G89.2247 Lecture 2 Regression as paths and covariance structure Alternative “saturated” path models Using matrix notation to write linear models Multivariate Expectations Mediation G89.2247 Lect 2
Question: Does exposure to childhood foster care (X) lead to adverse outcomes (Y) ? Example of purported "causal model" X Y Y = B0 + B1X + e Regression approach B0 and B1 can be estimated using OLS Estimates can be expressed in terms of the sample variance of X (S2X), the sample covariance of X and Y (SXY), and the means of the two variables (MY and MX) e B1 G89.2247 Lect 2
Question: Does exposure to childhood foster care (X) lead to adverse outcomes (Y) ? Regression approach, continued In addition to estimating the structural coefficients, we will be interested in estimating the amount of variation in Y that is not explained by the model. .i.e., Var(Y|X) = Var(e). The correlation, rXY = SXY/SXSY, can be used to estimate the variance of the residual, e, V(e). S2e = S2Y(1-r2XY) = S2Y - S2XY/S2X G89.2247 Lect 2
A Covariance Structure Approach If we have data on Y and X we can compute a covariance matrix This estimates the population covariance structure, s2Y can itself be expressed as B21s2X + s2e Three statistics in the sample covariance matrix are available to estimate three population parameters G89.2247 Lect 2
Covariance Structure Approach, Continued A structural model that has the same number of parameters as unique elements in the covariance matrix is "saturated". Saturated models always fit the sample covariance matrix. G89.2247 Lect 2
Another saturated model: Two explanatory variables The first model is likely not to yield an unbiased estimate of foster care because of selection factors (Isolation failure). Suppose we have a measure of family disorganization (Z) that is known to have an independent effect on Y and also to be related to who is assigned to foster care (X) Y X Z e b2 b1 rXZ G89.2247 Lect 2
Covariance Structure Expression The model: Y=b0+b1X+b2Z+e If we assume E(X)=E(Z)=E(Y)=0 and V(X) = V(Z) = V(Y) = 1 then b0=0 and b's are standardized The parameters can be expressed When sample correlations are substituted, these expressions give the OLS estimates of the regression coefficients. G89.2247 Lect 2
Covariance Structure: 2 Explanatory Variables In the standardized case the covariance structure is: Each correlation is accounted by two components, one direct and one indirect There are three regression parameters and three covariances. G89.2247 Lect 2
The more general covariance matrix for two IV multiple regression If we do not assume variances of unity the regression model implies G89.2247 Lect 2
More Math Review for SEM Matrix notation is useful G89.2247 Lect 2
A Matrix Derivation of OLS Regression OLS regression estimates make the sum of squared residuals as small as possible. If Model is Then we choose B so that e'e is minimized. The minimum will occur when the residual vector is orthogonal to the regression plane In that case, X'e = 0 G89.2247 Lect 2
When will X'e = 0? When e is the residual from an OLS fit. G89.2247 Lect 2
Multivariate Expectations There are simple multivariate generalizations of the expectation facts: E(X+k) = E(X)+k = mx+k E(k*X) = k*E(X) = k*mx V(X+k) = V(X) = sx2 V(k*X) = k2*V(X) = k2*sx2 Let XT=[X1 X2 X3 X4], mT=[m1 m2 m3 m4] and let k be scalar value E(k*X) = k*E(X) = k*m E(X+k* 1) = {E(X) + k* 1} = m + k*1 G89.2247 Lect 2
Multivariate Expectations In the multivariate case Var(X) is a matrix V(X)=E[(X-m) (X-m)T] G89.2247 Lect 2
Multivariate Expectations The multivariate generalizations of V(X+k) = V(X) = sx2 V(k*X) = k2*V(X) = k2*sx2 Are: Var(X + k*1) = S Var(k* X) = k2S Let cT = [c1 c2 c3 c4]; cT X is a linear combination of the X's. Var(cT X) = cT S c This is a scalar value If this positive for all values of c then S is positive definite G89.2247 Lect 2
Partial Regression Adjustment The multiple regression coefficients are estimated taking all variables into account The model assumes that for fixed X, Z has an effect of magnitude bZ. Sometimes people say "controlling for X" The model explicitly notes that Z has two kinds of association with Y A direct association through bZ (X fixed) An indirect association through X (magnitude bXrXZ) G89.2247 Lect 2
Pondering Model 1: Simple Multiple Regression Y X Z e b2 b1 rXZ The semi-partial regression coefficients are often different from the bivariate correlations Adjustment effects Suppression effects Randomization makes rXZ = 0 in probability. G89.2247 Lect 2
Mathematically Equivalent Saturated Models Two variations of the first model suggest that the correlation between X and Z can itself be represented structurally. Y X Z eY b2 b1 eZ b3 Y X Z eY b2 b1 eX b3 G89.2247 Lect 2
Representation of Covariance Matrix Both models imply the same correlation structure The interpretation, however, is very different. G89.2247 Lect 2
Model 2: X leads to Z and Y X is assumed to be causally prior to Z. eY b2 b1 eZ b3 X is assumed to be causally prior to Z. The association between X and Z is due to X effects. Z partially mediates the overall effect of X on Y X has a direct effect b1 on Y X has an indirect effect (b3b2) on Y through Z Part of the bivariate association between Z and Y is spurious (due to common cause X) G89.2247 Lect 2
Model 3: Z leads to X and Y Z is assumed to be causally prior to X. eY b2 b1 eX b3 Z is assumed to be causally prior to X. The association between X and Z is due to Z effects. X partially mediates the overall effect of Z on Y Z has a direct effect b2 on Y Z has an indirect effect (b3b2) on Y through X Part of the bivariate association between X and Y is spurious (due to common cause Z) G89.2247 Lect 2
Choosing between models Often authors claim a model is good because it fits to data (sample covariance matrix) All of these models fit the same (perfectly!) Logic and theory must establish causal order There are other possibilities besides 2 and 3 In some instances, X and Z are dynamic variables that are simultaneously affecting each other In other instances both X and Z are outcomes of an additional variable, not shown. G89.2247 Lect 2
Mediation: A theory approach Sometimes it is possible to argue on theoretical grounds that Z is prior to X and Y X is prior to Y The effect of Z on Y is completely accounted for by the indirect path through X. This is an example of total mediation If b2 is fixed to zero, then Model 3 is no longer saturated. Question of fit becomes informative Total mediation requires strong theory G89.2247 Lect 2
A Flawed Example Someone might try to argue for total mediation of family disorganization on low self-esteem through placement in foster care Baron and Kenny(1986) criteria might be met Z is significantly related to Y Z is significantly related to X When Y is regressed on Z and X, b2 is significant but b1 is not significant. Statistical significance is a function of sample size. Logic suggests that children not assigned to foster care who live in a disorganized family may suffer directly. G89.2247 Lect 2
A More Compelling Example of Complete Mediation If Z is an experimentally manipulated variable such as a prime X is a measured process variable Y is an outcome logically subsequent to X It should make sense that X affects Y for all levels of Z E.g. Chen and Bargh (1997) Are participants who have been subliminally primed with negative stereotype words more likely to have partners who interact with them in a hostile manner? G89.2247 Lect 2