SOCY7706: Longitudinal Data Analysis

Instructor: Natasha Sarkisian

Two Wave Panel Data Analysis

In any longitudinal analysis, we can distinguish between analyzing trends vs individual change –
that is, model the actual level of DV (Y) vs model the change in DV (ΔY). The predictors also
can be either actual levels (X=time-varying, Z=time-invariant) or measures of change (ΔX;
because ΔZ=0), as well as time itself (T).

We turn to the main approaches of explaining change in two wave panel dataset. We will review
four main approaches.

Lagged dependent variable model: X, Z  Y

Difference score model: X, Z  ΔY
First difference model: ΔX  ΔY
Cross-lagged model: X, Z Y and Y, Z  X

Lagged Dependent Variable (LDV) approach

This approach is also known as regressor variable approach. The idea is to predict time 2
outcome using time 1 independent variables while controlling for stability in the outcome
variable by including the dependent variable from time 1 into the model.
. reg rworkhours80 l. rworkhours80

Source | SS df MS Number of obs = 5897

-------------+------------------------------ F( 1, 5895) = 6301.45
Model | 1609174.94 1 1609174.94 Prob > F = 0.0000
Residual | 1505380.87 5895 255.365712 R-squared = 0.5167
-------------+------------------------------ Adj R-squared = 0.5166
Total | 3114555.82 5896 528.248951 Root MSE = 15.98

------------------------------------------------------------------------------
rworkhours80 | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
rworkhours80 |
L1. | .7368788 .0092827 79.38 0.000 .7186812 .7550763
|
_cons | 5.339778 .3551734 15.03 0.000 4.643507 6.036048
------------------------------------------------------------------------------

. reg rworkhours80 l. rworkhours80 l. rallparhelptw

Source | SS df MS Number of obs = 5767

-------------+------------------------------ F( 2, 5764) = 3081.77
Model | 1573387.1 2 786693.548 Prob > F = 0.0000
Residual | 1471395.53 5764 255.27334 R-squared = 0.5167
-------------+------------------------------ Adj R-squared = 0.5166
Total | 3044782.63 5766 528.058034 Root MSE = 15.977

------------------------------------------------------------------------------
rworkhours80 | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
rworkhours80 |
1
 L1. | .7345166 .0094029 78.12 0.000 .7160834 .7529498
|
rallparhel~w |
L1. | -.1601855 .0719849 -2.23 0.026 -.3013029 -.0190681
|
_cons | 5.483749 .3637186 15.08 0.000 4.770724 6.196774
------------------------------------------------------------------------------

. reg rworkhours80 l. rworkhours80 l. rallparhelptw l. rpoorhealth l. rmarried l.

rtotalpar l. rsiblog l.hchildlg raedyrs female age minority

Source | SS df MS Number of obs = 5457

-------------+------------------------------ F( 11, 5445) = 582.89
Model | 1557155.96 11 141559.633 Prob > F = 0.0000
Residual | 1322370.75 5445 242.859642 R-squared = 0.5408
-------------+------------------------------ Adj R-squared = 0.5398
Total | 2879526.71 5456 527.772491 Root MSE = 15.584

------------------------------------------------------------------------------
rworkhours80 | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
rworkhours80 |
L1. | .668734 .010576 63.23 0.000 .6480009 .6894672
|
rallparhel~w |
L1. | -.0942385 .0734988 -1.28 0.200 -.2383254 .0498485
|
rpoorhealth |
L1. | -4.44369 .5954816 -7.46 0.000 -5.611072 -3.276308
|
rmarried |
L1. | .4209347 .612163 0.69 0.492 -.7791495 1.621019
|
rtotalpar |
L1. | .2755657 .2905194 0.95 0.343 -.2939684 .8450998
|
rsiblog |
L1. | -.42027 .374524 -1.12 0.262 -1.154487 .3139468
|
hchildlg |
L1. | -.5223844 .400087 -1.31 0.192 -1.306715 .2619461
|
raedyrs | .1235686 .0776308 1.59 0.111 -.0286189 .2757561
female | -3.392911 .46171 -7.35 0.000 -4.298048 -2.487775
age | -.7810018 .0711669 -10.97 0.000 -.9205174 -.6414862
minority | -.7411717 .5320883 -1.39 0.164 -1.784278 .3019342
_cons | 52.52523 4.385398 11.98 0.000 43.92809 61.12236

We can do the same thing in wide format:

. reshape wide
(note: j = 1 2)

Data long -> wide

-----------------------------------------------------------------------------
Number of obs. 13182 -> 6591
Number of variables 20 -> 26
j variable (2 values) wave -> (dropped)
xij variables:
rworkhours80 -> r1workhours80 r2workhours80

2
 rpoorhealth -> r1poorhealth r2poorhealth
rmarried -> r1married r2married
rtotalpar -> r1totalpar r2totalpar
rsiblog -> r1siblog r2siblog
hchildlg -> h1childlg h2childlg
rallparhelptw -> r1allparhelptw r2allparhelptw
-----------------------------------------------------------------------------

. reg r2workhours80 r1workhours80 r1allparhelptw r1poorhealth r1married r1totalpar

r1siblog h1childlg age minority female raedyrs

Source | SS df MS Number of obs = 5457

-------------+------------------------------ F( 11, 5445) = 582.89
Model | 1557155.96 11 141559.633 Prob > F = 0.0000
Residual | 1322370.75 5445 242.859642 R-squared = 0.5408
-------------+------------------------------ Adj R-squared = 0.5398
Total | 2879526.71 5456 527.772491 Root MSE = 15.584
------------------------------------------------------------------------------
r2workhou~80 | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
r1workhou~80 | .668734 .010576 63.23 0.000 .6480009 .6894672
r1allparhe~w | -.0942385 .0734988 -1.28 0.200 -.2383254 .0498485
r1poorhealth | -4.44369 .5954816 -7.46 0.000 -5.611072 -3.276308
r1married | .4209347 .612163 0.69 0.492 -.7791495 1.621019
r1totalpar | .2755657 .2905194 0.95 0.343 -.2939684 .8450998
r1siblog | -.42027 .374524 -1.12 0.262 -1.154487 .3139468
h1childlg | -.5223844 .400087 -1.31 0.192 -1.306715 .2619461
age | -.7810018 .0711669 -10.97 0.000 -.9205174 -.6414862
minority | -.7411717 .5320883 -1.39 0.164 -1.784278 .3019342
female | -3.392911 .46171 -7.35 0.000 -4.298048 -2.487775
raedyrs | .1235686 .0776308 1.59 0.111 -.0286189 .2757561
_cons | 52.52523 4.385398 11.98 0.000 43.92809 61.12236
------------------------------------------------------------------------------

This format also allows us to examine interactions of the effects of each of the variables of
interest with the lagged DV.

Difference score approach

This approach is also known as the change score approach. There has been a lot of controversy
surrounding this approach.

. gen diff= r2workhours80- r1workhours80

(694 missing values generated)

. reg diff r1allparhelptw

Source | SS df MS Number of obs = 5767

3
-------------+------------------------------ F( 1, 5765) = 0.04
Model | 10.7404403 1 10.7404403 Prob > F = 0.8475
Residual | 1674892.93 5765 290.527828 R-squared = 0.0000
-------------+------------------------------ Adj R-squared = -0.0002
Total | 1674903.67 5766 290.479304 Root MSE = 17.045

------------------------------------------------------------------------------
diff | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
r1allparhe~w | -.0147277 .076598 -0.19 0.848 -.1648885 .1354331
_cons | -2.792029 .2297434 -12.15 0.000 -3.242412 -2.341645
------------------------------------------------------------------------------

. reg diff r1allparhelptw r1poorhealth r1married r1totalpar r1siblog h1childlg

raedyrs female age minority

Source | SS df MS Number of obs = 5457

-------------+------------------------------ F( 10, 5446) = 6.05
Model | 17340.4376 10 1734.04376 Prob > F = 0.0000
Residual | 1560639.79 5446 286.566249 R-squared = 0.0110
-------------+------------------------------ Adj R-squared = 0.0092
Total | 1577980.23 5456 289.21925 Root MSE = 16.928

------------------------------------------------------------------------------
diff | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
r1allparhe~w | -.0267496 .0798046 -0.34 0.737 -.1831985 .1296994
r1poorhealth | .2642639 .6259046 0.42 0.673 -.9627592 1.491287
r1married | 1.383919 .6641307 2.08 0.037 .0819573 2.68588
r1totalpar | .0906871 .3155152 0.29 0.774 -.5278488 .7092229
r1siblog | -.7903476 .406629 -1.94 0.052 -1.587503 .0068077
h1childlg | -.4283254 .4345873 -0.99 0.324 -1.28029 .4236395
raedyrs | -.1313198 .0838629 -1.57 0.117 -.2957246 .033085
female | 1.381293 .4734211 2.92 0.004 .4531982 2.309387
age | -.4761804 .0765798 -6.22 0.000 -.6263073 -.3260534
minority | -.578333 .5779601 -1.00 0.317 -1.711366 .5546998
_cons | 25.22486 4.668661 5.40 0.000 16.07242 34.3773
------------------------------------------------------------------------------
For many years, difference scores were criticized. One reason is their presumed unreliability – if
the DV for time 1 and time 2 are positively correlated (which is pretty much always the case),
then the difference score will have lower reliability than each of the time points individually, and
if the correlation across time is high, that decrease in reliability will be substantial.

But Paul Allison (1990) has argued that it is not a problem – “low reliability results from the fact
that in calculating the change score we differ out all the stable between-subject variation.” He
showed that what matters is measurement error, not unreliability – the same amount of error
variance that was contained in the individual scores just appears to be more prominent once the
stable component is removed, but in fact it has not changed.

Furthermore, change score models control for any unobserved factors as long as their effects are
stable over time, while the lagged dependent variable models do not, so this is a big advantage of
change score models.

The second critique is that difference score models do not account for the regression to the mean
effect—the trend wherein extremely low initial scores will be followed by an increase, and

4
extremely high scores – by a decrease. So the initial level might shape change, but if we add the
lagged DV to this change score model, we are back to the LDV model, so this strategy is not
useful:
. reg diff r1allparhelptw r1poorhealth r1married r1totalpar r1siblog h1childlg
raedyrs female age minority r1workhours80

Source | SS df MS Number of obs = 5457

-------------+------------------------------ F( 11, 5445) = 95.68
Model | 255609.477 11 23237.2252 Prob > F = 0.0000
Residual | 1322370.75 5445 242.859642 R-squared = 0.1620
-------------+------------------------------ Adj R-squared = 0.1603
Total | 1577980.23 5456 289.21925 Root MSE = 15.584

------------------------------------------------------------------------------
diff | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
r1allparhe~w | -.0942385 .0734988 -1.28 0.200 -.2383254 .0498485
r1poorhealth | -4.44369 .5954816 -7.46 0.000 -5.611072 -3.276308
r1married | .4209347 .612163 0.69 0.492 -.7791495 1.621019
r1totalpar | .2755657 .2905194 0.95 0.343 -.2939684 .8450998
r1siblog | -.42027 .374524 -1.12 0.262 -1.154487 .3139468
h1childlg | -.5223844 .400087 -1.31 0.192 -1.306715 .2619461
raedyrs | .1235686 .0776308 1.59 0.111 -.0286189 .2757561
female | -3.392911 .46171 -7.35 0.000 -4.298048 -2.487775
age | -.7810018 .0711669 -10.97 0.000 -.9205174 -.6414862
minority | -.7411717 .5320883 -1.39 0.164 -1.784278 .3019342
r1workhou~80 | -.331266 .010576 -31.32 0.000 -.3519991 -.3105328
_cons | 52.52523 4.385398 11.98 0.000 43.92809 61.12236
------------------------------------------------------------------------------
But Allison argued that regression to the mean does not always happen (although it is common)
– mostly if there are ceiling and/or floor effects (e.g., if the variable was measured in such a way
that it cannot go below above a certain value and above a certain value – that is usually the case
with scales, by the way); the correlation between the initial score and the increase does not have
to be negative – it can be positive and then the variance of scores increases with time. Allison
argues that regression to the mean is not a problem when we compare stable groups, and in such
cases, difference score approach may produce better results (less bias) than LDV approach.

Evaluating regression to the mean empirically by examining a group with high scores (above 75th
percentile) at time 1 and examining their distance from the mean at time 1 and time 2:
. for var r1workhours80: sum X, det \ scalar Xmean1=r(mean) \ gen sample=1 if
X>r(p75)\ sum X if X>r(p75)\di r(mean)-Xmean1

-> sum r1workhours80, det

1 rworkhours80
-------------------------------------------------------------
Percentiles Smallest
1% 0 0
5% 0 0
10% 0 0 Obs 6548
25% 0 0 Sum of Wgt. 6548

50% 40 Mean 30.73396

Largest Std. Dev. 22.52788
75% 45 80

5
90% 57 80 Variance 507.5055
95% 63 80 Skewness -.175734
99% 80 80 Kurtosis 1.930742

-> scalar r1workhours80mean1=r(mean)

-> gen sample=1 if r1workhours80>r(p75)

(5020 missing values generated)

-> sum r1workhours80 if r1workhours80>r(p75)

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
r1workhou~80 | 1528 57.37304 9.33897 46 80

-> di r(mean)-r1workhours80mean1
26.639072

. for var r2workhours80: sum X \ scalar Xmean1=r(mean) \ sum X if sample==1\di

r(mean)-Xmean1

-> sum r2workhours80

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
r2workhou~80 | 5929 28.22078 23.02388 0 80

-> scalar r2workhours80mean1=r(mean)

-> sum r2workhours80 if sample==1

Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
r2workhou~80 | 1429 46.80476 19.63145 0 80

-> di r(mean)-r2workhours80mean1
18.583979

These individuals moved closer to the mean. So we conclude that regression to the mean is a
problem for our data, so LDV will be better, especially if we want to document interactions
between the starting level of DV and the IVs.

First difference model

. for any poorhealth married totalpar siblog allparhelptw: gen Xdiff=r2X-r1X

-> gen poorhealthdiff=r2poorhealth-r1poorhealth

(627 missing values generated)

-> gen marrieddiff=r2married-r1married

(625 missing values generated)
6
-> gen totalpardiff=r2totalpar-r1totalpar
(691 missing values generated)

-> gen siblogdiff=r2siblog-r1siblog

(325 missing values generated)

-> gen allparhelptwdiff=r2allparhelptw-r1allparhelptw

(864 missing values generated)

. for any childlg: gen Xdiff=h2X-h1X

-> gen childlgdiff=h2childlg-h1childlg

(1132 missing values generated)

. reg diff allparhelptwdiff poorhealthdiff marrieddiff totalpardiff siblogdiff

childlgdiff

Source | SS df MS Number of obs = 5229

-------------+------------------------------ F( 6, 5222) = 2.88
Model | 4995.34416 6 832.55736 Prob > F = 0.0084
Residual | 1510362.04 5222 289.230571 R-squared = 0.0033
-------------+------------------------------ Adj R-squared = 0.0022
Total | 1515357.38 5228 289.854129 Root MSE = 17.007

------------------------------------------------------------------------------
diff | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
allparhelp~f | -.0796341 .0596778 -1.33 0.182 -.1966276 .0373593
poorhealth~f | -2.450367 .6794682 -3.61 0.000 -3.782409 -1.118325
marrieddiff | -.8902544 1.360583 -0.65 0.513 -3.557567 1.777058
totalpardiff | .5724302 .494059 1.16 0.247 -.3961321 1.540993
siblogdiff | -1.649011 2.908561 -0.57 0.571 -7.351007 4.052985
childlgdiff | 1.415648 1.658858 0.85 0.393 -1.836407 4.667703
_cons | -2.515716 .260116 -9.67 0.000 -3.025652 -2.00578
------------------------------------------------------------------------------

Once we created a first difference model, can we introduce time-invariant variables as well? We
can; by doing that, we are assuming that the effect of this time-invariant variable is not stable
over time, and interpret the resulting coefficient as an interaction term for time and that variable,
although since main effects are not in the model, it is difficult to interpret such results; such
models are usually not used.

. reg diff allparhelptwdiff poorhealthdiff marrieddiff totalpardiff siblogdiff

childlgdiff raedyrs female age minority

Source | SS df MS Number of obs = 5227

-------------+------------------------------ F( 10, 5216) = 6.43
Model | 18452.1386 10 1845.21386 Prob > F = 0.0000
Residual | 1496890.64 5216 286.980567 R-squared = 0.0122
-------------+------------------------------ Adj R-squared = 0.0103
Total | 1515342.77 5226 289.962261 Root MSE = 16.941

------------------------------------------------------------------------------
diff | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
allparhelp~f | -.0779137 .0595081 -1.31 0.190 -.1945744 .038747
poorhealth~f | -2.417475 .6781968 -3.56 0.000 -3.747025 -1.087926
marrieddiff | -.7896093 1.355911 -0.58 0.560 -3.447763 1.868545

7
totalpardiff | .4298372 .4928697 0.87 0.383 -.5363938 1.396068
siblogdiff | -1.740446 2.905442 -0.60 0.549 -7.436328 3.955437
childlgdiff | 1.10057 1.654093 0.67 0.506 -2.142146 4.343286
raedyrs | -.0944175 .0800286 -1.18 0.238 -.2513072 .0624722
female | 1.262989 .4708219 2.68 0.007 .3399806 2.185997
age | -.4535023 .0760188 -5.97 0.000 -.602531 -.3044735
minority | -.9362349 .5703079 -1.64 0.101 -2.054277 .1818075
_cons | 23.36358 4.426971 5.28 0.000 14.68486 32.0423
------------------------------------------------------------------------------

Cross-lagged panel model

This type of model, in many ways similar to LDV (in that it models level rather than change), is
useful if you are interested in mutual effects of two variables on one another:

. reg r2workhours80 r1workhours80 r1allparhelptw

Source | SS df MS Number of obs = 5767
-------------+------------------------------ F( 2, 5764) = 3081.77
Model | 1573387.1 2 786693.548 Prob > F = 0.0000
Residual | 1471395.53 5764 255.27334 R-squared = 0.5167
-------------+------------------------------ Adj R-squared = 0.5166
Total | 3044782.63 5766 528.058034 Root MSE = 15.977
------------------------------------------------------------------------------
r2workhou~80 | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
r1workhou~80 | .7345166 .0094029 78.12 0.000 .7160834 .7529498
r1allparhe~w | -.1601855 .0719849 -2.23 0.026 -.3013029 -.0190681
_cons | 5.483749 .3637186 15.08 0.000 4.770724 6.196774
------------------------------------------------------------------------------

. reg r2allparhelptw r1allparhelptw r1workhours80

Source | SS df MS Number of obs = 5697
-------------+------------------------------ F( 2, 5694) = 151.12
Model | 3376.80486 2 1688.40243 Prob > F = 0.0000
Residual | 63615.6175 5694 11.1723951 R-squared = 0.0504
-------------+------------------------------ Adj R-squared = 0.0501
Total | 66992.4223 5696 11.7613101 Root MSE = 3.3425
------------------------------------------------------------------------------
r2allparhe~w | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
r1allparhe~w | .261863 .0151334 17.30 0.000 .2321957 .2915302
r1workhou~80 | -.0008848 .0019782 -0.45 0.655 -.0047629 .0029932
_cons | 1.129847 .0764159 14.79 0.000 .9800425 1.279651
------------------------------------------------------------------------------

8
To establish causal predominance, we can compare standardized effects:
. reg r2allparhelptw r1allparhelptw r1workhours80, beta
Source | SS df MS Number of obs = 5697
-------------+------------------------------ F( 2, 5694) = 151.12
Model | 3376.80486 2 1688.40243 Prob > F = 0.0000
Residual | 63615.6175 5694 11.1723951 R-squared = 0.0504
-------------+------------------------------ Adj R-squared = 0.0501
Total | 66992.4223 5696 11.7613101 Root MSE = 3.3425
------------------------------------------------------------------------------
r2allparhe~w | Coef. Std. Err. t P>|t| Beta
-------------+----------------------------------------------------------------
r1allparhe~w | .261863 .0151334 17.30 0.000 .2240261
r1workhou~80 | -.0008848 .0019782 -0.45 0.655 -.0057909
_cons | 1.129847 .0764159 14.79 0.000 .
------------------------------------------------------------------------------

. reg r2workhours80 r1workhours80 r1allparhelptw, beta

Source | SS df MS Number of obs = 5767
-------------+------------------------------ F( 2, 5764) = 3081.77
Model | 1573387.1 2 786693.548 Prob > F = 0.0000
Residual | 1471395.53 5764 255.27334 R-squared = 0.5167
-------------+------------------------------ Adj R-squared = 0.5166
Total | 3044782.63 5766 528.058034 Root MSE = 15.977

------------------------------------------------------------------------------
r2workhou~80 | Coef. Std. Err. t P>|t| Beta
-------------+----------------------------------------------------------------
r1workhou~80 | .7345166 .0094029 78.12 0.000 .7171015
r1allparhe~w | -.1601855 .0719849 -2.23 0.026 -.0204278
_cons | 5.483749 .3637186 15.08 0.000 .

A better way of modeling these same relationships is to perform simultaneous estimation with
correlated residuals. We can do this with structural equation modeling (SEM). [For a more in-
depth exploration of SEM, take my class next year.]
. sem (r1workhours80 -> r2workhours80, ) (r1workhours80 -> r2allparhelptw, )
(r1allparhelptw -> r2workhours80, ) (r1allparhelptw -> r2allparhelptw, ), cov(
r1allparhelptw*r1workhours80 e.r2workhours80*e.r2allparhelptw) nocapslatent

Structural equation model Number of obs = 5651

Estimation method = ml
Log likelihood = -78231.18
--------------------------------------------------------------------------------------
| OIM
| Coef. Std. Err. z P>|z| [95% Conf. Interval]
---------------------+----------------------------------------------------------------
Structural |
r2workhours80 <- |
r1workhours80 | .7377979 .0095035 77.63 0.000 .7191713 .7564244
r1allparhelptw | -.1516555 .0723852 -2.10 0.036 -.2935278 -.0097831
_cons | 5.366347 .3672892 14.61 0.000 4.646473 6.086221
-------------------+----------------------------------------------------------------
r2allparhelptw <- |
r1workhours80 | -.0008713 .0019916 -0.44 0.662 -.0047748 .0030323
r1allparhelptw | .2616838 .0151696 17.25 0.000 .2319519 .2914157
_cons | 1.133529 .0769721 14.73 0.000 .9826666 1.284392
---------------------+----------------------------------------------------------------
Mean |
r1workhours80 | 30.77579 .2983912 103.14 0.000 30.19096 31.36063
r1allparhelptw | .6459817 .039176 16.49 0.000 .569198 .7227653
9
---------------------+----------------------------------------------------------------
Variance |
e.r2workhours80 | 255.4756 4.8062 246.2272 265.0714
e.r2allparhelptw | 11.22017 .2110823 10.81399 11.6416
r1workhours80 | 503.1497 9.465631 484.9353 522.0483
r1allparhelptw | 8.672941 .1631619 8.358973 8.998701
---------------------+----------------------------------------------------------------
Covariance |
e.r2workhours80 |
e.r2allparhelptw | -1.446145 .7124758 -2.03 0.042 -2.842572 -.0497185
-------------------+----------------------------------------------------------------
r1workhours80 |
r1allparhelptw | -4.739734 .8810168 -5.38 0.000 -6.466495 -3.012972

We can also request standardized coefficients in SEM by using the “standardized” option.
. sem (r1workhours80 -> r2workhours80, ) (r1workhours80 -> r2allparhelptw, )
(r1allparhelptw -> r2workhours80, ) (r1allparhelptw -> r2allparhelptw, ), cov(
r1allparhelptw*r1workhours80 e.r2workhours80*e.r2allparhelptw) nocapslatent stand

Structural equation model Number of obs = 5651

Estimation method = ml
Log likelihood = -78231.18
--------------------------------------------------------------------------------------
| OIM
Standardized | Coef. Std. Err. z P>|z| [95% Conf. Interval]
---------------------+----------------------------------------------------------------
Structural |
r2workhours80 <- |
r1workhours80 | .718444 .0064574 111.26 0.000 .7057877 .7311003
r1allparhelptw | -.0193887 .0092543 -2.10 0.036 -.0375268 -.0012505
_cons | .2329623 .0172625 13.50 0.000 .1991284 .2667962
-------------------+----------------------------------------------------------------
r2allparhelptw <- |
r1workhours80 | -.0056853 .0129958 -0.44 0.662 -.0311567 .0197861
r1allparhelptw | .2241885 .012667 17.70 0.000 .1993617 .2490153
_cons | .3297509 .0227067 14.52 0.000 .2852467 .3742552
---------------------+----------------------------------------------------------------
Mean |
r1workhours80 | 1.372021 .0185342 74.03 0.000 1.335694 1.408347
r1allparhelptw | .2193497 .0134617 16.29 0.000 .1929653 .2457341
---------------------+----------------------------------------------------------------
Variance |
e.r2workhours80 | .4814634 .009224 .4637198 .4998858
e.r2allparhelptw | .9495243 .0056756 .9384651 .9607137
r1workhours80 | 1 . . .
r1allparhelptw | 1 . . .
---------------------+----------------------------------------------------------------
Covariance |
e.r2workhours80 |
e.r2allparhelptw | -.0270108 .0132929 -2.03 0.042 -.0530644 -.0009572
-------------------+----------------------------------------------------------------
r1workhours80 |
r1allparhelptw | -.07175 .0132341 -5.42 0.000 -.0976885 -.0458116
--------------------------------------------------------------------------------------
We can also test for the equivalence of coefficients to determine causal predominance. It is
important to compare the standardized coefficients for this test, since the units for the b
coefficients are not identical.

. estat stdize: test

(_b[r2workhours80:r1allparhelptw]==_b[r2allparhelptw:r1workhours80])
10
 ( 1) [r2workhours80]r1allparhelptw - [r2allparhelptw]r1workhours80 = 0

chi2( 1) = 0.74
Prob > chi2 = 0.3909

Here, neither standardized coefficient is significantly larger than the other - we cannot reject the
null hypothesis that they are equal (p = 0.39).

There are a number of advantages to using SEM for two-wave analysis (e.g., construction of
latent variables, direct modeling of mediation, management of missing data via MLMV, etc.),
but one of the most practical for us here is the diagramming of paths. Stata allows you to specify
SEM models not only with syntax, but also by using path diagrams via its SEM Builder. (Many
SEM software packages will produce path diagrams as output along with results tables; Stata
only produces path diagram outputs if you specify the model using the SEM Builder.)

Using the dropdown menus, select: Statistics  SEM (structural equation modeling)  Model
building and estimation. As a note, in SEM measured variables are represented using rectangles,
while latent variables are represented by ellipses. Ordinary regression only uses measured
variables, so for our purposes here all you need to know is that our variables will be represented
using rectangles. In the SEM Builder, we can specify a model that matches the path diagram in
the notes above:

Click the “Estimate” button in the upper right-hand corner and hit “OK” for Maximum
likelihood estimation, and Stata will perform this simple cross-lagged SEM model.
Structural equation model Number of obs = 5651
Estimation method = ml
Log likelihood = -78231.18
--------------------------------------------------------------------------------------
| OIM
| Coef. Std. Err. z P>|z| [95% Conf. Interval]
---------------------+----------------------------------------------------------------
Structural |
r2workhours80 <- |
r1workhours80 | .7377979 .0095035 77.63 0.000 .7191713 .7564244
r1allparhelptw | -.1516555 .0723852 -2.10 0.036 -.2935278 -.0097831
_cons | 5.366347 .3672892 14.61 0.000 4.646473 6.086221
-------------------+----------------------------------------------------------------
r2allparhelptw <- |
r1workhours80 | -.0008713 .0019916 -0.44 0.662 -.0047748 .0030323
r1allparhelptw | .2616838 .0151696 17.25 0.000 .2319519 .2914157
_cons | 1.133529 .0769721 14.73 0.000 .9826666 1.284392
---------------------+----------------------------------------------------------------
Mean |
r1workhours80 | 30.77579 .2983912 103.14 0.000 30.19096 31.36063
r1allparhelptw | .6459817 .039176 16.49 0.000 .569198 .7227653

11
---------------------+----------------------------------------------------------------
Variance |
e.r2workhours80 | 255.4756 4.8062 246.2272 265.0714
e.r2allparhelptw | 11.22017 .2110823 10.81399 11.6416
r1workhours80 | 503.1497 9.465631 484.9353 522.0483
r1allparhelptw | 8.672941 .1631619 8.358973 8.998701
---------------------+----------------------------------------------------------------
Covariance |
e.r2workhours80 |
e.r2allparhelptw | -1.446145 .7124758 -2.03 0.042 -2.842572 -.0497185
-------------------+----------------------------------------------------------------
r1workhours80 |
r1allparhelptw | -4.739734 .8810168 -5.38 0.000 -6.466495 -3.012972

Note that the results are exactly the same whether the model is estimated using syntax or the
SEM Builder.

SEM also allows for standardized coefficients to be reported in the SEM Builder diagram, even
when the “standardized” option wasn’t requested at estimation. You can report the standardized
coefficients on the paths in the diagram by selecting View  Standardized estimates.

Assumptions of this type of analysis:

Finite causal lag corresponding to our measurement: In such models, we are assuming that causal
process happens with a specific lag, and the distance between time points in our dataset reflects,
or closely approximates that lag.

Continuity of causal process: This model assumes that the causal processes are continuous and
ongoing so we can observe that at any time.

Equality of causal lags: We assume that AB and BA causal lag is of the same length.

Diagnostics for longitudinal data with two time points:

Since the vast majority of the models we discussed can be estimated using OLS regression,
diagnostics should be conducted the same way as they are for OLS.

12