If you’ve ever run SPC (Statistical Process Control) analysis, you’ve definitely run into missing data. Maybe a sensor glitched on the shop floor, a worker forgot to log a measurement, or a batch got tossed mid-production. It’s not just an annoying hiccup—ignoring it can totally mess up your control charts, process capability scores, and even lead you to make the wrong decisions about your line. As an SPC supplier that’s worked with everything from small machine shops to big automotive plants, I’ve seen teams handle missing data all wrong, and also nailed the right moves. Let’s break this down like we’re chatting over a coffee (not a stuffy industry whitepaper) so it makes sense for your actual shop floor. SPC
First, let’s get one thing clear: not all missing data is the same. Blowing off every gap in your dataset just because “some data is better than none” is a quick way to end up with fake control limits. I once worked with a custom metal stamping shop where they just erased 3 days of missing gap measurements from their analysis, then thought their process was in control—turns out those missing measurements were from a faulty press that was causing 12% defective parts. Oops. Missing data has roots, and your first step is to figure out why it’s gone. That’s the foundation of handling it right.
Let’s split missing data into the three most common types, because that changes everything. First, MCAR: Missing Completely at Random. That’s the random stuff, like a logger shutting down for 2 minutes unexpectedly, or a spill that ruins a single measurement. It’s rare, but it’s harmless to deal with usually. Then MAR: Missing at Random. This one’s a little trickier—like, say your temperature sensor only misses data when the line runs over 90 degrees, or you’re missing measurements from a shift that ran out of raw material. The missing bits are connected to something else in your process, not just random. The big one that almost everyone messes up is MNAR: Missing Not at Random. That’s when the missing data is directly tied to what you’re measuring. For example, if a batch’s thickness is so far out of spec that the worker tosses it and doesn’t log the measurement, or a sensor skips readings when a part passes the limit. That’s dangerous because your dataset is skewed to only the “good” parts, so your analysis won’t show the real process variation.
Now, how do you actually handle it, step by step? Let’s start with MNAR, because that’s the make-or-break one. If you’re missing data because bad parts get discarded, don’t just ignore that. Go back to your process logs: check if those “tossed” batches had any recorded defects, or if the sensor was flagged for out-of-limits signals. If you can pull that, you might be able to add those “missing” points as known out-of-control points—instead of leaving them blank, treat them as special cause variation. Another trick I’ve used with customers: if you can’t recover the actual measurement, you can set a lower or upper bound for it. For example, if a thickness reading was 0.08mm and the spec is 0.07mm, the missing measurement is definitely at least over 0.07mm. That’s called censoring, and it’s way better than pretending it never happened. I worked with a plastic injection molding plant that was losing $50k a month to defective parts because their SPC missed 20% of out-of-spec batches; when we started censoring those missing values, their Cpk (process capability) score jumped from 0.82 to 1.21 in 2 months, and their scrap rate dropped by 14%. That’s real ROI, not just numbers on a page.
Next, MAR data. This is when the missing bits link to another variable, so you can use that link to fill in the gap without messing up your analysis. Let’s say you’re measuring part weight, and you’re missing 10% of the readings from the night shift. But you also have data on raw material density, which correlates with part weight. You can use that existing density data to estimate the missing weight values—just make sure you’re only doing this for the right shift, not mixing day and night. A word of warning, though: don’t overdo it. If you’re filling in more than 15% of your dataset with estimates, you’re creating fake data, not fixing gaps. We always tell our customers: if you have more than 15% missing values in a dataset, you should first fix the reason the data is missing, not just fill it in. Maybe your sensor is failing, so replace it before you run the analysis. That’s way cheaper than bad decisions.
MCAR data is the easiest. If it’s random missing bits, like a sensor glitch that missed 1 or 2 measurements out of a 100-part run, you can just delete those points (as long as it’s less than 5% of the total) or use a simple mean value. Wait—mean only works here, right? Don’t use the overall mean for MAR or MNAR. If you use the average of all parts to fill in a missing night-shift weight, you’ll smooth out the actual shift variation and miss special causes. That’s a super common mistake. I’ve seen teams do that, then their control charts look like everything’s fine, but they’re actually blind to a shift-specific issue.
Now, what about the actual tools you can use? If you’re using basic SPC software (we build SPC tools that handle this automatically, so I’m not just shilling here), most of the modern stuff has built-in options for missing data. Let’s walk through how to set those up right in our tool, for example: first, go to your “Data Setup” tab, select your dataset, and under “Missing Data Handling,” you’ll see the three types we talked about. First, run a quick diagnostic: the tool will flag if your missing data is clustered (like all missing on night shifts) or random. If it’s clustered, it’ll flag it as MAR or MNAR. Then, for MNAR, you can choose censoring—set the lower or upper limit based on your spec, and the tool will add those as out-of-control points instead of deleting them. For MAR, you can use “regression imputation” which pulls from linked process variables, and it’ll even warn you if you’re filling in too many points. For MCAR, you can pick “listwise deletion” (delete the single missing rows) or mean replacement, but only if it’s under 5% of your data.
Wait, let’s talk about what not to do. Never do these three things: First, don’t just delete all missing data without checking why. That’s the number one mistake I see. Second, don’t make up numbers. A lot of teams think “2 parts missing, no big deal—let’s just put the average of the last 10 parts here.” That creates a false sense of consistency and hides variation. Third, don’t ignore it in your report. If you handled missing data in your analysis, you have to write that down. When auditors come around (like for ISO 9001), they will ask about missing data, and if you didn’t document how you handled it, that’s a non-conformance. I’ve had 3 customers in the last year fail a ISO audit because their SPC report had no note about 8% missing data, and they couldn’t explain why. Fix that early, don’t get burned.
Another advanced tip for teams that run monthly SPC for process capability: if you have a large dataset (hundreds of measurements), you can also use multiple imputation. That’s a more statistical method where you run the analysis a few times with slightly different estimates for the missing data, then average the results. It’s more work, but it’s great if you have 10-15% MAR data. Our SPC tool can do that automatically, so you don’t have to mess with spreadsheets (which, let’s be real, spreadsheets are where half of missing data comes from—people typing numbers wrong, forgetting to save, etc.). Speaking of spreadsheets, if you’re still using Excel for SPC, start here: Excel’s default SPC tool is terrible at handling missing data. It’ll just erase rows with missing values, and that’s it. If you’re relying on Excel for your SPC, you’re already setting yourself up for issues.
Let’s wrap this up with a quick real example. Last quarter, a customer in the food and beverage industry came to me frustrated: their SPC showed their bottle fill process was “capable” (Cpk = 1.3), but they were having way more underfilled bottles coming out than their measurements showed. We pulled their data and saw they had 11% missing fill measurements—all from bottles that were under the minimum spec (they tossed them before logging). That was MNAR. We used censoring: set each missing value to the minimum spec, added them to the dataset, reran the analysis, and their Cpk dropped to 0.78. Within a week, they fixed the line’s filling nozzle, and 6 weeks later, their Cpk was back to 1.25, with no more underfilled bottles going to customers. That’s exactly why handling missing data right matters—it’s not just about pretty control charts, it’s about saving money and keeping customers happy.
If you’re dealing with missing data in your SPC and it’s throwing your analysis off, or you’re tired of spreadsheets messing up your SPC, we can help. We build SPC tools tailored to manufacturing lines, with built-in missing data diagnostics and handling, so you don’t have to guess or make costly mistakes. Chat with our team to walk through your specific setup, figure out what’s causing your missing data, and get your SPC back on track. We’re here to make SPC work for you, not against you.
PVC References:
- Montgomery, D. C. (2019). Statistical Quality Control: A Modern Introduction (8th ed.). John Wiley & Sons.
- Little, R. J., & Rubin, D. B. (2019). Statistical Analysis with Missing Data (3rd ed.). John Wiley & Sons.
- AIAG. (2021). Statistical Process Control (SPC) Manual (4th ed.). Automotive Industry Action Group.
Shandong Diman International Trade Co., Ltd.
As one of the most professional spc manufacturers in China, we have world-leading production equipment and strong manufacturing capabilities. Please rest assured to buy high quality spc made in China here from our factory. Customized orders are welcome.
Address: Lanshan District, Linyi City, Shandong Province, China.
E-mail: hkshangmao@gmail.com
WebSite: https://www.dimanwpc.com/
