Dependence Analysis
Dependence analysis examines relationships among variables. Banasiewicz organizes causal assessment around sequence, association, possible confounding, and a plausible explanation. Statistical dependence can support an investigation, but it does not by itself show that one factor causes another.
THE IDEA
Dependence Analysis
If two variables move together, does one cause the other? Dependence analysis can help investigate the relationship, but answering the causal question also requires attention to how the data were produced and what else could explain the pattern.
Banasiewicz organizes that inquiry around four considerations:
- Temporal sequentiality: the proposed cause must occur before the effect it is supposed to produce.
- Associative variation: changes in the proposed cause should be related to changes in the outcome under the conditions being studied.
- Non-spuriousness: a third factor should not account for the apparent relationship. Such a factor is a confounder; a moderator, by contrast, changes the size or direction of a relationship.
- Theoretical support: a plausible account should explain how the proposed cause could produce the effect. A plausible story helps frame a test but is not itself proof.
Consider a discount coupon. A customer receives it before redeeming it, so the order is clear. Repeated redemptions show that customers use coupons, but they do not show how many additional purchases the coupons caused. Some customers might have bought the product at full price. Others might be especially likely both to seek coupons and to buy. An experiment comparing otherwise comparable customers offered and not offered the incentive can help estimate its effect. Other research designs can also provide causal evidence when their assumptions are justified.
Within his 3E framework, Banasiewicz calls temporal sequence and association ‘hard’ substantiations because they are relatively tangible, and non-spuriousness and theoretical support ‘soft’ substantiations because they are harder to establish. He treats the former pair as sufficient evidence of causality. That conclusion does not follow from sequence and association alone: confounding and selection can produce both. The labels should not make those alternative explanations optional to examine.
Regression, discriminant analysis, structural equation modeling, and analysis of variance offer ways to model relationships. Calling a variable independent or dependent identifies its role in a model; it does not establish which way causation runs. A causal interpretation needs a research design and assumptions that support it.
Even within a family such as regression, methods serve different purposes. Linear, logistic, polynomial, quantile, ridge, lasso, and ordinal regression differ in their outcomes, assumptions, or fitting procedures. Choosing a more elaborate method does not repair a poor causal comparison. Analysts need to check the data, explain their assumptions, assess uncertainty, and distinguish a useful prediction from an estimate of what an intervention would change.
When to use it
Evaluating whether a new training program actually improves employee performance (causal assessment) using statistical dependence analysis.
Determining if a price reduction leads to higher sales volume while accounting for confounding factors like seasonality.
When assessing a treatment’s effect, examine the study design, comparison group, and potential confounding; sequence and association are starting points for the causal question.
Analyzing customer loyalty programs to distinguish between true causal effects and mere coincidence or self-selection bias.
FROM THEORY TO EVERYDAY LIFE
In practice
Purchase Incentive and Coupon Redemption
A customer must receive a coupon before redeeming it, but repeated redemptions do not establish how many extra purchases the coupon caused. Compare comparable customers who were and were not offered the incentive, and check whether selection or other differences could explain the result.
A MOMENT TO REFLECT
Put it to work.
Think of a real-world situation where you suspect a cause-effect relationship. Identify which of the four conditions (temporal sequentiality, associative variation, non-spuriousness, theoretical support) are easiest and hardest to verify in that case.
How would you design a simple experiment to demonstrate non-spuriousness for a claimed relationship between a marketing campaign and increased sales?
Why might a behavioral economist and a traditional economist offer different theoretical support for the same observed association between a discount and a purchase?
Often confused with
CONTINUE THE CONVERSATION
Source texts
- Evidence-Based Decision-Making How to Leverage Available Data and Avoid Cognitive Biases (Andrew D. Banasiewicz)