Covariate-Adjusted Deep Causal Learning for Heterogeneous Panel Data Models
Organizations: University of Notre Dame
Abstract
This paper studies the task of estimating heterogeneous treatment effects in causal panel data models with covariate effects. We propose a novel Covariate-Adjusted Deep Causal Learning (CoDEAL) framework that cohesively deals with the underlying heterogeneity and nonlinearity of both panel units and covariate effects. CoDEAL integrates nonlinear covariate effect components (parameterized by a feed-forward neural network) with nonlinear factor structures (modeled by a multi-output autoencoder) to form a heterogeneous causal panel model. The nonlinear covariate component flexibly captures complex covariate influences on outcomes, and the nonlinear factor decomposition enables CoDEAL to effectively capture both cross-sectional and temporal dependencies inherent in the data panel. This latent structural information is subsequently integrated into a customized matrix completion algorithm, thereby facilitating more accurate counterfactual imputation. Moreover, the use of a multi-output autoencoder explicitly accounts for heterogeneity across units and enhances the model interpretability of the latent factors. We establish theoretical guarantees on the convergence of the estimated counterfactuals and demonstrate the compelling performance of the proposed method using extensive simulation studies and real-world data applications.
Figures & tables
| ? | |||||
| Factor | Method | Config. 1 | Config. 2 | Config. 3 | Config. 4 | ||||
| Linear | Nonlinear | Linear | Nonlinear | ||||||
| Linear | CoDEAL | 0.646(0.009) | 0.701(0.008) | 0.609(0.004) | 0.644(0.005) | 0.614(0.007) | 0.625(0.005) | 0.532(0.003) | 0.560(0.004) |
| Single AE | 0.821(0.010) | 0.851(0.008) | 0.777(0.007) | 0.795(0.007) | 0.789(0.011) | 0.795(0.008) | 0.693(0.005) | 0.714(0.006) | |
| AEMC-NE | 0.838(0.009) | 0.862(0.009) | 0.759(0.007) | 0.771(0.007) | 0.731(0.010) | 0.730(0.007) | 0.647(0.005) | 0.660(0.005) | |
| DiD | 0.892(0.010) | 0.916(0.009) | 0.810(0.007) | 0.828(0.008) | 0.801(0.011) | 0.802(0.008) | 0.711(0.005) | 0.726(0.006) | |
| Vert-Reg | 0.757(0.006) | 0.798(0.007) | 0.761(0.004) | 0.770(0.005) | 0.940(0.005) | 0.865(0.023) | 0.691(0.003) | 0.750(0.016) | |
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
| SEQ | JOINT | |||||
| Mean | 1.9010 | 0.0409 | 1.9417 | 0.2814 | 1.6465 | 1.9279 |
| SD | 1.0508 | 0.1043 | 1.0934 | 0.1433 | 1.0283 | 1.1063 |
| MSE | 0.9832 | 0.9573 | 0.0033 | 2.3166 | 2.2593 | 0.0069 |
| Sim | True value | SEQ | JOINT | ||||||
| 1 | 1.6327 | -1.1332 | 0.4995 | 0.4476 | -0.0081 | 0.4395 | 0.0671 | 0.2076 | 0.2748 |
| 2 | 1.8617 | 1.2469 | 3.1086 | 2.8624 | 0.2424 | 3.1048 | 0.1713 | 2.9729 | 3.1442 |
| 3 | 1.8485 | -0.5985 | 1.2500 | 1.3981 | -0.1415 | 1.2567 | 0.1582 | 1.2026 | 1.3609 |
| 4 | 1.8246 | 0.0237 | 1.8483 | 1.8629 | 0.0875 | 1.9504 | 0.3391 | 1.5944 | 1.9335 |
| 5 | 1.8295 | 0.0629 | 1.8924 | 1.8260 | -0.0135 | 1.8125 | 0.4222 | 1.4780 | 1.9002 |
| Factor Effect Generation ( ) | Covariate Effect Generation ( ) | ||
| Linear | , | Matrix Linear | , |
| Sine | Vector Linear | , | |
| Polynomial | , | tanh | , |
| ReLU MLP | poly | ||
| , | log | , | |
| , | ReLU | ||
| Covariate | Method | LR , Linear Factor | NN , Linear Factor | LR , Sine Factor | NN , Sine Factor | ||||
| MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | ||
| No covariate | CoDEAL | 0.560(0.009) | 0.578(0.028) | 0.560(0.009) | 0.578(0.028) | 0.933(0.015) | 1.722(0.066) | 0.933(0.015) | 1.722(0.066) |
| Single AE | 0.724(0.009) | 1.081(0.034) | 0.724(0.009) | 1.081(0.034) | 1.071(0.013) | 2.131(0.056) | 1.071(0.013) | 2.131(0.056) | |
| AEMC-NE | 0.734(0.009) | 1.142(0.039) | 0.734(0.009) | 1.142(0.039) | 1.090(0.013) | 2.172(0.056) | 1.090(0.013) | 2.172(0.056) | |
| DiD | 0.804(0.010) | 1.364(0.043) | 0.804(0.010) | 1.364(0.043) | 1.141(0.013) | 2.343(0.059) | 1.141(0.013) | 2.343(0.059) | |
| Vert-Reg | 0.601(0.002) | 0.574(0.004) | 0.601(0.002) | 0.574(0.004) | 1.024(0.012) | 2.228(0.069) | 1.024(0.012) | 2.228(0.069) | |
| Covariate | Method | LR , Linear Factor | NN , Linear Factor | LR , Sine Factor | NN , Sine Factor | ||||
| MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | ||
| No covariate | CoDEAL | 0.507(0.004) | 0.425(0.010) | 0.507(0.004) | 0.425(0.010) | 0.779(0.007) | 1.173(0.027) | 0.779(0.007) | 1.173(0.027) |
| Single AE | 0.622(0.006) | 0.761(0.021) | 0.622(0.006) | 0.761(0.021) | 0.963(0.010) | 1.772(0.038) | 0.963(0.010) | 1.772(0.038) | |
| AEMC-NE | 0.647(0.006) | 0.850(0.023) | 0.647(0.006) | 0.850(0.023) | 1.004(0.009) | 1.884(0.036) | 1.004(0.009) | 1.884(0.036) | |
| DiD | 0.713(0.007) | 1.043(0.028) | 0.713(0.007) | 1.043(0.028) | 1.085(0.010) | 2.163(0.041) | 1.085(0.010) | 2.163(0.041) | |
| Vert-Reg | 0.651(0.002) | 0.671(0.004) | 0.651(0.002) | 0.671(0.004) | 0.865(0.007) | 1.422(0.034) | 0.865(0.007) | 1.422(0.034) | |
| Factor | Method | Config. 1, linear covariate | Config. 1, nonlinear covariate | Config. 2, linear covariate | Config. 2, nonlinear covariate | ||||
| MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | ||
| Linear | CoDEAL | 0.646(0.009) | 0.722(0.023) | 0.701(0.008) | 0.871(0.031) | 0.609(0.004) | 0.622(0.010) | 0.644(0.005) | 0.704(0.015) |
| Single AE | 0.821(0.010) | 1.320(0.039) | 0.851(0.008) | 1.383(0.033) | 0.777(0.007) | 1.147(0.024) | 0.795(0.007) | 1.190(0.027) | |
| AEMC-NE | 0.838(0.009) | 1.394(0.041) | 0.862(0.009) | 1.445(0.036) | 0.759(0.007) | 1.093(0.025) | 0.771(0.007) | 1.120(0.026) | |
| DiD | 0.892(0.010) | 1.587(0.045) | 0.916(0.009) | 1.642(0.037) | 0.810(0.007) | 1.262(0.029) | 0.828(0.008) | 1.309(0.030) | |
| Vert-Reg | 0.757(0.006) | 0.957(0.019) | 0.798(0.007) | 1.075(0.024) | 0.761(0.004) | 0.936(0.011) | 0.770(0.005) | 0.968(0.015) | |
| Factor | Method | Config. 1, linear covariate | Config. 1, tanh covariate | Config. 1, poly covariate | |||
| MAE | MSE | MAE | MSE | MAE | MSE | ||
| Linear | CoDEAL | 0.709(0.007) | 0.914(0.022) | 0.746(0.007) | 1.009(0.025) | 0.751(0.007) | 1.018(0.024) |
| SCM | 0.804(0.011) | 1.306(0.045) | 0.909(0.010) | 1.240(0.027) | 1.126(0.012) | 1.785(0.034) | |
| GSC | 0.484(0.006) | 0.370(0.009) | 0.847(0.011) | 0.997(0.022) | 1.088(0.013) | 1.590(0.032) | |
| SDID | 1.584(0.013) | 4.422(0.075) | 1.089(0.010) | 1.995(0.038) | 1.282(0.011) | 2.559(0.042) | |
| Sine | CoDEAL | 1.116(0.011) | 2.232(0.050) | 1.156(0.011) | 2.362(0.052) | 1.155(0.011) | 2.356(0.048) |
| CoDEAL | Single AE | CoDEAL | Single AE | CoDEAL | Single AE | |
| 0.0 | 1.476 | 2.287 | 1.476 | 2.303 | 1.509 | 2.379 |
| 0.2 | 1.488 | 2.262 | 1.473 | 2.271 | 1.504 | 2.304 |
| 0.4 | 1.452 | 2.261 | 1.496 | 2.287 | 1.501 | 2.325 |
| 0.6 | 1.446 | 2.246 | 1.451 | 2.250 | 1.475 | 2.312 |
| 0.8 | 1.418 | 2.181 | 1.417 | 2.189 | 1.443 | 2.273 |
| Setting | Metric | Method | Linear | Sine | Polynomial | ReLU-MLP |
| Config 1, no covariates | MSE | Single AE | 0.263 | 0.262 | 0.314 | 0.857 |
| CoDEAL | 0.285 | 0.277 | 0.365 | 0.451 | ||
| MAE | Single AE | 0.409 | 0.409 | 0.442 | 0.702 | |
| CoDEAL | 0.424 | 0.428 | 0.464 | 0.525 | ||
| Config 1, with covariates | MSE | Single AE | 0.656 | 0.796 | 0.927 | 1.299 |
| CoDEAL | 0.623 | 0.709 | 0.973 | 0.963 |
| # Never-treated units | MAE | MSE |
| 40 | ||
| 60 | ||
| 80 | ||
| 100 | ||
| 120 | ||
| 140 |
| Treatment start time | MAE | MSE |
| 15 | ||
| 20 | ||
| 25 | ||
| 30 | ||
| 35 |
| Learning rate | 0.0005 | 0.001 | 0.002 | 0.003 | 0.004 | 0.005 |
| MSE | 1.624 | 1.284 | 1.173 | 1.078 | 1.098 | 1.122 |
| Hidden layers | 1 | 2 | 3 | 4 | 5 | |
| MSE | 1.268 | 1.189 | 1.173 | 1.269 | 1.413 | |
| Bottleneck | 2 | 3 | 4 | 5 | 6 | |
| MSE | 1.464 | 1.173 | 1.108 | 1.063 | 1.023 |
| Config. 1 | Config. 2 | |||
| Linear | Sine | Linear | Sine | |
| CoDEAL | 1.295 | 1.286 | 2.429 | 2.495 |
| Single AE | 0.025 | 0.026 | 0.025 | 0.027 |
| AEMC-NE | 0.115 | 0.118 | 0.125 | 0.129 |
| DiD | 0.000 | 0.001 | 0.002 | 0.001 |
| Vert-Reg | 0.034 | 0.037 | 0.083 | 0.083 |
| Method | Confirmed Cases | Deaths |
| Ground Truth | 26.6 | 0.39 |
| CoDEAL | 29.8 | 0.52 |
| Single AE | 34.1 | 0.61 |
| AEMC-NE | 25.3 | 0.35 |
| DiD | 31.5 | 0.45 |
| Vert-Reg | 28.3 | – |
Causal Effect Estimation
Causal effect estimation measures how outcomes would change under interventions, using experiments or observational data and explicit assumptions.
See all 287 papers