What-if analysis for business decisions: beyond the spreadsheet scenario
A what-if analysis is only as good as the relationships behind it, and most spreadsheet scenarios assume those relationships instead of modeling them.
5 min read
What a what-if analysis is for
A what-if analysis asks how an outcome changes if you change a decision: what happens to defects if we train more operators, to attrition if advisors call more clients, to cost if we add a documentation step. It is the working form of scenario planning, and it is where most investment decisions are really made.
Why spreadsheet scenarios fall short
In a typical spreadsheet scenario, someone changes an input cell and a formula recalculates the output. The weak point is the formula. It encodes an assumed relationship, often a single multiplier, and nothing in the sheet tells you where that assumption came from or what it ignores.
That is how a scenario can look rigorous while resting on a coefficient someone remembered. It also cannot tell you that the thing you are changing is influenced by something else you are not modeling.
What a causal what-if adds
A causal what-if analysis models the system rather than a single formula, which changes four things.
- The relationships are explicit: a graph shows what influences what.
- Each probability records its source, so you can tell evidence from assumption.
- The tool can report that a question is not answerable from the model, instead of returning a number.
- The result comes with how far it moves if your assumptions are wrong.
Run the analysis in four steps
Frame the decision: name the action, the outcome you care about and the options you are comparing. Build the model of how the relevant variables influence each other and mark which are measured. Ask the question and check it is answerable, and if it is not, fix the model rather than the data. Then read the answer as a range, and write down the assumptions it depends on.
Read the range, not the point
A $40k effect with a range from −$10k to $90k supports “probably helps, but we cannot rule out harm”. That is a different decision from a clean $40k, and a point estimate hides the difference. The range is not a statistical confidence interval, because a model built from elicited knowledge has no sample to draw sampling error from. It is a measure of how far the answer moves as the assumptions move.
Written by the AnantCausal team. More in the guides library.
Run a what-if you can defend.
Bring a scenario you currently model in a spreadsheet and we will rebuild it as a causal model in the walkthrough.