|
Size: 1440
Comment:
|
← Revision 8 as of 2025-10-15 16:34:56 ⇥
Size: 1848
Comment: Rewrite
|
| Deletions are marked like this. | Additions are marked like this. |
| Line 3: | Line 3: |
| The fundamental goal of econometrics as a field is '''causal inference'''. | '''Causal inference''' is an experimental design used to isolate causation and then make use of predictive statistics. |
| Line 11: | Line 11: |
| == Causation == | == Description == |
| Line 13: | Line 13: |
| To study the impact of some '''treatment variable''' on an '''outcome variable''', all other variables must be held constant. If a relationship can be observed ''ceteris paribus'', then causation has been identified. === In Social Science === The unit of measurement in social science is almost always people. It isn't possible to perfectly recreate the conditions which led to a person interacting with an experiment, just to apply a treatment. As a result, true causal experiments are impossible in social sciences. Instead, '''statistical inference''' is applied. In a random sample, observed statistics reflect the actual population parameters. When two random samples are drawn, and an experiment is conducted applying a treatment to one, any statistically significant difference in outcome statistics reflect causation. A further complication: there are non-negligible ethical barriers to conducting many experiments when the unit of measurement is people. This is mitigated only institutionally, i.e. by use of '''Institutional Review Boards''' ('''IRBs'''). |
To study the impact of a '''treatment''' on an '''outcome''', all other variables must be held constant. If a relationship can be observed ''ceteris paribus'', then '''causation''' has been identified. |
| Line 29: | Line 19: |
| == Measurement == | == Natural Experiments == |
| Line 31: | Line 21: |
| Quantitative methods for analysing causation (esp. econometrics) rely on precise measurement of both treatment and outcome variables. | A '''natural experiment''' or '''true causal experiment''' involves random assignment of a treatment such that the treatment and control groups are identical. Methods for causal inference are then applied. The unit of measurement in social science is almost always ''people''. It isn't possible to form groups of ''identical people''. As a result, causal inference is generally inapplicable. Instead, methods for '''statistical inference''' are applied. A tolerance for Type I error ''(false negatives)'' is set as a '''critical value'''. Experimental data is measured and, if the [[Statistics/TestStatistic|test statistic]] is greater than the critical value, the null hypothesis can be rejected. ---- == Quasi-Experiments == A '''quasi-experiment''' is distinguished by the assignment of treatment. When the assignment was not completely random, but differences are expected to be controllable, quasi-experimental methods are applied. ---- == Observational Studies == An '''observational study''' does not feature any real assignment to treatment. Instead, subjects self-assign a 'treatment' through their own decisions and behavior. There is no reason to expect treatment and control groups to be identical. There are non-redundant '''confounders''' which are related to the treatment and/or the outcome that must be controlled for. |
Causal Inference
Causal inference is an experimental design used to isolate causation and then make use of predictive statistics.
Description
To study the impact of a treatment on an outcome, all other variables must be held constant. If a relationship can be observed ceteris paribus, then causation has been identified.
Natural Experiments
A natural experiment or true causal experiment involves random assignment of a treatment such that the treatment and control groups are identical. Methods for causal inference are then applied.
The unit of measurement in social science is almost always people. It isn't possible to form groups of identical people. As a result, causal inference is generally inapplicable. Instead, methods for statistical inference are applied. A tolerance for Type I error (false negatives) is set as a critical value. Experimental data is measured and, if the test statistic is greater than the critical value, the null hypothesis can be rejected.
Quasi-Experiments
A quasi-experiment is distinguished by the assignment of treatment. When the assignment was not completely random, but differences are expected to be controllable, quasi-experimental methods are applied.
Observational Studies
An observational study does not feature any real assignment to treatment. Instead, subjects self-assign a 'treatment' through their own decisions and behavior.
There is no reason to expect treatment and control groups to be identical. There are non-redundant confounders which are related to the treatment and/or the outcome that must be controlled for.