Mediation and Moderation Walk Into a Tea Stall

It’s a Wednesday afternoon. The kettle hisses. The fan turns slowly above the counter. Viji has spent ₹2,400 on four wooden benches, and for 80 days he has been writing down how many of them he puts out each morning.

Some mornings all 4 go out. Some mornings he leaves them stacked against the wall and puts out 1, or none.

The number he cares about is biscuits. He ran the numbers this morning:

Each bench out is worth about 3.1 more packets of biscuits a day.

His cousin, who has been told this twice already, is not satisfied: “Benches do not eat biscuits, brother.”

Viji types back, “The number says 3.1.”

“I believe the number. I am asking what the bench actually does. And I’ll tell you something else — on the wet days it does nothing at all. Check.”

Viji puts the phone down. He says, “A bench is a bench. It either works or it doesn’t.”

The awning rustles.

A man steps in under the awning — neat, deliberate, deep violet shirt, walking like a man who has somewhere to be. In his right hand he carries a smooth wooden relay baton. He does not sit. He stands exactly halfway between Viji’s stack of benches and his biscuit tin, and holds the baton up.

He says, “I am Mediation. I am the answer to how. Your benches do not reach the biscuits on their own. They hand the effect to me, and I carry it the rest of the way.”

A woman walks in behind him — quick, amused, deep teal shirt, carrying a heavy brass dial the size of a plate, the kind bolted to a machine. Its face is marked from 0 at one end to 10 at the other. She turns it a little, back and forth, without looking at it.

She says, “I am Moderation. I am the answer to when. I never carry anything. I do not stand on his road at all. I sit beside it, and I decide how strong the whole thing is today.”

Viji looks at them. “So one of you is in the middle of the arrow, and the other is…”

They reply, together, “Beside it.”

Mediation speaks first

Mediation sets the baton across the counter, pointing from the bench stack towards the biscuit tin.

He says, “You have measured the two ends. Benches, then biscuits. 3.1 packets per bench. That number is the total effect, and it is correct. It is also silent about the middle, which is the part your cousin is asking about.”

He says, “So let me ask about the middle. Do you have any column that is not benches and not biscuits?”

Viji turns the pages. He does. Since the benches arrived he has been noting, roughly, how long an average customer stays — he calls it minutes stayed.

Mediation picks up the pencil.

“Two more regressions. First, benches against minutes stayed. Each bench you put out adds about 4.2 minutes to how long the average customer stays. That is the first leg of the relay, and it is called path a.”

“Second, minutes stayed against biscuits, holding benches constant. Every extra minute a customer stays is worth about 0.62 of a packet. That is the second leg, path b.”

He multiplies. “4.2 × 0.62 ≈ 2.60. That is the indirect effect — the part of your 3.1 that travels through me.”

Viji says, “And the rest?”

“3.10 − 2.60 = 0.50. That is the direct effect, written c′ — whatever benches do to biscuits that has nothing to do with people staying longer. It is small. Most of your bench effect is not a bench effect at all. It is a sitting-down effect.”

He picks the baton back up.

“And here is what matters for your money. You did not really buy benches. You bought minutes. If somebody offered you a cheaper way to make people stay 17 minutes longer — a shade cloth, a radio, a place to put a cup down — it would sell the same biscuits. Benches are one delivery van. The cargo is the minutes.”

He adds, “One honest warning about myself. I only exist if you measured the middle. If you had never written down minutes stayed, I could not have been found, and no amount of staring at benches and biscuits would have revealed me.”

Moderation speaks

Moderation sets her dial on the counter with a thump.

She says, “Now the other question, which is the one your cousin actually tested. Is 3.1 the same number every day?”

Viji says, “It’s the average.”

“It is the average, which means it is nobody’s Wednesday. Split your 80 days by whether it rained.”

He does. 31 wet days, 49 dry ones. He runs the bench effect inside each pile.

Dry days: +4.4 packets per bench. Wet days: +0.6 packets per bench.

Viji stares. “On wet days the benches do almost nothing.”

“Of course they do nothing. Nobody sits on a wet bench.” She turns her dial from 10 down towards 0. “That is all I am. I am not on the road between benches and biscuits — rain is not a step in the journey. I sit beside the road with my hand on the strength of it. On a dry day I turn it up. On a wet day I turn it down to almost nothing.”

She says, “In a model I show up as a multiplication, not an addition. You do not write benches plus rain. You write benches × rain, an interaction term, and the size of that term is me.”

Mediation says, “And notice her 4.4 and 0.6 do not contradict my 3.1. His average is her two numbers blended in the proportion the weather happened to supply. Change the weather and the average moves, while neither of her two numbers has moved at all.”

Viji says, slowly, “So my 3.1 is not a property of benches. It is a property of benches in this city, in this season.”

Moderation says, “Now you have it.”

What just happened

Viji pours himself a glass of tea, sits down on his own counter, and writes the two rows out.

He looks at the middle row for a while. “Minutes are a step in the journey. Rain is the weather the journey happens in.”

The same chat, in a chart

That picture is the same conversation, drawn. The first panel is the relay: the effect leaves the benches, travels through the minutes, and arrives at the biscuits, with only 0.50 of 3.10 taking the short road. The second panel is the dial: the same benches on the same x-axis, one line steep and one line almost flat, and nothing between them except whether it rained.

One last warning before they leave

Mediation tucks the baton under his arm.

He says, “Three traps. The first two are mine, and journals reject papers over them every week.”

“One. A mediator is a causal claim wearing a correlation’s clothes. All I actually computed was two regressions. For me to mean what you want me to mean, the benches have to come before the minutes, and the minutes before the biscuits — and nothing unmeasured can be quietly causing both the minutes and the biscuits. If you collect all three columns in one questionnaire on one afternoon, you have measured three things at the same moment and drawn arrows between them out of hope.”

“Two. Do not test me the way the textbooks from 1986 say. The old Baron and Kenny procedure asks for a significant total effect first, then calls the result full or partial mediation. Modern practice tests the indirect effect ab directly, with a bootstrap confidence interval, because ab is usually not normally distributed and the Sobel test assumes it is. An indirect effect can be real even when the total effect is not significant. And full versus partial is language worth dropping — it mostly reports your sample size.”

Moderation says, “Three, and this one is the bridge back to the last time you had visitors. Never control for a mediator when the total effect is what you want. Put minutes stayed into the model as a control and the bench effect collapses from 3.1 to 0.5 — and a careless reader will announce that benches do not work. Benches work. You just put the answer in the model as a covariate and then asked whether it was still there.”

She picks up the dial. “And once you are comfortable with both of us, you will meet us together. If the strength of his relay depends on my dial — if benches buy more minutes on dry days than wet ones — that is moderated mediation, and it has its own index and its own bootstrap interval.”

Viji writes it down. Mediators sit on the path. Moderators sit beside it. Never control for the first one by accident.

The bill

They left the way people leave a tea stall on a working afternoon. Mediation went first, baton in hand, walking a straight line out past the awning. Moderation followed, dial under one arm, and paused at the edge of the shade to look at the sky before deciding which way to go.

Viji did two things that week. He stopped describing the benches as a biscuit machine and started describing them as a minutes machine — and once he thought of it that way, he spent ₹300 on a length of shade cloth, because shade also buys minutes, and it buys them on the days when benches cannot.

Then he wrote a small note at the back of the notebook for wet days: benches stacked, no loss. On those days the benches were never doing anything, and now he knew it in advance instead of discovering it at closing time.

His cousin, predictably: “So I was right about the rain.”

Viji typed back: “You were right about the rain. You were wrong about why the benches work at all.”

He turned to a fresh page and wrote the sentence he wanted to keep:

Mediation tells you what the cause actually does. Moderation tells you when it bothers to do it.

For the math-curious

The three equations. With X the treatment, M the mediator and Y the outcome: $$ M = i_1 + aX + e_1 $$ $$ Y = i_2 + c’X + bM + e_2 $$ $$ Y = i_3 + cX + e_3 $$ The indirect effect is ab, the direct effect is c′, and for continuous outcomes with no interaction they decompose exactly: c = c′ + ab. Viji’s numbers: a = 4.2, b = 0.62, ab = 2.60, c′ = 0.50, c = 3.10.

Testing the indirect effect. Use a percentile bootstrap confidence interval on ab — 5,000 or 10,000 resamples — and report the interval, not a p-value from the Sobel test, whose normality assumption ab rarely satisfies. This is PROCESS Model 4, or lavaan with := to define the product and se = "bootstrap". The old requirement that c be significant before testing ab is obsolete; inconsistent mediation, where ab and c′ have opposite signs, can leave c near zero with a real mechanism underneath.

Moderation. With W the moderator: $$ Y = i + b_1X + b_2W + b_3(X \cdot W) + e $$ b₃ is the moderation. Mean-centre X and W before forming the product, so that b₁ reads as the effect of X at the average W rather than at W = 0, which is often meaningless. Probe a significant b₃ with simple slopes at ±1 SD, or with the Johnson–Neyman technique, which reports the range of W where the effect of X is significant rather than testing arbitrary points. This is PROCESS Model 1. A categorical moderator such as rain is the same equation with a dummy, and equals the two subgroup slopes Viji computed.

Both at once. When W changes the size of a path in the mediation chain, you have moderated mediation (Model 7 if W moderates path a, Model 14 if it moderates path b). Report the index of moderated mediation with a bootstrap interval rather than comparing two separately significant indirect effects — a difference between a significant and a non-significant result is not itself a significant difference.

Assumptions worth stating out loud. Correct temporal order, no unmeasured confounding of X–M, M–Y and X–Y, no measurement error in M (error there biases b downward and c′ upward), and correct functional form. The causal-inference literature restates mediation in counterfactual terms, splitting the effect into natural direct and natural indirect components, which makes those assumptions explicit instead of implied.

One of them is a step on the road. The other is the weather the road is in.