Application I: Unemployment Durations
Discrete time models are applied on a data set of the NEPS panel study (Blossfeld et al. 2011). The data was obtained as retroperspective interviews. Analyzed are unemployment durations. As unemployment spells can be repeated, only the first unemployment
- PDF / 1,324,120 Bytes
- 8 Pages / 419.528 x 595.276 pts Page_size
- 95 Downloads / 212 Views
yesian Analysis of Failure Time Data Using P-Splines, BestMasters, DOI 10.1007/978-3-658-08393-9_5, © Springer Fachmedien Wiesbaden 2015
62
5 Application I: Unemployment Durations Table 5.1: Model assessment
grouped Cox probit
LPML
DIC
pD
deviance
-2063.6 -2085.7
4093.5 4138.1
33.4 32.1
4060.1 4105.9
• registered binary variable indicating registration of unemployment at begin of unemployment These covariates were used for modeling of smooth effects via P-splines: • casmin education years coded via casmin classification • time process time • historical time historical time, ranging from 2001 to 2011 • age age in years at beginning of unemployment spell Furthermore, seasonality with perior length 12 was included, leading to the model g(P(T = t|T ≥ t, η)) =β0 + β1 ∗ sex + β2 ∗ f oreigner + β3 ∗ training+ β4 ∗ registered + f0 (time) + f (age) + f (casmin)+ f (season) + f (historicaltime) A grouped Cox and a probit model were used. For the grouped Cox, sampling scheme 3 was used, additionally a completion Gibbs sampler with latent Poisson variables as described in section 4.2.3 in combination with sampling scheme 3 was tried. Truncated Poisson variables were sampled via the inverse cdf method using standard R-functions for the Poisson distribution. For the grouped Cox sampler there were very high correlations between the seasonal and the baseline hazard coefficients causing the sampler to get stuck (see figure 5.1), those parameters were drawn in one block. This was done by combining the design, penalty matrices and restriction matrices into blockmatrices. This resulted in lower acceptance rates (around 20%) which was found to be acceptable for a vector of size 171.
5 Application I: Unemployment Durations
63
parameter
1.0 0.5 0.0 í0.5 í1.0 0
2500
5000
iteration
7500
10000
Figure 5.1: Convergence problems for unblocked parameter.
Sampling scheme 3 was found to behave somewhat erratically: There were long stretches where no draws were accepted, often after convergence seemed to have been reached see figure 5.6. Furthermore there were numerical problems due to the exponential functions involved in the complementary log-log link. It might be the case that sampling was not done long enough, for the completion Gibbs sampler this behaviour was not observed however, as such the results of this sampler are reported here. All in all, this sampler was much more stable. For other parameters, results of the samplers were very similar. Acceptance rates were around 50% for all parameters excluding baseline hazard coefficients. For the probit model, no further adjustments were made and the standard sampler was used, using sampling scheme 4. The probit model achieves more indepedent draws than both sampler. A fair comparison would furthermore include execution times, as the sampler for the probit model is about 2 times faster although this depends on the implementation. For both models, 12000 iterations were used, the first 2000 were taken as burnin. Convergence could probably be accelerated by using better start value
Data Loading...
