WARNING: This function is still in development and has not been tested throughly.
Following Aral et al. (2009), netmatch computes matching
estimators for network data. The function netmatch_prepare, which
prepares the data to be used with matchit from
the MatchIt package, is called by netmatch.
netmatch_prepare(
dat,
graph,
timevar,
depvar,
covariates,
treat_thr = rep(1L, length(graph)),
adopt_thr = rep(1L, length(graph)),
expo_pcent = FALSE,
expo_lag = 0L
)
netmatch(
dat,
graph,
timevar,
depvar,
covariates,
treat_thr = rep(1L, length(graph)),
adopt_thr = rep(1L, length(graph)),
expo_pcent = FALSE,
expo_lag = 0L,
...
)data.frame with dynamic data. Must be of
nrow(dat)==nslices(graph)*nnodes(graph).
List with sparse matrices.
Character scalar. Name of time variable
Character scalar. Name of the dependent variable
Character vector. Name(s) of the control variable(s).
Either a numeric scalar or vector of length nslices(graph).
Sets the threshold of exposure at which it is considered that an
observation is treated.
Either a numeric scalar or vector of length nslices(graph).
Sets the threshold of depvar at which it is considered that an observation
has adopted a behavior.
Logical scalar. When TRUE, exposure is computed
non-normalized (so it is a count rather than a percentage).
Integer scalar. Number of lags to consider when computing
exposure. expo_lag=1 defines exposure in T considering behavior and
network at T-1.
Further arguments to be passed to matchit.
In the case of netmatch_prepare
A data.frame with the original data (covariates), plus the
following new variables: treat, adopt, exposure.
A formula to be passed to netmatch
netmatch returns the following:
A numeric vector of length \(N_1\) (number of treated used in the matching process). Treatment effects on the treated at the individual level
The output from matchit.
In Aral et al. (2009), the matching estimator is used as a response to the fact that the observed network is homophilous. Essentially, using exposure as a treatment indicator, which is known to be endogenous, we can apply the same principle of matching estimators in which, after controlling for characteristics (covariates), individuals from the treated group (exposed to some behavior) can be compared to individuals from the control group (not exposed to that behavior), as the only difference between the two is the exposure.
As pointed out in King & Nielsen (2015), it is suggested that, contrary to what Aral et al. (2009), the matching is not performed over propensity score since it is know that the later can increase imbalances in the data and thus obtaining exactly the opposed outcome that matching based estimators pursue.
A couple of good references for matching estimators are Imbens and Wooldridge (2009), and Sekhon (2008).
Aral, S., Muchnik, L., & Sundararajan, A. (2009). Distinguishing influence-based contagion from homophily-driven diffusion in dynamic networks. Proceedings of the National Academy of Sciences of the United States of America, 106(51), 21544–21549. doi:10.1073/pnas.0908800106
Imbens, G. W., & Wooldridge, J. M. (2009). Recent Developments in the Econometrics of Program Evaluation. Journal of Economic Literature, 47(1), 5–86. doi:10.1257/jel.47.1.5
King, G., & Nielsen, R. (2015). Why Propensity Scores Should Not Be Used for.
Sekhon, J. S. (2008). The Neyman-Rubin Model of Causal Inference and Estimation Via Matching Methods. The Oxford Handbook of Political Methodology. doi:10.1093/oxfordhb/9780199286546.003.0011