Prediction Errors: Comparing Objective And Subjective Re-Employment Probabilities DRAFT ONLY. January Abstract

Transcription

1 Prediction Errors: Comparing Objective And Subjective Re-Employment Probabilities Sonja C. Kassenboehmer MIAESR, University of Melbourne January 2012 Abstract Sonja G. Schatz University of Bochum We investigate several misconceptions of people with respect to their reemployment probability which might alter their job search behaviour or their reservation wages in a sub-optimal way. The biggest misconception was found for male professionals who significantly underestimte their re-employment probability. Men with many years of tenure are more likely to overestimate their re-employment probability indicating that people overestimate the value of the acquired firm specific human capital while they underestimate the value of the acquired qualifications and job titles. JEL classification: J6, J64, J01, D8, D84. Keywords: Subjective Re-employment Probability, Prediction Errors, Labour Force Status. DRAFT ONLY Address corresponding author: University of Melbourne, Level 7, Alan Gilbert Building, 161 Barry Street, Victoria 3010, Australia, sonja.kassenboehmer@unimelb.edu.au We would like to thank John Haisken-DeNew for very helpful comments on this paper.

2 1 Introduction Several studies have shown that unemployment is associated with a decrease in well-being, whereas re-employment is associated with an increase in welfare as measured by subjective self-evaluated life satisfaction questions in surveys. Interestingly, very little is known about the divergence in subjective and objective re-employment probabilities for the unemployed, although such a divergence might also have significant implications for the well-being of the unemployed, their search behaviour, their reservation wages and might alter these in a sub-optimal way. Misconceptions about the re-employment probability might result into an insufficient job search effort or is related to biased reservation wages. This paper therefore sets out to close this gap and will answer the following questions: Are unemployed people able to predict precisely their re-employment probabilities or is there a divergence between subjective and objective re-employment probability? Who are the people that make prediction errors? How big are the prediction errors? In past labour economics research, subjective expectations have mainly been used in the context of job insecurity such as in Campbell et al. (2007), who compare objective and subjective probabilities of becoming unemployed or in Green et al. (2000) in which job insecurity and the difficulty of becoming re-employed are analyzed. Other studies on job insecurity using subjective expectations are Dickerson and Green (2009), Green et al. (2001) and Manski and Straub (2000).Other research in labour economics has focused on subjective expectations about future income such as the studies by Pistaferri and Jappelli (2010), Dominitz and Manski (1997) or Kaufmann and Pistaferri (2009) or Jappelli and Pistaferri (2000) who looked at expectations about income and inflation. To our knowledge, there is no comparable study that explicitly looks at re-employment expectations such as ours. Using data from the German Socio-Economic Panel (SOEP), a model-based re-employment probability is estimated (correlated random effects models and conditional logit) and compared with subjective predictions of finding a new job. This shows us whether the unemployed are able to predict accurately their own reemployment probability or whether they under- or overestimate their chances of finding a new job. We compare several error indicators and furthermore subdivide our analysis by gender, which will impact on the interpretation of the results. 1

3 We find several misconceptions of people with respect to their re-employment probability. The biggest misconception is found for male professionals with a 44 percentage points higher probability of underestimating their re-employment probability and a prediction error that is 27 percentage points bigger compared to the elementary occupations. Also tenure of 10 or more years was found to have a significant impact of the direction and the size of the prediction error. For women, past unemployment experience is the biggest predictor of a misconception with respect to the re-employment probability. Female machine operators also underestimate their re-employment probability with a prediction error that is 20 percentage points bigger (more negative) than the one for the elementary occupations. The remainder of this paper is organized as follows: Section 2 outlines the empirical framework, section 3 describes the data and section 4 presents the results. Section 5 concludes. 2 Empirical Framework 2.1 Statistical Re-employment Probability To obtain an estimation for the objective re-employment probability for the unemployed, the following labour force status regression is estimated: P r(employed i+1,i+2 = 1 X) = Λ(x it β + α i + ɛ it ) α i = x i γ + υ i i = 1,..., N k = 1,..., K (1) where Λ is the cumulative distribution function (CDF) of the logistic function and x it consists of the explanatory variables. The vector β is the coefficient vector to be estimated by maximum likelihood, ɛ it consists of the the error terms and α i are the individual random effects where we allow for some correlation of the random effect and the explanatory variables of the form α i = x i γ + υ i ala Mundlak (1978). The dependent variable 2

4 employed i+1,i+2 is the labor force status of the respondent in the following two years (t+1, t+2) indicating whether one is employed (=1) or unemployed (=0): 1, employed in t+1 and/or t+2 employed i+1,i+2 = 0, remaining unemployed in t+1 and t+2 (2) The predictions of employed i+1,i+2 represent the statistical re-employment probablities for every individual which are based on the actually observed labour force states of the individuals in t+1 and t+2. These statistical re-employment probabilities are then compared to the reported subjective re-employment probabilities in time t in order to determine who are the people that make prediction errors. Because the reported subjective re-employment probability is a discrete variable but the statistical actual re-employment probability is a continous variable, the variables have to be made comparable and some assumptions have to be imposed. 2.2 Subjective Re-employment Probability For the statistical re-employment probability, we obtain a continuous variable from the predictions based on the correlated random effects logit model as described above. However, the reported subjective re-employment probability is an ordinal variable on a 11- point scale. Apart from the simple straightforward comparison of these two variables, it is reasonable to transform the reported subjective re-employment probability into a continuous variable so that both variables are on the same scale and thus directly comparable. A two-sided correlated random effects Tobit regression model is estimated with the reported subjective probability as the dependent variable (limited between 0 percent and 100 percent). The prediction serves as a continuous approximation of the true underlying subjective re-employment probability. The Tobit model estimates the following regression equation: ly i = x it β + α i + ɛ it ɛ it N(0, σ 2 ) (3) 3

5 where x it consists of the explanatory variables. The vector β is the coefficient vector to be estimated by maximum likelihood, ɛ it consists of the the error terms and α i are the individual random effects where we allow for some correlation of the random effects and the explanatory variables of the form α i = x i γ + υ i ala Mundlak (1978). It is assumed that there is a latent variable y i such that: y i = y i if 0 < y i < if y i if y i 100 (4) The prediction for y i of the Tobit model serves as a linear approximation for the discrete reported subjective re-employment probability. 2.3 Prediction Errors Three types of prediction errors which will serve as the dependent variables in a regression analysis are calculated: 1. Error1 = Discrete Reported Subjective Probability - Model based Statistical Probability 2. Error2 = Approximated Continous Subjective Probability - Model based Statistical Probability 3. Error3 = Error2/Model based Statistical Probability The first error compares the originally reported ordered subjective re-employment probability with the predicted objective re-employment probability from the correlated random effects labour force status regression of the previous section. The second error compares the predicted continuous subjective re-employment probability from the correlated random effects Tobit regression with the predicted statistical re-employment probability from the correlated random effects labour force status regression of the previous section. The third error simply divides the second error by the model based statistical probability, to put the absolute prediction error in relation to the size of the objective re-employment probability. 4

6 3 Data For the analysis in this paper the German Socio-Economic Panel (SOEP) is used which is a longitudinal representative panel dataset of private households in Germany starting in The SOEP re-interviews the same private households annually and thereby approximately households and people are sampled every year. It consists of 8 subsamples with extra oversampling for certain minority groups such as foreigners. The SOEP collects information on objective information like education, health or labor force status as well as subjective informatio like opinions on several domains or life satisfaction 1. The focus in this paper is on the question concerning subjective expectation about reemployment of the unemployed: How likely is it that you start paid work within the next two years?. The responses range on a 11-point scale from 0 percent to 100 percent. Our sample is based on ages 16 to 64 for the years 1985 until People have to be unemployed by definition which reduces our sample significantly. Furthermore, we will only use the waves 1999, 2001, 2003, 2005 and 2007 since the subjective re-employment probability is only asked every two years. We will use the two subsequent years to estimate the objective probability that someone is employed within the next two years after his initial unemployment status. Thus we are left with 2312 observations. As explanatory variables, we will use several additional socio-economic and labor market specific variables such as age, health status, labour force experience (part time and full-time work, unemployment, tenure), education, last occupation, socio-economic status measured by the international Socio-Economic Index of Occupational Status 2, marital status, home ownership, children in household and last month s labour income. Furthermore we control for the unemployment rate to capture the macro-economic influences on one s subjective re-employment probability. The variables in our analysis are described in more detail in the appendix. 3 1 For more information see 2 For more details see Ganzeboom et al. (1992) 3 All variables except for the unemployment rate, which comes from the Federal Statistical Office, have been extracted from the SOEP using PanelWhiz. For more information see Haisken-DeNew and Hahn (2006). 5

7 4 Regression Analysis 4.1 Statistical Re-employment Probability Table 2 shows the results for the labour force status model with re-employment in t+1 and/or t+2 as the dependent variable. The first three columns (1) to (3) of Table 2 show the marginal effects of regressions assuming no individual random effect. The results for model (1) to (3) are quite stable regardless of whether one includes as an additional explanatory variable the subjective re-employment probability or the unemployment rate. Being still unemployed two years after unemployment is more likely for older persons, persons with more unemployment experience and with children. On the other hand, being a home owner and high last labour income are positively related to re-employment in the future. The occupations that are related to a higher re-employment probability compared to the elementary occupations are the probably highly versatile occupations clerks and service/shop workers. The subjective re-employment probability is still significantly related to the statistically observable re-employment probability once many other characteristics are controlled for as columns 2 and 3 show. Gross last labour income becomes insignificant though, once the subjective re-employment probability and additionally the unemployment rate are controlled for as in columns (2) and (3) indicating that these variables are highly correlated and significantly influence future employment status. Hence, subjective re-employment probabilities are based on the one hand on past income and on the unemployment rate and on the other hand additionally refelect certain unobserved labour market relevant characteristics of the person, such as motivation or persuasiveness. Column (4) and (5) of Table 2 furthermore controls for other unobserved individual specific random effects as described above. It is important to control for unobserved effects in the most optimal way to reduce the bias of the estimated coefficients that might result from a correlation of the individual effects and the explanatory variables if both are related to the dependent variable. Controlling for correlated random effects indeed renders some of the coefficients that were significant before insignificant (age, years in education, home owner). Now the coefficients for part-time and full-time experience become significantly negative, 6

8 meaning that an additional year of part-time experience decreases the probability of reemployment by 4 percentage points. Assuming a certain correlation structure of the random effects with the explanatory variables might not be the most ideal way to deal with unobserved heterogeneity. Therefore as a sensitivity test, a conditional logit model was also estimated in column (6) which relaxes the assumed correlation structure of the random effects and allows for unobserved fixed effects instead with any correlation between the fixed effects and the explanatory variables. However, the problem with this estimator is that we lose a lot of information because not only the explanatory variables have to vary for each individual in order to be able to be used in the estimation but also the dependent variable has to vary for every individual, otherwise the individual is dropped from the analysis. This leads to a special sample that is considered in the estimation which only consists of people who were at least unemployed twice and became at least once re-employed and once not re-employed within the next 2 years. Thus not surprisingly, fixed effects estimation shows very different results from the correlated random effects estimation. That the different results between fixed effects estimation and random effects estimation are not due to the different estimation techniques but due to the quite different samples is shown when the correlated random effects estimator is applied to the smaller sample of 563 and the same results as with the conditional logit model emerge as seen in column (7). Another drawback of the conditional logit fixed effects estimator is that it does not allow us to predict the estimated probabilities because the fixed effects are not calculated by the model. We will therefore focus in the subsequent analysis on the correlated random effects estimator. Because male and female labour supply behaviour is in general quite different from one another, Table 3 further splits the estimation of column (4) by gender. It is interesting to note that the negative effects of previous part-time experience seem to stem solely from females. This effect probably reflects that women who have been in part-time work prior to unemployment might not be so attached to the labour market and therefore will remain unemployed within the next two years. Also prior unemployment experience seems to be only negatively related to employment status for women. Furthermore, women s re-employment is significanly influenced by the unemployment rate. An increase in the 7

9 unemployment rate of 1 percentage point is associated with a decrease in female s reemloyment probability of 3 percentage points. Women with children have a 11 percentage points lower probability of re-employment in the next two years. 4.2 Subjective Re-employment Probability In order to make the reported subjective re-employment probability comparable to the statistically derived continous re-employment probability, the prediction for the subjective re-employment probability based on the correlated random effects tobit model is used as a continous approximation for the true underlying subjective re-employment probability. The regressions results are shown in Table 4. It can bee seen that both men and women base their subjective re-employment probability on their information about their age, full-time work experience and unemploment experience. Women with more past tenure additionally expect a lower re-employment probability. Men base their subjective reemployment expectations also on the unemployment rate, although as seen from the labour force status regressions before, the unemployment rate did not influence their actual re-employment probability whereas it influenced the women s re-employment probability indiacating some sort of discrimination of women in the labour market if current economic conditions are bad. Figure 1 compares the predictions from these tobit regressions with the statistical reemployment probabilities. Indeed the approximation by the tobit regressions seems to be quite close to the original ordered variable on the 11-point scale. Percent Reported Subj Re Empl Prob Percent SubProb_neu_det Figure 1: Histograms of Reported and Estimated Re-Employment Probability 8

10 4.3 Prediction Errors In a first step, an ordered logit regression is estimated to determine those people who exactly predict their re-employment probability and those who over- and underestimate. Table 1 shows the percentage of people under-, over- or exactly estimate their re-employment probability by year. The prediction error is here the originally reported subjective reemployment probability compared to the objective model-based re-employment probability rounded to the nearest 10. It becomes obvious that in this case around 9 to 15 percent are estimated correctly and about 40 to 50 percent over-and under-estimate their re-employment probability. Table 1: Error by Year (rounded model based Empl Prob), (in %) Year (4-digit) Subjective under-, exact or overestimation Total % % % % % % underestimated estimated correctly overestimated Total Table 5 shows the results of the ordered logit regressions for males. Columns (2) to (4) show the marginal effects for under-, exact and overestimation respectively. Five characteristics seem to play a role. The most important one is the occupation of being a professional (44 percentage points higher probability than the elementary occupations to underestimate) followd by technicians and associate professionals (24 percentage points higher probability to underestimate). Hence professionals get a job faster than expected. Having had tenure for 10 years or more is associated with a higher probability of overestimation than having had tenure of less than 3 years (13 percentage points). Obviously people do not seem to realize that they have acquired a high amount of firm specific human capital which is not directly transferable to other companies. Although people with high amount of tenure are in general paid more than people with low tenure, the higher earnings usually only compensate the previous payments below productivity for low tenure and are nor representative of the current productivity level of the employee. Home ownership is associated with a higher probability of re-employment which on the one hand might reflect unobserved characteristics of the home owners which they might not themselves be aware of and on the other hand could also be due to the higher pressure 9

11 of home owners to regain employment because of the financial commitment to owning a house. Having children is associated with an increased probability of overestimation (7 percentage points). The last charactersitic that is associated with a higher probability of overestimation is the occupational status (measured by the last ISEI status). This points out that given the actual human capital acquired (controled for by the profession and education level), there is not the additional benefit of having a high occupational status as people with this status seem to expect. This is probably due to the fact that occupational status is attached to a firm and not necessarily to the person once the person becomes unemployed. For women there are not so big and strong associations between the characteristics and the prediction error. The biggest effects are found for the unemployment experience. Women with an unemployment experience of about 1 year have a 30 percentage points higher probability to underestimate their re-employment probability. Similarly to men being a home owner is associated with a 8 percentage points higher probability of overestimating the re-employment probability. Female machine operators have a 26 percentage points higher probability to underestimate their re-employment probability. If there is a high unemployment rate, women are more likely to underestimate their re-employment probability. Not only is it interesting to determine who makes prediction errors and in what direction, but also to determine the extent to which the subjective re-employment probability and the statistical re-employment probability differ for the people who over- and underestimate their re-employment probability. Figure 2 shows kernel density estimates of the prediction errors 2 and 3, both centering around 0 which will serve as the dependent variables in the next step. Table 7 and Table 8 therefore show the estimation results for the three types of dependent variables, the different prediction errors as described in section 2.3, each column reporting results from OLS regression and fixed effects regression. We focus here on those people where the previous analysis showed that their characteristics were associated with an over- or under-estimation. As was shown before, male professionals and technicians and associated professionals had a 43 and 24 percentage points higher probabiliy of underestimating their re-emplozment probability. Column (1) 10

12 Density Kernel density estimate Density Kernel density estimate Prediction Error kernel = epanechnikov, bandwidth = Prediction Error/Obj Re Empl Prob kernel = epanechnikov, bandwidth = Figure 2: Kernel Density Estimates of Prediction Error and Error/Objective Probability of Table 7 shows that they have a prediction error which is 28 and 21 percentage points lower than the prediction errror of elementary occupations. Men with tenure of 10 or more years had a 13 percentage points higher probability of overestimation and the exact amount of overestimation can now be determined to be 10 percentage points higher than those those with less than three years of tenure. Home ownership was associated with a 11 percentage points higher probability of underestimation than non home owners and the prediction error is now determined to be 10 percentage points bigger (more negative) than the one of non home owners. Children were associated with an increased probability of overestimation (7 percentage points). It is now seen that the error is 5 percentage points bigger than the error of those without children. The occupational was associate with a 0.7 percentage points higher probability of overestimation and the prediction error is now found to be 4 percentage points bigger. Comparing the coefficients of these variables from column (1) with the coefficients of column (2) where the dependent variable is prediction error 2, is becomes obvious that the results are not significantly different from each other. Hence the results are robust to different specifications of the prediction error. It is very interesting that the OLS results in columns (1) and (3) and the fixed effects results in columns (2) and (4) differ substantially, indicating that unobserved effects are correlated with some of the explanatory variables. 11

13 The last two columns show the results for Error3 as the dependent variable. It can be seen that the results are robust to the location of the statistical prediction probability in its distribution. Now moving on to the prediction error models for women in Table 8, the variable to be mostly (positively) associated to the probability of underestimation (by 30 percentage points), unemployment experience of one year, is found to be associated with a prediction error of 16 percentage points lower than no unemployment experience. Machine operators having a probability of 26 percentage points of underestimation are found to have a prediction error which is 20 percentage points lower than for the elementary occupations. Smaller effects are found for female home owners who had a higher probability of underestimation although very low (8 percentage points). The prediction error of female home owners is not different of non home owners. One additional year in education which also only had a slightly higher probability of underestimation (3 percentage points) is associated with a prediction error that is 1.5 percentage points smaller compared to one year less of education. A higher unemployment rate associated with a higher probabilitz of underestmation (1 percentage points) is also associated with a prediction error 0.5 percentage points lower. Again comparing the results for the two different prediction errors, the coefficients of the variables of interest associated with an under or overestimation of the re-employment probability, are not significantly different from each other. Also the coefficients for the regressions with prediction error 3 as the dependent variable confirm these results (last two columns). Again the results of the fixed effects resgressions are quite different indicating unobserved heterigeneity. 5 Conclusions In this study, we found several misconceptions of people with respect to their reemployment probability which might alter their job search behaviour or their reservation wages in a sub-optimal way. If people underestimate their re-employment probability as was found for male professionals, this might result for example into an insufficient job search effort or is related to biased reservation wages. The biggest misconception was 12

14 found for male professionals with a 44 percentage points higher probability of underestimating their re-employment probability and a prediction error that is 27 percentage points bigger compared to the elementary occupations. Also tenure of 10 or more years was found to have a significant impact of the direction and the size of the prediction error. People with many years of tenure are more likely (by 13 percentage points) to overestimate their re-employment probability and the pedicition error is 11 percentage points bigger than for those with tenure of less than 3 years. This indicates that people overestimate the value of the acquired firm specific human capital while they underestimate the value of the aquired qualifications and job titles. For women, past unemployment experience is the biggest predictor of a misconception with respect to the re-employment probability. Women with an unemployment experience of one year have a 30 percentage points higher probability of underestimating their reemployment probability with a prediction error that is 16 percentage points bigger (more negative) than the one of those without previous unemployment experience. Female machine operators also underestimate their re-employment probability with a prediction error that is 20 percentage points bigger (more negative) than the one for the elementary occupations. 13

15 Table 2: Labour Force Status Models Logit Models, Marginal Effects (ME) CRE Logit Cond.Logit CRE Logit (1) (2) (3) (4) (5) (6) (7) Model 1 Model 2 Model 3 ME Coeff. Coeff. Coeff.,Sample=(6) main Age (0.0026) (0.0025) (0.0025) (0.0087) (0.0552) (0.1216) (0.0907) Yrs in Education (0.0067) (0.0066) (0.0068) (0.0203) (0.1281) (0.3425) (0.1979) Part Time Exp (0.0041) (0.0040) (0.0040) (0.0181) (0.1166) (0.4410) (0.2638) Full Time Exp (0.0025) (0.0025) (0.0025) (0.0116) (0.0751) (0.2949) (0.1308) 1.Unempl. Exp. 0,1-0,9yrs (0.0432) (0.0421) (0.0424) (0.0506) (0.3172) (1.0212) (0.8657) 1.Unempl. Exp. 1yr (0.0688) (0.0690) (0.0687) (0.0817) (0.5012) (1.2455) (1.1204) 1.Unempl. Exp yrs (0.0374) (0.0371) (0.0374) (0.0484) (0.3174) (1.1055) (0.8348) 1.Unempl. Exp yrs (0.0427) (0.0432) (0.0435) (0.0627) (0.3816) (1.2390) (0.8865) 1.Unempl. Exp. 5+yrs (0.0470) (0.0485) (0.0491) (0.0868) (0.4796) (1.3939) (0.9906) 1.Tenure last Job 3-9 Years? (0.0274) (0.0266) (0.0264) (0.0400) (0.2769) (0.9027) (0.5063) 1.Tenure last Job 10 or more Years? (0.0384) (0.0381) (0.0380) (0.0515) (0.4062) (2.1205) (0.6728) 1.Handicapped (0.0440) (0.0427) (0.0428) (0.0645) (0.3986) (0.8743) (0.6344) 1.Married (0.0242) (0.0239) (0.0239) (0.0487) (0.3067) (0.7507) (0.4682) 1.Home Owner (0.0229) (0.0224) (0.0224) (0.0528) (0.3370) (0.9348) (0.5789) 1.Children (0.0232) (0.0229) (0.0228) (0.0353) (0.2266) (0.4503) (0.3565) Log Last Income (Gross) (0.0141) (0.0136) (0.0139) (0.0241) (0.1517) (0.3638) (0.2448) Last ISEI Status (0.0017) (0.0017) (0.0017) (0.0035) (0.0218) (0.0729) (0.0476) 1.Managers (0.0878) (0.0811) (0.0793) (0.1597) (0.9674) (2.9633) (2.1484) 1.Professionals (0.0816) (0.0810) (0.0795) (0.1727) (1.1368) (3.5117) (2.3950) 1.Techn./Assoc. Profess (0.0601) (0.0588) (0.0579) (0.1228) (0.7965) (2.6844) (1.6616) 1.Clerks (0.0506) (0.0508) (0.0504) (0.1030) (0.6760) (2.0450) (1.2921) 1.Service/Shop Workers (0.0412) (0.0406) (0.0406) (0.0925) (0.5610) (2.1452) (1.1242) 1.Agricult. Workers (0.0519) (0.0505) (0.0493) (0.0808) (0.6337) (1.1326) (0.8856) 1.Craft Workers (0.0377) (0.0375) (0.0372) (0.0722) (0.4581) (1.1020) (0.7755) 1.Machine Operators (0.0423) (0.0423) (0.0414) (0.0842) (0.5238) (1.5432) (0.9708) Reported Subj Re-Empl Prob (0.0003) (0.0003) (0.0007) (0.0044) (0.0055) (0.0058) Unempl Rate (0.0021) (0.0063) (0.0401) (0.0855) (0.0675) Pseudo R Number of Observations Marginal effects p < 0.1, p < 0.05, p < 0.01 Note: SOEP 1999, 2001, 2003, 2005,2007. LHS variable is Reemployment in the following two years t+1 or t+2 14

16 Table 3: Labour Force Status Model by Gender, CRE Logit (1) (2) Males Females Age (0.0141) (0.0111) Yrs in Education (0.0280) (0.0294) Part Time Exp (0.0498) (0.0200) Full Time Exp (0.0171) (0.0177) 1.Unempl. Exp. 0,1-0,9yrs (0.0683) (0.0730) 1.Unempl. Exp. 1yr (0.1021) (0.1355) 1.Unempl. Exp yrs (0.0705) (0.0640) 1.Unempl. Exp yrs (0.0855) (0.0825) 1.Unempl. Exp. 5+yrs (0.1010) (0.1077) 1.Tenure last Job 3-9 Years? (0.0505) (0.0651) 1.Tenure last Job 10 or more Years? (0.0666) (0.0792) 1.Handicapped (0.0813) (0.0983) 1.Married (0.0692) (0.0670) 1.Home Owner (0.0710) (0.0776) 1.Children (0.0515) (0.0475) Log Last Income (Gross) (0.0348) (0.0325) Last ISEI Status (0.0057) (0.0043) 1.Managers (0.1825) (0.2319) 1.Professionals (0.1887) (0.2425) 1.Techn./Assoc. Profess (0.1278) (0.1683) 1.Clerks (0.1454) (0.1396) 1.Service/Shop Workers (0.1741) (0.1090) 1.Agricult. Workers (0.0947) (0.1452) 1.Craft Workers (0.0917) (0.1329) 1.Machine Operators (0.0971) (0.1783) Reported Subj Re-Empl Prob (0.0009) (0.0011) Unempl Rate (0.0092) (0.0085) Pseudo R 2 Number of Observations Marginal effects p < 0.1, p < 0.05, p < 0.01 Note: SOEP 1999, 2001, 2003, 2005,

17 Table 4: Subj Re-Empl Tobit Model by Gender (1) (2) Males Females Age (0.6958) (0.5843) Yrs in Education (1.4430) (1.6204) Part Time Exp (2.3657) (1.1059) Full Time Exp (0.8504) (0.9497) 1.Unempl. Exp. 0,1-0,9yrs (3.2351) (3.7012) 1.Unempl. Exp. 1yr (5.3244) (6.6409) 1.Unempl. Exp yrs (3.4830) (3.7234) 1.Unempl. Exp yrs (4.6077) (4.6631) 1.Unempl. Exp. 5+yrs (6.1790) (5.9067) 1.Tenure last Job 3-9 Years? (2.8186) (3.5665) 1.Tenure last Job 10 or more Years? (4.6218) (5.2651) 1.Handicapped (4.4878) (5.4254) 1.Married (3.6770) (3.7317) 1.Home Owner (3.6613) (4.3518) 1.Children (2.6996) (2.6708) Log Last Income (Gross) (1.7828) (1.7816) Last ISEI Status (0.2842) (0.2283) 1.Managers ( ) ( ) 1.Professionals ( ) ( ) 1.Techn./Assoc. Profess (9.7207) (8.9246) 1.Clerks (8.3227) (7.3078) 1.Service/Shop Workers (8.9636) (5.6224) 1.Agricult. Workers (6.2398) (9.0978) 1.Craft Workers (4.7015) (7.0332) 1.Machine Operators (4.9318) (9.6720) Unempl Rate (0.4653) (0.4625) Pseudo R 2 Number of Observations Marginal effects p < 0.1, p < 0.05, p < 0.01 Note: SOEP 1999, 2001, 2003, 2005,

18 Table 5: For Males: Ordered Logit Model: Underestimate, Exactly Estimate, Overestimate (rounded) (1) (2) (3) (4) ordered logit underestimation exact estimation overestimation main Age (0.0212) (0.0047) (0.0004) (0.0051) Yrs in Education (0.0421) (0.0093) (0.0008) (0.0101) Part Time Exp (0.0408) (0.0090) (0.0008) (0.0097) Full Time Exp (0.0210) (0.0046) (0.0004) (0.0050) 1.Unempl. Exp. 0,1-0,9yrs (0.2055) (0.0452) (0.0040) (0.0490) 1.Unempl. Exp. 1yr (0.3776) (0.0832) (0.0070) (0.0902) 1.Unempl. Exp yrs (0.2172) (0.0478) (0.0042) (0.0518) 1.Unempl. Exp yrs (0.2599) (0.0573) (0.0048) (0.0619) 1.Unempl. Exp. 5+yrs (0.2877) (0.0634) (0.0053) (0.0687) 1.Tenure last Job 3-9 Years? (0.1520) (0.0335) (0.0028) (0.0363) 1.Tenure last Job 10 or more Years? (0.2293) (0.0501) (0.0047) (0.0542) 1.Handicapped (0.2349) (0.0515) (0.0045) (0.0559) 1.Married (0.1646) (0.0363) (0.0031) (0.0393) 1.Home Owner (0.1412) (0.0306) (0.0032) (0.0330) 1.Children (0.1365) (0.0299) (0.0028) (0.0324) Log Last Income (Gross) (0.0959) (0.0211) (0.0018) (0.0229) Last ISEI Status (0.0114) (0.0025) (0.0002) (0.0027) 1.Managers (0.5600) (0.1229) (0.0111) (0.1332) 1.Professionals (0.6191) (0.1347) (0.0141) (0.1462) 1.Techn./Assoc. Profess (0.4326) (0.0948) (0.0090) (0.1026) 1.Clerks (0.3573) (0.0783) (0.0072) (0.0850) 1.Service/Shop Workers (0.3604) (0.0791) (0.0071) (0.0859) 1.Agricult. Workers (0.3094) (0.0681) (0.0059) (0.0739) 1.Craft Workers (0.2259) (0.0497) (0.0043) (0.0538) 1.Machine Operators (0.2560) (0.0564) (0.0048) (0.0611) Unempl Rate (0.0127) (0.0028) (0.0003) (0.0030) cut1 Constant (0.8221) cut2 Constant (0.8206) Pseudo R Number of Observations Number Cluster p < 0.05, p < 0.01, p < Note: SOEP 1999, 2001, 2003, 2005,

19 Table 6: For Females: Ordered Logit Model: Underestimate, Exactly Estimate, Overestimate (rounded) (1) (2) (3) (4) ordered logit underestimation exact estimation overestimation main Age (0.0155) (0.0036) (0.0002) (0.0033) Yrs in Education (0.0414) (0.0095) (0.0007) (0.0089) Part Time Exp (0.0256) (0.0059) (0.0004) (0.0055) Full Time Exp (0.0160) (0.0037) (0.0002) (0.0035) 1.Unempl. Exp. 0,1-0,9yrs (0.2583) (0.0596) (0.0038) (0.0559) 1.Unempl. Exp. 1yr (0.5877) (0.1344) (0.0092) (0.1266) 1.Unempl. Exp yrs (0.2519) (0.0581) (0.0036) (0.0546) 1.Unempl. Exp yrs (0.2811) (0.0649) (0.0040) (0.0609) 1.Unempl. Exp. 5+yrs (0.2864) (0.0659) (0.0043) (0.0620) 1.Tenure last Job 3-9 Years? (0.1888) (0.0434) (0.0028) (0.0408) 1.Tenure last Job 10 or more Years? (0.3162) (0.0728) (0.0045) (0.0685) 1.Handicapped (0.2754) (0.0634) (0.0040) (0.0596) 1.Married (0.1608) (0.0370) (0.0024) (0.0348) 1.Home Owner (0.1561) (0.0358) (0.0025) (0.0337) 1.Children (0.1509) (0.0346) (0.0023) (0.0326) Log Last Income (Gross) (0.0941) (0.0217) (0.0014) (0.0204) Last ISEI Status (0.0100) (0.0023) (0.0001) (0.0022) 1.Managers (0.5691) (0.1313) (0.0080) (0.1233) 1.Professionals (0.5427) (0.1250) (0.0076) (0.1176) 1.Techn./Assoc. Profess (0.3739) (0.0863) (0.0053) (0.0810) 1.Clerks (0.3382) (0.0777) (0.0049) (0.0732) 1.Service/Shop Workers (0.2791) (0.0644) (0.0040) (0.0605) 1.Agricult. Workers (0.3395) (0.0783) (0.0052) (0.0733) 1.Craft Workers (0.3900) (0.0898) (0.0056) (0.0844) 1.Machine Operators (0.4655) (0.1065) (0.0074) (0.1004) Unempl Rate (0.0145) (0.0033) (0.0003) (0.0031) cut1 Constant (0.8294) cut2 Constant (0.8283) Pseudo R Number of Observations Number Cluster p < 0.05, p < 0.01, p < Note: SOEP 1999, 2001, 2003, 2005,

20 Table 7: For Males: Prediction Error Models Error (Rep Sub Prob - Model Based) Error (Tobit Est Sub Prob - Model Based) Error/Model Based (1) (2) (3) (4) (5) (6) OLS FE OLS FE OLS FE Age (0.3008) (0.9708) (0.1361) (0.1205) (0.0073) (0.0155) Yrs in Education (0.5482) (2.8187) (0.2480) (0.3499) (0.0132) (0.0450) Part Time Exp (0.6857) (4.8714) (0.3103) (0.6047) (0.0166) (0.0777) Full Time Exp (0.2928) (1.6494) (0.1325) (0.2047) (0.0071) (0.0263) Unempl. Exp. 0,1-0,9yrs (3.4498) (6.2702) (1.5609) (0.7783) (0.0833) (0.1001) Unempl. Exp. 1yr (5.4072) (9.2928) (2.4465) (1.1535) (0.1306) (0.1483) Unempl. Exp yrs (3.3960) (5.9179) (1.5365) (0.7346) (0.0820) (0.0944) Unempl. Exp yrs (3.8759) (6.5978) (1.7537) (0.8189) (0.0936) (0.1053) Unempl. Exp. 5+yrs (4.2593) (8.1180) (1.9271) (1.0076) (0.1029) (0.1296) Tenure last Job 3-9 Years? (2.1345) (5.1625) (0.9657) (0.6408) (0.0516) (0.0824) Tenure last Job 10 or more Years? (2.8694) (9.4289) (1.2983) (1.1704) (0.0693) (0.1505) Handicapped (2.7873) (7.5326) (1.2611) (0.9350) (0.0673) (0.1202) Married (2.0925) (6.1389) (0.9467) (0.7620) (0.0506) (0.0980) Home Owner (1.9160) (6.8238) (0.8669) (0.8470) (0.0463) (0.1089) Children (1.8904) (4.2545) (0.8553) (0.5281) (0.0457) (0.0679) Log Last Income (Gross) (1.2848) (2.6551) (0.5813) (0.3296) (0.0310) (0.0424) Last ISEI Status (0.1739) (0.5615) (0.0787) (0.0697) (0.0042) (0.0090) Managers (7.6490) ( ) (3.4608) (3.2886) (0.1848) (0.4228) Professionals (9.1446) ( ) (4.1375) (5.1121) (0.2209) (0.6573) Techn./Assoc. Profess (6.3150) ( ) (2.8572) (2.4870) (0.1526) (0.3198) Clerks (5.5577) ( ) (2.5146) (2.0647) (0.1343) (0.2655) Service/Shop Workers (5.7310) ( ) (2.5930) (1.8910) (0.1385) (0.2431) Agricult. Workers (5.2285) ( ) (2.3656) (1.2454) (0.1263) (0.1601) Craft Workers (3.0591) (7.6372) (1.3841) (0.9480) (0.0739) (0.1219) Machine Operators (3.4169) (8.7967) (1.5460) (1.0919) (0.0825) (0.1404) Unempl Rate (0.1736) (0.7257) (0.0785) (0.0901) (0.0042) (0.0116) Pseudo R 2 Number of Observations Marginal effects p < 0.1, p < 0.05, p < 0.01 Note: SOEP 1999, 2001, 2003, 2005,

21 Table 8: For Females: Prediction Error Models Error (Rep Sub Prob - Model Based) Error (Tobit Est Sub Prob - Model Based) Error/Model Based (1) (2) (3) (4) (5) (6) OLS FE OLS FE OLS FE Age (0.2030) (0.9231) (0.0983) (0.1165) (0.0043) (0.1165) Yrs in Education (0.5227) (2.6917) (0.2532) (0.3396) (0.0111) (0.3396) Part Time Exp (0.2867) (3.2836) (0.1388) (0.4143) (0.0061) (0.4143) Full Time Exp (0.2071) (2.7635) (0.1003) (0.3486) (0.0044) (0.3486) 1.Unempl. Exp. 0,1-0,9yrs (3.8273) (9.0765) (1.8536) (1.1451) (0.0813) (1.1451) 1.Unempl. Exp. 1yr (6.9510) ( ) (3.3664) (2.4359) (0.1477) (2.4359) 1.Unempl. Exp yrs (3.6322) (7.9507) (1.7591) (1.0031) (0.0772) (1.0031) 1.Unempl. Exp yrs (3.9219) (8.6337) (1.8994) (1.0893) (0.0834) (1.0893) 1.Unempl. Exp. 5+yrs (3.9428) (9.5279) (1.9095) (1.2021) (0.0838) (1.2021) 1.Tenure last Job 3-9 Years? (2.5736) (7.7083) (1.2464) (0.9725) (0.0547) (0.9725) 1.Tenure last Job 10 or more Years? (3.6440) ( ) (1.7648) (1.3365) (0.0774) (1.3365) 1.Handicapped (4.2517) (8.7496) (2.0591) (1.1039) (0.0904) (1.1039) 1.Married (2.0803) (6.9568) (1.0075) (0.8777) (0.0442) (0.8777) 1.Home Owner (2.0691) (8.6539) (1.0021) (1.0918) (0.0440) (1.0918) 1.Children (1.9829) (4.3357) (0.9603) (0.5470) (0.0421) (0.5470) Log Last Income (Gross) (1.1900) (2.9755) (0.5763) (0.3754) (0.0253) (0.3754) Last ISEI Status (0.1374) (0.4255) (0.0665) (0.0537) (0.0029) (0.0537) 1.Managers (7.9047) ( ) (3.8283) (2.6329) (0.1680) (2.6329) 1.Professionals (7.8165) ( ) (3.7856) (2.6993) (0.1661) (2.6993) 1.Techn./Assoc. Profess (5.2323) ( ) (2.5340) (2.2319) (0.1112) (2.2319) 1.Clerks (4.6672) ( ) (2.2603) (1.6126) (0.0992) (1.6126) 1.Service/Shop Workers (3.6795) (8.9335) (1.7820) (1.1271) (0.0782) (1.1271) 1.Agricult. Workers (4.9794) ( ) (2.4116) (1.8874) (0.1058) (1.8874) 1.Craft Workers (4.9255) ( ) (2.3855) (1.4840) (0.1047) (1.4840) 1.Machine Operators (5.8570) ( ) (2.8366) (2.4361) (0.1245) (2.4361) Unempl Rate (0.1918) (0.7383) (0.0929) (0.0931) (0.0041) (0.0931) Pseudo R 2 Number of Observations Marginal effects p < 0.1, p < 0.05, p < 0.01 Note: SOEP 1999, 2001, 2003, 2005,

22 References Campbell, D., Carruth, A., Dickerson, A., and Green, F. (2007). Wages. The Economic Journal, 117: Job Insecurity and Dickerson, A. and Green, F. (2009). Fears and Realisations of Employment Insecurity. Department of Economics, The University of Sheffield, Working Paper, Dominitz, J. and Manski, C. F. (1997). Using Expectations Data To Study Subjective Income Expectations. Journal of the American Statistical Association, 92(439): Ganzeboom, H. B. G., Graaf, P. M. D., and Treiman, D. J. (1992). A Standard International Socio-Economic Index of Occupational Status. Social Science Research, 21(1):1 56. Green, F., Dickerson, A. P., Carruth, A. A., and Campbell, D. (2001). An Analysis of Subjective Views of Job Insecurity. Department of Economics, University of Kent, Studies in Economics. Green, F., Felstead, A., and Burchell, B. (2000). Job Insecurity and the Difficulty of Regaining Employment: An Empirical Study of Unemployment Expectations. Oxford Bulletin of Economics and Statistics, 62(0): Haisken-DeNew, J. P. and Hahn, M. (2006). PanelWhiz: A flexible Modularized Stata Interface for Accessing Large Scale Panel Data Sets. Jappelli, T. and Pistaferri, L. (2000). Using Subjective Income Expectations to Test for Excess Sensitivity of Consumption to Predicted Income Growth. European Economic Review, 44(2): Kaufmann, K. and Pistaferri, L. (2009). Disentangling Insurance and Information in Intertemporal Consumption Choices. American Economic Review, 99(2): Manski, C. F. and Straub, J. D. (2000). Worker Perceptions of Job Insecurity in the Mid- 1990s: Evidence from the Survey of Economic Expectations. The Journal of Human Resources, 35(3): Pistaferri, L. and Jappelli, T. (2010). (forthcoming). The consumption response to income changes. 21

23 6 Appendix Table 9: Summary Statistics of Key Variables Variable Mean Std. Dev. Min. Max. Employed in t+1 or t Model Based Re-Empl Prob Reported Subj Re-Empl Prob Age Handicapped Part Time Exp Full Time Exp Unempl. Exp. 0,1-0,9yrs Unempl. Exp. 1yr Unempl. Exp yrs Unempl. Exp yrs Unempl. Exp. 5+yrs Yrs in Education Tenure last Job 3-9 Years? Tenure last Job 10 or more Years? Married Home Owner Children Log Last Income (Gross) Managers Professionals Techn./Assoc. Profess Clerks Service/Shop Workers Agricult. Workers Craft Workers Machine Operators Last ISEI Status Reported Subj Re-Empl Prob Unempl Rate N