Abstract
This chapter sketches out how Structured Additive Distributional Regression relates to other regression models, like classical linear models, quantile regression models and conditional transformation models. In addition, the chapter entails some remarks on covariate selection, model complexity and state space issues.
Essentially, all models are wrong, but some are useful.
George Box (1987)
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
Interestingly enough, the advancement of stochastic concepts ran parallel to the appearance of the fictional novel on the literary scene (see, Esposito 2007).
- 2.
For the cross-sectional data predominantly used in this book, OLS generally requires independent and identically distributed error terms that are exogenous. Moreover, the design matrix needs to be void of multicollinearity. Often homoscedasticity and normality are also assumed for inferential purposes. Despite frequently supposing a normality, OLS does not necessarily require any assumption on the nature of the distribution of the error terms though.
- 3.
Best in this context means giving the lowest variance of the estimate.
- 4.
It should be stressed that the requirement that the estimator is unbiased is quintessential to the theorem and other biased estimators have been found which can have better mean-squared error (MSE) properties as they feature lower variance, e.g. estimators from ridge regression.
- 5.
Note that alternatively one could also minimise the negative of the likelihood or related forms thereof like the deviance, where one basically subtracts the likelihood of a given model from another benchmark value thought to represent a saturated model.
- 6.
It should be noted that the term convergence is not used in a strict mathematical sense, like almost sure convergence or stochastic convergence, but rather in a heavily heuristic sense. For a more formally inclined discussion of convergence and estimator properties, see among others White (2001).
- 7.
Fisher criticised Bayesians for assuming that uncertainties can be expressed in form of probabilities (see Gigerenzer et al. 1989, p. 93). While it may be argued if that were true this would prompt inconsistencies all over inferential statistics, it is a fair point that the operationalisation of implicit assumptions can be highly challenging, to say the least.
- 8.
Varying coefficients and more generally mixed models can be seen as a model class that arguably bridges the divide between the frequentist and the Bayesian paradigm. For more information on mixed models the reader is referred to Safken (2015).
- 9.
The direct quote from Marx (1983, p. 189) reads: “Die Gesellschaft besteht nicht aus Individuen, sondern drückt die Summe der Beziehungen, Verhältnisse aus, worin diese Individuen zueinander stehen”.
- 10.
I thus focus on potential labour market experience rather than actual labour market experience. This choice is grounded in the belief that experience is not only derived out of employment spells but also from other life experiences such as caring for children.
- 11.
The reason for not using a continuous variable as in Mincer is that such a continuous variable would feature high point masses rather than a continuous spectrum of the distribution.
- 12.
For example, relating a normal distribution with two parameters to two variables linearly would require the estimation of 6 parameters when including a constant, while even a coarse grid of only ten analogously specified conditional quantiles to approximate the conditional distribution would already require the estimation of 30 parameters.
- 13.
For notational brevity I have excluded a distinct time-specific effect, \(b_t\). Since most economic panel databases for income analysis feature relatively few time periods, this effect can be captured by few more linear effects in the second term of Eq. (3.17).
- 14.
ISEs can either be seen as a fixed effect or a random effect. A fixed effect, in a frequentist setting, is conceived as an ultimately deterministic effect to which the estimator is thought to converge. A random effect, more akin to the Bayesian mode of thought, is conceived as a realisation from a random variable and thus ultimately stochastic in their nature. For sake of simplicity and clarity of contrast, I will assume the estimation of ISEs in the form of fixed effects in the frequentist setting in the following. This is warranted by the fact that fixed effects analysis is without a doubt much more frequent in panel data applications in economics than random effects analysis. The main reason for this is the emphasis put on unbiased estimators and the resultant application of the Hausman test (see Hausman 1978) which, in practice, almost invariably rejects random effect specification on the basis that the expected effects differ significantly from a fixed effect specification for the ISEs. Additionally, it may be argued that random effects are actually somewhere in between including fixed effects and outright excluding all ISEs. In a random effects specification, where the random effects are thought to follow a distribution converging towards a Dirac delta function, the results would be converging towards the results obtained from a specification without ISEs.
- 15.
Most software use transformations of the data, like the within-transformation or first differences to speed up the estimation process dramatically (see StataCorp 2011). By virtue of the transformation the direct estimation of the ISE is evaded and the other linear and nonlinear effects can easily be estimated. However, by evading the estimation of the ISEs, the model complexity is reduced but rather major parts of it are shifted outside the estimation process and potentially forgotten about.
- 16.
In some way, the inclusion of ISEs can be seen as a form of kitchen sink regression, whereby any conceivably relevant variable is lumped into the regression.
- 17.
Especially in the context of income analysis, the possibility of controlling for otherwise unobserved/unobservable factors must seem like divine aid sent by Athena herself to aid the valiant economists given the herculean task of tackling the hydra of the labour market. Accounting for ISEs thus seems to tie down all but one of the biting and hissing heads that influence labour market outcomes such as income. Yet, anyone studying the Greek mythology would know of the often twofold nature of divine aid and should be weary of the potential variety of its implications. Analogously, I believe that one needs to be weary of conditioning on covariates as complex and opaque as ISEs.
- 18.
As I point out in Sect. 2.3, I believe that this iterative loop does not only include analyses with quantitative methods but should explicitly also contemplate insights from qualitative research.
- 19.
Note that I do not conceive state space to be constrained to temporally varying states as is often done in the literature, but rather as a more general concept which allows to capture different economic functions for different populations across time, space or any dimension by which populations may be differentiated. Nonetheless, like in most state space models, I conceive time to be the pivotal dimension. Indeed, based on my research on the development of professorial salaries (Sohn 2016), I am a strong believer in temporal dependence of income structures. This dependence may be conceived in the framework of state space models in general and hidden Markov models in specific.
- 20.
This is very similar to the problem of samples consisting only of WEIRD (Western, Educated, Industrialised, Rich, Democratic Countries) in psychology (see Bellemare et al. 2008).
References
Aitken AC (1936) IV.–on least squares and linear combination of observations. Proc R Soc Edinb 55:42–48. doi:10.1017/S0370164600014346
Arnold BC (2008) Pareto and generalized pareto distributions. In: Chotikapanich D (ed) Modeling income distributions and lorenz curves. Springer, New York, pp 119–145
Belitz C, Brezger A, Klein N, Kneib T, Lang S, Umlauf N (2015) BayesX - software for Bayesian inference in structured additive regression models: version 3.0. http://www.bayesx.org
Bellemare C, Kröger S, van Soest A (2008) Measuring inequity aversion in a heterogeneous population using experimental decisions and subjective probabilities. Econometrica 76(4):815–839
Bondell HD, Reich BJ, Wang H (2010) Noncrossing quantile regression curve estimation. Biometrika 97(4):825–838. doi:10.1093/biomet/asq048
Bourdieu P (1995) Sozialer Raum und “Klassen”. Leçon sur la leçon: 2 Vorlesungen, vol 500, 3rd edn. Suhrkamp, Frankfurt am Main
Bourdieu P, Passeron JC (2007) Die Erben: Studenten, Bildung und Kultur. UVK-Verl.-Ges, Konstanz
Box GEP (1976) Science and statistics. J Am Stat Assoc 71(356):791–799
Box GEP (1987) Empirical model-building and response surfaces. Wiley, New York
Box GEP, Tiao G (1973) Bayesian inference in statistical analysis. Addison-Wesley, Rading
Brezger A, Lang S (2006) Generalized structured additive regression based on Bayesian P-splines. Comput Stat Data Anal 50(4):967–991. doi:10.1016/j.csda.2004.10.011
Chernozhukov V, Fernandez-Val I, Melly B (2013) Inference on counterfactual distributions. Econometrica 81(6):2205–2268. doi:10.3982/ECTA10582
Cox DR (1997) The current position of statistics: a personal view. Int Stat Rev 65(3):261–290
Cox DR, Fitzpatrick R, Fletcher AE, Gore SM, Spiegelhalter DJ, Jones DR (1992) Quality-of-life assessment: can we keep it simple. J R Stat Soc A 155(3):353–393
Dagum C (1977) A new model of personal income distribution: specification and estimation. Economie Applicée 30:413–437
Daston L (2001) Wunder, Beweise und Tatsachen: Zur Geschichte der Rationalität, orig.-ausg. edn. Fischer-Taschenbuch-Verl., Frankfurt am Main
Diebold FX (2013) No hesitations 2013: a blog book. http://www.ssc.upenn.edu/~fdiebold/papers/paper117/NoHesitations2013.pdf
Dobb M (1973) Theories of value and distribution since Adam Smith: ideology and economic theory. Cambridge University Press, Cambridge
Duesenberry JS (1949) Income, saving and the theory of consumer behaviour. (1967) Galaxy, Book edn. Oxford University Press, New York
Eilers PHC, Marx BD (1996) Flexible smoothing with B-splines and penalties. Stat Sci 11(2):89–102
Esposito E (2007) Die Fiktion der wahrscheinlichen Realität. Suhrkamp, Frankfurt am Main
Fahrmeir L, Kneib T, Lang S, Marx BD (2013) Regression: models, methods and applications. Springer, Berlin
Fortin NM, Lemieux T, Firpo S (2011) Decomposition methods in economics. In: Ashenfelter O, Card DE (eds) Handbook of labor economics, vol 4A. North-Holland, Amsterdam, pp 1–102
Galton F (1889) Natural inheritance. Macmillan, London
Gigerenzer G, Swijtink Z, Porter T, Daston L, Beatty J, Krüger L (1989) The empire of chance: how probability changed science and everyday life. Cambridge University Press, Cambridge
Gillham NW (2001) A life of sir francis galton: from African exploration to the birth of Eugenics. Oxford University Press, New York
Gneiting T, Katzfuss M (2014) Probabilistic forecasting. Annu Rev Stat Appl 1:125–151
Hambuckers J, Kneib T, Langrock R, Sohn A (2016) A Markov-switching generalized additive model for compound poisson processes, with applications to operational losses models. ZfS Working Paper 09/2016
Hausman JA (1978) Specification tests in econometrics. Econometrica 46(6):1251–1271. doi:10.2307/1913827
Herrnstein RJ (1971) IQ Atl Mon 228(3):43–64
Hothorn T, Kneib T, Bühlmann P (2014) Conditional transformation models. J R Stat Soc Ser B (Stat Methodol) 76(1):3–27
Jensen A (1969) How much can we boost IQ and academic achievement. Harv Educ Rev 39(1):1–123
Kleiber C (1996) Dagum versus Singh-Maddala income distributions. Econ Lett 57:39–44
Klein N, Kneib T, Lang S (2015a) Bayesian generalized additive models for location, scale and shape for zero-inflated and over-dispersed count data. J Am Stat Assoc 110(509):405–419. doi:10.1080/01621459.2014.912955
Klein N, Kneib T, Lang S, Sohn A (2015b) Bayesian structured additive distributional regression with an application to regional income inequality in Germany. Ann Appl Stat 9(2):1024–1052. doi:10.1214/15-AOAS823
Koenker R (2005) Quantile regression. Cambridge University Press, Cambridge
Koenker R, Bassett G (1978) Regression quantiles. Econometrica 46(1):33–50
Lancaster T (2004) An introduction to modern Bayesian econometrics. Blackwell, Oxford
Langrock R, Kneib T, Sohn A, DeRuiter S (2015a) Nonparametric inference in hidden markov models using P-Splines. Biometrics 71(2):520–528
Langrock R, Michelot T, Sohn A, Kneib T (2015b) Semiparametric stochastic volatility modelling using penalized splines. Comput Stat 30(2):517–537
Machado J, Mata J (2005) Counterfactual decompostion of changes in wage distributions using quantile regression. J Appl Econ 20(4):445–465
Marx K (1983) Ökonomische Manuskripte 1857/1858. In: Institut für Marxismus-Leninismus beim ZK der SED (ed) MEW, vol 42. Dietz, Berlin, pp 3–769
Melly B (2005) Public-private sector wage differentials in Germany: evidence from quantile regression. Empir Econ 30(2):505–520
Mincer J (1974) Schooling, experience, and earnings. National Bureau of Economic Research and distributed by Columbia University Press, New York
Oexle OG (2007) Krise des Historismus, Krise der Wirklichkeit: Wissenschaft, Kunst und Literatur 1880–1932. Vandenhoeck & Ruprecht, Göttingen
Pareto V (1897) Cours d’Economie Politique. In: Bousquet GH, Busino G (eds) New Edition (1964). Librairie Droz, Geneva
Peters J, Langbein J, Roberts G (forthcoming) Policy evaluation, randomized controlled trials and external validty - a systematic review. Econ Lett
Petty W (1899) Political arithmetic. In: Hull CHH (ed) The economic writings of Sir William Petty, together with The Observations upon Bills of Mortality, more probably by Captain John Graunt. Cambridge University Press, Cambridge, pp 233–313
Poincaré H (1902) La Sience et l’Hypothèse. Flammarion, Paris
Pudney S (1999) On some statistical methods for modelling the incidence of poverty. Oxf Bull Econ Stat 61(3):385–408
R Core Team (2012) R: a language and environment for statistical computing. http://www.R-project.org
Rennies H, Kneib T (2015) Structural equation models for dealing with spatial confounding. In: Proceedings of the 30th international workshop on statistical modelling, volume 2, pp 231–234
Reulen H (2015) Modelling combined transition-type effects in multi-state models. PhD thesis, Georg-August-Universität Göttingen, Göttingen
Rigby RA, Stasinopoulos DM (2005) Generalized additive models for location, scale and shape. J R Stat Soc Ser C (Appl Stat) 54(3):507–554
Robert CP (2007) The Bayesian choice: from decision-theoretic foundations to computational implementation, 2nd edn. Springer texts in statistics. Springer, New York
Rothe C, Wied D (2013) Misspecification testing in a class of conditional distributional models. J Am Stat Assoc 108(501):314–324. doi:10.1080/01621459.2012.736903
Rue H, Held L (2005) Gaussian Markov random fields: theory and applications. Chapman & Hall/CRC, Boca Raton
Säfken B (2015) Model choice and variable selection in mixed and semiparametric models. PhD thesis, Georg-August-Universität Göttingen, Göttingen. http://d-nb.info/1069664928/34
Schulze Waltrup L, Sobotka F, Kneib T, Kauermann G (2015) Expectile and quantile regression - David and Goliath? Stat Modell 15(5):433–456. doi:10.1177/1471082X14561155
Singh SK, Maddala GS (1976) A function for size distribution of incomes. Econometrica 44:963–970
Sobotka F, Kneib T (2012) Geoadditive expectile regression. Comput Stat Data Anal 56(4):755–767. doi:10.1016/j.csda.2010.11.015
Sohn A (2016) Poor university professors? The relative earnigns decline of German professors during the 20th century. Scand Econ Hist Rev 64(2):84–102. doi:10.1080/03585522.2016.1175374
StataCorp (2011) Stata statistical software. http://www.stata.com
Tableman M, Kim JS (2004) Survival analysis using S: analysis of time-to-event data. Chapman & Hall/CRC, Boca Raton, Fla
Vianelli S (1983) The family of normal and lognormal distributions of order r. Metron 41:3–10
Waldmann E, Kneib T, Yue YR, Lang S, Flexeder C (2013) Bayesian semiparametric additive quantile regression. Stat Modell 13(3):223–252. doi:10.1177/1471082X13480650
Waldmann E, Sobotka F, Kneib T (forthcoming) Bayesian regularisation in geoadditive expectile regression. Stat Comput
White H (2001) Asymptotic theory for econometricians. Academic Press, San Diego
Wood SN (2006) Generalized additive models. Chapman & Hall, Boca Raton
Wooldridge J (2011) Introductory econometrics, 5th edn. South-Western Cengage Learning, Mason
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
Copyright information
© 2017 The Author(s)
About this chapter
Cite this chapter
Silbersdorff, A. (2017). Estimating and Assessing Distributional Regression. In: Analysing Inequalities in Germany. SpringerBriefs in Statistics. Springer, Cham. https://doi.org/10.1007/978-3-319-65331-0_3
Download citation
DOI: https://doi.org/10.1007/978-3-319-65331-0_3
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-65330-3
Online ISBN: 978-3-319-65331-0
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)