-
Notifications
You must be signed in to change notification settings - Fork 0
/
README.Rmd
193 lines (138 loc) · 12.1 KB
/
README.Rmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
---
output: github_document
---
<!-- README.md is generated from README.Rmd. Please edit that file -->
```{r, include = FALSE}
knitr::opts_chunk$set(
collapse = TRUE,
comment = "#>",
fig.path = "man/figures/README-",
out.width = "100%"
)
```
# PAGFL
<!-- badges: start -->
[](https://cran.r-project.org/package=PAGFL)
[](https://cran.r-project.org/package=PAGFL) [](https://www.gnu.org/licenses/gpl-3.0.html)
[](https://github.com/Paul-Haimerl/PAGFL/actions/workflows/R-CMD-check.yaml)
[](https://app.codecov.io/gh/Paul-Haimerl/PAGFL)
<!-- badges: end -->
Unobservable group structures are a common challenge in panel data analysis. Disregarding group-level heterogeneity can introduce bias. Conversely, estimating individual coefficients for each cross-sectional unit is inefficient and may lead to high uncertainty.
This package efficiently addresses the issue of unobservable group structures by implementing the pairwise adaptive group fused Lasso (*PAGFL*) by Mehrabani ([2023](https://doi.org/10.1016/j.jeconom.2022.12.002)). *PAGFL* is a regularizer that identifies latent group structures and estimates group-specific coefficients in a single step. On top of that, we extend the PAGFL to time-varying functional coefficients.
The `PAGFL` package makes this powerful procedure easy to use. On top of that, we extend the `PAGFL` to time-varying functional coefficients.
## Installation
Always stay up-to-date with the development version of `PAGFL` (1.1.2) from [GitHub](https://github.com/):
```{r message=FALSE, warning=FALSE}
# install.packages("devtools")
devtools::install_github("Paul-Haimerl/PAGFL")
library(PAGFL)
```
The stable version (1.1.2) is available on CRAN:
```
install.packages("PAGFL")
```
## Data
The `PAGFL` package includes a function that automatically simulates a panel data set with a group structure in the slope coefficients:
```{r}
# Simulate a simple panel with three distinct groups and two exogenous explanatory variables
set.seed(1)
sim <- sim_DGP(N = 20, n_periods = 150, p = 2, n_groups = 3)
data <- sim$data
```
$$y_{it} = \beta_i^\prime x_{it} + \eta_i + u_{it}, \quad i = 1, \dots, N, \quad t = 1, \dots, T,$$ where $y_{it}$ is a scalar dependent variable, $x_{it}$ a $p \times 1$ vector of explanatory variables, and $\eta_i$ reflects a fixed effect. The slope coefficients are subject to the group structure
$$\beta_{i} = \sum_{k = 1}^K \alpha_k \boldsymbol{1} \{i \in G_k \},$$ with $\cup_{k = 1}^K G_k = \{1, \dots, N \}$, and $G_k \cap G_j = \emptyset$ as well as $|| \alpha_k \neq \alpha_j ||$ for any $k \neq j$, $k,j = 1, \dots, K$ (see Mehrabani [2023](https://doi.org/10.1016/j.jeconom.2022.12.002), sec. 2).
`sim_DGP` also nests, among other, all DGPs employed in the simulation study of Mehrabani ([2023](https://doi.org/10.1016/j.jeconom.2022.12.002), sec. 6). I refer to the documentation of `sim_DGP` or Mehrabani ([2023](https://doi.org/10.1016/j.jeconom.2022.12.002), sec. 6) for more details.
## Applying PAGFL
To execute the PAGFL procedure, pass the dependent and independent variables, the number of time periods, and a penalization parameter $\lambda$.
```{r}
estim <- pagfl(y ~ X1 + X2, data = data, n_periods = 150, lambda = 20, verbose = F)
summary(estim)
```
`pagfl()` returns an object of type `pagfl`, which holds
1. `model`: A `data.frame` containing the dependent and explanatory variables as well as individual and time indices (if provided).
2. `coefficients`: A $K \times p$ matrix of the post-Lasso group-specific parameter estimates.
3. `groups`: A `list` containing (i) the total number of groups $\hat{K}$ and (ii) a vector of estimated group memberships $(\hat{g}_1, \dots, \hat{g}_N)$, where $\hat{g}_i = k$ if $i$ is assigned to group $k$.
4. `residuals`: A vector of residuals of the demeaned model.
5. `fitted`: A vector of fitted values of the demeaned model.
6. `args`: A list of additional arguments.
7. `IC`: A `list` containing (i) the value of the IC, (ii) the employed tuning parameter $\lambda$, and (iii) the mean squared error.
8. `convergence`: A `list` containing (i) a logical variable if convergence was achieved and (ii) the number of executed *ADMM* algorithm iterations.
9. `call`: The function call.
Furthermore, `pagfl` objects can be used in a variety of useful generic methods like `summary()`, `fitted()`, `resid()`, `df.residual`, `formula`, and `coef()`.
```{r}
estim_fit <- fitted(estim)
```
Selecting a $\lambda$ value a priori can be tricky. For instance, it seems like `lambda = 20` is too high since the number of groups $K$ is underestimated. We suggest iterating over a comprehensive range of candidate values to trace out the correct model. To specify a suitable grid, create a logarithmic sequence ranging from 0 to a penalty parameter that induces an entirely homogeneous model (i.e., $\widehat{K} = 1$). The resulting $\lambda$ grid vector can be passed in place of any specific value, and a BIC IC selects the best-fitting parameter.
Furthermore, it is also possible to supply a `data.frame` with named variables and choose a specific formula that selects the variables in that `data.frame`. If the explanatory variables in `X` are named, these names also appear in the output.
```{r}
colnames(data)[-1] <- c("a", "b")
lambda_set <- exp(log(10) * seq(log10(1e-4), log10(10), length.out = 10))
estim_set <- pagfl(y ~ a + b, data = data, n_periods = 150, lambda = lambda_set, verbose = F)
summary(estim_set)
```
When, as above, the specific estimation method is left unspecified, `pagfl` defaults to penalized Least Squares (*PLS*) `method = 'PLS'` (Mehrabani, [2023](https://doi.org/10.1016/j.jeconom.2022.12.002), sec. 2.2). *PLS* is very efficient but requires weakly exogenous regressors. However, even endogenous predictors can be accounted for by employing a penalized Generalized Method of Moments (*PGMM*) routine in combination with exogenous instruments $\boldsymbol{Z}$.
Specify a slightly more elaborate endogenous and dynamic panel data set and apply *PGMM*. When encountering a dynamic panel data set, we recommend using a Jackknife bias correction, as proposed by Dhaene and Jochmans ([2015](https://doi.org/10.1093/restud/rdv007)).
```{r}
# Generate a panel where the predictors X correlate with the cross-sectional innovation,
# but can be instrumented with q = 3 variables in Z. Furthermore, include GARCH(1,1)
# innovations, an AR lag of the dependent variable, and specific group sizes
sim_endo <- sim_DGP(
N = 20, n_periods = 200, p = 2, n_groups = 3, group_proportions = c(0.3, 0.3, 0.4),
error_spec = "GARCH", q = 2, dynamic = FALSE
)
data_endo <- sim_endo$data
Z <- sim_endo$Z
# Note that the method PGMM and the instrument matrix Z need to be passed
estim_endo <- pagfl(y ~ ., data = data_endo, n_periods = 200, lambda = 2, method = "PGMM", Z = Z, bias_correc = TRUE, max_iter = 50e3, verbose = F)
summary(estim_endo)
```
Furthermore, `pagfl` lets you select a minimum group size, adjust the efficiency vs. accuracy trade-off of the iterative estimation algorithm, and modify a list of further settings. Visit the documentation `?pagfl()` for more information.
## The Time-varying PAGFL
The package also includes the functions `sim_tv_DGP()`and `tv_pagfl()`, which generate and estimate grouped panel data models with the time-varying coefficients $\beta_i (t/T)$. Just like in the static case, the functional coefficients admit to a group structure $\beta_{i} (t/T) = \sum_{k = 1}^K \alpha_k (t/T) \boldsymbol{1} \{i \in G_k \}$. Following Su et al. ([2019](https://doi.org/10.1080/07350015.2017.1340299)), the time-varying coefficients are estimated using polynomial B-spline functions employing a penalized sieve estimation (*PSE*).
```{r}
# Simulate a time-varying panel with a trend and a group pattern
N <- 20
n_periods <- 100
tv_sim <- sim_tv_DGP(N = N, n_periods = n_periods, sd_error = 1, intercept = TRUE, p = 1)
tv_data <- tv_sim$data
tv_estim <- tv_pagfl(y ~ 1, data = tv_data, n_periods = n_periods, lambda = 5, verbose = F)
summary(tv_estim)
```
`tv_pagfl()` returns an object of class `tvpagfl`, which contains
1. `model`: A `data.frame` containing the dependent and explanatory variables as well as individual and time indices (if provided).
2. `coefficients`: A list holding (i) a $T \times p^{(1)} \times \hat{K}$ array of the post-Lasso group-specific functional coefficients and (ii) a $K \times p^{(2)}$ matrix of time-constant parameter estimates (when running a mixed time-varying panel data model).
3. `groups`: A `list` containing (i) the total number of groups $\hat{K}$ and (ii) a vector of estimated group memberships $(\hat{g}_1, \dots, \hat{g}_N)$, where $\hat{g}_i = k$ if $i$ is assigned to group $k$.
4. `residuals`: A vector of residuals of the demeaned model.
5. `fitted`: A vector of fitted values of the demeaned model.
6. `args`: A list of additional arguments.
7. `IC`: A `list` containing (i) the value of the IC, (ii) the employed tuning parameter $\lambda$, and (iii) the mean squared error.
8. `convergence`: A `list` containing (i) a logical variable if convergence was achieved and (ii) the number of executed *ADMM* algorithm iterations.
9. `call`: The function call.
Again, `tvpagfl` objects have generic `summary()`, `fitted()`, `resid()`, `df.residual`, `formula`, and `coef()` methods.
In empirical applications, it is commonplace to encounter unbalanced panel data sets. In such instances, time-varying coefficient functions can be estimated nonetheless. The nonparametric spline functions simply interpolate missing periods. However, when using unbalanced datasets it is required to provide explicit indicator variables that declare the cross-sectional individual and time period each observation belongs to.
Lets delete 30% of observations, add indicator variables, and run `tv_pagfl()` again.
```{r}
# Draw some observations to be omitted
delete_index <- as.logical(rbinom(n = N * n_periods, prob = 0.7, size = 1))
# Construct cross-sectional and time indicator variables
tv_data$i_index <- rep(1:N, each = n_periods)
tv_data$t_index <- rep(1:n_periods, N)
# Delete some observations
tv_data <- tv_data[delete_index, ]
# Apply the time-varying PAGFL to an unbalanced panel
tv_estim_unbalanced <- tv_pagfl(y ~ 1, data = tv_data, index = c("i_index", "t_index"), lambda = 5, verbose = F)
summary(tv_estim_unbalanced)
```
Furthermore, `tv_pagfl` lets you specify a lot more optionalities than shown here. For example, it is possible to adjust the polyomial degree and the number of interior knots in the spline basis system, or estimate a panel data model with a mix of time-varying and time-constant coefficients. See `?tv_pagfl()` for details.
## Future Outlook
The package is still under active development. Future versions are planned to include
- Global coefficients
- Un-penalized individual coefficients
- Inference methods
You are not a R-user? Worry not - An equivalent Python library is in the works.
Feel free to reach out if you have any suggestions or questions.
## References
- Dhaene, G., & Jochmans, K. (2015). Split-panel jackknife estimation of fixed-effect models. *The Review of Economic Studies*, 82(3), 991-1030. DOI: [10.1093/restud/rdv007](https://doi.org/10.1093/restud/rdv007)
- Mehrabani, A. (2023). Estimation and identification of latent group structures in panel data. *Journal of Econometrics*, 235(2), 1464-1482. DOI: [10.1016/j.jeconom.2022.12.002](https://doi.org/10.1016/j.jeconom.2022.12.002)
- Schumaker, L. (2007). Spline functions: basic theory. *Cambridge university press*. DOI: [10.1017/CBO9780511618994](https://doi.org/10.1017/CBO9780511618994)
- Su, L., Wang, X., & Jin, S. (2019). Sieve estimation of time-varying panel data models with latent structures. *Journal of Business & Economic Statistics*, 37(2), 334-349. DOI: [10.1080/07350015.2017.1340299](https://doi.org/10.1080/07350015.2017.1340299)