My Machine Learning model show in different ways, high correlation between parents' education and student's outcomes. I 'm trying to identify a causal relationship in a separate model since the analysis can be completely confounded. In addition to these variables, there are other important variables available, such as the school infrastructure and student's background. My dataset is from a large-scale educational assessment over time, that can be analyzed at each year or more than one. The dataset is cross-sectional in the student level, but could be a panel at school level (using the average or mode for student transformation measures). Until then, I thought some econometrics approach, but I can't decide and design it. Does anyone have any idea?
Correlation does not imply causation, in general. I recommend:
Article From Association to Causation: Some Remarks on the History o...
andChapter Statistical Models and Causal Inference A Dialogue with the ...
Best of luck, David Booth
For identifiying a causal relationship you should really think in terms of Economics/Econometrics theory.
(1) Can you rule out time-constant heterogeneity? You can address with issue with fixed-effects.
(2) Can you rule out time-varying heterogeneity? If you can, go ahead with fixed-effects. If you cannot, you should look for an instrumental-variable-like strategy.
In principle, with your data, I would start with a FE model at the school level and controlling for reasonably exogenous variables. You should convince the referees/readers that there is no selection bias or two-way causality. This issue may depends even on the institutional framework (freedomo to choose school, etc.).
I hope it helps.
Correlation does not imply causation, in general. I recommend:
Article From Association to Causation: Some Remarks on the History o...
andChapter Statistical Models and Causal Inference A Dialogue with the ...
Best of luck, David Booth
Try showing the causal model (e.g., a directed acyclical graph) to help identify those confounds and if there are ways to account for them and still reach some causal conclusions. As Nancy Cartwright succinctly says, "no causes in; no causes out"
ps. I talk a bit about using graphs to help reach causal inference (or as is typical among stats-folks more time is spent on how to identify where there are problems in the quest of reaching causal inference) in:
www.tandfonline.com/doi/full/10.1080/2330443X.2017.1294037
Hello Rogerio Filho. I'm in the middle of my Sunday chores and have not had time to read your document yet. But the general subject reminded me of this web-page on Karl Wuensch's website--you might find it helpful.
- http://core.ecu.edu/psyc/wuenschk/stathelp/Correlation-Causation.htm
Cheers,
Bruce
I am really glad for these answers.
Causality is an old dilemma in science, and for many, it is impossible to infer them without experiments. However, I have studied causality in observational data for a while, and I going to share a short overview of my thoughts.
I could split causality into observational data in two worlds. Both required strong assumptions, that vary among techniques. However, it's clear that they are even more important in the second world.
The first world is regarding José-Ignacio Antón reply, which turns around econometrics, but also is present in other traditional fields such as psychology and medicine. The majority of these studies are entailed in causality inference (That is, we want to explore the effect of X on Y) under the Potential Outcomes framework. There's a set of canonical techniques used for (Matching, Diff-diff, Instrumental Variables, Discontinuity...). There's a variety of possibilities for longitudinal data, as fixed and random effect as mentioned byJosé-Ignacio Antón. It worth mentioning the significant impact of machine learning algorithms on the growth of methodological literature as well the recent interest of industry on that.
Ex:
https://github.com/microsoft/dowhy
https://github.com/uber/causalml
https://fentechsolutions.github.io/CausalDiscoveryToolbox/html/index.html
The second one is close to Daniel Wright answer, and turns around Computer Science field. When algorithms carried out multiple independence tests throughout data looking for patterns in order to seek better causal graphs (DAG) that fit the data. They are used for either causal inference then causality discovery (that is, discover X's which cause Y, or Y's ). Bayesian Network plays a central role in that and Markov assumption is necessary. The search strategies differ a lot. But also, there's a recent and interesting strategy in this field, that seeks invariability among environments as a signal of causality (I really like it).
Few traditional researchers had considered the potential of DAG, those who did, realize the increase in interpretability, entailing in better strategies to avoid bias. However, they highlight the importance of domain knowledge to build them instead of the x-squared test.
Actually, there is a great scientific disagreement between these two worlds. Traditional researchers claim for better design approaches but have been accepting the enhancement of canonical methods by applying machine learning power. On the other hand, they have been naming bayesian network approaches as simple creative mathematics and little useful to social science problems. But there're controversies.
Look at this really recent discussion:
Econ view: https://arxiv.org/abs/1907.07271v1
The reply from CS: http://causality.cs.ucla.edu/blog/?s=imbens
Thanks again.
Rogério
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining