Luxi Han

[rickecon]

I love the topic. Hidden Markov income process. I assume you've looked at the work by Guvenen and Kaplan. How is this different from their work? Is it the Bayesian estimation that is new? Rough finish. Need a more polished conclusion. But I love the project.

[jamesallenevans]

Estimation of complex Income process that accounts for how people hedge against those shocks Lots of jargon on the data...so not sure what you are estimating effects from... The Bayesian procedure seem like it could be reasonable...EM is interesting...what about alternative estimation procedures? What do the empirical shocks look like? (nonsmooth income?)

[magicahan]

@rickecon Hi Rick! Yes I have seen their work. First of all, we are using a different income process where we allow for correlation between transitory and permanent shock, which was assumed to be independent in the past literature (for example, the earnings record shows after a recession ,people's earnings record don't return to its original value but permanent low, this can indicate relationship between these two shocks). Moreover Kaplan use a moment method and relies on two assumptions: short memory and no advance information. We don't have that assumption, moreover we are estimating the nonparametric distribution.

The Bayesian method is new, it's a variant based on importance sampling which sort of has the MCMC backbone, for example from Arellano, Blundell, Bonohomme 2016. This method is more accurate and efficient since we are taking into account of people's whole earnings record when estimating the posterior. Normally, we would only do a forward filtering where we draw the particles from the transition density. But now we are taking into account of earnings in next period which is more accurate and efficient.

So one of the critical part is that we are allowing for a nonparametric distribution for permanent and transitory shocks with correlation. The previous literature don't.

[magicahan]

@rickecon And also I'm more interested in seeing how the estimated nonparametric distribution will affect the income process. More importantly allowing for correlation can have new insights into how budget constraint will affect the insurability (People experience high permanent income shock vs. people with high wealth). I'm interested in tracing out the partial insurance coefficient for different cohort of the people instead of a overall partial insurance coefficient. This can be done following the basic concepts from Bonhomme's 2016 paper. Most importantly, the data I got approval from (thanks to Prof. Larry Schmidt) has considerably more records on earnings and they record earnings during unemployment as well. Which can potentially allows us even discuss the effect on unemployment. But I have to discuss more in detail with Larry.

[ChenAnhua]

Hi Luke! Thanks for presenting! I am actually quite interested in the data part of your project. You mentioned a main datasource: SIAB and also two altrnatives: PSID and CEX. I was wondering what are the pro/cons using these three different dataset to non-parametrically identify transitory/permanent shocks? (e.g., what are certain variables of interest, and for variables measuring similar thing in these three datasets, what are their differences? specifically for your identification purpose) Thanks!

[ariboyarsky]

Is there a reason you decided to use a particle filter for updating over a Kalman filter?

[magicahan]

@ariboyarsky Hi Ari! The Kalman filter would normally have the normal assumption. Basically particle filtering method doesn't make any specific distribution assumption. The quantile regression is basically estimate the quantile, imagining draw dots on the distribution, and then use cubic spline to estimate it.

[magicahan]

@ChenAnhua The SIAB has 1.5 million workers with 40 million plus observations that is much larger than PSID data. And we have earnings during unemployment as well. The downside of this data is that we don't have assets and consumption data and this is why I use a simulation method. But Larry and I are also considering using the Norwegian data which does have asset information.