Journal of Econometrics 100(2001)7}10
Essays
Achievements and challenges in econometric
methodology
David F.Hendry
Department of Economics,Oxford Uni v ersity,Oxford,UK
Abstract
Disputes about econometric methodology have abounded in econometrics,yet their attempted resolution has attracted only a small proportion of rearch e !ort.Recently,computer-automated general-to-speci "c reductions have been shown to perform well in Monte Carlo experiments,recovering the DGP speci "cation from a much larger general model with size and power clo to commencing from the DGP itlf.Thus,future developments appear promising,with many ideas awaiting implementation and both theoretical and simulation evaluation. 2001Published by Elvier Science S.A.All rights rerved.
Keywords:Econometric methodology;Computer automation;General-to-speci "c;Model lection;Data mining
As in many disciplines,computers have played an esntial role in making econometrics operational.Software for data management,modelling,estimation,inference,simulation and graphics,has underpinned most of the great empirical strides in our subject (and many other social sciences },Coppock,1999).A new generation of programs now bids fair to resolve some of the hotly debated issues about model lection in macro-econometrics }and &data mining '}by demonstrating the success of general-to-speci "c (henceforth denoted Gets ).
There can be little dispute that econometric methodology lacks a connsus:witness the debates in Granger (1990)and Hendry et al.(1990),the diversity of approaches in d 'Autume and Cartelier (1996)and Magnus and Morgan (1999),as well as the gulf in views on Business Cycle Empirics in the Economic Journal (November,1995)and in Backhou and Salanti (2000),among many other possible citations.Yet,there has been a paucity of evidence,as against argument,on how well alternative approaches actually perform in realistic 0304-4076/01/$-e front matter 2001Published by Elvier Science S.A.All rights rerved.PII:S 0304-4076(00)00045-2
8 D.F.Hendry/Journal of Econometrics100(2001)7}10
ttings.A notable exception is the t of Monte Carlo experiments in Lovell (1983)}who found that none of the methods which he implemented worked well.There has recently,however,been a dramatic change with the computer implementation of veral model-arch algorithms}e Hoover and Perez (1999),Hendry and Krolzig(1999),Hann(1999),Sullivan et al.(1998)and White(1997)inter alia.We focus on the"rst of the.云南介绍
血府逐瘀软胶囊
To program their Monte Carlo simulations for evaluating Gets,Hoover and Perez(1999)make veral important innovations.Firstly,they arch many reduction paths,not just successively smallest t values}so their algorithm avoids gettings&stuck'in a poor choice.Secondly,each lection remains congruent.Thirdly,they choo between any resulting(non-nested)terminal models by"t.Finally,they u sub-sample reliability to asss the lection. Re-running the experiments in Lovell(1983)revealed excellent performance by Gets in many experiments.
怎样食用灵芝Krolzig and Hendry(2000)develop an improved algorithm(written in Ox:e Doornik,1998)called PcGets,which adds block-reduction F-tests,calibrates the diagnostic tests,applies encompassing tests between terminal contenders,re-peats arches with non-unique outcomes,and us the Schwartz(or BIC) criterion for"nal-model lection if required.Thus,PcGets embodies all the principles discusd in Hendry(1995).In the Lovell Monte Carlo experiments with highly over-parame
terized ,40irrelevant variables when arching for a null model),PcGets improves substantially over the Hoover and Perez(1999),correctly locating that null model more than97%of the time:e Hendry and Krolzig,1999).In their own Monte Carlo experiments,Krolzig and Hendry(1999)report that PcGets re-covers the data-generation process(DGP)speci"cation with an accuracy clo to what one would expect when commencing from the DGP and conducting equivalent tests.Thus,the costs of structured arches are much lower than currently perceived}repeated testing need not induce over-sized models. Applied to the macro-data in Davidson et al.(1978)and Hendry and Ericsson (1991),PcGets lects models at least as good as tho developed over veral years by their authors.Thus,this operational version of the methodology con"rms its power}and we have barely scratched the surface of automated model-lection methodology.
不敢告诉你
Writing programs requires preci concepts and operations,and often high-lights lacunae in existing theories(e the estimator-generating equation in this journal prompted by programming FIML,Hendry,1976).Automating model lection is doing likewi.The multiple-path arch strategy plays a funda-mental role in improving performance over single-path&step-wi'arches;and ensuring congruence throughout avoids incorrect inferences from inappropriate standard errors.An i
mportant distinction between the costs of inference (the subject of extensive rearch under the rubric&pre-testing'e inter alia Bock et al.,1973;Judge and Bock,1978)and the costs of arch has been clari"ed:the
D.F.Hendry/Journal of Econometrics100(2001)7}109 latter is measured by the di!erence in Gets performance between commencing from the general model and from the DGP speci"cation,whereas the former is the di!erence between the lected model and the DGP starting point.The statistical properties of cumulative F-tests(which also perform well in Hann (1999)prior to the u of BIC),of many-path arches,of split-sample reliability and recursive testing,repeated reduction rounds with encompassing choices, and the interactions of diagnostic testing and lection,are all undergoing rigorous analys to clarify why such good lections result.Analysis of the number of paths to arch is required:one is too small,as the algorithm can get stuck,but arching every possible path ems to risk over-lecting adventi-tious e!ects.
The speci"cation of the initial general congruent model is central to the arch process.Careful prior analysis remains esntial,concerning the choice of variables,lag lengths,functional forms,orthogonality,parameterizations,(in-cluding asonals)etc.The larger the initial regressor t,the more likely adventitious e!ects will be retained,suggesting a key role for economic theory in &
prior simpli"cation'.But the smaller the t,the more likely that important variables will be omitted:given the low costs of arch,relatively generous initial parameterizations em nsible,which runs counter to much of modern macro-econometrics where&reprentative-agent,tightly parameterized'models are in vogue.
Many generalizations are both needed and feasible.For example,the impact of forced arch paths(to ensure the prence of some variables);system and instrumental-variables generalizations;combined diagnostic tests as suggested by Godfrey and Veale(1999),as well as studies of model revision following diagnostic-test rejections under local alternatives(not fal constructivism, but using the correct test for the problem);non-linearities;and cointegration reductions,to name a few already under way.Moreover,existing automations do not embody any&expert systems'to guide formulation and lection.&Hor races'between the approaches and ,White,1997)would be valu-able.Little is known about cas where programs like PcGets might#ounder (perhaps extremely collinear problems where many small e!ects matter).Never-theless,the entrance to this particular gold mine is now wide open,putting model lection"rmly on the rearch agenda with prospects of success,and suggesting that&data mining'might yet end being a compliment for an empirical study.
Acknowledgements
I am grateful to Jurgen Doornik,Neil Ericsson,and Hans-Martin Krolzig for comments.Financial support from the UK Economic and Social Rearch Council under grant R000234954is gratefully acknowledged.
10 D.F.Hendry/Journal of Econometrics100(2001)7}10
References
Backhou,R.,Salanti,A.(Eds.),2000.Macroeconomics and the real world.Oxford University Press,Oxford.
Bock,M.E.,Yancey,T.A.,Judge,G.C.,1973.Statistical conquences of preliminary test estimators in regression.Journal of the American Statistical Association68,109}116.
善良风俗Coppock,T.(Ed.),1999.Information Technology and Scholarship.Oxford University Press, Oxford.
d'Autume,A.,Cartelier,J.(Eds.),1996.L'Economie Devient-elle une Science Dure?Paris:Economi-ca.Reprinted in English as:Is Economics Becoming a Hard Science?Edward Elgar,Cheltenham, UK,1997.实习方案
Davidson,J.E.H.,Hendry,D.F.,Srba,F.,Yeo,J.S.,1978.Econometric modelling of the aggregate time-ries relationship between consumers'expenditure and income in the United Kingdom.
Economic Journal88,661}692.
Doornik,J.A.,1998.Object-Oriented Matrix Programming using Ox2.0.Timberlake Consultants Press,London.
Godfrey,L.G.,Veale,M.R.,1999.Alternative approaches to testing by variable addition.Mimeo, York University,UK.
Granger,C.W.J.(Ed.),1990.Modelling Economic Series.Clarendon Press,Oxford.羊昌河
Hann,B.E.,1999.Discussion of&Data mining reconsidered'.Econometrics Journal2,26}40. Hendry,D.F.,1976.The structure of simultaneous equations estimators.Journal of Econometrics4, 51}88.
Hendry,D.F.,1995.Dynamic Econometrics.Oxford University Press,Oxford.
Hendry,D.F.,Ericsson,N.R.,1991.An econometric analysis of UK money demand in&Monetary Trends in the United States and the United Kingdom by Milton Friedman and Anna J.
Schwartz'.American Economic Review81,8}38.
Hendry,D.F.,Krolzig,H.-M.,1999.Improving on&Data mining reconsidered'by K.D.Hoover and S.J.Perez.Econometrics Journal2,41}58.
Hendry,D.F.,Leamer,E.E.,Poirier,D.J.,1990.A conversation on econometric methodology.
Econometric Theory6,171}261.
华为手机忘记锁屏密码怎么办Hoover,K.D.,Perez,S.J.,1999.Data mining reconsidered:Encompassing and the general-to-speci"c approach to speci"cation arch.Econometrics Journal2,1}25.
Judge,G.G.,Bock,M.E.,1978.The Statistical Implications of Pre-Test and Stein-Rule Estimators in Econometrics.North-Holland,Amsterdam.
Krolzig,H.-M.,Hendry,D.F.,2000.Computer automation of general-to-speci"c model lection procedures.Forthcoming,Journal of Economic Dynamics&Control.
Lovell,M.C.,1983.Data mining.Review of Economics and Statistics65,1}12.
Magnus,J.R.,Morgan,M.S.(Eds.),1999.Methodology and Tacit Knowledge:Two Experiments in Econometrics.Wiley,Chichester.
Sullivan,R.,Timmermann,A.,White,H.,1998.Dangers of data-driven inference:The ca of calendar e!ects in stock returns.Mimeo,Economics Department,University of California at San Diego.
White,H.,1997.Data snooping with con"dence.Mimeo,Economics Department,University of California at San Diego.
