Suggested reading
Guelke (1978).
log-linear modelling Procedures for analys-
ing data at the nominal level ofmeasurement
- that is, when the variables comprised unor-
dered categories (such as a binary division –
urban/rural – or a set of political parties –
Conservative/Labour/Liberal-Democrat etc.).
The goal – as inregression(hence the alter-
native termlogistic regression) – is to fit an
equation that estimates the values in the cells
of a contingency table, where the independent
variables are also measured at either the nom-
inal (categorical) or ordinal level only; models
can also be fitted if one or more of the inde-
pendent variables are measured at interval or
ratio level. The terms of themodel, presented
in logarithmic form, indicate the deviations
from the individual cell values of the contin-
gency table from the grand mean for the entire
sample being analysed, and can be interpreted
as the odds of getting a particular outcome
(e.g. the likelihood that electors in urban
areas are more likely to vote for one political
party than those in rural areas). (See also
categorical data analysis). rj
Suggested reading
O’Brien (1992).
logit regression models These models bel-
ong to the family of techniques forcategorical
data analysis. They are used when the response
variable is either binary (e.g. a person either
moved or stayed) or a proportion (the propor-
tion who have moved). The nature of this
response variable, such that it cannot exceed
0 and 1, means that the standardregression
model should not be used. Instead, the logit
(the logarithm of the odds of moving) is mod-
elled and the random part of the model takes on
a binomial form so that the stochastic variation
(seestochastic process) around the underlying
relationship is structured to be least when the
values of 0 and 1 are approached. Although
the modelling is undertaken on the logits, it is
simple to transform these to both probabilities
and odds for interpretation. The popularity of
this form of the binary response model stems
from the ability when the predictor is categorical
of giving a relative odds interpretation, generat-
ing, for example, the relative odds of a person
aged over 35 moving in comparison to someone
aged under 35.
The multinomial form of the logit model is
used when there are more than two possible
outcomes; the conditional form is used when
analysing choice alternatives and the predictor
variables may include attributes of the choice
alternatives (e.g. cost) as well as characteristics
of the individuals making the choices (such as
income); the multi-level logit (see multi-
level modelling) assesses the effects of
variables measured at different levels (such as
individual, household and neighbourhood
effectson an individual moving); and the
nested logit model is used when there is a
hierarchical structure to the outcomes (with
movesbrokendownintoshort-andlong-
distance, for example). kjSuggested reading
Hensher and Greene (2004); Hosmer and
Lemshow (2000).longitudinal data analysis (LDA) A set of
quantitative methodsthat involve measures
over time. In contrast to time-series analysis,
in which there tends to be one entity measured
a large number of times (cf.sequence analy-
sis), LDA is usually concerned with a large
number of entities (e.g. people, places or
firms) measured a relatively few times. Data
for such analysis come from extensive designs
(see extensive research) such as panel or
cohort studies. LDA is increasing in import-
ance because it allows the study of develop-
ment and change, including: the transition
from one state to another (e.g. from single to
married to separated); the time spent in a
particular state (e.g. the length of unemploy-
ment); and the determinants of such duration.
LDA is fundamental to adopting a life-course
perspective in which people are affected by
cumulative stimuli over a long period.
The value of such an approach can be seen
by contrasting the data analysis of cross-
sections with that of a panel. In the former,
because we are measuring different people at
each time point, we can only assess net or
aggregate change – we can only know, for
example, that the percentage of the population
below the poverty line has increased from 10
to 12. But with panel data we can assess the
volatility of micro-social change as individuals
move in and out of poverty, thereby tackling
questions about the permanent nature of an
underclass. Such questions are of particular
importance inevidence-based policywhen
we are concerned with either affecting change
or removing barriers to change. Cross-
sectional analysis can also be misleading
about the direction of causality, for without
repeated measures over time we cannot distin-
guish between, say, unemployment causingGregory / The Dictionary of Human Geography 9781405132879_4_L Final Proof page 431 31.3.2009 2:44pmLONGITUDINAL DATA ANALYSIS (LDA)