English Summary

Small knowledge firms are struggling to keep their best employees: highly-skilled workers are needed for growth and success, but small firms cannot hold onto them. Advanced knowledge increases the productivity of skilled workers. Because large firms are more innovative and more technological, this knowledge-skill complementarity may be different for small and large firms. We study how the complementarity affects job duration, and how firm size influences this relationship. We find that skilled workers in large knowledge firms see a premium in terms of longer durations. In small knowledge firms skilled workers suffer a penalty instead. We also test how personnel management practices mediate these effects and find that even small knowledge firms paying higher wages find it hard to keep workers. Our study shows that small firms struggle to retain skilled workers, limiting their ability to handle knowledge. This calls for policy mechanisms to diminish barriers to knowledge, recognizing that human resources are the main assets of new ventures.

This is a preview of subscription content, access via your institution.

Fig. 1
Fig. 2
Fig. 3
Fig. 4

Data availability

The datasets analyzed during the current study are available from Statistics Portugal on request for research.

Code availability

Not applicable.

Notes

  1. For extensive reviews of technological change’s impact on employment, see Vivarelli (2014), Calvino and Virgillito (2018) and Barbieri et al. (2020).

  2. For example, Baffour et al. (2020), Balsmeier and Woerter (2019), Barbieri et al. (2019), Van Roy et al. (2018).

  3. See Hall and Khan (2003) for extensive reviews of technology adoption and firm size. Stoneman and Battisti (2010) review the literature on technology adoption and diffusion at the international, inter-industry, intra-industry, intra-firm, and household levels.

  4. See also the recent discussion by Kotey and Koomson (2021) on the heterogeneous effects of different flexible work arrangements across firm size.

  5. In Portugal, a worker may retire as soon as the age of 55, under some contractual schemes.

  6. The same argument has also been used to obtain evidence on the consequences of job loss on the family environment (e.g., parental and childhood health) and education outcomes (Mörk et al., 2020), the probability of divorce (Doiron & Mendolia, 2012), and body mass index and alcohol consumption (Deb et al., 2011).

  7. Dauth et al. (2021) provide a recent example of this identification strategy when studying wages and globalization.

  8. See Table 5 in the Appendix for the Eurostat classifications according to industry codes.

  9. See Table A1 in Baptista et al. (2012) for a description of each level according to Decree-Law 121/78 of July 2.

  10. The sample for the fixed effects wage regression covers each company’s history since 1996 up to the year before hiring a worker belonging to our main sample. The regression controls for workers’ years of education, gender, age, hierarchical level, occupation, and tenure in firm, the firm’s region, industry, sales, legal structure, presence of foreign equity, and year dummies.

  11. To our knowledge, an unbiased discrete-time proportional hazards fixed effects estimator does not exist. Including an individual-level fixed effect into our binary model would result in a selected sample which excludes all individuals who do not experience a job separation during the period of analysis (35.5% of workers in our sample), because of no variation in the dependent variable. Additionally, a fixed effects method would exclude workers with one-year durations (49.5% of workers in our sample) because of demeaning. Allison and Christakis (2006) discuss the consequences of using conditional logistic regression with fixed effects for single-event analysis and propose an alternate method. This method, however, is impractical if researchers are interested in studying the effect of multiple variables, cannot be used to estimate the effect of covariates that are monotonic with time, and seems very sensible to omitted variables. For these reasons we prefer a random effects (frailty) approach.

  12. We accumulate all tenures equal to or above five years in a single level of the tenure variable for the sake of parsimony. In our sample, only about 8% of job separations occur on the fifth year of tenure or later, compared to 61% in the first year, or 18% in the second year.

  13. See Sub-section 3.2 for more detail on this variable.

  14. See results in Tables 9 through 11 in Appendix C.

  15. Because the college education and high-skilled jobs variables are highly correlated (\(\rho =0.53\)), the marginal effects are obtained from specifications where these variables are included separately. See last column of Table 9 where both variables are included in the same specification — conclusions for high skills still hold, but the marginal effect of college education is reduced and is no longer significant in knowledge-intensive firms with at least 50 employees.

  16. See Sub-section 3.2 for more details on the fixed effects.

  17. Results for robustness checks beyond those in Appendix D are available upon request.

References

Download references

Acknowledgments

Our thanks are due to Fundação para a Ciência e a Tecnologia for funding (grant PTDC/EGE-ECO/32557/2017). We are indebted to the GEP—Gabinete de Estratégia e Planeamento (Office for Strategy and Planning) of the Portuguese Ministry of Labor, Solidarity and Social Security and to Statistics Portugal for access to the data. The datasets analyzed during the current study are available from Statistics Portugal on request for research. The views expressed in this article are those of the authors and do not necessarily reflect the official views of authors' institutions.

Funding

This work was supported by Fundação para a Ciência e a Tecnologia through grant PTDC/EGE-ECO/32557/2017 and project UIDB/00097/2020.

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Francisco Lima.

Ethics declarations

Conflict of interest

The authors declare no competing interests.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendices

Appendices

A. Supporting tables

Tables 4, 5, 6,

Table 4 Knowledge-intensive industries
Table 5 Descriptive statistics for all new job relationships
Table 6 Descriptive statistics for all new job relationships with previous experience
Table 7 Wage regression with firm fixed effects

7

B. Employment and firm size: probit estimates

We estimate a probit model for the probability of being hired to a small firm (versus being hired to a large firm) to obtain the weights for the duration models. The probit model describes how the attributes and abilities of displaced individuals contribute to finding a job in firm with fewer than 50 workers. Our choice of independent variables follows previous works such as those by Evans and Leighton (1989) and Idson and Feaster (1990).

Results (Table 8) show that less educated workers coming from smaller and non-knowledge-intensive firms have a higher probability of finding a job in a small firm. Women are just as likely as men of finding employment in small firms and age positively contributes (with diminishing returns) to employment in small firms. Time to find a job since displacement decreases the probability of being employed in a small firm. The discussion of this particular effect is out of the scope of our paper but deserves further exploration (with a more complete model). Hiring practices of each kind of firm in conjunction with the worker’s specific attributes (e.g., prone to more precarious job relationships) might drive this result.

Table 8

Table 8 Marginal effects for probability of starting job in a firm with fewer than 50 employees

C. Regression tables

Tables

Table 9 Average marginal effects of college education and high-skilled job on the hazard of job separation: Models with interactions with knowledge intensity and size category

9,

Table 10 Average marginal effects of tenure on the hazard of job separation: Model with interactions with knowledge intensity and size category

10,

Table 11 Average marginal effects of firm fixed effects quartile on the hazard of job separation: Model with interactions with knowledge intensity and size category

11

D. Robustness checks

The assumption of independence between displaced workers and firm closure is not necessarily clear-cut and that our main sample might include a larger share of lower ability individuals than in the workforce. In our original analysis, we allow a three-year window after firm closure for workers to find a job, giving workers of different skill levels and different job-finding rates some time to find a new job such that our final sample has considerable variance in the ability distribution. Nonetheless, one could argue that despite this, we oversample individuals who have lower ability levels compared to the workforce.

To further address this issue, and in the lines of Dustmann and Meghir (2005) and Schmidpeter and Winter-Ebmer (2021), we ran several robustness checks that in general reinforce our main findings. The robustness analyses fall into two main families:

  1. a)

    In the first family of analyses, we apply restrictions to our main sample (workers displaced by firm closure) aiming to exclude lower ability individuals and thus lessen the possible left-skewness in the distribution. Namely, we built the following sub-samples and estimated the same models as before:

  2. b)

    A sub-sample that includes only workers who were displaced by firm closure and found employment in a new firm in less than a year. The main sample allowed three-year window after displacement. Workers who find employment shortly after firm closure are expected to be more able in both observed and unobserved characteristics.

  3. c)

    A sub-sample that excludes all workers who had a management position (top and middle managers) in the firm that closed. Below-average managers might drive firm closure. Excluding those workers makes the assumption of independence between worker ability and firm closure stricter.

  4. d)

    A sub-sample that includes only workers who were displaced from older firms (firm age above median of seven years). As Dustmann and Meghir (2005) suggest, the market exit rates of older firms are lower. Firms that close will be disproportionally younger. Younger firms are likely to have, on average, less experienced and less capable workers. Excluding workers displaced from younger firms should result in a less left-skewed distribution.

  5. e)

    A sub-sample that includes only older displaced workers (age before displacement above median of 30 years). On average, older workers will tend to have more experience, be more productive and possibly acquired more knowledge. Also note that, as mentioned in Section 3.1, we already previously excluded workers with ages above 55 (to avoid job separations caused by retirement). This ensures that our sub-sample of older workers does not include aging workers with declining productivities.

  6. f)

    A sub-sample that includes only workers who at the time of displacement had tenures great or equal to the 75th percentile of tenure (three years). Workers with longer tenures should be more capable and more productive, as evidenced by their longer and better matches (Jovanovic 1979). We also considered restricting to workers with above-median pre-displacement tenures and found that our conclusions still held (results available on request). However, as most jobs end early (Farber, 1999), the pre-closure median tenure was of just one year which we believe is too low of a threshold to ensure that the resulting sub-sample was less left-skewed.

  7. g)

    An extended sample that includes not only the workers displaced by firm closure but also workers who left firms one year before closure and find new employment in a three-year window. As we cannot identify the reasons for job separations, this early-leaver group might result, on one hand, from firms selectively laying off lower ability individuals in an attempt to stave off closure (Gibbons & Katz, 1991), and on the other hand from a group of higher ability individuals, with more employment alternatives who, anticipating firm closure, opt to find another employer. Dustmann and Meghir (2005) find that the wages of those who are in the firm one year or more before closure are higher (especially among the unskilled) than those who stay until closer to firm shutdown. Their findings suggest that the early leavers should be more productive than those who remain. We expect the ability distribution of extended sample to be less left-skewed than our main sample. However, because it includes voluntary switchers, the sample likely suffers from the endogeneity that we attempt to avoid with our identification strategy. Arguably, the presence of endogeneity in this case is likely to be smaller than if we were looking at all new job relationships if those who choose to leave one year before firm closure are not doing so in a completely voluntary way.

  8. h)

    In the second family of robustness analyses, we drop the displaced by firm closure restriction and apply our duration models to a broader set of workers, aiming to assert if any divergences from our main results exist.

  9. i)

    An extended sample of workers starting a new job spell after leaving a previous employment relationship. This extended sample includes the workers who were displaced by firm closure as well as all others who left a previous employer either voluntarily or were forced to leave (e.g., fired) and found new a new employer. This unrestricted sample of workers with experience may be more representative of the workforce but does not allow us to mitigate selection effects that might result from voluntary firm switches. Given the size of this sample, we estimate our models to a 30% sub-sample of randomly selected workers.

  10. j)

    An extended sample of workers starting a new job spell, without necessarily having had a previous employment experience. This sample includes all individuals in sample 7, and individuals starting their first (registered in the dataset) employment spell (mostly younger workers). This extended sample closely represents the population of new job spells, but again might suffer from selectivity from voluntary job switches. Because not all individuals have a previous employment experience, the models we estimate are modified versions of our main models, excluding control variables that relate to previous experience (previous firm size class, previous knowledge intensity category, moved to different sector, came from non-employment). For this reason, the results might not be directly comparable to our main results but are nonetheless relevant. Given the size of this sample, we estimate our models to a 10% sub-sample of randomly selected workers.

We provide results of the robustness analyses and a comparison with the main results below.

The results for our college education model applied to robustness samples 1 through 8 are presented in Figure

Figure 5
figure 5

Average marginal effects of college education on hazard of job separation, by knowledge intensity and firm size (95% confidence interval).

5, and the comparison is to be made with the results plotted in Figure 1. In all cases the marginal effects of college education in small knowledge-intensive firms are not negative meaning a college education in these firms does not translate to lower job separation hazards. By contrast, the returns to college education in larger knowledge-intensive firms is the largest (in absolute terms) of all four categories of firms. The marginal effects of college education in less knowledge-intensive firms are generally negative and do not differ between small and larger firms (exceptions to the displaced from long tenures and the unrestricted with previous experience samples). As with Figure 1, the knowledge-college complementarity in terms of job duration is only observed in larger firms while smaller knowledge-intensive firms struggle to keep college educated individuals. This analysis shows that our main results related to college education are persistent even if we consider sub-samples of more able workers and hold in less restricted samples as well.

In Figure

Figure 6
figure 6

Average marginal effects of high-skilled jobs on hazard of job separation, by knowledge intensity and firm size (95% confidence interval).

6, we show the marginal effects from our high skilled jobs model applied to robustness samples 1 through 8, to be compared with Figure 2. As with Figure 2, the marginal effect of a high skilled job is not significant at 5% or positive but small (sample of displaced from older firms). The marginal effect of a high skilled job in a knowledge-intensive firm with 50 or more employees is much larger (in absolute) than in the other categories in all the displaced sub-samples (1–6). In the unrestricted samples the marginal effect in that category is also larger, but not significantly different from the marginal effect in small less knowledge-intensive firms. These results point to knowledge-skill complementarity only in firms with 50 or more employees. The high skilled workers in small knowledge-intensive firms experience no returns to their skills. Our main findings regarding the high skilled jobs hold both when we drop the displaced restriction and in sub-samples of more able workers.

The robustness analysis regarding tenure and the accumulation of firm-specific human capital are shown in Figures

Figure 7
figure 7

Average marginal effects of tenure on hazard of job separation, by knowledge intensity and firm size (95% confidence interval).

7 and

Figure 8
figure 8

Average marginal effects of tenure on hazard of job separation, by knowledge intensity and firm size (95% confidence interval).

8. These results should be compared with those presented in Figure 3. As with the main results, we see that in small firms there is either no significant difference between knowledge-intensity categories in the marginal effects of tenure, or the marginal effects are weaker when knowledge-intensity is higher (especially in the unrestricted samples). Thus, in small firms we observe no knowledge-skill complementarity with regards to the accumulation of firm-specific human capital. In some cases, the reduction in the hazard rate of separation from accumulation of human capital is smaller in knowledge firms with less than 50 workers. When looking at larger firms, and considering the sub-samples of displaced workers, the marginal effects of tenure often diverge (especially for longer tenures) with each additional year. The divergence occurs because the marginal effects in knowledge-intensive firms with 50 or more employees tend to become more negative at a faster rate than in less knowledge-intensive firms. In the unrestricted samples, though the divergence is not as clear, the accumulation of firm-specific capital always brings greater decreases of the hazard of separation in larger knowledge-intensive firms. In both cases, this can be seen as evidence of varying degrees of knowledge-skill complementarity in firms with 50 or more employees: either workers in knowledge-intensive firms have a somewhat persistent advantage over those in less knowledge-intensive firms related to how tenure influences separation rates, or that advantage progressively increases with tenure. Our main conclusions on the accumulation of firm-specific human capital hold in most sub-samples of displaced workers as well as in the more general samples.

Finally, Figures

Figure 9
figure 9

Average marginal effects of firm fixed effects quartile, relative to 1st quartile, on hazard of job separation, by knowledge intensity and firm size (95% confidence interval).

9 and

Figure 10
figure 10

Average marginal effects of firm fixed effects quartile, relative to 1st quartile, on hazard of job separation, by knowledge intensity and firm size (95% confidence interval).

10 show the marginal effects of the quartiles of the firm fixed effect across the four categories of firms, that should be compared with the results in Figure 4. Generally speaking, small knowledge firms in the two top-paying quartiles do not exhibit any significant advantage in worker retention over less knowledge-intensive firms. When looking at larger firms, and similarly to our main results, the marginal effect of the fourth quartile of the firm fixed effect is more negative when knowledge-intensity is higher. While knowledge-intensive firms with 50 or more employees may resort to top personnel practices to increase retention, the same does not appear to be true for small firms. As with the previous set of analyses, our conclusions from the main sample of displaced workers hold in most of the robustness checks. The exceptions—the displaced from older firms and the displaced from long tenures sub-samples—come from rather selected group of workers and may point to some sort of match effects.

These broad set of robustness analyses suggest that our main findings may be generalizable to the workforce starting new job relationships and are persistent even when considering samples of more able displaced workers.

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Castro-Silva, H., Lima, F. The struggle of small firms to retain high-skill workers: job duration and the importance of knowledge intensity. Small Bus Econ (2022). https://doi.org/10.1007/s11187-022-00602-z

Download citation

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1007/s11187-022-00602-z

Keywords

  • Knowledge intensity
  • Technology
  • Firm size
  • Small firms
  • Job duration
  • Skills

JEL Classification

  • J24
  • J63
  • M51
  • O33
  • L26