|
|
Anglo-French
Workshop
In a state?
Although the use of
cross-section survey data is now commonplace in applied economics, the
evaluation of models based on such data has only very recently attracted
the attention it deserves. This new emphasis on model evaluation has
occurred almost simultaneously in both France and the UK; it was
therefore natural for the Centre to bring together these researchers at
a workshop on 13/14 March, organized by ATE Co-director Richard
Blundell.
Models based on cross-section data are very different from those
involving aggregate time-series data: they rely more heavily on
assumptions concerning unobservable attributes or characteristics of
individuals in the sample. Policy prescriptions based on estimated
cross-section models will only be reliable if the assumptions concerning
these unobservable attributes provide a good approximation to reality.
The papers presented at the March workshop focussed on simple methods by
which the researcher could assess whether these underlying assumptions
were reasonable.
In cross-section data, individuals are typically observed to be in
different 'states', e.g. employed or unemployed, a purchaster or a
non-purchaser of a particular good. An individual is either 'in' or
'out' of a particular state and this form of data is therefore
'discrete' or 'zero-one' in nature. Now suppose we observe an individual
to be in a particular state. We can then record their 'level of
activity' within this state. This level of activity is a more
conventional variable, which can assume a range of values, although it
is generally taken to be greater than zero. For example, we can record
the number of hours worked for an individual who is in employment, or
the level of expenditure of those who are purchasers of a particular
good. Cross-section models are distinctive because they combine
'discrete' variables (whether in or out of a particular state) with
'continuous censored' variables (the positive level of activity of those
in the state). This requires researchers to specify carefully the
factors influencing the probability that an individual is in a
particular state and the factors determining the level of activity of
individuals in that state. Some assumptions must be made in order to
model these probabilities and activity levels, yet these assumptions are
often left untested, even though they are are crucial to the policy
conclusions which are drawn from the model.
The first paper of the workshop, by Richard Blundell (University
College, London, and CEPR) and Costas Meghir (UCL), underlined
the importance of these considerations. In 'Bivariate Limited Dependent
Variable Models of Individual Labour Supply and Commodity Demand
Behaviour', Blundell and Meghir developed a model in which the
individual's level of activity once in a particular state was determined
separately from their probability of occupying that state. They
illustrated this approach using two examples: the first from the labour
market, where the state of employment or unemployment was determined
separately from the hours worked when in employment; and the second from
consumer behaviour, where the decision to buy or not to buy was modelled
separately from the decision of how much to buy. Blundell and Meghir
discussed simple methods for testing the underlying assumptions of these
models. They also presented procedures for testing that the hypothesis
that the same behavioural model could explain both the probability of
being in a state and the level of activity once in that state.
The discussion focussed on the desirability of model selection
procedures which relied on a sequence of 'reliability' tests. The tests
in the sequence were calculated from the same data set and were
therefore related in a complex fashion. It was therefore difficult to
determine the overall properties of the evaluation procedure: papers
presented by Alain Trognon and Richard Smith later in the workshop
returned to this important issue. Some workshop participants questioned
whether, even for the large data sets common in cross-section analysis
(often 2000 or more observations), Blundell and Meghir were justified in
their use of asymptotic tests.
Francois Bourguignon (Centre d'Economie Quantitative et
Comparative, Paris) presented the workshop's second paper, entitled
'Internal Labour Markets: An Econometric Analysis'. Bourguignon used a
cross-section of work histories to describe the transition between
states in the labour market. He noted that analyses of work histories
are often based on the records of an individual firm and are therefore
unreliable: individuals leave the firm, and the investigator loses track
of their work history. Those who leave the firm are usually not
representative of all employees, and the data available from those who
remain are not a random sample of the firm's workforce: it is therefore
difficult to retrieve an unbiased picture of transition probabilities.
The data used by Bourguignon and his co-author Pierre-Andre Chiappori
were drawn, however, from the records of a large single firm. Work
histories were relatively complete, because few individuals left the
firm and transitions occurred almost wholly within the firm, which
constituted an almost perfect internal labour market. Bourguignon and
Chiappori modelled the probability of an individual being observed at
different 'states' or stages of seniority within the firm. The
transition probabilities, that is, the probabilities of promotion were
influenced by both individual and career characteristics. Workshop
participants noted that fluctuations in the demand for labour, assumed
constant in Bourguignon's model, could have led to biased estimates of
the effects of age and length of service on promotion.
The random components in many cross-section models are assumed to behave
as if drawn from a normal distribution. Papers presented by Manuel
Arrellano (Institute of Economics and Statistics, Oxford) and Tim
Fry (Manchester) proposed interesting alternatives to the normality
assumption. In his paper, 'Estimating Contaminated Limited Dependent
Variable Models', Arrellano analyzed models in which the random
components were formed from a mixture of normal errors with different
variances. This distribution offered a computationally simple
alternative to the usual assumption of normality, Arrellano argued. As a
by- product, it also led to a simple test for the appropriateness of the
conventional normality assumption.
Almost every conventional evaluation procedure for econometric models is
based on an analysis of the model's residuals, the differences between
the predictions of the model and the observed data. The papers presented
by Alain Trognon (Institut National de la Statistique et des
Etudes Economiques, Paris) and Andrew Chesher (Bristol and CEPR)
attempted to develop the theory of residuals analysis for cross-section
models. The combination of discrete and continuous variables makes such
models very different from their time-series counterparts: residuals
cannot be analysed in the conventional time-series fashion, and new
approaches are needed. Trognon and Chesher developed the notion of
'generalized residuals' whose statistical properties are relatively
simple to establish and which can therefore serve as the basis for tests
of model 'reliability'. These generalized residuals are defined as the
expectations of the underlying or 'latent' residuals, conditional upon
the observed data and the structure of the model.
Trognon's paper, entitled 'Using Simulated and Generalized Residuals'
and written jointly with Alain Monfort, Christian Gourieroux and Eric
Renault (INSEE), developed a general theory of such residuals which
could be applied to many of the commonly used cross-section models.
Generalized residuals could be used, for example, to test the
assumptions that the random components were distributed normally, with a
constant variance and independently of each other. The investigator
would not need to undertake further estimation of the model in order to
test these assumptions. Trognon discussed an alternative procedure, in
which, given the parameters of an estimated model, simulated data can be
generated and simulated residuals calculated. Trognon showed how these
simulated residuals can then be used to plot residuals and to calculate
model reliability tests in much the same way as residuals are used
conventionally. In the discussion of the paper it was pointed out that
this simulation procedure did not make model estimation easier, but only
simplified the formulae for specification tests. Since these tests
depended on simulated residuals, this simplification was at the expense
of introducing additional 'noise' by virtue of the simulation itself.
The problems of using 'generalized residuals' were further outlined by Andrew
Chesher (Bristol and CEPR) in his paper 'Graphical Analysis of
Censored Data', written jointly with Margaret Irish (Bristol).
Generalized residuals do not possess the same statistical distribution
as conventional residuals: their analysis is therefore somewhat more
complicated. Chesher illustrated the problems involved by considering
the omission of variables from a cross-section model explaining the
duration of unemployment. Variables that are correctly omitted from the
model can appear as if they belong in the model, if correlated with
variables which are included in the regression. To overcome this problem
and to make plots of 'generalized residuals' easier to interpret,
Chesher introduced a method of smoothing the residuals which provided a
clearer picture of potential misspecification.
Blundell and Meghir had argued in their opening paper of the workshop
that many cross-section models involved random components which were
best analysed as mixtures of discrete and continuous dependent
variables. Richard Smith (Manchester) in his paper, 'Testing the
Normality Assumption in Multivariate Simultaneous Limited Dependent
Variable Models', described a methodology for testing a key assumption
underlying such models. Smith's test statistic involved higher order
moments of the generalized residuals and bore a close resemblance to the
popular tests based on the Information Matrix. The subsequent discussion
focussed on whether this result held generally for all cross- section
models. It was argued that Smith's exact correspondence between the
'generalized residual' and Information Matrix tests disappears in models
where there is selection bias, i.e. where data are not observed for some
part of the underlying population. Participants concluded that the
elements of tests based on 'generalized residuals' were always present
in the Information Matrix test. The choice of test statistic therefore
depended on whether it was non-normality or some broader alternative,
like parameter variation, which was thought to be the potential source
of misspecification.
When dealing with large data sets and large numbers of decision
variables, the ability to group decisions in some hierarchical ordering
is crucial for tractability in estimation. Francois Laisney
(Toulouse), in his paper 'A Method for the Description of Separability
Properties of Large Empirical Demand Systems', described a simple
algorithm for grouping consumer decisions in such a hierarchy. This
procedure was applied to a model of consumer behaviour in France with
some success. However, it was concluded that the resulting grouping
depends rather heavily on the broad assumptions about preferences or
technology that are made at the outset.
Conventional econometric theory is based on the estimation of
parameters, that is, numbers which characterise particular probability
distributions: the normal distribution, for example is characterised by
its mean and variance. The disadvantage of such parametric techniques is
their reliance on the assumed form of the underlying distribution; other
techniques have been devised which avoid parameter estimation and its
reliance on distributional assumptions. In the final paper of the
workshop, Tony Lancaster (Hull) described the properties of a
'semi- parametric' model for cross-section data on unemployment
durations. Although Lancaster's model did not need strong distributional
assumptions, the test statistics he had developed behaved quite poorly
even for the large samples typical of cross- section data. This
confirmed what many participants feared: that non-parametric modelling
could result in a considerable loss of efficiency in estimation and
testing. But for the particular test that Lancaster was considering, a
transformation was suggested that seemed at first sight to improve the
properties of the test considerably.
As a result of this workshop, there was general agreement that a useful
set of procedures for assessing the adequacy of cross- section models
was now available. But the properties of these test statistics, when
applied to the data sets typically used in applied microeconomics, still
remained somewhat uncertain and required further research.
|
|