Title: Causal Relationships with measurement error in the data
1Causal Relationships with measurement error in
the data
-
- A brief introduction
- by
- Willem E.Saris
2Basic concepts
Direct effect
y
x
y
Indirect effect
z
x
y
Spurious relation
z
x
x
z
Joint effect
w
y
3An example of a model
- How can these effects be estimated ?
4Decomposition rule
- The correlation between two variables is equal to
the sum of - - the direct effect,
- - indirect effects,
- - spurious relationships and
- - joint effects between these variables.
5Expression for the different components
- The indirect effect, spurious relations and joint
effects are equal to the products
of the coefficients
along the path going from one variable to the
other while one can not pass the same variable
twice and can not go against the direction of the
arrows.
6Derivations
- These derivations can also be used to estimate
the parameters of this model. How ?
7A second example
8A Structural Equations Model
9Derivations
10The Proof
11The correlations between the variables
- The effects are equal to the correlations with x1
12What if x1 is not observed ? Can we still
estimate the effects ?
13What happens if we have 4 observed variables ?
14Identification
- Of these three equations we need only one to
- determine the value of b41 when we have solved
- b11 and the other coefficients from the first
three - correlation coefficients
- This model is called overidentified or the
- degrees of freedom or df 2
- df correlations - parameters to be estimated
15A test is possible
- If we know that b11 .7 and that
- r(y1y4) b11b41 .35 it follows that b41 .5
- Now we know all coefficients and two correlations
are not used yet and can be used to test the
model - r(y2y4) b21b41 r(y3y4) b31b41
- r(y2y4) - r(y2y4) r(y2y4) - b21b41.3- .6x.5
.0 - r(y3y4) - r(y3y4) r(y3y4) - b31b41.5 - .8x.5
.1 - These differences are called residuals.
- If these residuals are big the model must be
wrong.
16Identification again
- With 3 observed variables df0 and no test is
possible - With 2 observed variables df-1 and no test is
possible but even the effects can not be
estimated - If dflt0 the model is not identified
17Estimation
- The decomposition rules only hold for the
population correlations and not for the sample
correlations - But , normally, we know only the sample
correlations - It is easily shown that the solution is different
depending of the equations used - So an efficient estimation procedure is needed.
18Estimation
- There are several general principles.
- We will discuss
- - the Unweighted Least Squares (ULS) procedure
- - the Weighted Least Squares (WLS) procedure.
- Both procedures are based on the residuals
between the sample correlations and the expected
values of the correlations.
19Estimation
- The expected correlations are a function of the
parameters fij(p) - where p represents the set of parameters of the
model - and fij the specific function which gives the
link between the population correlations and the
parameters for the variables i and j.
20ULS estimators
- The ULS procedure suggests to look for the
parameter values that minimize the unweighted sum
of squared residuals - FULS S(rij fij(p))2
- where the summation is over all unique elements
of the correlation matrix.
21Estimation in this specific case
The program looks for the values of all the
parameters that minimize the function Fuls
22WLS estimators
- The WLS procedure suggests to look for the
parameter values that minimize the weighted sum
of squared residuals - FWLS Swij(rij fij(p))2 where the summation
is also over all unique elements of the
correlation matrix. - These weights can be chosen in different ways.
23ADF estimator
- Using weights derived from the Variance
Covariances of the covariances the Asymptotic
Distribution Free estimator is specified. - For any distribution of the observed variables
this estimator is consistent and provides
standard errors and a test statistic - The problem is that it requires very large
samples
24ML estimator
- The most commonly used procedure, the Maximum
Likelihood (ML) estimator, can be specified as a
special case of the WLS estimator. - The ML estimator provides standard errors for the
parameters and a test statistic for the fit of
the model for much smaller samples - but this estimator is developed under the
assumption that the observed variables have a
multivariate normal distribution.
25Standard Procedure for testing S E Models
- Testing is essential for S E Models
- The test statistic t used is the value of the
fitting function at its minimum - If the model is correct, t is c2 (df) distributed
- Normally the model is rejected if t gt Ca
- where Ca is the value of the c2 for which
- pr(c2df gt Ca) a
- We come back to this issue later
26LISREL input
- estimation and testing a factor model
- data ni4 no400 makm
- km
- 1.0
- .42 1.0
- .56 .48 1.0
- .35 .30 .40 1.0
- model ny4 ne1 lyfu,fi tedi,fi psdi,fi
- free ly 1 1 ly 2 1 ly 3 1 ly 4 1
- free te 1 1 te 2 2 te 3 3 te 4 4
- value 1 ps 1 1
- out ULS
27LISREL estimates of the effects of the latent
factor
28LISREL estimates of the error variances
29Goodness of fit test
30LISREL input for different correlation matrix
estimation and testing a factor model data ni4
no400 makm km 1.0 .42 1.0 .56 .48 1.0 .35 .50
.50 1.0 model ny4 ne1 lyfu,fi tedi,fi
psdi,fi free ly 1 1 ly 2 1 ly 3 1 ly 4 1 free te
1 1 te 2 2 te 3 3 te 4 4 value 1 ps 1 1 out ULS
31Estimates of the effects of the latent variable
estimation and testing a factor model
Number of
Iterations 9 LISREL Estimates (Unweighted
Least Squares) LAMBDA-Y
ETA 1 -------- VAR
1 0.64 (0.05)
14.18 VAR 2 0.67 (0.04)
15.43 VAR 3 0.79
(0.05) 15.75 VAR 4
0.64 (0.05) 14.28
32Goodness of fit test of the model on the new
correlation matrix
Goodness of Fit Statistics W_A_R_N_I_N_G
Chi-square, standard errors, t-values and
standardized residuals are calculated under the
assumption of multi-variate normality.
Degrees of Freedom 2 Normal Theory Weighted
Least Squares Chi-Square 19.62 (P
0.00) Estimated Non-centrality Parameter (NCP)
17.62 90 Percent Confidence Interval for NCP
(6.96 35.72)
33General Approach
- A model is specified with observed and latent
variables - Correlations (covariances) between the observed
variables can be expressed in the parameters of
the model (decomposition rules) - If the model is identified the parameters can be
estimated - A test of the model can be performed if dfgt0
- Eventual misspecifications can be detected
- Corrections in the models can be introduced
34Important Result
- The distinction between observed and latent
variables makes the estimation of error
variances possible - The errors in social science survey data can be
quite large. - These errors will bias the estimates if not taken
into account - So the SEM approach has important advantages