Sune Karlsson

Numerical Methods for Estimation and Inference in Bayesian VAR-Models

In Bayesian analysis of vector autoregressive models, and especially in forecasting applications, the Minnesota prior of Litterman is frequently used. In many cases other prior distributions provided better forecasts and are preferable from a theoretical standpoint. Several of these priors require numerical methods in order to evaluate the posterior distribution. Different ways of implementing Monte Carlo integration are considered. It is found that Gibbs sampling performs as well as, or better, then importance sampling and that the Gibbs sampling algorithms are less adversely affected by model size. We also report on the forecasting performance of the different prior distributions

On the power and interpretation of panel unit root tests

We demonstrate that panel unit root tests can have high power when a small fraction of the series are stationary and may lack power when a large fraction is stationary. The acceptance or rejection of the null is thus not sufficient evidence to conclude that all series have a unit root or that all are stationary.

Forecast Combination and Model Averaging Using Predictive Measures

We extend the standard approach to Bayesian forecast combination by forming the weights for the model averaged forecast from the predictive likelihood rather than the standard marginal likelihood. The use of predictive measures of fit offers greater protection against in-sample overfitting when uninformative priors on the model parameters are used and improves forecast performance. For the predictive likelihood we argue that the forecast weights have good large and small sample properties. This is confirmed in a simulation study and in an application to forecasts of the Swedish inflation rate, where forecast combination using the predictive likelihood outperforms standard Bayesian model averaging using the marginal likelihood.

Forecasting with Bayesian Vector Autoregression

This chapter reviews Bayesian methods for inference and forecasting with VAR models. Bayesian inference and, by extension, forecasting depends on numerical methods for simulating from the posterior distribution of the parameters and special attention is given to the implementation of the simulation algorithm.

Bayesian Forecast Combination for VAR Models

We consider forecast combination and, indirectly, model selection for VAR models when there is uncertainty about which variables to include in the model in addition to the forecast variables. The key difference from traditional Bayesian variable selection is that we also allow for uncertainty regarding which endogenous variables to include in the model. That is, all models include the forecast variables, but may otherwise have differing sets of endogenous variables. This is a difficult problem to tackle with a traditional Bayesian approach. Our solution is to focus on the forecasting performance for the variables of interest and we construct model weights from the predictive likelihood of the forecast variables. The procedure is evaluated in a small simulation study and found to perform competitively in applications to real world data.

Foreign Firms and Chinese Employment

This paper examines the effect of foreign direct investment (FDI) on employment in the Chinese manufacturing sector. As one of the worlds largest recipients of FDI, China has arguably benefited from foreign multinational enterprises in various respects. However, one of the main challenges for China, and other developing countries, is job creation, and the effect of FDI on employment is uncertain. The effect depends on the amount of jobs created within foreign firms as well as the effect of FDI on employment in domestic firms. We analyse FDI and employment in China using a large sample of manufacturing firms for the period 1998–2004. Our results show that FDI has positive effects on employment growth. The relatively high employment growth in foreign firms is associated with their firm characteristics and their high survival rate. Employment growth is also relatively high in private domestic Chinese firms. There also seems to be a positive indirect effect of FDI on employment in private domestically-owned firms, presumably caused by spillovers.

Forecasting the Swedish unemployment rate VAR vs. transfer function modelling

Abstract The Swedish unemployment rate is forecast using three time series methods: the ARIMA, transfer function and Vector Autoregressive (VAR) models. Within this context, the choice of modelling strategy is discussed. It is found that the forecasting performance of VAR models is improved by explicitly taking account of cointegration between the variables in the model, despite the fact that unemployment is not cointegrated. However, the more parsimonious ARIMA and transfer function models have lower RMSE for all forecasting horizons. It is also found that the additional variables in the VAR models are important for predicting the turning points in the unemployment rate.

Lag-length selection in VAR-models using equal and unequal lag-length procedures

SummaryIt is well known that inference in vector autoregressive models depends crucially on the choice of lag-length. Various lag-length selection procedures have been suggested and evaluated in the literature. In these evaluations the possibility that the true model may have unequal lag-length has, however, received little attention. In this paper we investigate how sensitive lag-length estimation procedures, based on assumptions of equal or unequal lag-lengths, are to the true model structure. The procedures used in the paper are based on information criteria and we give results for AIC, HQ and BIC. In the Monte Carlo study we generate data from a variety of VAR-models with properties similar to macro-economic time-series. We find that the commonly used procedure based on equal lag-length together with AIC and HQ performs well in most cases. The procedure (due to Hsiao) allowing for unequal lag-lengths produce reasonable results when the true model has unequal lag-length. The Hsiao procedure tend to do better than equal lag-length procedures in models with a more complicated lag structure.

Bayesian simultaneous determination of structural breaks and lag lengths

The detection of structural change and determination of lag lengths are long-standing issues in time series analysis. This paper demonstrates how these can be successfully married in a Bayesian analysis. By taking account of the inherent uncertainty about the lag length when deciding on the number of structural breaks and vice versa we avoid some common pitfalls and are able to draw more robust conclusions. The approach is illustrated using both real data and a simulation study.

Polymorphisms in the CLDN1 and CLDN7 genes are related to differentiation and tumor stage in colon carcinoma

Tight junction is composed of transmembrane proteins important for maintaining cell polarity and regulating ion flow. Among these proteins are the tissue‐specific claudins, proteins that have recently been suggested as tumor markers for several different types of cancer. An altered claudin expression has been observed in colon, prostatic, ovarian, and breast carcinoma. The aim of this study was to analyze the allele frequencies of three common single nucleotide polymorphisms (SNPs) in the genes for claudin 1 and claudin 7 in colon cancer (CC) patients and in a control population of healthy blood donors. Pyrosequencing was used to genotype the CLDN1 SNP rs9869263 (c.369C>T), and the CLDN7 SNPs rs4562 (c.590C>T) and rs374400 (c.606T>G) in DNA from 102 formalin fixed paraffin embedded (FFPE) colon cancer tissue, and 111 blood leukocyte DNA from blood/plasma donors. These results were correlated with clinical parameters such as TNM stage, tumor localization, tumor differentiation, complexity index, sex, and age. We found that there was a significant association between the CLDN1 genotype CC in tumor samples and a higher risk of colon cancer development (OR 3.0, p < 0.001). We also found that the CLDN7 rs4562 (c.590C>T) genotype CT had a higher risk of lymph node involvement (p = 0.031) and a lower degree of tumor differentiation (p = 0.028). In the control population, the allele frequencies were very similar to those in the HapMap cohort for CLDN7. The CLDN1 rs9869263 genotype (c.369C>T) was related to increased risk of colon cancer, and the CLDN7 rs4562 genotype (c.590C>T) was related to tumor differentiation and lymph node involvement in colon carcinoma. Further studies are warranted to ascertain their potential uses as biomarkers predicting tumor development, proliferation, and outcome in this disease.