Intergenerational Persistence and Misallocation. Section 3

3 EmpiricalEvidence In this section we document the degree of occupational persistence across generations, as well as test the key predictions (Empirical Prediction 1) of the model developed in the previous section. To this end, we use the British Household Panel Survey (BHPS), and in particular we use the same dataset constructed by Lo Lo Bello & Morchio (2013). 3.1 The Data The BHPS is a yearly survey taken by about 10,000 individuals per year in the UK. It started in 1991, and the last availablewave for this study is 2008. The survey is characterizedby a quite high followup rate, with more than 90% of the individuals being interviewed also in the subsequent year. Besides this, each year a number of new households enters the sample. In total, 32,377 individuals were interviewed in the BHPS in the period 1991-2008. We restrict our sample to males aged between 16 and 65, and eventually we are left with 12,982 individuals, for a total of 1,023,888 monthly observations.Individuals report every year a detailed job history of the previous year, listing all the employment/unemploymentspells, along with several characteristicsof the job (among which, the occupational group). In this way, we can construct long time series for each individual (potentiallyup to 216 months) and, more importantly , we can identify transitions at the monthly frequency. Along with a detailed job history, for each individual a large amount of information is typically available, including sex, age, education, 22 occupation, race, marital status, region of residence and much more. One key feature of this dataset is that it allows us to connect individuals to their fathers and to follow them together over time. The jobnding rate is dened as the probability of transiting from unemployment to employment, at the monthly frequency. Wages are calculated dividing the total monthly labor income by the number of hours normally worked per week multiplied by four (the information on hours worked is only available for the current job at the moment of the interview, that is it is recorder yearly). 3.2 Intergenerational Occupational Persistence Our data allow us to study the extent of occupational persistence across generations. We compute the occupational distribution of workers and study the probability with which father and ospring belong to the same occupational group. In order to account for the unequal distribution of workers across occupations, we construct likelihood ratios. For the sake of concreteness, we dene our Pj as the ratio between the probability of belonging to a given occupation j persistence index conditional on the father being also aliated to the same occupation and the unconditional probability of belonging to occupation j. Pj = P (o = j|oF = j) P (o = j) We dene nine occupational groups, following the SOC aggregation by major group (the 1digit level), as established by the Employment Department Group and the Oce of Population Censuses and Surveys. Results are shown in Table 1. We nd a large degree of occupational persistence. Our estimates indicate that likelihood ratios of occupational persistence are always above 1, indicating that workers are more likely to belong to a given occupation if their father is also in that occupation rather than somewhere else. The average likelihood ratio is estimated to be 1.76, meaning that individuals have 76% more chances to be in a 23 Occupational group Managers & Administrators 1.29 Professional 2.60 Associate Professional & Technical 1.62 Clerical & Secretarial 1.26 Craft & Related 1.55 Personal & Protective Service 1.58 Sales 1.34 Plant & Machine 1.94 Agriculture & Elementary 2.67 Average (unweighted) 1.76 Average (weighted) 1.72 Table 1. Occupational Persistence Indexes (Likelihood Ratios). given occupation if their father belongs to that. the occupation. Likelihood Ratio This ranges between 29% and 167%, depending on Interestingly, we do not nd a specic trend of persistence across occupations (that 11 is, persistence is neither higher or lower in highpaying or highskill occupations) . Repeating the same exercise at the 2digit level, we nd that the average unweighted likelihood ratio is 5.69, and the weighted one is 4.71 (see Table 14 in the Appendix). Part of this persistence might be explained by usual socioeconomic variables such as age, education or the region of residence. In order to account for this, we estimate linear probability models in which we regress the probability of belonging to a given occupation (against any other) adding a number of covariates. We estimate such a model for each occupation, with the results shown in Table 2. The estimates reveal that workers are between 1.59 and 15.1 p.p. more likely to belong to the same occupation as their father, coeteris paribus. These are large probability dierences, highly statistically signicant in each of the 9 occupations. 11 This seems to suggest that borrowing constraints are not playing a large role in the occupational choice. We also checked whether, within an occupational group, likelihood ratios vary by father's income. Our results show that occupational persistence decreases only slightly. For instance, if we only consider the top 1/3 earners in each of the occupations, the average likelihood ratio is 1.61 (Table 15 in the Appendix). 24 25 ∗∗∗ 62114 0.196 0.073 (0.003) 0.0365 62114 (0.003) Occ. 2 Occ. 1 ∗∗∗ 0.0335 0.079 62114 (0.005) ∗∗∗ 0.0550 Occ. 3 (3) ∗∗∗ 0.047 62114 (0.005) 0.0411 Occ. 4 (4) ∗∗∗ 0.106 62114 (0.004) 0.151 Occ. 5 (5) ∗∗∗ 0.053 62114 (0.005) 0.0398 Occ. 6 (6) 0.076 62114 (0.006) ∗∗∗ 0.0159 Occ. 7 (7) ∗∗∗ 0.084 62114 (0.003) 0.127 Occ. 8 (8) ∗∗∗ 0.089 62114 (0.006) 0.105 Occ. 9 (9) marital status, ethnicity, quarter. linear probability models, and include a third-degree polynomial in age and dummies for education, region of residence, smoking behavior, Table 2. Regressions of Occupational Choice (dummy that takes value 1 if the ospring is in a given occupation, 0 otherwise); coecient for father in a given occupation (dummy variable, 0 if the father is in some other occupation), standard errors. All models are Standard errors in parentheses ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01 N R2 Father in occupation 9 Father in occupation 8 Father in occupation 7 Father in occupation 6 Father in occupation 5 Father in occupation 4 Father in occupation 3 Father in occupation 2 Father in occupation 1 (2) (1) We also perform some more checks in order to investigate whether occupational persistence is mostly related to regional factors. For instance, belonging to a poor region (or a region where the variety of jobs supplied in the market is limited) might mechanically increase the likelihood ratio. In this sense, we would not be comparing the conditional probability to the correct unconditional probability. We plot the regionspecic weighted average likelihood ratio against the average regional wage and the Herndahl index of occupations, and nd that neither of these two dimensions can predict persistence (results are shown in the Appendix, Figures 4 and 5). This is reassuring that regional factors are not playing too much of a role in our results 12 . 3.3 Occupational Persistence and Occupational Attachment So far we have shown that workers' occupational aliation tends to be highly correlated with the one of their fathers. In this section we investigate whether the phenomenon of occupational persistence is persistent over the lifecycle. It is important to study this, as young workers who are potentially sampling dierent occupations might be the main responsible behind the likelihood ratios estimated in the previous section. More importantly, these workers might be using the father's occupation in a steppingstone fashion (possibly to avoid unemployment), for then rapidly moving on with their career. If this was the case, then occupational persistence would mainly be a shortrun phenomenon, with limited consequences for the allocation of workers to occupations. First, we document that likelihood ratios are not decreasing over the lifecycle: for instance, the average (unweighted) likelihood ratio is 1.88 for workers younger than 20, as opposed to 1.91 for workers aged 2530 and 2.03 for workers older than 30 (see Table 16 in the Appendix). Second, we study the length of the occupational spells of followers, as opposed to the ones of movers. 13 The average occupational tenure is 2.16 years for the followers, and 1.73 years for the movers 12 One . This is important caveat is that we only have 19 regions. It is plausible that the relevant level for the parentchild transmission is ner than that. However, our sample size does not allow us to estimate occupationspecic indexes of persistence at the regional level. 13 When we restrict our attention only to the spells that we observe from the start, we nd again that followers tend to 26 true also for the occupations chosen very early in the career. Figure 1 plots the share of workers still in the rst occupation against the years of labor market experience, separately for followers and movers (this latter variable is here dened at the moment of the start of the spell). Figure 1. Share of workers still in the rst occupation, by years of labor market experience. Source: BHPS. We can see that those workers who start their career in the same occupation as their father's are substantially less likely to experience occupational mobility. For instance, after 2 years from the start of their rst employment spell, 60% of occupational followers will not have experienced yet an occupational 14 change , against 49% of occupational movers. degree of hysteresis At the same time, these statistics reveal that a large characterizes the occupational transitions. In other words, the initial occupation is a good predictor of the occupational aliation even years after the start of the spell. In this sense, the father's inuence on the initial occupational choice may have longlasting consequences for the workers' outcomes and the aggregate allocation. As an additional piece of evidence, we study whether the contemporaneous presence of the father be more attached to their occupation (average tenure is 1.84 years against 1.69). 14 In Figure 1, we do not count ows back into the original occupation as still in nd a dierence between followers and movers that is slightly larger. 27 . If we do that, we the same occupation in the same occupation has an impact on the odds of changing occupations. We run the following regression: OCi,t = α + βπi,t−1 + γXi,t + i,t , where: t−1 OCi,t (32) t is a dummy taking the value 1 if the occupation at time 15 (there is an occupational switch 1 if the occupation of ospring i ) and 0 otherwise; πi,t is dierent from the one at is a dummy variables that takes the value and his father coincide at time t, and 0 otherwise; Xi,t is a vector of control variables that include a third degree age polynomial, dummies for educational categories and occupational groups (observed for the employed, imputed for the unemployed), marital status, ethnic group, smoking behavior (to capture the health level), region of residence and quarter dummies; i,t is the idiosyncratic error term. We estimate Equation 33 with Pooled OLS, Random Eects and Fixed Eects, with the estimates of β shown in Table 3. Dependent Variable: Occupational Change Father in same occupation (πi,t−1 ) (1) (2) (3) POLS ∗∗∗ -0.00811 RE FE -0.00794 ∗∗∗ -0.00872 ∗∗∗ (0.002) (0.002) (0.002) Average insample OC rate 0.0265 0.0265 0.0265 N R2 53208 53208 53208 0.015 - 0.014 - 938 938 Number of pairs Standard errors in parentheses ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01 Table 3. Regressions of Occupational Change (transition from one occupation to another); coecient for father in same occupation last month (dummy variable), standard errors and average occupational change rate in the regression sample. Model 1 is a pooled OLS regression, model 2 is a random eects GLS regression, model 3 is a xed eects regression. All models include a third-degree polynomial in age and dummies for education, gender, region of residence, smoking behavior, marital status, ethnicity, quarter, occupation of employment. We nd that having the father employed in the same occupation is associated with a substantial 15 The occupational switch can take place through unemployment (where we compare the previous and subsequent occupation) or not (direct employmenttoemployment switch). 28 reduction in the likelihood of changing occupation. The estimated impact is in the region of -0.8/0.9% p.p., which represents about one third of the average insample occupational change rate at the monthly frequency (2.65 p.p.). One possible interpretation could be that some workers are more mobile than others in general, and therefore they will happen to be less often in the same occupation as their father, mechanically generating a correlation between the two variables. However, notice that: i) we are exploiting the exact timing of the transitions (using the lagged persistence variable), making this interpretation less likely; ii) in Column 3 we are controlling for individual xed eect, denitely ruling out this kind of explanations. The estimated coecient quite stable across specications suggests indeed that workers are less reluctant to switch occupation when this coincides with the one of the father, even on top of any unobserved xed heterogeneity. In the following subsections, we turn to testing the two main predictions of our model, namely that occupational followers 16 have higher jobnding rates but lower wages, on average. 3.4 Occupational Persistence and Job Finding Rate The rst prediction of our model is that occupational followers, ospring who are in the same occupation as their father, will have on average a higher job nding rate. Given that we observe the employment status and the occupational aliation, we can directly test this prediction. We run the following regression: JFi,t = α + βπi,t + γXi,t + i,t , where: JFi,t otherwise; (33) is dened only for the unemployed, taking the value 1 if a job is found at time πi,t is a dummy variables that takes the value 1 if the occupation of ospring coincide at time t17 , and 0 otherwise; Xi,t t and 0 i and his father is a vector of control variables that include a third degree age 16 Our working denition of occupational follower is based on πi,t . However, what we document in the following for JF rates and wages hold for more stringent denitions of occupational follower (for instance, starting the spell as a follower and then staying there for at least 6 or 12 months). 17 The occupation of the unemployed ospring is assumed to be equal to the one in which the job will be found at the end of the unemployment spell. Moreover, this variable is dened only for those with an employed father. 29 polynomial, dummies for educational categories and occupational groups (observed for the employed, imputed for the unemployed), marital status, ethnic group, smoking behavior (to capture the health level), region of residence and quarter dummies; i,t is the idiosyncratic error term. We estimate Equation 33 with Pooled OLS, Random Eects and Fixed Eects, with the estimates of β shown in Table 4. Dependent Variable: Job Finding Rate (1) POLS ∗∗∗ Father in same occupation (πi,t ) 0.0546 (2) (3) RE 0.0531 FE ∗∗ 0.0546 ∗∗ (0.016) (0.021) (0.026) Average insample JF 0.125 0.125 0.125 N R2 4142 4142 4142 0.057 - 0.046 - 401 401 Number of pairs Standard errors in parentheses ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01 Table 4. Regressions of Job Finding Rate (transition from Unemployed to Employed); coecient for father in same occupation (dummy variable), standard errors and average job nding rate in the regression sample. Model 1 is a pooled OLS regression, model 2 is a random eects GLS regression, model 3 is a xed eects regression. All models include a third-degree polynomial in age and dummies for education, gender, region of residence, smoking behavior, marital status, ethnicity, quarter, occupation of search/employment. We nd that occupational followers rate (+5.4 p.p.) relative to have, on average, a substantially higher monthly jobnding occupational movers. Given that the unconditional probability of nding a job is estimated to be 12.5%, having the father in the same occupation increases the chances of nding employment by about 42%. Importantly, the eect is robust to the inclusion of individual xed eects (column 3), that control for unobserved heterogeneity across individuals. The identication of the individual xed eects is allowed by the panel structure of our data. The coecient in column (3) of Table 4 is estimated by exploiting variation in or not), within the ospring's worklife. π (the presence of the father in the same occupation This is in line with our model, that predicts that even after controlling for the occupation and the xed type (τ ), the father's occupational aliation is still an 30 important determinant of the individual job nding rate. We also nd that these eects are robust to the exclusion of the selfemployed from the estimating sample (Table 17, in the Appendix). Finally, we study whether the impact of πi,t changes with age. We nd that the eect is particularly high (up to +12 p.p.) among the youngest workers, and then monotonically declines thereafter (Figure 6 in the Appendix). This piece of evidence lends support to our interpretation: young workers, who lack experience in the market, are the ones who are expected to depend more heavily on their father's contacts. In a dynamic setting, such as the one that we will develop in Section 4, workers accumulate contacts themselves, and therefore the inuence of their father will fade out over time. 3.5 Occupational Persistence and Wages The second key prediction of our model is that occupational followers will have on average lower wages. A share of those who choose the same occupation as their father's do not pursue their comparative advantage, and therefore they will have on average lower wages. However, according to our model this is entirely due to selection. In fact, upon choosing an occupation, the wage only depends on the individual productivity level in that occupation. That is, the father's occupational aliation does not matter anymore. Thus, the dierence between occupational followers and occupational movers is predicted to disappear after controlling for individual xed eects (that capture the productivity level). In order to test for this hypothesis, we construct for each individual an index of how much time he spent in the same occupation as his father, over his worklife. We do this for two reasons: rst, we want to construct groups that do not change over the lifecycle of individuals; second, it is generally unclear whether we should look at the father's occupation at the moment of the wage observation or at the start of the job spell 18 In 18 . Thus, we dene qi as the fraction of employed worklife spent by individual the appendix (Figure 7) we show that dierences in wages are robust to alternative denitions of . followers 31 i in the occupational according to their father's contacts' availability, then it becomes symptomatic of an underlying mismatch. In this sense, unless we are able to decompose the sources of occupational persistence across generations, it is unclear whether the latter is socially desirable or not. In Section 4, we will construct a structural model of intergenerational transmission and occupational choice, so that we will be able to perform the decomposition exercise in Section 5. 3 Empirical Evidence In this section we document the degree of occupational persistence across generations, as well as test the key predictions (Empirical Prediction 1) of the model developed in the previous section. To this end, we use the British Household Panel Survey (BHPS), and in particular we use the same dataset constructed by Lo Lo Bello & Morchio (2013). 3.1 The Data The BHPS is a yearly survey taken by about 10,000 individuals per year in the UK. It started in 1991, and the last available wave for this study is 2008. The survey is characterized by a quite high followup rate, with more than 90% of the individuals being interviewed also in the subsequent year. Besides this, each year a number of new households enters the sample. In total, 32,377 individuals were interviewed in the BHPS in the period 1991-2008. We restrict our sample to males aged between 16 and 65, and eventually we are left with 12,982 individuals, for a total of 1,023,888 monthly observations. Individuals report every year a detailed job history of the previous year, listing all the employment/unemployment spells, along with several characteristics of the job (among which, the occupational group). In this way, we can construct long time series for each individual (potentially up to 216 months) and, more importantly, we can identify transitions at the monthly frequency. Along with a detailed job history, for each individual a large amount of information is typically available, including sex, age, education, 22 occupation, race, marital status, region of residence and much more. One key feature of this dataset is that it allows us to connect individuals to their fathers and to follow them together over time. The jobnding rate is dened as the probability of transiting from unemployment to employment, at the monthly frequency. Wages are calculated dividing the total monthly labor income by the number of hours normally worked per week multiplied by four (the information on hours worked is only available for the current job at the moment of the interview, that is it is recorder yearly). 3.2 Intergenerational Occupational Persistence Our data allow us to study the extent of occupational persistence across generations. We compute the occupational distribution of workers and study the probability with which father and ospring belong to the same occupational group. In order to account for the unequal distribution of workers across occupations, we construct likelihood ratios. For the sake of concreteness, we dene our Pj as the ratio between the probability of belonging to a given occupation j persistence index conditional on the father being also aliated to the same occupation and the unconditional probability of belonging to occupation j. Pj = P (o = j|oF = j) P (o = j) We dene nine occupational groups, following the SOC aggregation by major group (the 1digit level), as established by the Employment Department Group and the Oce of Population Censuses and Surveys. Results are shown in Table 1. We nd a large degree of occupational persistence. Our estimates indicate that likelihood ratios of occupational persistence are always above 1, indicating that workers are more likely to belong to a given occupation if their father is also in that occupation rather than somewhere else. The average likelihood ratio is estimated to be 1.76, meaning that individuals have 76% more chances to be in a 23 Occupational group Managers & Administrators 1.29 Professional 2.60 Associate Professional & Technical 1.62 Clerical & Secretarial 1.26 Craft & Related 1.55 Personal & Protective Service 1.58 Sales 1.34 Plant & Machine 1.94 Agriculture & Elementary 2.67 Average (unweighted) 1.76 Average (weighted) 1.72 Table 1. Occupational Persistence Indexes (Likelihood Ratios). given occupation if their father belongs to that. the occupation. Likelihood Ratio This ranges between 29% and 167%, depending on Interestingly, we do not nd a specic trend of persistence across occupations (that 11 is, persistence is neither higher or lower in highpaying or highskill occupations) . Repeating the same exercise at the 2digit level, we nd that the average unweighted likelihood ratio is 5.69, and the weighted one is 4.71 (see Table 14 in the Appendix). Part of this persistence might be explained by usual socioeconomic variables such as age, education or the region of residence. In order to account for this, we estimate linear probability models in which we regress the probability of belonging to a given occupation (against any other) adding a number of covariates. We estimate such a model for each occupation, with the results shown in Table 2. The estimates reveal that workers are between 1.59 and 15.1 p.p. more likely to belong to the same occupation as their father, coeteris paribus. These are large probability dierences, highly statistically signicant in each of the 9 occupations. 11 This seems to suggest that borrowing constraints are not playing a large role in the occupational choice. We also checked whether, within an occupational group, likelihood ratios vary by father's income. Our results show that occupational persistence decreases only slightly. For instance, if we only consider the top 1/3 earners in each of the occupations, the average likelihood ratio is 1.61 (Table 15 in the Appendix). 24 25 ∗∗∗ 62114 0.196 0.073 (0.003) 0.0365 62114 (0.003) Occ. 2 Occ. 1 ∗∗∗ 0.0335 0.079 62114 (0.005) ∗∗∗ 0.0550 Occ. 3 (3) ∗∗∗ 0.047 62114 (0.005) 0.0411 Occ. 4 (4) ∗∗∗ 0.106 62114 (0.004) 0.151 Occ. 5 (5) ∗∗∗ 0.053 62114 (0.005) 0.0398 Occ. 6 (6) 0.076 62114 (0.006) ∗∗∗ 0.0159 Occ. 7 (7) ∗∗∗ 0.084 62114 (0.003) 0.127 Occ. 8 (8) ∗∗∗ 0.089 62114 (0.006) 0.105 Occ. 9 (9) marital status, ethnicity, quarter. linear probability models, and include a third-degree polynomial in age and dummies for education, region of residence, smoking behavior, Table 2. Regressions of Occupational Choice (dummy that takes value 1 if the ospring is in a given occupation, 0 otherwise); coecient for father in a given occupation (dummy variable, 0 if the father is in some other occupation), standard errors. All models are Standard errors in parentheses ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01 N R2 Father in occupation 9 Father in occupation 8 Father in occupation 7 Father in occupation 6 Father in occupation 5 Father in occupation 4 Father in occupation 3 Father in occupation 2 Father in occupation 1 (2) (1) We also perform some more checks in order to investigate whether occupational persistence is mostly related to regional factors. For instance, belonging to a poor region (or a region where the variety of jobs supplied in the market is limited) might mechanically increase the likelihood ratio. In this sense, we would not be comparing the conditional probability to the correct unconditional probability. We plot the regionspecic weighted average likelihood ratio against the average regional wage and the Herndahl index of occupations, and nd that neither of these two dimensions can predict persistence (results are shown in the Appendix, Figures 4 and 5). This is reassuring that regional factors are not playing too much of a role in our results 12 . 3.3 Occupational Persistence and Occupational Attachment So far we have shown that workers' occupational aliation tends to be highly correlated with the one of their fathers. In this section we investigate whether the phenomenon of occupational persistence is persistent over the lifecycle. It is important to study this, as young workers who are potentially sampling dierent occupations might be the main responsible behind the likelihood ratios estimated in the previous section. More importantly, these workers might be using the father's occupation in a steppingstone fashion (possibly to avoid unemployment), for then rapidly moving on with their career. If this was the case, then occupational persistence would mainly be a shortrun phenomenon, with limited consequences for the allocation of workers to occupations. First, we document that likelihood ratios are not decreasing over the lifecycle: for instance, the average (unweighted) likelihood ratio is 1.88 for workers younger than 20, as opposed to 1.91 for workers aged 2530 and 2.03 for workers older than 30 (see Table 16 in the Appendix). Second, we study the length of the occupational spells of followers, as opposed to the ones of movers. 13 The average occupational tenure is 2.16 years for the followers, and 1.73 years for the movers 12 One . This is important caveat is that we only have 19 regions. It is plausible that the relevant level for the parentchild transmission is ner than that. However, our sample size does not allow us to estimate occupationspecic indexes of persistence at the regional level. 13 When we restrict our attention only to the spells that we observe from the start, we nd again that followers tend to 26 true also for the occupations chosen very early in the career. Figure 1 plots the share of workers still in the rst occupation against the years of labor market experience, separately for followers and movers (this latter variable is here dened at the moment of the start of the spell). Figure 1. Share of workers still in the rst occupation, by years of labor market experience. Source: BHPS. We can see that those workers who start their career in the same occupation as their father's are substantially less likely to experience occupational mobility. For instance, after 2 years from the start of their rst employment spell, 60% of occupational followers will not have experienced yet an occupational 14 change , against 49% of occupational movers. degree of hysteresis At the same time, these statistics reveal that a large characterizes the occupational transitions. In other words, the initial occupation is a good predictor of the occupational aliation even years after the start of the spell. In this sense, the father's inuence on the initial occupational choice may have longlasting consequences for the workers' outcomes and the aggregate allocation. As an additional piece of evidence, we study whether the contemporaneous presence of the father be more attached to their occupation (average tenure is 1.84 years against 1.69). 14 In Figure 1, we do not count ows back into the original occupation as still in nd a dierence between followers and movers that is slightly larger. 27 . If we do that, we the same occupation in the same occupation has an impact on the odds of changing occupations. We run the following regression: OCi,t = α + βπi,t−1 + γXi,t + i,t , where: t−1 OCi,t (32) t is a dummy taking the value 1 if the occupation at time 15 (there is an occupational switch 1 if the occupation of ospring i ) and 0 otherwise; πi,t is dierent from the one at is a dummy variables that takes the value and his father coincide at time t, and 0 otherwise; Xi,t is a vector of control variables that include a third degree age polynomial, dummies for educational categories and occupational groups (observed for the employed, imputed for the unemployed), marital status, ethnic group, smoking behavior (to capture the health level), region of residence and quarter dummies; i,t is the idiosyncratic error term. We estimate Equation 33 with Pooled OLS, Random Eects and Fixed Eects, with the estimates of β shown in Table 3. Dependent Variable: Occupational Change Father in same occupation (πi,t−1 ) (1) (2) (3) POLS ∗∗∗ -0.00811 RE FE -0.00794 ∗∗∗ -0.00872 ∗∗∗ (0.002) (0.002) (0.002) Average insample OC rate 0.0265 0.0265 0.0265 N R2 53208 53208 53208 0.015 - 0.014 - 938 938 Number of pairs Standard errors in parentheses ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01 Table 3. Regressions of Occupational Change (transition from one occupation to another); coecient for father in same occupation last month (dummy variable), standard errors and average occupational change rate in the regression sample. Model 1 is a pooled OLS regression, model 2 is a random eects GLS regression, model 3 is a xed eects regression. All models include a third-degree polynomial in age and dummies for education, gender, region of residence, smoking behavior, marital status, ethnicity, quarter, occupation of employment. We nd that having the father employed in the same occupation is associated with a substantial 15 The occupational switch can take place through unemployment (where we compare the previous and subsequent occupation) or not (direct employmenttoemployment switch). 28 reduction in the likelihood of changing occupation. The estimated impact is in the region of -0.8/0.9% p.p., which represents about one third of the average insample occupational change rate at the monthly frequency (2.65 p.p.). One possible interpretation could be that some workers are more mobile than others in general, and therefore they will happen to be less often in the same occupation as their father, mechanically generating a correlation between the two variables. However, notice that: i) we are exploiting the exact timing of the transitions (using the lagged persistence variable), making this interpretation less likely; ii) in Column 3 we are controlling for individual xed eect, denitely ruling out this kind of explanations. The estimated coecient quite stable across specications suggests indeed that workers are less reluctant to switch occupation when this coincides with the one of the father, even on top of any unobserved xed heterogeneity. In the following subsections, we turn to testing the two main predictions of our model, namely that occupational followers 16 have higher jobnding rates but lower wages, on average. 3.4 Occupational Persistence and Job Finding Rate The rst prediction of our model is that occupational followers, ospring who are in the same occupation as their father, will have on average a higher job nding rate. Given that we observe the employment status and the occupational aliation, we can directly test this prediction. We run the following regression: JFi,t = α + βπi,t + γXi,t + i,t , where: JFi,t otherwise; (33) is dened only for the unemployed, taking the value 1 if a job is found at time πi,t is a dummy variables that takes the value 1 if the occupation of ospring coincide at time t17 , and 0 otherwise; Xi,t t and 0 i and his father is a vector of control variables that include a third degree age 16 Our working denition of occupational follower is based on πi,t . However, what we document in the following for JF rates and wages hold for more stringent denitions of occupational follower (for instance, starting the spell as a follower and then staying there for at least 6 or 12 months). 17 The occupation of the unemployed ospring is assumed to be equal to the one in which the job will be found at the end of the unemployment spell. Moreover, this variable is dened only for those with an employed father. 29 polynomial, dummies for educational categories and occupational groups (observed for the employed, imputed for the unemployed), marital status, ethnic group, smoking behavior (to capture the health level), region of residence and quarter dummies; i,t is the idiosyncratic error term. We estimate Equation 33 with Pooled OLS, Random Eects and Fixed Eects, with the estimates of β shown in Table 4. Dependent Variable: Job Finding Rate (1) POLS ∗∗∗ Father in same occupation (πi,t ) 0.0546 (2) (3) RE 0.0531 FE ∗∗ 0.0546 ∗∗ (0.016) (0.021) (0.026) Average insample JF 0.125 0.125 0.125 N R2 4142 4142 4142 0.057 - 0.046 - 401 401 Number of pairs Standard errors in parentheses ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01 Table 4. Regressions of Job Finding Rate (transition from Unemployed to Employed); coecient for father in same occupation (dummy variable), standard errors and average job nding rate in the regression sample. Model 1 is a pooled OLS regression, model 2 is a random eects GLS regression, model 3 is a xed eects regression. All models include a third-degree polynomial in age and dummies for education, gender, region of residence, smoking behavior, marital status, ethnicity, quarter, occupation of search/employment. We nd that occupational followers rate (+5.4 p.p.) relative to have, on average, a substantially higher monthly jobnding occupational movers. Given that the unconditional probability of nding a job is estimated to be 12.5%, having the father in the same occupation increases the chances of nding employment by about 42%. Importantly, the eect is robust to the inclusion of individual xed eects (column 3), that control for unobserved heterogeneity across individuals. The identication of the individual xed eects is allowed by the panel structure of our data. The coecient in column (3) of Table 4 is estimated by exploiting variation in or not), within the ospring's worklife. π (the presence of the father in the same occupation This is in line with our model, that predicts that even after controlling for the occupation and the xed type (τ ), the father's occupational aliation is still an 30 important determinant of the individual job nding rate. We also nd that these eects are robust to the exclusion of the selfemployed from the estimating sample (Table 17, in the Appendix). Finally, we study whether the impact of πi,t changes with age. We nd that the eect is particularly high (up to +12 p.p.) among the youngest workers, and then monotonically declines thereafter (Figure 6 in the Appendix). This piece of evidence lends support to our interpretation: young workers, who lack experience in the market, are the ones who are expected to depend more heavily on their father's contacts. In a dynamic setting, such as the one that we will develop in Section 4, workers accumulate contacts themselves, and therefore the inuence of their father will fade out over time. 3.5 Occupational Persistence and Wages The second key prediction of our model is that occupational followers will have on average lower wages. A share of those who choose the same occupation as their father's do not pursue their comparative advantage, and therefore they will have on average lower wages. However, according to our model this is entirely due to selection. In fact, upon choosing an occupation, the wage only depends on the individual productivity level in that occupation. That is, the father's occupational aliation does not matter anymore. Thus, the dierence between occupational followers and occupational movers is predicted to disappear after controlling for individual xed eects (that capture the productivity level). In order to test for this hypothesis, we construct for each individual an index of how much time he spent in the same occupation as his father, over his worklife. We do this for two reasons: rst, we want to construct groups that do not change over the lifecycle of individuals; second, it is generally unclear whether we should look at the father's occupation at the moment of the wage observation or at the start of the job spell 18 In 18 . Thus, we dene qi as the fraction of employed worklife spent by individual the appendix (Figure 7) we show that dierences in wages are robust to alternative denitions of . followers 31 i in the occupational same occupation as his father: P t qi = P πi,t Ei,t , (34) t where Ei,t is a dummy taking value 1 if the individual is employed in period unemployed. The index qi t and 0 if he is ranges from 0 to 1, and measure for how many months (out of all those that an individual has spent employed) his occupation coincided with the one of this father. In Figure 2, we plot the wage proles of three groups: those with Figure 2. qi = 0, qi >= 0.5 and qi = 1. Wage proles by proportion of employed worklife spent in the same occupation as the father. Source: BHPS. We can see that those who spent more time in the same occupation as their father experience wages that are on average lower by up to 20%, throughout their entire life. It is remarkable how this dierence appears to be constant over the lifecycle. We investigate whether these dierences depend on observable heterogeneity across workers. particular, we estimate the following regressions: 32 In log(wi,t ) = α + βqi + γXi,t + i,t , where: in (34); (35) log(wi,t ) is the natural logarithm of the wage (observed at the yearly frequency); qi Xi,t is dened is a vector of control variables that include a third degree age polynomial, dummies for educational categories and occupational groups, a second order degree occupational tenure and potential labor market experience polynomial, marital status, ethnic group, smoking behavior (to capture the health level), rm size (proxy for rm productivity), region of residence and year dummies; i,t is the idiosyncratic error term. We estimate Equation 35 by POLS and RE. Given the impossibility of controlling for unobserved heterogeneity in this regression model, we estimate again the same regression substituting the independent variable qi with πi,t , the timespecic persistency dummy variable: log(wi,t ) = α + βπi,t + γXi,t + i,t , (36) The rst two columns of Table 5 show tell us that individuals who have spent more time in the same occupation as their father tend to earn lower wages, even after adding a number of controls. The discount associated with being in the same occupation is estimated to be about 14.5-15.2%, large and strongly statistically signicant. Columns 3 to 5 of the same table shows estimates of β in Equation (36). We nd an average discount of 7.6% associated to the presence of the father in the same occupation. However, this eect decreases to 2.4% (barely statistically signicant at the 90% condence level) when we allow for RE in the regression model, and it decreases further (it is neither statistically nor economically signicant) when including individual xed eects in the regression. Overall, these results lend strong support to our theory. In the Appendix (Table 18), we show that these results are robust to trimming our sample (cutting top and bottom 1% and 5% of our wage observations). 33 Dependent Variable: Log Hourly Wage (1) Share of time in same occ. as father (qi ) (2) (3) (4) (5) POLS RE FE -0.0763*** -0.0239* -0.000290 (0.014) (0.013) (0.014) POLS RE -0.145*** -0.152*** (0.017) (0.035) Father in same occ. (πi,t ) N R2 Number of pairs 6485 6485 4776 4776 4776 0.623 - 0.604 - 0.624 - 922 - 850 850 Standard errors in parentheses * p < 0.1, ** p < 0.05, *** p < 0.01 Table 5. Regressions of Log Hourly Wage; coecient for share of time spent in same occupation as father (from 0 to 1), standard errors and father in same occupation (dummy variable). Models 1 and 3 are pooled OLS regressions, models 2 and 4 are random eects GLS regressions, model 5 is a xed eects regression. All models include a third-degree polynomial in age and dummies for education and occupation, a second order degree occupational tenure and potential labor market experience polynomial, rm size, region of residence, smoking behavior, marital status, ethnicity, year. 3.6 Unemployment Risk and Wages So far, we have established that occupational followers: i) among the unemployed, tend to spend less time in unemployment; ii) among the employed, tend to earn lower wages. However suggestive, these two pieces of evidence per se do not imply that individuals actually face the tradeo (between employment prospects and wages) that we depicted in our model. Hence, here we try to produce some more compelling evidence, exploiting the entire career of the workers in our sample. For each worker i, we compute the share of time spent employed (a measure of their employment prospects) and the average monthly wage career W̄i (a measure of lifetime labor earnings). 20 consider observations from age 25 onwards 19 Ēi = P P t t Ei,t Ei,t +Ui,t earned throughout his In order to compute these lifetime statistics, we . 19 This measures incorporates the unemployment risk margin, too. We construct it in the following way: rst, we multiply the monthly wage by the number of months that the individual is employed, for each year; then, we sum all of them up over the years, and nally we divide the total by the number of wage observations (to correct for the unbalancedness of our panel). 20 The rationale behind this choice is that we want to make sure not be capturing things that have to do with dierent entry ages in the labor market. 34 We nd that Ēi is positively related to qi , while the opposite is true for W̄i (Figure 3). Occupational followers appear to have at the same time better employment prospects and lower wages. Figure 3. Locally weighted linear polynomial regression (degree 1, bandwidth 0.5) of share of lifetime employed and log average mean wage against the share of time spent as a follower. Source: BHPS. In Table 6, we show the residual partial correlation between the aforementioned variables, controlling for xed characteristics of individuals (education and race). Column 1 and 2 are a simple reection of Figure 3: occupational followers tend to have lower wages (by up to -24%) but better employment prospects (on average, they spend as employed 5.2 p.p. more of their total time spent in the labor force). Interestingly, employment prospects and wages are generally positively correlated (Column 3), but their correlation with qi has dierent signs. The sign of both of these correlations is robust to the introduction of the other variables as a control. That is, conditional on lifetime employment prospects, followers tend to have lower wages (Column 4); and conditional on the average lifetime wage, followers tend to spend more time employed (Column 5). Taken as a whole, the empirical evidence presented here directly speaks to the tradeo predicated in our simple model, strongly suggesting that workers face such a tradeo indeed. In the next sections, we build a dynamic quantitative model of occupational choice, through which 35 Dependent Variable (1) (2) W̄i Share of time in same occ. as father (qi ) -0.239 Ēi ∗∗∗ (0.069) 0.0526 (3) (4) Ēi W̄i ∗∗ Log avg. mean wage (W̄i ) Ēi -0.290 (0.025) (5) ∗∗∗ (0.061) 0.103 ∗∗∗ ∗∗∗ (0.014) 0.108 (0.009) Share of liteme employed (Ēi ) 0.0514 ∗∗∗ (0.009) 2.006 ∗∗∗ (0.169) Controls (educ, race) N R2 X X X X X 524 601 524 524 524 0.064 0.014 0.209 0.267 0.228 Standard errors in parentheses * p < 0.1, ** p < 0.05, *** p < 0.01 Table 6. Regressions of Log average mean wage (W̄i ) and Share of lifetime employed (Ēi ); coecient for share of time spent in same occupation as father (from 0 to 1), log average mean wage and share of lifetime employed; standard errors in parenthesis. All models include dummies for education and race. we perform the decomposition exercise and study the welfare implications of occupational persistence. 36