This R package implements the dynamic panel data modeling framework described by Allison, Williams, and Moral-Benito (2017). This approach allows fitting models with fixed effects that do not assume strict exogeneity of predictors. That means you can simultaneously get the robustness to confounding offered by fixed effects models and account for reciprocal causation between the predictors and the outcome variable. The estimating approach from Allison et al. provides better finite sample performance in terms of both bias and efficiency than other popular methods (e.g., the Arellano-Bond estimator).
These models are fit using structural equation models, using maximum
likelihood estimation and offering the missing data handling and
flexibility afforded by SEM. This package will reshape your data,
specify the model properly, and fit it with lavaan
.
If a result doesn’t seem right, it would be a good idea to
cross-reference it with xtdpdml
for Stata. Go to
https://www3.nd.edu/~rwilliam/dynamic/ to learn about xtdpdml
and
the underlying method. You may also be interested in the article by Paul
Allison, Richard Williams, and Enrique Moral-Benito in Socius,
accessible
here.
dpm
is now on CRAN and can be installed like other R packages.
install.packages("dpm")
This package assumes your data are in long format, with each row
representing a single observation of a single participant. Contrast this
with wide format in which each row contains all observations of a
single participant. For help on converting data from wide to long
format, check out the
tutorial that
accompanies the panelr
package.
First we load the package and the WageData
from panelr
.
library(dpm)
data("WageData", package = "panelr")
This next line of code converts the data to class panel_data
, which is
a class specific to the panelr
that helps to simplify the treatment of the long-form panel data. You
don’t have to do this, but it saves you from providing id
and wave
arguments to the model fitting function each time you use it.
wages <- panel_data(WageData, id = id, wave = t)
The formula syntax used in this package is meant to be as similar to a typical regression model as possible.
The most basic model can be specified like any other: y ~ x
, where y
is the dependent variable and x
is a time-varying predictor. If you
would like to include time-invariant predictors, you will make the
formula consist of two parts, separated with a bar (|
) like so:
y ~ x | z
where z is a time invariant predictor, like ethnicity.
One of the innovations of the method, however, is the notion of
pre-determined, or sequentially exogenous, predictors. To specify a
model with a pre-determined variable, put the variable within a pre
function, y ~ pre(x1) + x2 | z
. This tells the function that x1
is
pre-determined while x2
is strictly exogenous by assumption. You could
have multiple pre-determined predictors as well (e.g.,
y ~ pre(x1) + pre(x2) | z
).
You may also fit models with lagged predictors. Simply apply the lag
function to the lagged predictors in the formula:
y ~ pre(lag(x1)) + lag(x2) | z
. To specify more than 1 lag, just
provide it as an argument. For instance,
y ~ pre(lag(x1, 2)) + lag(x2) | z
will use 2 lags of the x1
variable.
This will replicate the analysis of the wages data in the Socius article that describes these models.
Note that to get matching standard errors, set
information = "observed"
to override lavaan
’s default,
information = "expected"
.
fit <- dpm(wks ~ pre(lag(union)) + lag(lwage) | ed, data = wages,
error.inv = TRUE, information = "observed")
summary(fit)
MODEL INFO:
Dependent variable: wks
Total observations: 595
Complete observations: 595
Time periods: 2 - 7
MODEL FIT:
𝛘²(76) = 138.476
RMSEA = 0.037, 90% CI [0.027, 0.047]
p(RMSEA < .05) = 0.986
SRMR = 0.025
| | Est. | S.E. | z val. | p |
|:------------------|-------:|------:|-------:|------:|
| union (t - 1) | -1.206 | 0.522 | -2.309 | 0.021 |
| lwage (t - 1) | 0.588 | 0.488 | 1.204 | 0.229 |
| ed | -0.107 | 0.056 | -1.893 | 0.058 |
| wks (t - 1) | 0.188 | 0.020 | 9.586 | 0.000 |
Model converged after 600 iterations
Any arguments supplied other than those that are documented within the
dpm
function are passed on to sem
from the lavaan
package.
The following arguments allow you to make changes to the default model specification:
y.lag
: By default the lag 1 value of the DV is included as a predictor (this is why they are dynamic models). You may choose a different value or multiple values instead, including 0 (no lagged DV at all).fixed.effects
: By default, the model is specified as a fixed effects model. If you set this to FALSE, you get a random effects specification instead.error.inv
: This constrains error variances to be equal in each wave. It is FALSE by default.const.inv
: This constrains the constants to be equal in each wave. It is FALSE by default, but if TRUE it eliminates cross-sectional dependence.y.free
: This allows the regression coefficient of the lagged DV to vary across time. It is FALSE by default and you can either set it to TRUE or to the specific lag number(s).x.free
: This allows the regression coefficients for the predictors to vary across time. It is FALSE by default and you can either set it to TRUE to set all predictors’ coefficients free over time or else pass a vector of strings of the predictors whose coefficients should be set free over time.alpha.free
: If TRUE, relaxes the constraint that the fixed effects are equal across time. Default is FALSE to be consistent with how fixed effects models normally work.partial.pre
: If TRUE (FALSE by default), predetermined lagged predictors will also be allowed to correlate with the contemporaneous error term as suggested by Paul Allison for scenarios when it’s not clear whether you have chosen the right lag structure.
You have most of the options available to you via lavaan
’s summary
method.
You can choose to omit any of: the z statistics (zstat = FALSE
), the
standard errors (se = FALSE
), or the p values (pvalue = FALSE
). You
may also add confidence intervals (ci = TRUE
) at any specified level
(ci.level = .95
). If you used bootstrapping for uncertainty intervals,
you can also specify the method (boot.ci.type = "perc"
).
The number of digits to print can be set via digits
or with the option
dpm-digits
. You may also standardize coefficients via lavaan
’s
method using standardize = TRUE
.
If you just want the lavaan
model specification and don’t want this
package to fit the model for you, you can set print.only = TRUE
. To
reduce the amount of output, I’m condensing wages
to 4 waves here.
dpm(wks ~ pre(lag(union)) + lag(lwage) | ed, data = wages[wages$t < 5,],
print.only = TRUE)
## Main regressions
wks_2 ~ en1 * union_1 + ex1 * lwage_1 + c1 * ed + p1 * wks_1
wks_3 ~ en1 * union_2 + ex1 * lwage_2 + c1 * ed + p1 * wks_2
wks_4 ~ en1 * union_3 + ex1 * lwage_3 + c1 * ed + p1 * wks_3
## Alpha latent variable (random intercept)
alpha =~ 1 * wks_2 + 1 * wks_3 + 1 * wks_4
## Alpha free to covary with observed variables (fixed effects)
alpha ~~ union_1 + union_2 + union_3 + lwage_1 + lwage_2 + lwage_3 + wks_1
## Correlating DV errors with future values of predetermined predictors
wks_2 ~~ union_3
## Predetermined predictors covariances
union_1 ~~ ed + lwage_1 + lwage_2 + lwage_3 + wks_1
union_2 ~~ ed + lwage_1 + lwage_2 + lwage_3 + union_1 + wks_1
union_3 ~~ ed + lwage_1 + lwage_2 + lwage_3 + union_1 + union_2 + wks_1
## Exogenous (time varying and invariant) predictors covariances
lwage_1 ~~ ed + wks_1
lwage_2 ~~ ed + lwage_1 + wks_1
lwage_3 ~~ ed + lwage_1 + lwage_2 + wks_1
ed ~~ wks_1
## DV error variance free to vary across waves
wks_2 ~~ wks_2
wks_3 ~~ wks_3
wks_4 ~~ wks_4
## Let DV variance vary across waves
wks_2 ~ 1
wks_3 ~ 1
wks_4 ~ 1
Alternately, you can extract the lavaan
model syntax and
wide-formatted data from the fitted model object to do your own fitting
and tweaking.
get_wide_data(fit)
get_syntax(fit)
The model is a special type of lavaan
object. This means most methods
implemented for lavaan
objects will work on these. You can also
convert the fitted model into a typical lavaan
object:
as(fit, "lavaan")
While you could convert the model to lavaan
model and apply any of
lavaan
’s functions to it (and you should!), as a convenience you can
use lav_summary()
to get lavaan
’s summary of the model.
Take advantage of lavaan
’s missing data handling by using the
missing = "fiml"
argument as well as any other arguments accepted by
lavaan::sem()
.
- CFI/TLI fit measures are much different than Stata’s and consistently more optimistic. For now, they are not printed with the summary because they are probably misleading.
You cannot use multiple lags of the same predictor (e.g.,(Fixed iny ~ x + lag(x)
).1.0.0
)- The function does not yet support input data that is already in wide format. (Not planning to fix)
You cannot apply arbitrary functions to variables in the formula like you can with regression models. For instance, a specification like(Works as ofy ~ scale(x)
will cause an error.1.1.0
)
Feature parity with xtdpdml
(Stata) is a goal. Here’s how we are doing
in terms of matching relevant xtdpdml
options:
-
alphafree
(asalpha.free
) -
xfree
(asx.free
) -
xfree(varlist)
(asx.free
) -
yfree
(added asy.free
argument in1.0.0
) -
yfree(numlist)
-
re
(added viafixed.effects
argument in1.0.0
) -
errorinv
(aserror.inv
) -
nocsd
/constinv
(asconst.inv
) -
ylag(numlist)
(added asy.lag
argument in1.0.0
; option to specify as 0 — no lagged DV — added in1.1.0
) -
std
(butstandardize
argument ofsummary
may suffice) -
dryrun
(asprint.only
)
Many and perhaps more SEM fitting options are implemented by virtue of
accepting any lavaan::sem()
argument.
- Get proper CFI/TLI statistics — this is a
lavaan
problem. - Allow full use of formula syntax, e.g.
y ~ scale(x)
(fixed in1.1.0
) - Add
broom
methods (tidy
,glance
) (addedtidy
in1.1.0
) - Create a
predict
method and perhaps some ability to plot predictions - Add
x.free
option to allow the coefficients of all predictors to vary across periods. This will make thesummary
output a pain, so it will take some time to implement. (added in1.1.1
)
Allison, P. (2022, October 24). Getting the lags right – a new solution. Statistical Horizons. https://statisticalhorizons.com/getting-the-lags-right-a-new-solution/
Allison, P. D., Williams, R., & Moral-Benito, E. (2017). Maximum likelihood for cross-lagged panel models with fixed effects. Socius, 3, 1–17. https://doi.org/10.1177/2378023117710578
Leszczensky, L., & Wolbring, T. (2022). How to deal with reverse causality using panel data? Recommendations for researchers based on a simulation study. Sociological Methods & Research, 51(2), 837–865. https://doi.org/10.1177/0049124119882473
Moral-Benito, E., Allison, P., & Williams, R. (2019). Dynamic panel data modelling using maximum likelihood: An alternative to Arellano-Bond. Applied Economics, 51, 2221–2232. https://doi.org/10.1080/00036846.2018.1540854
Williams, R., Allison, P. D., & Moral-Benito, E. (2018). Linear dynamic panel-data estimation using maximum likelihood and structural equation modeling. The Stata Journal, 18, 293–326. https://doi.org/10.1177/1536867X1801800201