We are constantly inundated by studies that show some medical supplement is good for you and a couple of years later another study showing it is bad for you.

There is a lot of bad methodologies in these studies let alone bias depending on who is funding the study.

The common example of the logical fallacy is a correlation between shark attacks and ice cream consumption.
A better example is a hypothetical study that shows a correlation between ice cream consumption and hay fever. Does ice cream cause hay fever? No ice cream consumption and pollen count are both correlated with warm weather. Warm weather causes higher pollen count which in turn causes hay fever.

These are a special case of Multicollinearity (also collinearity) where two or more predictor variables e.g. obesity and lack of exercise, in a multiple regression model are highly correlated. You can't tell if the dependent variable e.g. heart disease is caused by one or the other.
In the case of the hay fever example the problem was the real causal variable, pollen count, was not even included in the correlation model.

I took an econometrics course from one of the leading econometrics experts. He had a mission to reduce these kind of mistakes in research in a variety of fields, medicine, economics, business, ...

As a graduate student I was a research assistant for several professors who frequently misused statistics, but still got their papers published.

As data becomes more readily available via the Internet it is tempting to run multiple regressions on many data sets and then concoct a theory for two or more variables that are correlated.

Spurious correlation is especially likely to occur with time series data, where two variables trend upward over time because of increases in population, income, prices, or other factors.

There many factors in good experimental design. Books are written on the subject. You will see terms like,
Randomized control trial (RCT) - In a clinical trial, people are randomly assigned to a control group who receives a placebo and the group which receives the treatment under investigation. Double Blind
"How to Lie with Statistics" Darrell Huff, Irving Geis, 1954M
In the 1960s and 1970s, "How to Lie with Statistics" became a standard textbook introduction to the subject of statistics for many college students.

Examples for teaching: Correlation does not mean causation - Cross Validated
Marriage and Longevity
Introduction to Correlation and Regression Analysis
Experimental Design

last updated 20 Mar 2017