Causal Impact and Statistics I

How can we use statistics to get at causal impact?

Of course, correlation does not imply causation. Remember my lesson from a Dilbert's comic last week. But still, there are inferences to be drawn somehow.

It may be a fact that MacDonald's just spent a lot of money on an ad campaign involving firecrackers and tapdancing penguins. It may be, also, that they sold a million hamburgers worldwide yesterday. It does not follow that they sold the million hamburgers because of that ad campaign. What MacDonald's clearly wants is some idea of how many hamburgers they would have sold if they had had some other ad campaign in place, possibly some less expensive campaign. Something involving their familiar clown and his sidekicks.

Here's a social sciency thing about "Bayesian structural time-series models" that aims at getting at the problem. It's actually based on the experiences of Google, paid clicks versus free clicks, etc. -- I just prefer the MacDonald's example, maybe because I'm hungry.

That phrase has the word "model" in it, so I'm putting a photo of a "model" above. Yuck yuck yuck.

It also has the word "Bayesian" in it, which I think I kinda understand.

One of the questions about social science cause-effect that has given me much pause in recent years is this: what, if anything, is the connection between a firm's labor policy and its indebtedness? There is an intuitive case to be made that indebted firms would be less desirable employers, other things being equal, than equity financed firms. After all, firms that have financed themselves through debt have to pay the interest as it comes due. Firms with a lot of equity on the balance sheet don't really have to pay dividends in a given quarter, they have discretion as to pay-out. So paying back debt investors is a higher priority than, say, keeping the workplace safe, or paying cost-of-living allowances in inflationary periods. But paying equity investors ... not necessarily. That's one way of looking at it. There are other ways.

But how do we determine this as a matter of empirical science? What statistics could persuade is that there is (or isn't) a connection between a firm's labor policy and its indebtedness? Can Bayesian structural time-series models help here?

More next week.