Multilevel Binary Response Models

1
Multilevel Binary Response Models
Session 4
Damon Berridge
2
Multilevel Binary Response Models

• In all of the multilevel linear models
  considered so far, it was assumed that the
  response variable has a continuous distribution
  and that the random coefficients and residuals
  are normally distributed.
• These models are appropriate where the expected
  value of the response variable at each level may
  be represented as a linear function of the
  explanatory variables.
• The linearity and normality assumptions can be
  checked using standard graphical procedures.
• There are other kinds of outcomes, however, for
  which these assumptions are clearly not
  realistic. An example is the model for which the
  response variable is discrete.
• Important instances of discrete response
  variables are binary variables (e.g., success vs.
  failure of whatever kind) and counts (e.g., in
  the study of some kind of event, the number of
  events happening in a predetermined time period).

3
For a binary variable yij that has probability
mij for outcome 1 and probability 1-mij for
outcome 0, the mean is
and the variance is
4
The Two-Level Logistic Model

• We start by introducing a simple two-level model
  that will be used to illustrate the analysis of
  binary response data.
• Assume that there are

The total number of level-1 observations across
level-2 units is given by
and the level-2 model becomes
5
such that
6

• The distribution of yij is called a Bernoulli
  distribution with parameter mij , and can be
  written as

The functional form for mij

• The probit model is based upon the assumption
  that the disturbances eij are independent
  standard normal variates, such that

7
Logit and Probit Transformations

• Interpretation of the parameter estimates
  obtained from either the logit or probit
  regressions are best achieved on a linear scale,
  such that for a logit regression, we can
  re-express mij as

• The probit model may be rewritten as

• For both the logit and probit functions, any
  probability value in the range 0,1 is
  transformed so that the resulting values of log
  it(mij) and probit(mij) will lie between - and
  .

• A further transformation of the probability
  scale that is sometimes useful in modelling
  binomial data is the complementary log-log
  transformation. This function again transforms a
  probability mij in the range 0,1 to a value in
  (-, ), using the relationship log-log(1-mij).

8
General Two-Level Logistic Models
9
Residual Intraclass Correlation Coefficient
For binary responses, the intraclass coefficient
is often expressed in terms of the correlation
between the latent responses y . Since the
logistic distribution for the level-1 residual,
eij, implies a variance of p2/3 3.29 , this
implies that for a two-level logistic random
intercept model with an intercept variance of
, the intraclass coefficient is
For a two-level random intercept probit model,
this type of intraclass correlation becomes
10
Likelihood
where
and
11
Binary response model Example C3

• Raudenbush and Bhumirat (1992) analysed data on
  children repeating a grade during their time at
  primary school. The data were from a national
  survey of primary education in Thailand in 1988,
  we use a sub set of that data here.

Raudenbush, S.W., Bhumirat, C., 1992. The
distribution of resources for primary education
and its consequences for educational achievement
in Thailand, International Journal of Educational
Research, 17, 143-164
Number of observations (rows) 7185 Number of
variables (columns) 4 The variables include the
following schoolid school identifier sex 1 if
child is male, 0 otherwise pped1 if the child
had pre primary experience, 0 otherwise repeat1
if the child repeated a grade during primary
school, 0 otherwise
12

• We take as the binary response variable, the
  indicator whether a child has ever repeated a
  class (repeat 0 no , 1 yes).
• The level-1 explanatory variables are child
  gender (gender 0 girl , 1 boy)
• Child pre-primary education ( pped 0 no , 1
  yes).
• The probability that a child will repeat a grade
  during the primary years, mij, is of interest.

13
Estimate a model with just a constant
where
Estimate a multilevel model with
14
Sabre commands
15
As gender is a dummy variable indicating whether
the pupil is a girl or a boy, it can be helpful
to rewrite a pair of fitted models, one for each
gender. By substituting the values 1 for boy and
0 for girl in gender, we get the boy's constant

and we can write
girl
boy