Dependent and Independent Variables: What Is Explained, and What Does the Explaining
The dependent variable is the column being explained. The independent variable is the one doing the explaining. Swapping them does not give the same line read backwards. On ten paired months of the Nakshatra unit and the Vasant unit, both invented, the forward slope is 1.5000 and the reverse slope is 0.5039, not the 0.6667 that one over the forward slope would give.
Three things have to be in place before any of that means anything, and all three are covered separately. A fitted lineA straight line picked by arithmetic so that it sits as close as it can to a set of paired readings. How the picking is done was settled where regression is covered. and the arithmetic that chooses one. A correlation, and R squaredHow much of the wobble in the column being explained a line manages to account for, marked on a scale from nothing at all up to one. Settled where regression is covered., both already computed on the very ten months used below. And the plain observation that a line is chosen to make its misses small in one direction and pays no attention whatever to the other. Whether one column causes the other is a different argument, covered separately. Narrower and more practical questions come first. Which column does a line get fitted against, who decides, and what does the decision cost if it is taken carelessly?
What are the two roles, and what does each one mean?
Naming both at once: the dependent variable and the independent variable
The two names only exist relative to each other, so defining them one at a time is what makes them slippery. Take a rule, any rule, that reads something and returns a statement about something else. The column the rule reads is the independent variable. The column the rule returns a statement about is the dependent variable. The pair of definitions is the whole of it. One column is handed in, the other comes out.
Both names describe a job inside a rule, not a property of the numbers in a column. Missing that distinction is what makes everything later on look arbitrary. Nothing about a column marks it as the explaining sort or the explained sort. The answer does not live in the readings at all. No test run on a list of readings would say which slot a column belongs in. The answer lives in the question somebody asked.
Picture a small stationery shop at the end of a lane. The owner keeps two counts every day: how many people walked in, and how many rupees went through the till. Now ask two questions. First: given the number of people who walked in today, roughly what will the till show? Second: given what the till showed today, roughly how many people must have walked in? The two questions are genuinely different and have genuinely different answers, and they use exactly the same two columns. The columns did not change. The job each one is doing changed.
Which of the two roles belongs to the column being explained?
What actually decides which column takes which role?
Not the arithmetic. The arithmetic is perfectly happy either way round, and this is worth sitting with for a moment because it is the source of most of the confusion. Hand a fitting routine two columns in one order and it returns a line. Hand it the same two columns in the other order and it returns a different line, with no complaint, no warning and no sign anywhere that a decision was made. The routine does not know which question was meant. The routine cannot know. So it fits whatever it is given and hands the result back with the same confidence either way.
The roles are decided by which column is actually in hand first at the moment the rule gets used. Not which one is more interesting, not which one came first in the file, and not which one has the tidier numbers. If the reading in hand is the Nakshatra unit’s monthly change and the thing wanted is a statement about the Vasant unit, then the Nakshatra unit is the independent variable and the Vasant unit is the dependent one. Fit it the other way round and the result is unusable. Using it would require already knowing the answer being sought.
The stationery shop makes this concrete too. If the owner can see the footfall by lunchtime and wants to guess the evening’s till, footfall does the explaining. If instead the till is the only thing recorded and somebody is trying to work backwards to how busy the lane was, the till does the explaining. Same shop, same two counts, opposite assignments, and only the second one is usable by somebody holding a till slip and nothing else.
The Nakshatra unit’s monthly change is known before the Vasant unit’s. Which column is the dependent one?
What happens to the fitted line if the two columns are swapped?
The ten paired months below carry every figure that follows. The Nakshatra unit is a monthly change in per cent and so is the Vasant unit, and the time order matters and is kept exactly as it stands.
| Month | The Nakshatra unit, per cent | The Vasant unit, per cent |
|---|---|---|
| 1 | 1.00 | 3.00 |
| 2 | 6.00 | 18.50 |
| 3 | minus 4.00 | 2.50 |
| 4 | 11.00 | 17.00 |
| 5 | 1.00 | minus 3.00 |
| 6 | minus 9.00 | minus 13.00 |
| 7 | 6.00 | 6.50 |
| 8 | 1.00 | minus 1.00 |
| 9 | minus 4.00 | minus 7.50 |
| 10 | 1.00 | minus 3.00 |
| Average | 1.00 | 2.00 |
The forward slope is 1.5000. What does the reverse slope turn out to be?
Fit the Vasant unit against the Nakshatra unit and the line comes out at a slope of 1.5000 with an interceptThe height at which a line sits when the column doing the explaining reads zero. An intercept slides a line up or down without tilting it. of 0.5000. Now turn the question round. Fit the Nakshatra unit against the Vasant unit, on the same ten months, changing nothing else at all, and the line comes out at a slope of 0.5039 with an intercept of minus 0.0078.
| The question being asked | Slope | Intercept |
|---|---|---|
| Forward. The Vasant unit explained by the Nakshatra unit | 1.5000 | 0.5000 |
| Reverse. The Nakshatra unit explained by the Vasant unit | 0.5039 | minus 0.0078 |
The reverse fit is a different line, not the same line drawn on rotated axes. If it were the same line seen from the side, the two slopes would have to be reciprocals and they are not, as the identity below settles exactly. Both lines are correct. Neither is a mistake. The two lines are correct answers to two different questions, and the only way to tell which one is wanted is to know which question is being asked.
Is the reverse slope one over the forward slope?
No, and this is the trap. Inverting is such a natural thing to do. On a drawn line, inverting the slope is a completely legitimate move: a line that goes up 1.5 for every 1 across does go across 0.6667 for every 1 up, and nobody is doing arithmetic wrongly when they say so. One over 1.5000 is 0.6667, and that reciprocalOne divided by a number. The reciprocal of 4 is 0.25, and any number multiplied by its reciprocal comes to exactly one. is a real fact about the drawn line.
The reciprocal is simply not the answer to the reversed question. The reversed question asks for a new fitted line, chosen fresh against the other column, and that line has a slope of 0.5039. The difference between 0.6667 and 0.5039 is not rounding and it is not noise. The gap is structural. The gap will be there on any record where the fit is less than perfect, and it gets wider the worse the fit is.
The danger in that substitution is how quiet it is. Nothing errors. No routine refuses to run. The number 0.6667 is a perfectly ordinary size, has the right sign, and sits in the right neighbourhood. The figure looks like an answer. Somebody who inverts a fitted slope and reports the result has produced a figure that is not the answer to any question that was asked, and there is nothing in the output to tell them so.
Why do the two slopes multiply to the leftover share?
Multiply the two slopes together. 1.5000 times 0.5039 comes to 0.7559. The product is R squared on these ten months, exactly, and not a coincidence that happens to hold on this particular record. The forward slope times the reverse slope is R squared, always, on any paired record. It falls straight out of how each slope is built: one divides the shared movement by the spread of one column, the other divides the same shared movement by the spread of the other, and multiplying them puts the shared movement on top twice and both spreads underneath.
Read as an identity, it says something useful in plain words. Two numbers are reciprocals exactly when they multiply to one. So the forward and reverse slopes are reciprocals exactly when R squared is one. R squared reaches one only when the line passes through every single point and there is nothing left over. A fit that perfect never happens on a real record. And the further a fit sits from perfect, the smaller the product, and the further apart the two directions get.
Here the product is 0.7559, so inverting the forward slope overstates the reverse slope by a comfortable margin. On a record with a much weaker fit the two directions would be further apart still, and on one with an almost perfect fit they would be near enough that nobody would notice the error. The strength of the fit and the size of this trap are the same quantity.
The two slopes multiply to exactly 0.7559. When would the two of them be reciprocals of each other?
Why is correlation the same both ways when the line is not?
The correlation between the Nakshatra unit and the Vasant unit is 0.8694. Computed the other way round it is 0.8694. The correlation does not move. Nothing inside a correlation can move. A correlation has no column being explained at all, so there is nothing inside it for the question to turn round. It is symmetricA relationship that reads the same from either end. Sitting in the same room as somebody is symmetric. Being taller than them is not. by construction: it asks how tightly two columns move together, which is a question about the pair rather than about either one of them.
The symmetry is a strength and a limitation in the same breath, and the cleanest demonstration available that a slope and a correlation are genuinely different quantities rather than two dialects of the same idea. Because the correlation refuses to name a direction, it also cannot answer a question that has a direction in it. Asked how far the Vasant unit moves when the Nakshatra unit moves by one point, 0.8694 has no reply. A correlation is not the right shape of answer. Only a fitted line gives a figure in the units of the column being explained, and getting that figure is exactly what forces the analyst to nominate which column is being explained.
Think of it the household way. Two people always leave the house at the same time; that is a symmetric observation and either of them can state it. But asking how much later dinner runs for every ten minutes the first one is delayed has a direction built into it, and answering it requires somebody to decide which delay is being read and which is being predicted.
Correlation reads 0.8694 whichever way the question is asked, and the slope does not. What does that say about a correlation?
Why is the fit not symmetric, given where the misses are measured?
The mechanism is simpler than the arithmetic makes it look. A fitted line is chosen to make its misses small, and a miss is measured in the direction of the column being explained. Fit the Vasant unit and each miss is a verticalStraight up and down on the drawing. Its partner is horizontal, straight across, and between the same two points those are two different distances. distance: how far above or below the line the actual month sat. Fit the Nakshatra unit instead and each miss becomes a horizontal distance: how far to the left or right of the line the month sat.
The vertical misses and the horizontal misses are two different sets of numbers over the same ten points. Making the first set as small as it will go says nothing whatever about the second set, so the line that wins one competition is not the line that wins the other. Two different measurements of closeness produce two different winners, and that is the whole of why swapping the roles is not symmetric.
The two lines do agree about one thing. Both fitting procedures are built to put the line through the middle of the readings. Both lines therefore pass through the point where the two averages meet, at 1.00 per cent across and 2.00 per cent up. So the picture is not two lines flying off in unrelated directions. The picture is two lines pinned together at one point and separating steadily on either side of it.
How fast they separate is what R squared is measuring. The closer the fit, the more nearly the two lines lie on top of each other, until at a perfect fit they coincide exactly and the whole question stops mattering. At 0.7559 they are noticeably apart, and the further from the middle the reading sits, the more the gap between them costs.
Why does swapping the two roles produce a different line rather than the same line seen from the other side?
What does the asymmetry cost, on one stated reading?
A number can be put on it. Somebody supplies a Vasant reading of 9.50 per cent and asks what the Nakshatra unit was that month. There are two things that might be done, and they do not agree.
The quick move is to take the forward line already in hand and run it backwards. Subtract the intercept of 0.5000 from 9.50 and divide by the slope of 1.5000. The answer is 6.00 per cent. The honest move is to fit the reverse line and use it. Minus 0.0078 plus 0.5039 times 9.50 gives 4.78 per cent.
A gap of 1.22 per cent came out of nothing but a decision about which column went on which side. No reading changed. No month was added or dropped. Nobody made an arithmetic slip. The entire difference is the cost of taking a shortcut on a question of roles, and on a reading this far from the middle it is a wide gap by any standard.
Anybody who reads relationships out of paired records for a living settles the two roles before the arithmetic starts, and writes the decision down somewhere it can be checked. The reason is not tidiness. The reason is that a fitted line is built for one direction of question, and the moment the question turns round the line stops being the right tool, however good it was a minute ago.
The working habit is one sentence long: when the question turns round, refit. No inverting, no rearranging, no taking the line already in hand and reading it the other way. The refit is done from the readings. A refit takes a minute and removes the whole problem.
The reason the wrong thing gets done so often is worth naming plainly. Inverting is free and instant and needs nobody. Refitting needs the readings, a moment, and a small act of discipline. Whenever a wrong method is cheaper than the right one and nothing flags the difference, the cheap one wins by default. The defence has to be a habit rather than a judgement call made under time pressure.
Slide the reading and watch the two answers come apart
Both fitted lines are held exactly where they are. The only thing that moves is the Vasant reading being asked about. A marker slides along each line, the wedge between the two answers redraws, and the panel on the right keeps score. Travelling outward from the middle shows where the gap comes from.
Educational illustration. The Nakshatra unit, the Vasant unit and all ten paired readings were manufactured for teaching, and every figure is a monthly change in per cent.
Somebody supplies a fitted line and asks the question the other way round. What is the right move?
The failure: inverting a slope because inversion is the habitual move with a line
Somebody has a working line explaining the Vasant unit from the Nakshatra unit. The line was fitted properly, it has been used for months, and everybody trusts it. Then a question arrives from the other end: here is a Vasant reading, what was the Nakshatra unit doing? The analyst does the obvious thing, takes the slope of 1.5000, inverts it to 0.6667, and reads the line backwards. At a Vasant reading of 9.50 per cent they report a Nakshatra reading of 6.00 per cent, where a properly refitted reverse line gives 4.78 per cent.
Nothing in the output looks wrong, and that is the entire difficulty. Inverting a slope is a valid operation on a drawn line. No routine errors. No figure comes out an impossible size or an impossible sign. The number simply answers a question nobody asked, dressed in exactly the same clothes as an answer to the real one. No check on the output can separate the two. There is nothing in the output to check.
The habit that prevents it: settle the two roles from the question before a line is fitted, and when the question turns round, refit rather than invert. The habit is the whole defence, and it has to sit at the front of the work. There is nowhere further back the error could be caught.
An inverted slope produces a number that looks perfectly reasonable. What makes this failure so hard to catch?
Covered elsewhere. Whether one column causes the other is a different question with a different answer and is covered separately. So is the curve obtained by sweeping a threshold across a rule that returns a yes or a no, and so is what happens when a rule is tried on months it was never fitted to. Whether either of these two columns would be worth anything to anybody trading anything is dealt with in its own place.
Who backs the figures printed above?
No outside body does. Every quantity above is arithmetic on ten pairs of readings built to teach a shape, and redoing that arithmetic is the only check such readings admit. The table gives each figure, the arithmetic behind it, and the step where somebody redoing it by hand is most likely to slip. Fitting a straight line one way round or the other asks nobody’s permission and holds anywhere.
| The figure | The arithmetic behind it | The slip to watch for on a redo |
|---|---|---|
| Forward slope 1.5000, intercept 0.5000 | Shared movement divided by the spread of the Nakshatra unit, then the line put through both averages | Dividing by the spread of the wrong column, which lands on the reverse slope instead |
| Reverse slope 0.5039, intercept minus 0.0078 | The same shared movement divided by the spread of the Vasant unit | Reusing the forward intercept, which is not close to the reverse one |
| One over the forward slope, 0.6667 | A single division, correct about the drawn line and wrong as an answer to the reversed question | Nothing to redo. The slip is using it, not computing it |
| The product 0.7559 | 1.5000 multiplied by 0.5039, and separately R squared computed from scratch, to confirm the two land on the same figure | Rounding both slopes to four places first, which moves the last digit of the product |
| Correlation 0.8694 | The square root of 0.7559, computed both ways round to confirm it does not move | Losing the direction of the relationship, which the square root discards |
| 6.00 per cent and 4.78 per cent at a Vasant reading of 9.50 | The forward line rearranged, and the reverse line used as it stands | Rearranging the reverse line as well, when it is already pointing the right way |
| What the two role names mean | Ordinary usage in statistics, put into plain words here | Nothing to redo. The slip is reading the two names as properties of the columns rather than as jobs inside a rule |
The Nakshatra unit and the Vasant unit are invented.
Educational material. Not advice on any investment, tax, budget or market position.