tobiasraabe / locus-of-control Goto Github PK

• Explain how LoC is represented in the SOEP
• Explain what are the causal channels of LoC
• Describe predicted outcomes of our model

There are three traumatic events for which the current state could explain a lot of the effect: job loss, divorce and relationship breakup.
The effect of a job loss can be very different depending on whether you have a job at the time of the interview or not (reemployment), see Preuss, Hennecke (2017).
For the divorce and the relationship breakup, I think the same could be the case.
Preuss, Hennecke (2017) account for that by not only using a dummy for displacement, but using the interaction term displacement*employmentstate (employment states can be employed, unemployed or other states).
We wanted to control for the employment status, so we account for that, but we don't find out how the effect differs between the two groups. So we could think about including these interaction terms for these three variables as well.

Concerning unemployment and LOC in Preuss, Hennecke (2017):
"Following the theory of social identity, unemployment is a direct contradiction to the social norm of working, implying a loss in utility (Sch�ob, 2013; Hetschko et al., 2014). Hence, blaming fate
instead of blaming oneself `outsources' the responsibility of unemployment and may im-prove current individual well-being. Adapting ones own belief is then an active strategy to manage
unemployment and social desirability."
Shall we include this explanation as well? Because I think that if the unemployment is really not caused by oneself, this explanation is redundant, right?

• Why is LOC important
• Why is the stability assumption critical
• What changes LOC
• What are we doing/contributing

Following http://marcfbellemare.com/wordpress/wp-content/uploads/2019/02/BellemareWichmanIHSFebruary2019.pdf, the inverse hyperbolic sine transformation has the benefit of not using np.log(x + 0.01) or variants thereof to circumvent zeros in variables.

We decided to include "Any other family composition change" as a trauma. In the questionnaire there is a field to fill in which change it is. If we want to really use the variable we have to check each filled-in-field and evaluate whether the change is severe enough to set the dummy as one. That's a lot of work. Shall we do that or leave it out?

DEADLINE: tomorrow lunch

Imke:

• prepare three regressions
• clustering hh_id
• BREAK regressions review
• discussion fixed effects and clustering
• make regression with time dummies instead of continous variables [Tobias: No need to do that. I will take care of it when I have your third specification.]
• discuss inclusion of interaction terms etc.
• discuss internal vs exaternal types (Buddelmeyer, etc.)
• write conclusion
  • improvements for future research
  • examing the learning effect from context to stable LOC
  • critical time periods for establishment of LOC

Tobias:

• change disability als event
• create LAST_JOB_ENDED and LAST_JOB_ENDED_LIMITED to have the last job spec from preuss hennecke
• make index loc (drop ambigious item and cultural)
• pca with 7 and 10 items
• descriptives for individuals in the first period 2005, 2010
• nonparametric level test for those without and with events
• make graphic on heterogenous time effects

Later:

• make analysis regarding different age effects
• restrict to ind with no event in first period to exclude lagged effects.
• restrict sample to use fe
• sample restriction test for PReuss Hennecke

Problem 1

For almost all events, there are some people having multiple events in one period, meaning between 2005-2010 or 2010-2015. Looking at an individual having two divorces in one period, we thought that we might collapse this observation to one since we account for multiple events in our third specification. The problem is that some covariates like age, marital status are dependent on time and therefore we cannot simply reduce two observations in one period to one observation.

Solution 1

I would propose to include both observations and including another covariate into the regression stating whether the same shock has been experienced before. This is like Preuss & Hennecke who include a vector of parallel life events.

Problem 2

Another point: Preuss & Hennecke also remove all individuals from the sample who experience more than three job losses per period:

Overall, only the number of displacements is used as restriction, because consecutive job losses potentially correspond to an unusual environment. Individuals reporting more than three job losses between two LOC interviews are excluded from the analysis. (p. 11)

Is there a reason for us to do the same?

The meeting is scheduled on 19.01.2018 on eleven o'clock.

Talking points

Data preparation:

• how to handle legally handicapped in percentage points? The problem is that there positive as well as negative developments from the individuals perspective. Therefore, is it wise to use only jumps from 0-100%?
• How to deal with restrictions from Preuss, Hennecke?
  1. over 25 years old
  2. no more than three job losses during two loc interviews.
     Answer: Do not impose restrictions since the questoins differs. Selection bias in job markets does not apply here.
• Change in personal life because of death is currently missing but can be implemented (wont be implelmented)

Analysis:

• fixed effects on hh-level and individual-level cannot be used when there is only one observation per household or an individual is only observed once. We could restrict the sample to subpopulations to even use fe.
• What statistics or descriptives are necessary for LoC and PCA
• nonparametric result: groupby null or at least one event. mean of loc in first and in second period. ttest

Extensions Analysis:

• What nonparametric results would be interesting to show before the regression results?

• Graph with coefficients for time interval dummies to show depreciation effect

• Classifying ind as having internal vs external locus of control and comparing effects of shocks for those two groups.

• Make a plot where interaction between trauma and time interval dummy is on the x axis and the score of the coefficient on the yaxis. Show development of coefficient over longer time periods.

• What if all fails would strengthen our main point that locus of control is malleable. This relates to a nonparamtric result to motivate the topic.

• The observations before 2005 may have events which have lastlonging effects from events before 2005. A cleaner specification would select observations without events ten years before the sample period. We could sample all observations which have no event in 2005-2010 and look at effects in 2010-2015.

• We could use 0-1 dummies in the second specificatoin for disability_perc. More heaelth shock

• attenuation bias by bad measurement of LoC. Original results might be better.

The reasoning is that people having an extremely internal or external locus of control are not shifted by any event as their mindset is fixed to some extent. Maybe focusing on the middle of the distribution will provide better results.

Problem:
The month when the unplanned pregnancy was discovered is not identifiable from the data.

Solution:
Currently, we pick the beginning month of the pregnancy period.

I included the paragraph with the description of the LOC measure in the data section, because I think that it fits there best. The analysis of the causal channel of the LOC would better fit at the end of the estimation strategy section before the result part, I think.
For now, I included it at the end of the data section, but we can change that..

"Can having internal locus of control insure against negative shocks? Psychological evidence from panel data"
Their result: yes, for certain life events e.g. death of a close friend, illnesses.

How do we account for this evidence if we cannot control for the initial (2005) LOC?
Or is the biggest part already captured by controlling for age, education and gender?

How do we control for the time between the occurences of the traumatas and the loc measurement? We could ask Dohmen for that.
(As they discuss regularly with their professors in other projectmodules I also think that we can ask Dohmen for advice here.)

Problem:
The exact month when the disorder was discovered is not available in the data. There are references to under age x exams but there are a lot of missing values. Also the exact timing is still not available.

Solution:
We pick the month of birth as event date and assume that disorders will be discovered around that period.

Is available on person-level in the generated pgen.

Discuss whether to use years of education or dummies for levels of education.

• Basic description of SOEP, see Specht et al. (2013) or Preuss, Hennecke (2017)
• (Basic description of LOC, if not better explained in separate section)
• Detailed description of our encoding
• Results from PCA
• Document sample restrictions, missing values, etc.
• Document how many people suffered from how many traumata in the regarded period

Philipps comment was great. What if death of household member causes all the troubles.

Factor analysis can produce similar components (the columns of its loading matrix) to PCA. However, one can not make any general statements about these components (e.g. whether they are orthogonal).
The main advantage for Factor Analysis (over PCA) is that it can model the variance in every direction of the input space independently (heteroscedastic noise).
One can observe that with homoscedastic noise both FA and PCA succeed in recovering the size of the low rank subspace. The likelihood with PCA is higher than FA in this case. However PCA fails and overestimates the rank when heteroscedastic noise is present.

• Explain three strategies
• Separate section for common control variables/explanation for why we use these variables, see e.g. Sprecht et al. (2013), for explanation why we control for employment status: evidence from Preuss, Hennecke (2017) - important to differentiate between past and current unemployment spells
• Which effect is estimated?
• Explain SEs, other model specifications