Technology, Skills and Retirement PRELIMINARY & INCOMPLETE

Transcription

1 Technology, Skills and Retirement Federico Biagi Università degli Studi di Padova ECONPUBBLICA and S.D.A - Università Bocconi, Milano Danilo Cavapozzi Università degli Studi di Padova Ra aele Miniaci Università degli Studi di Brescia First draft: January 2007 PRELIMINARY & INCOMPLETE Abstract In our work we study the role of skills in the retirement decision by investigating whether individuals who are more technologically endowed, ceteris paribus, tend to retire later. Although more than a decade of research on skill biased technological change has proved that observed and unobserved skills are crucial determinants of wages and employment status, few empirical studies investigate the relation between workers technological skills and their retirement decisions. In our work we provide new evidence based on an Italian panel data set (the Bank of Italy Survey on Household s Income and Wealth). Our results show that better educated male employees who are able to use a computer tend to retire later, that this e ect is magni ed if the worker does use a PC on the job, it is stronger in a four years horizon rather than a two year horizon, and its magnitude is remarkably larger than what observed in US and Germany. We also show that the same skills seem not to play a crucial role in the retirement decision of women. Our results are robust to the estimation strategy adopted. 1 Introduction In our work we study the role of skills in the retirement decision. Our main research question is to investigate empirically whether individuals who are more technologically endowed, ceteris paribus, tend to retire later. After a decade or more of intense research on the importance of skill-biased technological change, we believe that it is possible to conclude that observed and unobserved skills are among the most important determinants of workers wages and employment 1

2 status (besides income inequality). While easier to model theoretically, the empirical counterpart of "skills" is not easily found. Workers productivity within occupations are likely to depend on many factors, some observables and some not. Talent (only very partially observable) is a major determinant, the other being education, training, experience, the structure of the rm s organization, the degree of technological progress experienced by the profession and the way in which the rm takes advantage of the latter. We believe that skills are likely to in uence the retirement choice as well: everything else constant, a worker endowed with poor technological skills will tend to become less and less productive, particularly in industries and professions characterized by rapid technological progress. At the same time, if human capital and technology are complementary, we cannot disregard the possibility that, everything else constant, a skill-biased technological change may favour more experienced workers because of their accumulation of human capital (see Weinberg, 2004). Hence, the prediction on the sign and the size of the e ect of technological skills on the probability to retire earlier (in one fashion or another) is mainly an empirical issue. In our work we proxy technological skills possessed by individuals with two variables: a (self-reported) measure of ability in the use of personal computers (Pc) and the actual use of a Pc on the job. Our preliminary results from the Bank of Italy Survey on Household s Income and Wealth (SHIW) indicate that male employees with higher ability in the use of a Pc and who use a Pc at work tend to retire later, while there is no evidence of any e ect for self-employed and females. Our work proceeds as follows: Section 2 reviews the literature on retirement choices and skill-biased technological change, while Section 3 provides some descriptive statistics on retirement and skills in the use of personal computers (by age, gender, education and employment status). Section 4 contains the estimation exercise: we have considered linear probability model (with OLS and IV estimators), standard logit regression and duration analysis, taking advantage of the panel structure of our data. Section 5 concludes our work. 2 Literature review The research on retirement behavior is wide and growing. Here we brie y review the main approaches found in the literature and then we discuss the evidence on the relationship between skill acquisition and retirement, which is at the heart of our research. The rst problem researchers encounter is in the de nition of retirement. Retirement is a process by which individuals abandon a productive role and withdraw from the labor market. For some individuals this is well represented by a dichotomous choice between working full time or abandoning the labor market altogether, while for others retirement is indeed a smooth process, for instance realized with a progressive reduction of hours worked. For some other individuals, retirement is obtained as a by-product of a down-sizing, which ultimately forces out of the labor force individuals who are unwilling or unable 2

3 to nd another job. Finally, when invalidity bene ts exists, as it is the case of Italy, many have noticed that invalidity pensions operate as substitutes for old age pensions. Hence, the rst conclusion we can draw is that the process by which individuals withdraw from the labor market is multifaceted. This also means that there exists an inherent di culty in trying to represent with a unique model individuals who are facing very di erent work experiences. For instance, we can observe a typical retirement pattern by which an individual, once eligible for Social Security bene ts, chooses to stop working altogether 1, but we can also observe somebody else who, satisfying the same eligibility criteria, chooses to continue working, either maintaining the same amount of hours worked or reducing it. While working, he could even receive pension bene ts (if the law allows the cumulation of labor and pension income), but we would not consider such an individual economically inactive (and hence retired from the labor force). Alternatively, we can have individuals who have been involved in restructuring processes and hence expelled from employment. Depending on the institutional arrangement in place, they could be unemployed or employed at zero hours (in this case a portion of their wage is usually publicly subsidized). If those individuals are older than 50 years old, there are good chances that they will end up being permanently out of the productive process (while perhaps resulting formally part of the labor force). These examples are meant to show that any de nition of retirement is bound to be somehow arti cial. In our empirical analysis we choose to work with two de nitions of retirement. Both are based on self-reporting and as such are exposed to the risk that the same word means di erent things across individuals. According to the rst de nition we consider retired only those who declare themselves retired ("pensionati"). According to the second one, we consider retired not only those who so self-report but also those who are unemployed and those who have exited the labor force for whatever reason 2. When trying to analyze retirement one is typically interested in the role played by the explanatory variables. From the previous examples it is evident that some variables, such as technological change or rm restructuring, a ect the labor-demand side, while others, such as personal and family conditions, pension bene ts, requirements for retirement, and presence of early retirement schemes, will a ect mostly the labor-supply side. Furthermore, factors like wages, training and incentive plans for retirement set at the rm or sector level will a ect both demand and supply. The contribution of these variables to the retirement outcome can be appreciated once a retirement model is speci ed, by which individuals choose whether and how much to work, taking into account the present value of the streams of bene ts and costs in both cases (for a recent survey see Lumsdaine and Mitchell, 1999; see also Lazaear, 1986). The variables usually considered relevant in the explanation of retirement behavior are: expected labor and non labor income, expected pension bene ts and expected pension 1 We could still observe individuals retiring even without having the characteristics necessary to receive pension bene ts, but this is much less frequent. 2 In both cases we look at the behaviour of individuals above age 45. 3

4 contribution, health conditions, preferences for leisure. As for the models, we have structural dynamic stochastic models (see Gustman and Steinmeier, 1986; Rust, 1989, Rust and Phelan, 1997), according to which, based on the expectations for the relevant explanatory variables, individuals choose the optimal consumption, saving and retirement paths. Alternatively, we have models in which, at each instant in time, the individual re-evaluates his/her labor supply choice, taking into account the option value of remaining in the labor market versus the bene ts of retiring (see Stock and Wise, 1990). This type of modeling strategy allows researches to take into account time-to-time changes in the explanatory variables. Finally we have reduced form models, in which researches do not estimate a full consumption/retirement model but simply look at the e ects on the retirement choice of the variables that would a ect retirement according to a fully structural model. This class of models typically include some variable capturing the e ects of Social Security legislation on the bene ts and costs of remaining active in the labor market 3. Given the focus on the conditional probability of retirement usually found in reduced form econometric models, the variables capturing the generosity of the Social Security system are generally representing the marginal bene t of remaining one more year in the labor market (researchers typically look at the accrual rate, at the implicit tax rate or at the option value). Typically, the econometric analysis of retirement behavior nds that individuals tend to retire at around the standard age of eligibility, with some retiring happening at an earlier age when early retirement is possible (see Blundell et al., 2002; Tanner 1998). Moreover most studies nd that social security legislation, both in term of eligibility and in terms on net bene ts, is an important determinant of retirement behavior (see Coile and Gruber, 2000; Coile and Gruber 2001; for a recent results on the Italian case see the special issue of Labour, August 2003, vol 17). Few empirical papers focus on the role of technological skills in the retirement decision. In this case, the empirical analysis in complicated by the fact that one has to take into consideration that the relation between skills and retirement is a ected also by the business cycle and by the training policy adopted by rms. In fact, workers (especially older and unskilled workers) might be severely a ected by technological shocks. But they might also be o ered the possibility of upgrading their skills through training, which might increase their future wages and employment probabilities. Though, training comes at a cost (usually shared by rms and workers) and generates bene ts, which are very much agerelated, since an older worker has a shorter time horizon in which he or she can appropriate (and share with the rm) the bene ts arising because of the higher productivity obtained through training. Bartel and Sicherman (1993) study the e ect of technological change on the career of older workers. They notice that technological change can a ect retirement, in uencing both the training decision and the depreciation of the stock 3 Notice that if it is possible to receive both public pension and labor income, the model needs to be enriched. 4

5 of human capital. They test two complementary hypotheses. According to the rst hypothesis, everything else constant, individuals retire later in industries in which technological change is particularly rapid 4. According to the second conjecture, an unexpected rise in the depreciation rate of human capital, for instance following an unexpected rise in the rate of technological change, should lead to earlier retirement. Both hypotheses are con rmed by the analysis of Bartel and Sicherman, whose main limitation is in the fact that uses sector data to measure technological change. Recent work by Friedberg (Friedberg, 2003) tries to provide an answer to a question very similar to the one analyzed in our paper. In fact, the author considers whether there exists evidence of a signi cant relationship between computer use and retirement. The basic intuition for this analysis is that computers might have a ected the demand for labor in various ways. First, they tend to be a substitute for unskilled labor for routine tasks. Second, they have altered the performance of non-routine tasks, mainly held by skilled workers. Finally, computerization alters the "bundle of skills and tasks that de ne a job". These changes can a ect the retirement choice of older individuals, given that older generations tend to be less educated and hence more likely to be assigned to routine jobs: for them training is generally less pro table (given the higher investment costs and the reduced time horizon over which these investment costs can be recouped). Friedberg studies the relationship between retirement and computer use at work with two datasets that refer to the U.S. The rst one is the CPS, in particular she takes into account the October surveys for the years 1984, 1989, 1993 and 1997, which report a question on the use of computers at work 5. The second dataset is the Health and Retirement Study, which, in the 1992 wave, asks how often workers are required to work with computers. She estimates a reduced form linear probability model, where the training choice and the retirement choice are modeled with simultaneous equations, with personal and rm characteristics as common explanatory variables 6. Friedberg takes into account that training and computer use are likely to be endogenous with respect to the retirement decision, and she chooses to instrument them using the percentage of computer users among prime-age workers in the same occupation and industry. She nds that, even controlling for many individual, rm and sector characteristics, computer use tend to induce delayed retirement. She concludes that "holding everything else constant, the median retirement age if everyone had used a computer would have occurred 12 months later". In a very similar framework Schleife (2006) uses the German Socio Economic 4 Notice that two e ects are at work: on the one hand, rapid technological change makes training more pro table, but, on the other one, technological change increases the depreciation rate of human capital and hence reduces the incentives to train especially for older workers, since a long period at work is required to make training bene cial for both parties. 5 The 1991 January survey asks questions about job training, including computer skills. 6 Notice that Friedberg recognizes that at least two are the causation chains that govern the training and retirement choice. First, it can be that individuals with low level of training, and hence low level of human capital, are displaced by the technological revolution and hence opt to retire early. Second, it can be that, since they plan to retire early, for instance due to technology shocks, older individuals choose not to train. 5

6 Panel to investigate the e ect of computer use at work on the employment status of older workers between 1997 and 1999, and between 1997 and Her twostage least square estimates of a linear probability model (in which computer use at work is instrumented with computer use at home) provide no evidence that Germans who use a computer at work tend to postpone their retirement. She therefore provides a rst evidence that US and European labor markets may di ers also with respect to this dimension. Our paper is strictly related to Friedberg (2003) and Schleife (2006) for what concerns the use of microdata and the focus on the e ect of the computer use at work on labour supply decisions. But we depart from their strategy when we take into account that the ability to use a computer is by its own a good proxy of the technological ability of the worker and when we choose as instrumental variables (for both computer use at work and computer skills) information on workers household characteristics. The rst choice is suggested by the observation that the estimates of the e ect of computer use on retirement decision are potentially a ected by an omitted variable problem. In particular, it might well be the case that computer use per se does not have any e ect, but it simply helps to identify those workers more able to cope with new sophisticated tasks. These workers are likely to react better to the innovations introduced by the rms and therefore to have a lower opportunity cost from keep on working. By directly controlling for individual ability in the use of a Pc (which may be fairly good even for those not using a Pc at work) we reduce the probability of having spurious results. For what concern the instrumental variables, due to the lack of reliable data we cannot follow Friedberg s approach, while we consider Schleife s choice as debatable. We instead use data on workers household composition in order to exploit the correlation between these exogenous variables and the probability that any family member is able to use a computer. The basic idea is the following: the more the other family members possess ability in the use of a PC, the higher is the probability that a worker is skilled and uses a computer at work. 3 Data description Data are drawn from the Bank of Italy Survey of Household s Income and Wealth, which every two years provides a sample of about 8000 households, representative of the Italian population. It contains detailed information on demographics, income and wealth, at the individual and family level (see Banca d Italia 2002). Half of the households participating to the sample belong to a longitudital survey, and we exploit data coming from the 2000, 2002 and 2004 waves. The 2000 wave provides us with information on the ability of individuals in the use of computers and, for those who were working, on their use of personal computers at work. Speci cally, the 2000 Survey, for each household s member, records self-rated computer skills on an increasing ve-step scale. We rearranged this scale in order to de ne a dichotomous variable which takes a value of one if the individual declares to have at least some ability in Pc utilization and zero 6

7 otherwise. For people at work, the 2000 Survey also collects the information concerning whether they use a Pc at work. These data provide us unique and valuable information for the question analyzed in this paper, since they permit us to estimate whether computer skills and computer use at work, which are to be considered as proxies for individual technological ability, are positively in uencing the probability of remaining attached to the labor market. 3.1 Pc skills and Pc utilization at work Before turning to our basic research question, it is useful to describe the distribution of computer skills and retirement choices in Italy in year 2000, as measured by our dataset (the Bank of Italy SHIW) and compare it to the most relevant source of information on ICT di usion among Italian population (the Multipurpose Survey, run by the Central Italian Statistical O ce -ISTAT), which has a sampling design similar to Bank of Italy SHIW, but also a di erent scope and sample size 7 (see Figure 1). According to the ISTAT survey, about 35% of Italians in the age class (5-10) were able to use a computer in This fraction is about 11% for pre-school children, it rises to 60% for high school students (15 to19 year old), it declines to 27% for the pre-retirement age class (45-54), it halves in the next age group and falls to 7% for the over 60. The survey documents a remarkable gap between the North and South of Italy, and a gender gap starting from the age of 35. As documented in Miniaci and Parisi (2004), if in SHIW we de ne an individual to be a computer user if he or she has at least some computer skills, then it can be observed that the Bank of Italy survey tends to underestimate the proportion of users among the young and to overestimate among people over 45 years old (see Figure 2). Finally, it is crucial to our aim to distinguish between workers who possess ability in the use of a computer and those who do use a Pc at work. As reported in Figure 3, according to the ISTAT survey, 25.1% of Italian Pc users do work with a computer at home, and 45.5% work with a computer away from home. We can compare these gures with those obtained from SHIW: 39.5% of Pc users do work with a computer, either at home or away. This means that, according to ISTAT, at least 13.5% of the population uses a computer when working (being at home or away from it), while, according to the Bank of Italy, this percentage amounts to 12.2%. Overall, the two surveys give a similar picture about the di usion of Pc use in Italy. We now focus on household s heads and their spouses who fall in the age range and declare themselves as either employee or self-employed in the 2000 wave of SHIW. We have 7,768 individuals, 2,914 of which are females. In this sub-sample, when we condition upon education, we see that the gender gap documented in previous tables disappears and it becomes an education gap. Consider Figure (4): the education gap in Pc skills is large and constant up to age 55. In the age interval the di erences between education groups widen. For females (5), the education gap is approximately constant up to age 7 There are 20,000 households in the Multipurpose Survey. 7

8 52 and it tends to shrink for older ages. But conditional on education, the proportion of users is almost equal across gender up to the age of 53. This evidence can be rationalized by the presence of a selection process which leads less technologically endowed individuals to stop working either voluntarily or not. It can be shown that there is no signi cant di erence in the fraction of skilled workers by type of occupation (employees vs. self-employed) once we control for age, education and gender. 3.2 Pc skills, Pc utilization at work and retirement choices In SHIW individuals are asked to declare their position with respect to the labor market. Using the 2002 wave we can construct the distribution of the age of retirement for those who report to be job pensioners in Looking at the distribution of the retirement age conditional on gender and education (see gure 6 and in Figure 7) it is easy to recognize peaks at the age of 55, 60 and 65 due to the retirement options "pensioni di anzianità" (retirement pension) and "pensioni di vecchiaia" (old age pension). Notice that the distribution of the age at retirement is more concentrated around the typical ages for individuals with lower education. Those with at least secondary education tend to have a more diversi ed behavior, as far as retirement age is concerned, since for this group bene ts and costs from retirement are more unequally distributed. Furthermore, data provide evidence that the peaks at the age of 65 are mainly due to the retirement behavior of the self employed. We now turn our attention to the transition to retirement. We restrict our attention to individuals who in 2000 are (i) household s heads or spouses, (ii) aged and (iii) employees. As before, we develop separate analyses for men and women and consider two de nitions of retirement. According to the rst, we consider retired those who declare themselves as job pensioners. According to the second, we consider retired all people who are not working for whatever reason. SHIW sampled households can be divided in two groups: those who are interviewed for the rst time and those who were already interviewed in previous waves. Hence, a subset of individuals is tracked over time and this allows us to observe transitions towards retirement. But this sampling structure also means that we have a smaller sample size when we focus on transitions. In Figure (8) and in Figure (9) we consider the two de nitions of retirement and distinguish between transitions occurring within 2002 and those occurring within First, we point out that the number of individuals followed up to 2002 is larger than that of those followed up to 2004 because of the sampling design. In fact, SHIW adopts a rotation of its panel section in order to maintain the representativeness of the overall sample across time. We further report the number of individuals who retire and their percentage with respect to the speci c group of interest. As reported in Figure (8), among male employees, 8% of those with Pc skills becomes job pensioner within This percentage rises to 20.06% when we consider individuals without Pc skills. At this pure descriptive level, being able to use a Pc results to be positively as- 8

9 sociated with the probability of remaining employed. Analogous considerations hold for all the groups of interest, even when we consider Pc utilization at work. 3.3 Nonparametric duration analysis We explore the relationship between the age of retirement and the variables indicating Pc use by means of a nonparametric models for continuous-time duration analysis. We estimate non-parametrically the survivor function in a sample strati ed by the variable of interest. Through the comparison of the results obtained across groups, we get a nonparametric estimate of the e ect of the variable of interest on the survivor function. Further, the log-rank test is conducted in order to check whether the hypothesis of equality of survivor functions is supported by data. In Figure (10) we consider the strict de nition of retirement and we estimate non-parametrically the e ect of having Pc skills on the males survivor functions. We note how those who have some ability in using a Pc tend to exhibit a higher probability of remaining in the initial state of employment up to age 64. The log-rank test shows that these di erences are statistically di erent from zero. In Figure (11) we conduct the same analysis focusing on males with high education. The previous results are entirely con rmed. The same does not hold when we focus only on males with low education, as reported in Figure (12). There are several crosses between the curves and no clear pattern emerges. Consistently, the log-rank test does not reject the null hypothesis of equality. As for the e ect of the Pc utilization at work, similar results are found. Up to age 64, males who use a Pc at work show higher probabilities of remaining at work. These di erences are supported by the results of the log-rank test which rejects the equality hypothesis. When the sample is strati ed by education, the results are con rmed for individuals with high education. For males with low education, although up to age 60 those who use a Pc at work present a higher probability of retiring later, these di erences are not statistically signi cant. When we consider the broader de nition of retirement, the previous ndings are valid. As for females, Figure (13) takes into account the strict de nition of retirement. Although those without Pc skills tend to have higher probabilities of remaining employed at all ages, the di erences between the groups are small and the log-rank test does not reject the null hypothesis of equality of the survivor functions. If we stratify the sample by education, a clearer pattern is not found and the log-rank test never rejects the null hypothesis of equality. Analogous considerations hold when the e ect of Pc utilization at work is considered. All these results do not vary with the de nition of retirement. 9

10 4 Estimation results 4.1 Discrete choice analysis Linear probability model In our rst estimation exercise we follow Friedberg (2003) and Schleife (2006), and use a linear probability model estimated by OLS and IV in order to describe how the probability of transition towards retirement is a ected by ICT utilization. We consider males and females who in 2000 are (i) household s heads or spouses, (ii) aged and (iii) employees. We distinguish between changes in the employment status occurring within 2002 and those occurring within By doing this, the e ect of the variables of interest is allowed to vary between the short and the long run. We still de ne retirement in two ways (in the former we consider individuals who stopped working to become job pensioners, in the latter individuals who stopped working for whatever reason). In all speci cations we control for age, age squared, age at the time of the rst job, years of contribution up to 2000, education, number of days spent at home for illness during 2000, region of residence, number of household s components, job characteristics, sector of employment, rm dimension, labor income, other household s income. We control for the age at the time of the rst job and for the number of years of contribution since they are useful to capture pension wealth and the incentives to retire faced by individuals. Moreover, we control for job characteristics, sector of employment and rm dimension in order to allow for variation in the rate of di usion of new technologies in the economy. We further allow the e ect of our main variables to vary across education by plugging in the linear speci cation an interaction term between having Pc skills and having low education (i.e a lowersecondary school degree). Hence, the parameter associated to Pc skills captures the e ect of Pc utilization for individuals with high education, whereas the sum between this parameter and the one on the interaction term individuates the e ect of interest for those with low education. Figure (14) reports the results concerning the e ect of having Pc skills on the transition towards retirement in the four-year period For males, we nd that having Pc skills is associated with a signi cant decrease in the probability of exiting the state of employment only for individuals with high education (-18 and -16 percentage points according to the de nition of retirement). For females no signi cant e ect is found. If we consider the e ect of Pc utilization at work alone, we obtain similar results, with a signi cant decrease in the probability of retirement found only for males with high education. When having Pc skills and using a Pc at work are considered jointly (Figure 15) we observe that the main e ect is played by skills. The overall e ect of using a Pc at work for individuals with high education is captured by the sum between the parameter capturing Pc skills and the one referring to Pc use at work 8. We nd 8 Analogously, for individuals with low education, the overall e ect of using a Pc at work is the sum of the parameters which refer to having Pc skills, its interaction with education and 10

11 that for males with high education there is a remarkable (negative) e ect of Pc skills on their probability of retirement, even after controlling for the their use of a Pc at work. On the other side, conditional on the level of Pc skills, Pc use at work is not signi cant. Further, the overall e ect of using a Pc at work is still signi cant only for men with high education. For women, no signi cant e ect is found. When we move to the probability of retirement in the two-year period (Figure 16), the overall e ect of using a Pc is now signi cant and negative for males irrespective of their education. For females signi cant results are obtained only in the case of the strict de nition of retirement: in this case using a Pc at work, conditional on the level of Pc skills, is associated with a rise in the probability of retirement. These OLS results might be biased due to a potential endogeneity of the variables concerning Pc skills and utilization. Endogeneity may arise from the fact that individuals who use Pc possess unobserved abilities which make them more likely to keep on working. Moreover, we may maintain the hypothesis that individuals who plan to retire later decide to use a Pc in order to increase the level of their skills and in turn their probability of being employed. Further, they have an incentive to improve their Pc knowledge because they expect a longer period of permanence in the labor market, during which capturing the bene ts from their investments in computer skills. Without allowing for endogeneity, the results reported above may be measures of the association between the Pc utilization and the retirement decisions. On the contrary, our empirical analysis is intended to assess how the probability of retirement is a ected by an exogenous change in Pc utilization skills. Hence, we estimate the linear probability model by means of a 2sls approach. The set of additional instruments we use in order to achieve the identi cation of the parameters of interest consists of the number of other family members with some Pc skill, its interaction with the number of household components, its interaction with education and nally the number of children at school in the household. In all speci cations we allow for the same set of control factors as before. Our exclusion restrictions can be summarized as follow: the skills of other household members may a ect the skills of the head of the households, but, once controlled for his/her actual PC abilities, other household members skills do not a ect his/her retirement decision. Miniaci and Parisi (2006) show that within household peer e ects are indeed relevant for the di usion of computer skills: considering the 2000 SHIW, 70% of individuals co-habitating with somebody skilled are skilled, this percentage falls to less than 10% if nobody else in the family is able to use a computer. Results of the IV procedure are reported in Figure (17) and they can be compared with those in Figure (14) for the period We conduct a test for the joint signi cance of the additional instruments in the rst stage equations and we always reject the null hypothesis of insigni cance. The test for the validity of the overidentifying restrictions (accounting for heteroschedasticity, see the one on Pc use at work. 11

12 Wooldridge 2002) always accept the null of validity of the exclusion restrictions. The results con rm that Pc skills lower the probability of retirement only for males with higher education. When the e ects of Pc utilization at work are considered, we obtain analogous results. In particular, only for males with high education we nd that the use of a Pc at work entails a lower probability of retirement. As in Friedberg (2003) IV estimates results in a higher reduction of the retirement probability with respect to the OLS estimates. In Figure (18) we consider the speci cations which allow for both having Pc skills and using a Pc at work. For both genders we do not nd any signi cant result. We note that for males the point estimates of the overall e ect of using a Pc at work do not change remarkably with respect to the OLS case but with IV the standard errors are larger and this fact leads to the insigni cance of the parameters. This mimics the results obtained by Schleife (2006), where IV estimates are imprecise. Schleife (2006) points out how, once controlling for endogeneity, Pc use at work seem not to a ect signi cantly the probability of retirement. Our ndings show that, for male employees with high education, having Pc skills and using Pc at work, taken separately, lead to a rise in the probability of continuing to work. These di erences may be produced by a variety of reasons. First, while Schleife (2006) uses a sample consisting of both employees and self-employed, we focus on employees only. Second, considering our results from nonparametric duration analysis, we allow the impact of Pc knowledge on retirement to vary with education. Our ndings suggest that for males the e ect of interest is actually a ected by educational attainments. This is consistent with the hypothesis that Pc skills and education are complements and that the di usion of personal computers in the economy leads to an increase of the demand for highly skilled labor. According to this view, only high education workers are able to fully exploit information technologies and hence more likely to nd the conditions (for instance in terms of expected wages) that make the postponement of retirement preferable. Using the linear probability model we obtain results similar to those found by Friedberg (2003) for the US. We have a signi cant reduction of the retirement probability for better educated males, stronger in the long run than in the short run, and the e ect seems to be underestimated if the potential endogeneity is neglected. We also have OLS evidence that almost all of the e ect for the PC users is driven by the skills e ect. This result is weakened when we switch to the IV estimates, where estimates are much less precise. We therefore wonder to what extent is necessary to use an IV approach in our case: if endogeneity is not a real issue, we might use standard logit and duration analysis to further investigate the relation between skills and retirement Exogeneity test and logit model We conduct the Rivers and Vuong (1988) exogeneity test in order to assess whether Pc utilization variables should be considered exogenous or not. This exogeneity test can be computed according to a two-step procedure: in the rst 12

13 step we compute the residuals of OLS regressions of the endogenous variables (PC skills and use of a PC at work) on the same sets of control factors and additional instruments as before; in the second step we estimate a logit model for the retirement decision on the set of control factors and the residuals obtained in the rst step regressions. If the hypothesis of joint insigni cance of the residuals is not rejected, then we can consider skills and computer use as exogenous. In Figure (19) and Figure (20) we report the results of the test in the long ( ) and in the short ( ) run, respectively: we never reject the null hypothesis of exogeneity. The results do not change if we consider also the interaction terms. If the Pc utilization variables can be considered exogenous, we can estimate more e ciently their e ect on the probability of retirement by means of standard logit regressions. Figure (21) reports the results for the speci cations which allow jointly for having Pc skills and using a Pc at work. Almost all the results of the linear probability models are con rmed: there is a signi cant reduction of the retirement probability for better educated males due to Pc utilization. As for females, in the case of the stricter de nition of retirement, being skilled seems to increase the retirement probability of the low educated. This result is not con rmed when we consider the broader de nition of retirement, and we believe it might be due to the small sample size and the few transitions observed in this case. When the shorter period is considered we get similar results. 4.2 Semiparametric duration analysis Finally, we develop a duration analysis conditioning on the same set of control factors used in Sections and It is worth noting that in this case we used covariates measured as of year 2000 for studying how labor market participation varies between 2000 and 2002, whereas we condition on the information contained in the covariates measured as of year 2002 for studying the transitions which may occur between 2002 and In other words, we use a duration model with time-varying covariates. We follow this strategy because it is important to observe how variables such as household s composition, labor income and other household s income vary across time, being aware of the role played by them in the individual retirement decision. For these reasons, we think it is useful to exploit the possibility provided by SHIW of updating all the control variables. Unfortunately, the information concerning Pc skills and Pc utilization at work is collected only in the wave 2000, and so it remains constant across time. Since in our framework time is represented by age, the control for this variable is implicit in the duration analysis, hence we do not include it among the explicative variables. We assume a proportional hazard model. In other words, we maintain the hypothesis that the individual retirement hazard rate is equal to the product between a baseline, which is common to all individuals, and an exponential function of the covariates of interest, which summarizes in a score the individual 13

14 observed characteristics. We estimate the e ect of the covariates on the hazard rate of interest by means of the Cox model. The Cox model estimates can also be obtained by logit regressions. Because of this fact, the results of the exogeneity test concerning the Pc utilization variables described in Section are still valid and we are allowed to consider exogenous these covariates. Figure (22) contains the results about the e ect of having Pc skills on the hazard rate of retirement. Irrespective of the retirement de nition, we nd signi cant results only for males. In particular we nd that for those with high education having Pc skills leads to a signi cant decrease in the hazard rate. Analogous results are found when we consider the e ects of the Pc utilization at work. In particular, males with high education who use a Pc at work still present a signi cant lower hazard rate of retirement. For low education males we do not nd any signi cant e ect. Contrarily to the evidence provided till now, when the two variables are jointly considered (Figure 23), the Pc skills parameter is no longer signi cant, while using a Pc at work leads to a signi cant decrease in the probability of leaving the initial state of employed for males. 5 Conclusions In this paper we have looked at the highly policy-relevant question concerning the role of technological skills on retirement choices. Using panel data from the Bank of Italy Survey of Household s Income and Wealth we are able to distinguish between having generic Pc skills and using them at work, and to proxy the built-up pension wealth in order to control for the incentives to retire faced by individuals. We focus on individuals who in 2000 are (i) household s heads or spouses, (ii) aged and (iii) employees. Our transition analysis by means of linear probability models show that males with at least a high school degree have a lower probability of retirement if they are able to use a computer, and this reduction is larger if they do use it at work. This result is robust to variation of the period considered (four years , or two years ), to the de nition of retirement, and to the estimation procedure considered (OLS, IV and MLE logit regression). No e ect is consistently detected for lower educated males and for females. Instrumental variables estimates tend to be imprecise, but we show that in our case a Rivers and Vuong (1988) test for the computer skills and use variables does not reject the null hypotheses of exogeneity. This leads us to develop a duration analysis by estimating a semiparametric Cox model, which con rms that high education males who use a Pc tend to retire later. 14

15 6 References Acemoglu, D. (1998), Why Do New Technologies Complement Skills? Directed Technological Change and Wage Inequality, Quarterly Journal of Economics, N.1, Acemoglu, D. (2001), Technical Change, Inequality, and The Labor Market, Journal of Economic Literature, Vol. 40(1), 7-72 Acemoglu, D. (2003), Patterns of Skill Premia, Review of Economic Studies, Vol. 70(2), Banca d Italia (2002), "Italian Household Budgets in 2000", Supplements to the Statistical Bulletin, XII, 6, Rome: Banca d Italia. Bartel A. and Lichtenberg, F. (1991), The Age of Technology and Its Impact on Employee Wages, Economics of Innovation and New Technology, Vol.1, Bartel, A. and Sicherman, N, (1993), "Technological Change and Retirement Decisions of Older Workers," Journal of Labor Economics, Vol. 11(1), Blau, F. and Khan, L.M. (1996) International Di erences in Male Wage Inequality: Institutions Versus Market Forces, Journal of Political Economy, 104, Blundell, R, Meghir, C., and Smith, S. (2002), "Pension Incentives and the Pattern of Early Retirement", Economic Journal, 112, Bound J. and Johnson, G. (1992), Changes in the Structure of Wages in the 1980 s: an Evaluation of Alternative Explanations, American Economic Review, Vol. 82 (3), Coile, C. and J. Gruber (2000), Social Security and Retirement, W.P , Center for Retirement Research, Boston College Coile, C. and J. Gruber (2001), The e ect of Social Security on Retirement in the U.S., mimeo Friedberg, L. (2003), The Impact of Technological Change on Older Workers: Evidence from Data on Computer Use, Industrial and Labor Relations Review, 56 (3), Gustman, A.L., and Steinmeier T.L. (1986), A Structural Retirement Model, Econometrica, 54 (3), Johnson, R. (2000), The e ect of Old-Age Insurance on Male Retirement: Evidence from Historical Cross-Country Data, W.P , Research Division, Federal Reserve Bank of Kansas City Juhn, C., Murphy K. and Pierce, B. (1993), Wage Inequality and the Rise in Returns to Skill, Journal of Political Economy, Vol.101 (3), Katz L. and Murphy, K. (1992), Changes in Relative Wages, : Supply and Demand Factors, Quarterly Journal of Economics, Vol. 107, Krueger, A.B. (1993), How Computers Have Changed the Wage Structure: Evidence from Microdata, , Quarterly Journal of Economics, Vol 108, Krueger, A.B. and Pischke, J.S. (1992), The E ect of Social Security on Labor Supply: a Cohort Analysis of the Notch Generation, Journal of Labor 15

16 Economics, 10 (4), Lazaear, E.P. (1986), Retirement from the Labor Force, in O. Ashenfelter and R. Layard, eds., Handbook of Labor Economics, Vol. I, North Holland, Amsterdam, Lumsdaine R and Mitchell, O.(1999); "New Developments in the Economics of Retirement", in O. Ashenfelter and D. Card eds., Handbook of Labor Economics, Vol. III, North-Holland, Amsterdam, Miniaci R. and M.L. Parisi (2004); "Educate the Parents by Subsidizing Their Children: Challenging the Digital Divide Through Social Interactions", University of Padova D.P Miniaci R. and M. L. Parisi (2006); "Social Interactions and the Digital Divide: Identi cation and Policy Implication", Contributions to Economic Analysis & Policy, 5(1), article 29. Murphy K. and F. Welch (1992), The Structure of Wages, Quarterly Journal of Economics, Vol. 107, Rivers, D. and Vuong, Q.H. (1988), Limited Information Estimators and Exogeneity Tests for Simultaneous Probit Models, Journal of Econometrics 39, Rust, J (1989), A Dynamic Programming Model of Retirement Behaviour, in D.A. Wise, ed., The Economics of Ageing, University of Chicago Press, Chicago, IL, Rust, J. and Phelan, C. (1997), How Social Security and Medicare A ect Retirement Behavior in a World of Incomplete Markets, Econometrica, 65 (4): Rust, J., M. Buchinsky and Benitez-Silva, H. (2001), An Empirical Model of Social Insurance at the End of the Life-Cycle, mimeo Schleife, K. (2006), "Computer Use and Employment Status of Older Workers. An Analysis Based on Individual Data", Labour, 20 (2) Stock, J.H. and Wise, D.A. (1990), Pensions, the Option Value of Work and Retirement, Econometrica, 58 (5), Tanner, S. (1998), "The Dynamics of Male Retirement Behaviour", Fiscal Studies, vol. 19, n. 2, Weinberg, B. A. (2004), "Experience and Technology Adoption", IZA D.P Wooldridge, J. M. (2002), Econometric Analysis of Cross Section and Panel Data, MIT Press, Cambridge, Mass. 16

17 7 Appendix Figure 1: Percentage of users, by age, area and age. Source: Istat Multiscope Survey,

18 Figure 2: Percentage of users, by area, gender and age. Source: Bank of Italy SHIW

19 Figure 3: Percentage of skilled workers using a PC at work, by age, gender, area and survey. 19

20 Italy, employed men Proportion of individuals with Pc skills Age At most intermediate school Source: SHIW Only household heads and their spouses are considered. At least secondary school Figure 4: Pc skills among employed men. Italy, employed women Proportion of individuals with Pc skills Age At most intermediate school Source: SHIW Only household heads and their spouses are considered. At least secondary school Figure 5: Pc skills among employed women. 20

21 Italy, men: density distribution of the age at retirement At most intermediate school At least secondary school Source: SHIW Only household heads and their spouses are considered. Figure 6: Age at retirement for men. Italy, women: density distribution of the age at retirement At most intermediate school At least secondary school Source: SHIW Only household heads and their spouses are considered. Figure 7: Age at retirement for women. 21

22 Figure 8: Strict de nition of retirement, descriptive statistics. 22

23 Figure 9: Broader de nition of retirement, descriptive statistics. 23

24 Italy, male employees aged in 2000 Nonparametric estimates of the survivor function of remaining employed The strict definition of retirement is adopted Age No PC skills PC skills Source: SHIW Only household heads and their spouses are considered. Note: According to the strict definition of retirement, an individual is retired if she declares herself as job pensioner. Figure 10: Surivor functions for men. Italy, male employees with high education aged in 2000 Nonparametric estimates of the survivor function of remaining employed The strict definition of retirement is adopted Age No PC skills PC skills Source: SHIW Only household heads and their spouses are considered. Note: High education means having at least a professional qualification. According to the strict definition of retirement, an individual is retired if she declares herself as job pensioner. Figure 11: Survivor functions for men. 24