Checking the Robustness and Calidity of Model Selection: An
Transcription
Checking the Robustness and Calidity of Model Selection: An
Introduction Macroeconomic contributions Econometric contributions Appendix Checking the Robustness and Validity of Model Selection: An Application to UK Wages Jennifer L. Castle Department of Economics Oxford University 17 January 2009 Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 1 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Motivation Contributions Build models of real wage growth over 125 years Macroeconomic application Compare feedforward and feedback models of wage in‡ation; Yield new insights into role of expectations over long-run. Econometric Methodology Evaluate robustness of model selection procedure; Parametric bootstrap to assess selection uncertainty. Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 2 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Motivation Models of wage growth 2 main schools of thought: Forward-looking models where expectations play a key role: rational expectations – ’NKPC’. Backward-looking models driven by causal factors no explicit role for expectations. Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 3 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Expectations Measures of expectations Standard Approach Replace Et [∆pt +1 ] with actual ∆pt +1 and instrument. Identi…cation problems Derive set of rational expectations of in‡ation using price in‡ation model to obtain recursive one-step ahead predictions (omitting contemporaneous variables). Equation Backward-looking expectations: Exponentially weighted moving average. Random walk. Jennifer L. Castle …gure RES Fourth PhD Presentation Meeting, UCL 4 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Automatic model selection Automatic model selection Gets – data-based, combining theory at the outset but relying on statistical characteristics of data to deliver …nal speci…cation. Initial model of real wage growth includes: 24 conditioning variables including economic and institutional factors; expectations measures; 4 dummies (wars, great depression, oil crises). Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 5 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Automatic model selection Non-linear price coe¢ cient Test of non-linearity: Castle and Hendry (2008). Non-linearity in real wages; cannot reject linearity in nominal test results wages; ) non-linear price coe¢ cient. Economic theory informed non-linear speci…cation: 1.0 0.9 0.8 0.7 0.6 fna fna = 1 1 +1000 (∆p t )2 0.5 0.4 0.3 Sticky information Mankiw and Reis (2002), slow dissemination of information. 0.2 0.1 −0.15 −0.10 −0.05 0.00 ∆p 0.05 0.10 0.15 0.20 0.25 As price in‡ation rises, workers are more attentive, act to prevent further erosion of real wages. Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 6 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Automatic model selection Macroeconomic conclusions Huge changes: technological innovation, legislative changes, social reforms, wars, policy regime shifts. Statistically well-speci…ed, constant parameter model of real wage growth over 125 years. No role for replacement ratio, TU power, not much for level of unemployment. Expectations, either instrumenting or direct expectations from price in‡ation forecasts, do not explain real wages. Model of real wage growth with unit labour costs, productivity growth, change in unemployment rate, and non-linear price reaction able to explain much of movement in real wages. Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 7 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Robustness Selection criteria All models selected based on some criteria but reported model is uncertain: variables retained by chance when irrelevant relevant variables omitted Select models using general-to-speci…c methodology. Objective: report reliability of selected model using 2 criteria: 1 robustness – parametric bootstrap to assess ability to recover chosen model in di¤erent draws using same selection algorithm 2 validity – selected model speci…cation is valid Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 8 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Checking the robustness of wage in‡ation models How reliable is the preferred real-wage model Use parametric bootstrap to assess robustness of model selection. What is the probability that the same model will be selected for a di¤erent draw using the same selection algorithm? Gauge – Number of irrelevant variables retained Potency – Number of relevant variables retained Results: Most explanatory variables retained with high probability; Uncertainty as to timing of e¤ects; Costs of search over costs of inference small; Invalid models detected through statistical mis-speci…cation. Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 9 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Conclusions Econometric conclusions Advances in econometric techniques, including: developments in automatic Gets selection; use of indicator saturation to identify shifts; test non-linearity in high dimension, collinear systems. Monte Carlo simulations demonstrate problems of instrumenting for forward-looking expectations. Parametric bootstrap to assess robustness. Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 10 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Conclusions Future Research Automate post-model selection bootstrap Use selected model as DGP augmented by additional variables with coe¢ cients drawn from normal distribution with variance corresponding to general model. Forecasting Extend data to see how model forecasts ex ante. Ex post use bootstrap procedure over forecast horizons to select forecasting models. Role of breaks in forward-looking models Learning models for expectations. How expectations are formed in periods of high uncertainty. Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 11 / 11 Introduction Macroeconomic contributions Econometric contributions Appendix Route Map 1 Economic theory 2 Data 3 Models of wage in‡ation 1 2 New Keynesian Phillips Curve Extensions to real wages and additional conditioning variables – the role of model selection 4 Monte Carlo evidence 5 Robustness testing 6 Conclusions Jennifer L. Castle RES Fourth PhD Presentation Meeting, UCL 12 / 11