SOCIAL INTERACTION AND SICKNESS ABSENCE. Assar Lindbeck, Mårten Palme and Mats Persson

Transcription

1 SOCIAL INTERACTION AND SICKNESS ABSENCE Assar Lindbeck, Mårten Palme and Mats Persson

2 Puzzle in sickness absence: The number of persons absent from work because of sickness, Sweden

3 Background Social interaction Social norms Consequences for moral hazard in social insurance Social multiplier

4 Background Difficult to measure norms directly Instead: indicators of the operations of norms

5 Background We study effects of group behavior on individual behavior Problem: Difficult to separate group effects from effects of selection and unobserved heterogeneity

6 Manski (1999, 2000) Correlated effect. Contextual effect. Social interaction.

7 Case study: Sickness absence in Sweden Variation across geographical areas Data set = LOUISE + National Social Insurance Board

8 Case study: Sickness absence in Sweden Variation across geographical areas Data set = LOUISE + National Social Insurance Board How to define geographical areas?

9 SAMS Small Area Market Statistics Geographical definition constructed by Statistics Sweden Idea is to create an area with the same type of houising within each church parish Around 1,000 individuals within each SAMS About 9,299 SAMS disstricts

10 N n S = S / N, n = 1,...,9003 n i= 1 in n Large standard deviation (σ = 5.4), large range (330). Can socio-economic factors explain the variation? Three types of variables.

11 Variables in the X vector for the individual age (all ages from 18 to 64, one dummy for each age, i.e, 46 dummies) education (seven levels, one dummy for each, i. E., six dummies) gender (one dummy) marital status (two dummies) having children aged 3 or younger (one dummy)

12 Variables in the X vector (cont d) for the workplace: industry (60 industries, i.e., 59 dummies) sector (central government authority, state-owned enterprise, local government authority, etc.), 10 dummies) size of workplace (21 dummies: 1 employee, 2-10, 11-20, 21-30,, , etc.

13 Variables in the X vector (cont d) for the geographical area: town or countryside (one dummy) life expectancy in the region (average life expectancy among the 21 counties in Sweden. Source: Statistical Yearbook of Sweden, 2004) local unemployment (expressed by the incidence of unemployment, 19 dummy variables, one for each 5-percent interval)

14 S = α + X β + ε in in in '

15 S = α + X β + ε in in in ' ε N n = ε / N, n = 1,..., 9003 n i in n Large remaining variation. σ R. about the same. Is this a random phenomenon?

16 S = α + X β + ε in n in in ' Neighborhood-specific intercepts. F= Significant neighborhood effects at 1 percent level.

17 What does this reflect?

18 What does this reflect? Unobserved permanent factors?

19 What does this reflect? Unobserved permanent factors? Should not influence changes.

20 What does this reflect? Unobserved permanent factors? Should not influence changes. Δ S = α + X β + ε in in in '

21 What does this reflect? Unobserved permanent factors? Should not influence changes. Δ S = α + X β + ε in in in ' Alternatively: specific α n. Still, significant neighborhood effects. F test : 1.559

22 Conclusion: There is something else.

23 Conclusion: There is something else. Could that be social norms?

24 Conclusion: There is something else. Could that be social norms? Is individual work absence causally related to average work absence in the local neighborhood?

25 One way of modelling social norms is ' S = + X + S + α β γ ε in in n in

26 One way of modelling social norms is ' S = + X + S + α β γ ε in in n in However: the reflection problem.

27 One way of modelling social norms is ' S = + X + S + α β γ ε in in n in However: the reflection problem. Four different appoaches to deal with this problem.

28 Four approaches:

29 Four approaches: 1. Public sector employees have higher work absence rate. We use the share of public sector employees as instrumental variable.

30 Four approaches: 2. We analyze the interaction between neighborhoods and workplaces. Are you more likely to be influenced by the absence behavior of your neighbors if they are also your workmates?

31 Four approaches: 3. We ask whether movers within the country tend to adjust their sickabsence to the area where they are moving to.

32 Four approaches: 4. We ask whether immigrants tend to adjust their sick-absence behavior to the behavior of natives in their neighborhood.

33 Four approaches:

34 Four approaches: Within each of these four approaches, we use alternative specifications and estimation techniques.

35 Four approaches: Within each of these four approaches, we use alternative specifications and estimation techniques. Each approach has specific weaknesses, but all approaches point in the same direction

36 Four approaches: Within each of these four approaches, we use alternative specifications and estimation techniques. Each approach has specific weaknesses, but all approaches point in the same direction

37 Public-sector vs. private-sector employees

38 Public-sector vs. private-sector employees Z proportion of population n in neighborhood n that works in the public sector. IV approach

39 ' S = a+ X b+ cz + e n in n n priv ' ˆ S = + X + S + α β γ ε in in n in

40 ' S = a+ X b+ cz + e n in n n priv ' ˆ S = + X + S + α β γ ε in in n in This gives us an estimate of γ in the equation priv ' S = + X + S + α β γ ε in in n in

41 Let s the do it the other way around: ' S = a+ X b+ c(1 Z ) + e publ ' ˆ S = α + X β + γ S + ε in in n in n in n n

42 Population All those who work in private sector Number of individuals and observations 2,839,410 ind. 14,556,753 obs. Regressor Share of population in neighborhood n that work in public sector ( Z ) nt Reduced form (μ in eq. (4) ) *** (0.0013) R 2 = First step in IV regression (c in eq. (3)) 6.670*** (0.0116) R 2 = IV estimate (γ in eq. (3)) 0.581*** (0.0199) R 2 = All those who work in public sector 1,956,740 ind. 10,502,405 obs. Share of population in neighborhood n that work in private sector ( 1 Z ) nt *** (0.0017) *** (0.0123) R 2 = 0026 R 2 = *** (0.0302) R 2 = All employees 4,796,150 ind. 25,059,158 obs. Share of population in neighborhood n that work in public sector ( Z ) nt *** (0.0011) R 2 = *** (0.0084) R 2 = *** (0.0173) R 2 =

43 Problems with the IV approach? Selection problem? Some individuals choose to live in specific neighborhoods?

44 Problems with the IV approach? Selection problem? Some individuals choose to live in specific neighborhoods? Control for gender, age, education, etc.

45 Problems with the IV approach? Selection problem? Some individuals choose to live in specific neighborhoods? Control for gender, age, education, etc. Unobservables that are uncorrelated with these?

46 Problems with the IV approach? Selection problem? Some individuals choose to live in specific neighborhoods? Control for gender, age, education, etc. Unobservables that are uncorrelated with these? Workplace fixed effects?

47 Refutability test priv EDU int = a + kt + X int β + μ1z nt + ' e nt publ EDU int = a + kt + X int β + μ 2Z nt + ' e nt μ = μ *** = ***

48 Interaction Model S ijk ' ijk ( CA ) jk Sk + λ j + μk + ϕcajk ε ijk = δ + X β + γ * + j sub-index for work place. k - sub-index for neighborhood. CA concentration of workers from the same neighborhood working at the same work place. Identification: through the interaction between the average work absence level in the neighborhood and the concentration of workers from the same neighborhood working at the same workplace. Note: controls for fixed workplace and neighborhood effects as well as the concentration of workers from the same workplace.

49 Interaction Model S ijk ' ijk ( CA ) jk Sk + λ j + μk + ϕcajk ε ijk = δ + X β + γ * + Interpretation: if there is an above average work absence in the neighborhood, the expected work absence increases with factor γ with the concentration of workers from the same neighborhood.

50 Interaction Model Corresponding model for the effect of the social interaction at the workplace. S ijk ' ijk ( CA ) jk S j + λ j + μk + ϕcajk εijk = δ + X β + γ * +

51 Results from the Interaction Model Table 4: The strength-of-network effect Number of obs. and ind. 24,449,603 obs. 4,693,560 ind. νˆ 2.146*** (0.0273) S S 2 iwn / n R = νˆ CA Note: The numbers of observations and individuals in this table are somewhat smaller than the corresponding numbers in Table 1. The reason is that for each individual, we have deleted the individual himself from the data when computing the averages S n. For some neighborhoods, there is only one individual who works in each workplace; these cases therefore do not appear in the regression.

52 Approach No. 3: Movers within the country. If an individual moves from a highabsence neighborhood to a low-absence one does he/she change work absence behavior?

53 Approach No. 3: Movers within the country. imnt mover non t n mover non m t t in imt t mover t in mover imt S S X X S S ε η β λ α = ) ( ) ' ' ( 1, 1, 1, 1, imnt mover non t n mover non m t it mover non t n mover non m t t in imt t mover t in mover imt S S D S S X X S S ε δ η β λ α = ) ( ) ( ) ' ' ( 1, 1, 1, 1, 1, 1, >. 0.,., 1 1, 1, otherwise S S if e i absensce higher with to an area time t at moved has i individual if D mover non t n mover non m t it

54 Approach No. 3: Movers within the country. Table 2: Movers within Sweden Symmetric specification (5) Asymmetric specification (5') No. of individuals and observations 1,551,059 ind. 2,202,466 obs. 1,551,059 ind. 2,202,466 obs. 2 R ηˆ δˆ *** ( *** (0.0094) (0.0777)

55 Approach No. 4: immigrants vs. natives m = α + β m + γ s + ε in in n in S X S

56 Table 3: Estimates of γ in equation (6) All immigrants Region Number of ind. and obs. All regions 720,742 ind. 3,376,753 obs. Nordic 210,059 ind. countries 1,088,923 obs. EU (except 77,982 ind. Nordic 358,797 obs. countries) Europe 154,378 ind. (except EU) 744,440 obs. Africa 38,422 ind. 163,554 obs. North 21,655 ind. America 92,321 obs. Latin 36,556 ind. America 167,644 obs. Asia 173,447 ind. 723,644 obs. Oceania 3,626 ind. Former Soviet Union 14,146 obs. 4,398 ind. 22,566 obs. Estimate of γ 0.629*** (0.0063) 0.651*** (0.0174) 0.461*** (0.0242) 0.126*** (0.0176) 0.091*** (0.0308) 0.278*** (0.0360) 0.345*** (0.0347) 0.237*** (0.0157) 0.222*** (0.0766) (0.0949)

57 We have excluded neighborhoods with less than 20 % natives. Sensitivity analysis: only recent immigrants.

58 Conclusions Is there evidence of social interaction? What is the magnitude of this effect? Four different strategies: Each captures a specific form of social interaction. Other forms are also possible: National level through the mass media. Social interactions outside the neighborhood: relatives, bowling clubs or unions.

59 Lower bound estimates: Interaction model: effect only through interaction between work and neighborhood networks. Movers within Sweden: Only partial effect since the entire adjustment is not likely to happen the first year and 0.050, respectively.

60 Upper bound estimates: IV model. Immigrants and 0.581, respectively.