Subjective Expectations and Income Processes in Rural India

Transcription

1 Subjective Expectations and Income Processes in Rural India Orazio Attanasio (UCL, IFS, NBER & BREAD) & Britta Augsburg (IFS) ASSA 2014, Philadelphia, Nature of Labor Income Dynamics

2 Motivation Beliefs and expectations play important role in decision making processes. Particularly income: Intertemporal optimization allocation of current income between consumption and savings Investment choices Such expectations however, rarely observed. No clear consensus of the nature of expected income processes. Most empirical studies infer expectations from data on actual income realizations appealing to some sort of rational expectations.

3 Subjective (Income) Expectations Collection and use of subjective expectations has become more widespread. Manski (2004) has advocated their use strongly. Examples of studies looking at expected income: Guiso, Jappelli and Terlizzes (1992, 1996) Italy Dominitz and Manski (1996, 1997), Dominitz (1998, 2001) US Das and van Soest (1997), Alessie et al. (1997) Netherlands Attanasio et al. (2005) Colombia Many other examples: inflation, effectiveness of contraception, HIV infection, stock market, rainfall, etc. An important feature of many of the new contributions: attempt to estimate the entire probability distribution.

4 Some word of caution The direct measurement of subjective expectations is not without problems. A standard question or methodology that works everywhere does not exist. Extensive piloting on the wording and ingenuity of the measurement tools is necessary. Common features: It is often easy to start eliciting the range of variation Eliciting the distribution of probability over the relevant range is hard. Examples to convey the concept of probability are important (rain, games of chance etc. ) Visual representations of the probability are useful (rulers, stones etc. )

5 Developing country context Many of these problems are aggravated in a developing country setting: Lower levels of education Lack of exposure to formal concept of probability Delavande, Giné and McKenzie (2011), Attanasio (2009) show that also in developing countries it is possible to elicit sensible data on the probability distribution of stochastic variables such as income. Other studies include: Attanasio et al (2005 Colombia, income), Luseno et al (2003 East Africa, rainfall), Lybbert et al (2007 East Africa, rainfall), McKenzie et al (2013 Tonga, income).

6 Data survey of MFI clients in Anantapur, Andhra Pradesh (evaluation) Sample includes clients and non-clients (in program and non-program villages) 64 villages, 2 survey rounds: (1) March/April 2008: 1,041 households (2) April-June 2009: 951 households (91%)

7 Characteristics of respondent households

8 Characteristics of respondent households

9 The expectations questions Followed by now well-established approach: 1. Elicit range (min and max): Imagine that you have a very good year, every member of working age in the household managed to have work, and there were no droughts or anything the like. What would be the maximum amount of income your household would receive in such a situation in one year? 1. Divide into 4 equally-spaced intervals (thresholds A, B, C). 2. Ask for value of the distribution function for each of these thresholds. Previously, concept of probability was introduced using example of rain and ruler (0-100) as an aid

10 Response rate & Validation High willigness to respond: - Response rate: ~97%

11 Response rate & Validation Low logical response error: - Violation to basic probability laws (monotonicity and wrong direction ): <1%

12 Response rate & Validation Bunching of percentages: Confidence in elicited range: very few 0 s and 100s Rounding to nearest 5% Whole range used

13 Distributional assumptions Interpret responses as points on the subjective cumulative distribution function of future income. Estimate moments of the entire distribution function by making functional form assumption. We assume a piecewise uniform distribution.

14 Summary statistics of income variables (2008) Able to distinguish individual uncertainty versus heterogeneity and uncertainty (here versus 0.780).

15 Comparison realized and actual income Realized and expected income are very similar. Moreover, strong correlation (>0.7). Both realizations and expectations vary sensibly with covariates.

16 A model for individual income We start with a very simple income process: Suppose: where is an iid innovation in the income process.

17 A model for individual income We start with a very simple income process: Suppose: where is an iid innovation in the income process. Under rational expectations:

18 A model for individual income We start with a very simple income process: Suppose: where is an iid innovation in the income process. Under rational expectations: Denote with Define: the expected value of This equation can be estimated from a single cross-section.

19 Estimation details (1) We allow for possible presence of fixed effects: Reason: It can induce important bias in the estimation of persistence. Approach: 1. Parametrize fixed effects as a function of observable variables ( normal income). 2. Take first difference. 3. Use Arellano and Bover (1995) approach.

20 Estimation details (2) We instrument current income with village level averages. Reason: Measurement error in current income induces downward bias in the estimated persistence. (3) We study the residuals from a first stage regression. Reason: Remove the effect of deterministic variables x i,t.

21 Level specification OLS specification: we can reject that the coefficient equal to 1.

22 Level specification OLS specification: we can reject that the coefficient equal to 1. IV specification: estimates of coefficients increase and we cannot reject that alpha1 =1 anymore.

23 Level specification Same conclusion holds when modelling the fixed effect as a function of observable variables ( normal income ). Same conclusion holds for 2009 and using pooled data.

24 First difference specification Same conclusions as in levels.

25 Arellano-Bover (1995) Use, where as an instrument for Coefficient marginally significantly different from 1. Adding normal income hardly changes the coefficient.

26 Conclusion Subjective expectations: by and large, internally consistent and sensible compute mean and s.d. of expected future income. Cross-sectional variance of income (=uncertainty + heterogeneity) much larger than variance computed from expected income (=individual uncertainty). Estimate time series model of future income: Income seems to be extremely persistent; Results remarkably consistent across specification; Expectation data can be used to estimate time series model for income that would not be identifiable in the absence of panel data.

27 ADDITIONAL SLIDES

28 The expectations questions 1. Elicit range: Minimum: Imagine that you have a very good year, every member of working age in the household managed to have work, and there were no droughts or anything the like. What would be the maximum amount of income your household would receive in such a situation in one year? Maximum: Now imagine the total opposite: the harvest is bad; animals get sick, finding work is not possible. What would be the monthly income of your household in such a situation?

29 The expectations questions 2. Divide interval in four equally sized intervals, determining three cut-off points, A, B, C. 3. Ask for value of the distribution function for each of these points: How likely do you think it is that your income in the coming year will be higher than (A/B/C) Rupees?

30 Timeline

31 Summary stats of income variables

32

33 1. Observable fixed effects : We use normal income (as asked about in the survey) 2. First difference: Under rational expectations the error term reflects only measurement error in the expectation, we can then use OLS

34 Results using 2009 data

35 Results using pooled data