Chapter 14: Statistical Process Control

Welcome to Last Minute Lecture.

This free chapter overview is designed to help students review and understand key concepts.

These summaries supplement not replaced the original textbook and may not be redistributed or resold.

For complete coverage, always consult the official text.

Okay, let's unpack this.

Think about it.

Our world is in constant motion.

Global temperatures shift.

Product quality varies.

Stock prices fluctuate.

We see data points moving up and down all the time.

Yeah, constantly.

But how do we really know if those changes are just random everyday noise or if something truly significant is happening beneath the surface, something that demands our attention?

It's a fundamental question, really, and it's one that statistical process control is uniquely designed to answer.

Without the right tools, it's remarkably easy to mistake a random fluctuation for a real signal.

Right.

And sometimes even worse, maybe a real signal for just more noise.

Precisely.

So today we're taking a deep dive into statistical process control, pulling key insights directly from Mario F.

Triola's Elementary Statistics.

Great source.

Our mission here is to equip you with the foundational understanding to determine if a process is statistically stable.

Which means it's predictable, behaving as expected.

Exactly.

Or if it's out of control and showing meaningful patterns that actually require action.

And this deep dive, it should give you a powerful lens, really, to see beyond just the raw numbers.

Whether you're curious about, say, climate trends,

analyzing business performance, or maybe just trying to cut through the information overload in daily life.

These methods offer a kind of shortcut to becoming truly well -informed.

You're going to uncover some, well, surprising facts, I think, and practical applications that turn complex statistics into those invaluable aha moments.

Love those aha moments.

So where do we begin?

Let's start with the very bedrock, right?

What exactly is process data, and why is that distinction so important?

Good starting point.

Yeah.

Process data isn't just any old collection of numbers.

It's data specifically arranged according to a time sequence.

Okay, time sequence.

Think of it as a series of measurements taken over time, reflecting some characteristic of goods or services.

It's generated by some combination of equipment, people, materials, methods, conditions, the whole system.

So it's not just what happened, but crucially, when it happened.

Our source gives a fantastic example.

Table 14 to 1.

Earth's global mean surface temperatures.

They're listed year by year, starting way back in 1880.

You can literally trace the chronological flow.

Precisely, and once we have that kind of sequential data, we can ask the next critical question.

What does it mean for a process to be statistically stable?

Right.

What does stable mean here?

When a process is statistically stable, it means it exhibits only what we call natural random variation.

Okay.

You won't see any discernible patterns, no obvious cycles, and no unusual individual points that sort of jump out.

It's just behaving within its expected inherent variability, like the gentle, unpredictable ripple on a calm pond, maybe.

Nice analogy.

So it's just noise, expected noise.

Expected noise, yeah.

And this concept of stability is vital because if a process is not statistically stable, if it's out of statistical control, that means something beyond random chances influencing it.

Ah, okay.

There's a specific non -random cause at play that needs to be identified.

This brings us to the two key types of variation we need to understand then.

First, there's random variation.

This is the inherent, unavoidable variability built into any process, right?

Absolutely.

No two products, no two services will ever be perfectly identical, even in the most controlled environment.

It's just part of the natural world, a slight tremor, even in the steadiest hand.

Okay.

That makes sense.

And the second type.

Then there's assignable variation.

Now, this is the variation that stems from specific causes.

Ah, identifiable.

Like imagine a machine component wearing out, or an employee needing more training, maybe a sudden unexpected change in raw materials.

These are issues you can pinpoint, investigate,

and ideally correct.

Okay.

So random is just background noise.

Assignable is like a specific problem signal.

You got it.

And this distinction is actually incredibly powerful because sometimes people intervene when they really shouldn't.

How so?

Well, our source has this great anecdote from Nashville Corp.

They had a paper -coding machine that was operating perfectly statistically stable.

But supervisors would periodically take samples.

And if the results were slightly off what they expected, even within that random variation, they'd make an adjustment.

They'd tweak the machine.

Oh, I see where this is going.

Exactly.

These over -adjustments, or tampering, as the quality expert W.

Edwards Deming called it, they actually increase the number of defects.

Wow.

So matching with a stable process made it worse.

Precisely.

The core insight here is that you're far better off not making adjustments unless there's a clear statistical signal that the process has truly shifted due to assignable variation.

Otherwise, you're just adding more random noise into the system.

Okay, got it.

So how do we actually see these patterns and distinguish the noise from a real signal?

That's one chart.

What are they and what's their immediate benefit?

A run chart is basically,

well, it's just a sequential plot of individual data values over time.

You can visualize it as a time series graph.

The horizontal axis tracks the chronological order time and the vertical axis shows the measured values themselves.

Simple plot.

So you're literally just plotting the data as they occur in order and the main purpose is a quick visual check.

Exactly.

It gives you a quick visual scan for any obvious patterns of change that suggest a process is not statistically stable.

Our source outlines several out -of -control criteria you can often spot just by looking.

Okay, like what should we look for?

Right.

You look for things like increasing or decreasing variation.

That's where the vertical spread of points noticeably widens or narrows over time.

Got it.

Then there are clear upward or downward trends where the points consistently rise or fall across the chart.

That's usually pretty obvious.

And shifts, you mentioned shifts.

Yeah, shifts too.

You might see an upward shift where points start noticeably lower at the beginning and then jump up and stabilize at a higher level.

Or a downward shift, the opposite.

Okay.

A single exceptionally high or low point, what we might call an exceptional value or an outlier that also immediately grabs your attention.

Makes sense.

And finally, a cyclical pattern, which is a distinct repeating up and down movement, like waves.

Okay.

Trends, shifts, outliers, cycles, changing spread.

Got it.

Let's apply this.

The global Earth's temperatures we mentioned earlier.

Perfect example.

If you plot the annual mean surface temperatures from 1880 onwards, like figure 14 to 1 in the source material does,

what immediately jumps out when you look at that chart?

Well, the visual pattern is striking, isn't it?

The points clearly rising from left to right almost relentlessly.

Yeah, it's hard to miss.

This isn't just random noise.

It's a distinct upward trend.

So the run chart tells us.

It suggests the Earth's temperature process is out of statistical control.

It provides compelling visual evidence consistent with, well, global warming.

It indicates that non -random factors are at play, pushing temperatures higher than natural variation alone would account for.

And it's not just climate data where this works.

Our source also briefly mentions the Flynn effect.

Ah yeah, the IQ scores.

The observed upward trend in IQ scores globally since around 1930.

The why is still debated, but it's another powerful example of identifying a significant upward trend using just a simple run chart.

Absolutely.

So run charts provide that initial intuitive visual check.

A great first step.

But to get more precise, right?

To really set those statistical alarm bells ringing, we often need something more robust.

Exactly.

That's where we turn to our second tool.

Control charts, sometimes called She -Hard charts, after Walter She -Hard, who developed them.

Control charts.

Okay, how do they differ from run charts?

They build on the run chart idea, but they add crucial statistical boundaries.

They're still sequential plots over time, but they include a distinct centerline, a lower control limit, LCL, and an upper control limit, UCL.

Centerline, LCL, UCL.

Got it.

The centerline represents the process's central value, its average behavior.

And the LCL and UCL act as statistical tripwires, basically.

They're calculated from the data itself to identify points that are significantly high or low.

Okay, tripwires.

I like that.

Now, there's a critical point here.

You emphasize these limits.

They're based on the actual process behavior.

Yes, absolutely critical.

They are based on the actual observed behavior of your process, how it's really performing day to day.

They are not based on your desired specifications or targets that may be marketing or engineering set.

So a process can be in control, but still making bad products.

Precisely.

A process can be in perfect statistical control, meaning it's stable and predictable, behaving consistently, but still produce items that don't meet your customer's desired quality specs.

That sounds counterintuitive.

Can you give an example of why that distinction matters so much?

Sure.

Think about the famous Ford Mazda transmission story.

It's a classic case study.

Both companies were producing transmissions to the exact same engineering specifications, same blueprints.

However,

Ford's transmissions had significantly more warranty repairs, way more problems.

Even though they met the specs.

Yes.

When Ford investigated, they found their transmissions were technically meeting specifications, but the variation within Ford's parts was much greater than Mazda's.

Mazda uses superior manufacturing process, particularly a better grinder, which resulted in much less internal variation, more consistency.

Ah, consistency.

Exactly.

So Ford realized they needed to focus on reducing that variation in their own process, even though they were already technically meeting specs.

And when they did, their quality improved dramatically.

Wow.

It shows that just hitting a target isn't enough.

Controlling the consistency, the variation of your process behavior is often the key to true quality.

That's a huge aha moment right there.

It's all about consistency, minimizing that variation, not just hitting a bullseye once in a while.

You got it.

And for control charts, the criteria for being out of control become much more specific, more statistically rigorous than just eyeballing a run chart.

Indeed.

The first criterion is still an obvious non -random pattern, trend or cycle, just like with run charts.

If it looks non -random, it probably is.

Second, and this is where the limits really come in, is having at least one point that falls above the upper control limit or below the lower control limit.

These are statistically improbable events if the process is truly stable and only random variation is present.

So those limits act like warning flags.

Definitely.

And then there's the famous run -of -eight rule.

Right, the run -of -eight.

What's that about and why is it so significant?

The run -of -eight rule states that a process is considered out of control if there are at least eight consecutive points that are all above the center line or all below the center line.

Eight in a row on one side.

Why is that a flag?

Because it's statistically very unlikely to happen just by random chance.

Think about it.

With a stable process, there's roughly a 50 -50 chance, .5 probability, that any given point will be above or below the center line.

Like flipping a coin?

Exactly.

The probability of getting eight heads in a row or eight tails in a row, well, it's .5 to the power of eight, which is one divided by 256.

Very small probability.

Ah, okay.

So if you see that run -of -eight, it's a strong statistical indicator that something has shifted, that it's not random anymore.

Precisely.

It signals a non -random shift in the process level.

So control charts give us these precise boundaries and rules.

Now when we're dealing with quantitative data actual measurements, like temperature or length or weight, there are two main types of control charts often used together.

Right.

We typically use R charts, which monitor the process variation or range.

Okay.

R for range, variation.

And X charts, X bar charts, which monitor the process center or mean.

X bar for the average.

Why use both?

Well, they're complementary because a process can become unstable in different ways.

It might be due to increased variation or maybe a shift in its average level or sometimes both happening at once.

Using both charts gives you a much fuller picture of the process's health.

Okay.

That makes sense.

Let's dive into R charts first for monitoring variation.

What's the core objective here for a listener trying to apply this?

The main objective of an R chart is to determine if the spread or the consistency of your process data is within statistical control.

Is the variation stable over time?

And what do you need to make one?

You need your process data collected in sequences of samples over time.

Importantly, all the samples need to be the same size.

We call that size N.

And ideally, the sample should be independent and the underlying data roughly normally distributed, though control charts are fairly robust to that.

Okay.

Samples of the same size.

And what are the key elements plotted on an R chart?

What goes where?

Okay.

On an R chart, the points you plot are the individual sample ranges.

For each sample, you find the range, highest value minus lowest value, and plot that.

Plot the ranges.

Got it.

Center line.

The center line is R bar, which is simply the average of all your individual sample ranges.

Average of the ranges.

Yeah.

And the limits.

UCL and LCL.

The upper control limit, UCL, is calculated as a constant D4 multiplied by R bar.

And the lower control limit, LCL, is another constant D3 multiplied by R bar.

D3 and D4.

Where do those come from?

Magic.

Not quite magic, but close.

Yeah.

No.

These D3 and D4 values are pre -calculated constants.

You look them up in statistical tables, like table 14 to 2 in our source material.

They depend on your sample size then.

Okay.

They are specifically chosen based on statistical theory to ensure the control limits capture nearly all the expected random variation, usually around 99 .7 % if the process is normal.

They effectively act as that statistical trip wire we talked about.

So if a sample range falls outside these D3, D4 derived limits, it's highly improbable that it's just random chance.

Exactly.

It's a strong signal that the variation itself might be out of control.

Let's go back to our global earth temperatures to see how an R chart works in practice.

The source uses the 14 decades of temperature data with 10 measurements per decade.

So N10.

What does that R chart reveal about the variation in temperatures?

Right.

So for each decade in sample of N10, you calculate the range of temperatures within that decade.

Okay.

Then you find the average of those 14 decade ranges that gives you your R bar, the center line, which comes out to 0 .4371 degrees Celsius in this case.

Then you find D3 and D4 for N10 in the table.

D4 is 1 .777 and D3 is 0 .223.

You multiply these by R bar.

So the upper control limit UCL is 1 .777 times 0 .4371, which equals 0 .7767.

And the lower control limit LCL is 0 .223 times 0 .4371, which is 0 .0975.

So we have our center line and our limits for the range.

Now we plot the 14 actual decade ranges.

And what do we see?

When you look at the 14 sample ranges on the R chart, the chart clearly shows that there is at least 1 .1 decades temperature range that lies beyond the upper control limit of 0 .7767.

Uh oh.

So what's the so what there?

What does that tell us?

The so what is critical.

It means the variation of the earth temperature process is itself out of statistical control.

Not just the average, but the swing.

Exactly.

The fluctuations in temperature, the range of temperatures within a decade seem to be becoming more extreme or at least behaving in a way that's beyond what random variability would predict based on the overall average range.

This isn't just about the average temperature going up.

It's also about the consistency or stability of the climate's variation changing.

Fascinating.

So the variation is out of control.

Now let's move to the other chart, X charts, which monitor the mean or average.

What's the core objective of this chart?

The objective of an X bar chart is to determine if the center or the average level of your process data is within statistical control.

Is the process average stable over time?

And the requirements are pretty much the same as for the R chart, right?

Consistent sample size N, et cetera.

Yes.

Same basic requirements.

Process data gathered in samples of size N over time, roughly normal distribution, independent samples.

You typically construct the R chart first to make sure variation is stable or at least understand it before interpreting the X bar chart.

Okay.

So instead of plotting ranges, here we're plotting the individual sample means, the averages from each sample.

Correct.

The points plotted are the individual sample means, X values.

And the center line and control limits here, how are they calculated?

The center line is X double bar, XX, which is simply the mean of all your individual sample means.

It's effectively the overall average of all your data points combined.

Okay.

Average of the average.

And the control limits, UCL and LCL, are calculated using R bar from the R chart.

The UCL is X double bar plus a constant A2 times R bar.

A2.

Another constant.

Yep.

Another constant from that same table, table 14 -2, which also depends on the sample size N.

And the LCL is X double bar minus A2 times R bar.

Okay.

So UCL plus A2R and LCL X A2R.

Oh, got it.

This sounds incredibly useful for spotting issues in say, manufacturing quality, product weight, fill volume.

But I found a truly fascinating, almost unexpected application in our source,

detecting bribery.

Oh, right.

The Jialai example.

It's a classic illustration of SPC's power.

Tell us about it.

How did control charts catch cheats?

Well, control charts were actually used as evidence in court to help convict someone who was bribing Florida Jialai players to deliberately lose certain games.

No way.

Yes.

An auditor noticed abnormally large sums of money being wagered on specific, usually low probability types of bets.

And looking at the player's performance, some contestants just weren't winning as much as their past performance would predict.

Something seemed off.

So they suspected cheating, but how did they prove it statistically?

Used control charts.

The R charts and X bar charts presented in court showed what the expert witness described as highly unusual patterns of betting.

Critically, there were points representing winnings or bet payoffs that fell well beyond the upper control limit.

Wow.

This powerfully indicated that the process of betting and payouts was way out of statistical control.

It wasn't just experiencing random luck or fluctuations.

The statistician could even use the charts to pinpoint the approximate date when this assignable variation that cheating seemed to stop, which correlated directly with the suspect's arrest.

That is absolutely wild.

A fantastic illustration of how statistics can provide compelling, actionable evidence in really unexpected domains.

Definitely shows the versatility.

Okay.

Back to our slightly less dramatic, but perhaps more globally significant temperature data.

Let's walk through the X bar chart for that.

What does it tell us about the average temperature over time?

Right.

So using the same global temperature data, we first calculate X double R, the overall mean of all 14 decade means.

That comes out to 14 .0802 degrees Celsius.

That's our center line for the X bar chart.

Center line, 14 .0802.

We already calculate R bar from our R chart work.

Right.

It was .4371 degrees Celsius.

Okay.

And for our sample size of N10, we look up the A2 constant in table 14 .2, it's .308.

A2 is .308.

Got it.

Now we plug those into the formulas.

The upper control limit is 14 .082 plus 0 .308 times .4371, which equals 14 .2148.

UCL.

14 .2148.

And the lower control limit is 14 .082 minus .308 times .4371, which equals 13 .9456.

LCL, 13 .9456.

And now we plot the 14 decade means against these limits.

What does the picture show?

Well, when you examine the resulting X bar chart, the conclusion is pretty stark.

The process mean is also unequivocally out of statistical control.

What criteria are met?

You see multiple signals.

There are points lying beyond both the upper and lower control limits at different times.

And critically, you also see evidence of the run of eight rule.

There are clearly at least eight consecutive points lying below the center line early on.

And then a very strong upward trend later with points eventually exceeding the UCL.

It fulfills multiple out of control criteria, all pointing strongly towards a non -random process, driving the average temperature.

So let's put it all together then.

The run chart visually showed the upward trend.

The R chart indicated the variation was increasing or out of control.

And now the X bar chart confirms the average temperature itself is shifting significantly out of its historical statistical bounds.

Exactly.

Taken together, the analysis consistently shows that the Earth's temperature process is out of statistical control according to these standard methods.

Which suggests, statistically speaking, that global warming isn't just random fluctuation, but a real phenomenon driven by identifiable non -random factors.

It provides a robust statistical argument consistent with that conclusion, yes, based on the data presented and the methods used.

Okay.

Now we've talked about measurements, temperature, ranges, means,

but SBC isn't limited to that, right?

We can also use control charts for attributes.

Yeah.

Things that are either bearer like defective or not defective.

Absolutely.

That's where P charts come in.

P charts.

What's the key concept here?

What are we tracking?

With P charts, we're monitoring the proportion, that's the P, of items in a sample that possess a specific attribute.

Usually that attribute is something like being defective or non -conforming to specifications, but it could be any binary characteristic.

We're tracking rates or percentages, not direct measurements.

Okay.

So the objective is to determine if that proportion of defects or whatever attribute is behaving predictably over time or if it's out of statistical control.

Precisely.

Is the defect rate stable or something causing it to change?

And what kind of data and setup do we need for a P chart?

You need process data collected over time in a sequence of samples.

Again, ideally the samples are the same size, N.

And for each item within a sample, it has to fall into one of just two categories like defective or not defective.

And as before, the samples should be independent.

Okay.

Two categories, same sample size N.

How do we plot it?

What goes on the chart?

The points plotted on a P chart are the proportions of the attribute found in each individual sample.

So if you have a sample of a hundred items and five are defective, you plot P hat equals a 5 ,100 equals 0 .05 for that sample.

Plot the sample proportions and the centerline.

The centerline is P bar.

This is the overall proportion of the attribute across all samples combined.

You calculate it by taking the total number of defects or items with the attribute found in all samples divided by the total number of items sampled across all batches.

Total defects divided by total items sampled.

Got it.

Now, the control limits for P charts.

You mentioned there's a crucial calculation distinction here, something to watch out for.

Yes.

This is a common point of confusion, so it's worth highlighting.

The formulas for the UCL and LCL use P bar, but they also involve N, the individual sample size.

The UCL is P bar plus three times the square root of P bar times Q bar divided by N.

Okay.

P bar plus three squared.

And Q bar is just one minus P bar?

Correct.

Q bar is the proportion without the attribute,

and the LCL is P bar minus three times that same square root term, P bar.

For three score cards.

Now, there are a couple of practical adjustments.

Since proportions can't be greater than one or less than zero, if your calculated UCL comes out greater than one, you just use one.

Cap it at one.

And if the calculated LCL comes out less than zero, you just use zero.

Floor at zero.

Makes sense.

But the key distinction you mentioned.

The key distinction is remembering that P bar, the centerline, is calculated using the total number of items sampled across all batches in the denominator.

But the square root term in the UCL and LCL formulas divides by N, the size of each individual sample.

Don't mix those up.

Subtle but critical.

Use the total for the center.

Use the individual sample size for the spread and the limits.

Got it.

It's exactly.

Let's look at the example from our source.

Defective aircraft altimeters made by the Orange Avionics Company.

Right.

They manufacture altimeters in batches of 100, so N100, and they recorded the number of defective altimeters found in 12 successive batches.

The numbers were like 2, 0, 1, 3, 1, 2, 2, 4, 3, 5, then it jumped to 12, then 7 defects.

Okay.

12 batches, N100 each.

First, the centerline, P bar.

Right.

You sum up all those defects, 2 plus 0 plus 1 plus 12 plus 12 plus 7.

That total comes to 42 defects.

Total defects equals 42.

And the total number of altimeters sampled is 12 batches times 100 altimeters per batch, which is 1200.

Total sample equals 100.

So P bar is 42 divided by 1200, which equals 0 .035.

That's our centerline.

The average defect rate is 3 .5%.

Centerline P bar is 0 .035.

So Q bar is 1 minus that, which is 0 .965, and N is 100.

Perfect.

Now we plug those into the control limit formulas.

UCL equals 0 .035 plus 3 score T.

0 .035, 0 .965, 100.

That calculates out to approximately 0 .090.

UCL 0 .090.

LCL 0 .035, 3 score T, 0 .035, 0 .965, 100.

That comes at negative.

So we adjust the LCL to point.

Correct.

LCL is 0.

So our control band is from 0 to 0 .090, or 0 % to 9 % defects per batch.

Now we plot the individual sample proportions, like 21000, 0 .020, 0 .030, go 1 .12, 7100, 0 .07 on this chart.

And what does that P chart for these altimeters reveal?

What's the story?

The P chart immediately tells a story.

First, you can visually see an upward trend in the proportion of defective altimeters over these 12 batches.

Not good.

Not good at all.

But more critically, there is at least one point specifically the batch with 12 defects, which corresponds to a proportion of 0 .12 that lies significantly beyond the upper control limit of 0 .090.

Ah, the tripwire was hit.

Definitely.

This is a clear statistical signal.

The process producing these altimeters is out of

chance or bad luck.

And when a process is out of control and something as critical as aircraft altimeters.

That means immediate corrective action is absolutely required.

You need to investigate why the defect rate is increasing and fix the assignable cost.

This isn't just theoretical number crunching, then.

It translates directly to real -world quality improvement and even cost savings, right?

Like the PerStorp components example.

Indeed.

That's a great practical case.

PerStorp components use computer -generated control charts, likely P charts or maybe X bar in our charts, to monitor the thicknesses of floor insulation they were making for Ford and Jeep vehicles.

By closely tracking their process performance with these charts, they were able to identify and address the assignable causes of variation when signals appeared.

The results?

They reduced their waste by over two -thirds.

Wow.

Over two -thirds reduction in waste?

That's huge.

Huge.

And the source notes that the $20 ,000 cost to the computer system they used for this was more than offset by a $40 ,000 savings in labor costs in just the first year.

Incredible ROI.

It's a perfect example of how applying these statistical methods diligently can have really tangible, significant financial and quality benefits for a business.

What a journey we've taken.

We've really dug into the incredible power of run charts for that first look, and then control charts are charts for watching the consistency or variation.

The spread.

X bar charts for monitoring the average or the center of the process.

The level.

And P charts for keeping an eye on the proportion of attributes, like defects.

Yep.

Covering the main tools for process monitoring.

And the key takeaway seems to be that these aren't just abstract statistical concepts buried in textbooks.

They are vital, practical tools.

Tools for understanding if any process, whether it's manufacturing goods, tracking website

managing hospital waiting times, or even looking at global climate patterns is truly stable and predictable.

Or if it's showing signals that demand our immediate attention.

Absolutely.

These methods empower us really.

They let us move beyond simply looking at isolated data points or reacting to every little up and down.

They allow us to actually interpret the underlying behavior of a system over time.

To separate the signal from the noise.

Exactly.

To distinguish with statistical confidence between that random noise.

The expected unavoidable ups and downs and meaningful non -random change.

Change that signifies an assignable cause.

Something that always requires investigation and often action.

So as you or listener reflect on this deep dive, here's something to maybe mull over.

What process in your own life or maybe in your work could you start monitoring?

It doesn't have to be complex.

Could be anything really.

How might applying some of these basic principles of statistical process control, maybe even just starting with a simple run chart, allow you to spot hitting trends, perhaps prevent problems before they escalate, or maybe just gain some unexpected insights from the data you see or the observations you make every day.