Experimental error is the difference between a measured value and the true or accepted value, and calculating it correctly starts with understanding that not all errors are the same or handled the same way. The most familiar formula, percent error, is just the surface. Underneath it sits a whole framework for quantifying uncertainty, propagating it through calculations, and presenting it so that other people can trust and use your results. Getting the calculation right matters less than most people think; getting the reporting right matters far more than most people realize.
Percent Error and Absolute Error
If you have an accepted or theoretical value to compare against, the simplest calculation is percent error: take the absolute value of the difference between your measured result and the accepted value, divide by the accepted value, and multiply by 100. A measured boiling point of 99.1 °C compared to the accepted 100.0 °C gives a percent error of 0.9%. This tells you how far off your result landed relative to the target.
Absolute error is even simpler: it is just the raw difference, with units attached. In the boiling-point example, the absolute error is 0.9 °C. Absolute error is useful when you want to know the size of the discrepancy in meaningful physical terms, while percent error lets you compare accuracy across measurements with different scales. A 0.9 °C error on a boiling point is trivial; a 0.9 °C error on a body temperature reading could be clinically significant.
These calculations only work when you have a known reference value. In many real experiments, you do not. You are measuring something new, and there is no “right answer” to compare against. That is where uncertainty estimation takes over, and the math shifts from simple subtraction to statistical reasoning about how much you can trust your own data.
Random Error and How to Quantify It
Random errors are the unpredictable fluctuations that cause repeated measurements to scatter around a central value. They come from tiny, uncontrollable variations in your equipment, your environment, or your technique. You cannot eliminate them, but you can measure how large they are and report that honestly.
The primary tool is the standard deviation, which tells you how spread out your individual measurements are around their average. A small standard deviation means your readings cluster tightly; a large one means they are scattered. Standard deviation describes the variability of your data itself.
If what you care about is how precisely you know the average, you use the standard error of the mean instead, which is the standard deviation divided by the square root of the number of measurements. Standard error shrinks as you take more data points, because averaging more readings gives you a better estimate of the true center. Standard deviation does not shrink with more data; it characterizes the natural scatter of your measurements regardless of how many you take.
This distinction matters enormously for reporting. Standard deviation answers “how much do individual measurements vary?” Standard error answers “how confident am I in the mean?”1PubMed Central. What to use to express the variability of data: Standard deviation or standard error of mean? Both are legitimate, but they answer different questions, and confusing them is one of the most common mistakes in scientific writing.
Systematic Errors and Why They Are Harder to Catch
Systematic errors push all your measurements in the same direction. A balance that reads 0.05 grams too high, a thermometer with a shifted calibration, a reaction that always loses a small amount of product to the walls of the flask: these do not average out. Take a hundred readings with a miscalibrated instrument and your average is still wrong by the same amount as a single reading.
That is what makes systematic errors dangerous. Random errors announce themselves through scatter in your data. Systematic errors hide, because your data can look beautifully precise and still be wrong. The only way to catch them is to compare your results against an independent method, a known standard, or a calibration reference. If your measured value of a known standard is consistently off, that offset is your systematic error, and you either correct for it mathematically or fix the source.
In practice, identifying systematic errors requires critical thinking about your experimental setup. Ask yourself: is there any step in this process that would always push my result in one direction? Is my equipment calibrated? Am I making an assumption that might not hold? These are the questions that separate careful experimenters from sloppy ones, and no amount of statistical analysis can substitute for them.
Error Propagation Through Calculations
Measurements rarely stand alone. You measure a mass, a volume, and a temperature, then combine them in a formula to get a density or a reaction rate. Each input measurement carries its own uncertainty, and those uncertainties combine in the final result. The question is how.
For addition and subtraction, the uncertainties add in quadrature: you square each absolute uncertainty, add the squares, and take the square root. For multiplication and division, you do the same thing with relative (percent) uncertainties. These rules come from a mathematical technique called propagation of uncertainty, which is based on how small changes in inputs produce changes in outputs.
The underlying math uses partial derivatives to figure out how sensitive your final result is to each input variable. If your formula depends strongly on one measurement and weakly on another, the uncertainty in the first measurement dominates. This is genuinely useful to know, because it tells you where to focus your effort. If the temperature measurement contributes 90% of the uncertainty in your final result, buying a better balance will not help. Improving the thermometer will.
These standard propagation rules assume your measurement model is roughly linear over the range of your uncertainties. For highly nonlinear equations, that assumption breaks down, and more advanced approaches like higher-order Taylor series expansions become necessary to capture how uncertainties actually flow through the math.2Measurement and Control. Uncertainty propagation on a nonlinear measurement model based on Taylor expansion For most laboratory work, though, the basic quadrature rules are sufficient and widely expected.
When the Math Gets Too Messy for Formulas
Some calculations are complex enough that writing out the propagation formula by hand is impractical or error-prone. Monte Carlo simulation offers an alternative that is conceptually simple even when the math is not. The idea is to let a computer do millions of pretend experiments. You tell it the uncertainty distribution for each of your input measurements, and it randomly draws values from those distributions, plugs them into your equation, and records the result. Repeat that a million times and you get a distribution of possible outcomes. The spread of that distribution is your propagated uncertainty.
This approach is especially useful when your equation involves nonlinear functions, when your input uncertainties are not symmetric, or when you have correlated inputs that make the analytical formulas unwieldy. The NIST Uncertainty Machine is a free online tool designed for exactly this purpose, allowing users to define their measurement equation and input distributions and then running the simulation automatically.3ACS Publications (“Journal of Chemical Education”). Monte Carlo Uncertainty Propagation with the NIST Uncertainty Machine For a student, this can be a powerful sanity check: run the Monte Carlo simulation and compare its result to your hand-calculated propagation. If they agree, you probably did both correctly.
Type A and Type B Uncertainty
The international framework for reporting measurement uncertainty, known as the GUM (Guide to the Expression of Uncertainty in Measurement), classifies uncertainty into two types based on how you estimate it, not on whether the underlying error is random or systematic.4International Organization for Standardization. ISO/IEC Guide 98-1:2009
Type A uncertainty is evaluated by statistical analysis of repeated measurements. You take multiple readings, calculate the standard deviation, derive the standard error, and that is your Type A component. Type B uncertainty is evaluated by any other means: manufacturer specifications for your instrument, calibration certificates, published reference data, or your own informed judgment about how large a particular error source could be.5International Journal of Metrology and Quality Engineering. An improved procedure for combining Type A and Type B components of measurement uncertainty
The distinction matters because many uncertainty sources cannot be evaluated by repeating the experiment. The resolution of a digital display, for example, creates an uncertainty that does not change no matter how many times you read it. That is a Type B component. In the GUM framework, you estimate it from the instrument’s specifications and combine it with your Type A components using quadrature, just like combining any other independent uncertainties.
This framework is the international standard used in calibration laboratories, accredited testing facilities, and serious metrology. Even if you are a student in an introductory lab, understanding the distinction between “I calculated this from my data” and “I estimated this from my instrument’s specs” makes your error analysis sharper and more honest.
Instrument Resolution as an Error Source
Every measuring instrument has a finite resolution, and that resolution sets a floor on your uncertainty. A ruler marked in millimeters cannot tell you anything about tenths of millimeters. A digital balance that reads to 0.01 grams has an inherent uncertainty of at least ±0.005 grams, because you cannot know whether the true value rounds up or down to the displayed digit.
This seems straightforward, but the interaction between resolution and random noise is subtler than it appears. When a measurement is affected by both Gaussian noise and finite resolution, the resulting distribution of readings depends on where the true value falls relative to the resolution step. There is no single simple formula that converts resolution into a confidence interval for all cases.6PubMed Central. Uncertainty Due to Finite Resolution Measurements In practice, the standard convention is to treat the resolution limit as a uniform (rectangular) distribution and calculate its standard uncertainty as the half-width of one resolution step divided by the square root of three. This is an approximation, but it works well for most lab situations.
The practical lesson is this: if your random scatter is much larger than your instrument’s resolution, the resolution contribution is negligible and you can focus on the statistical uncertainty from your repeated measurements. If your data barely scatter at all, the instrument’s resolution may be the dominant source of uncertainty, and reporting only a standard deviation from your data would understate the true uncertainty of your measurement.
Dealing With Outliers
Sometimes one measurement in your data set looks wildly different from the rest. The temptation is to throw it out, but doing so without justification is one of the fastest ways to compromise the integrity of your results. An outlier might be a legitimate data point from a process with more variability than you expected, or it might be a genuine mistake, like misreading a dial or recording a number in the wrong units.
Statistical tests exist to help make this decision more objectively. The most commonly taught is the Q-test (Dixon’s test), which compares the gap between the suspected outlier and its nearest neighbor to the overall range of the data. If the ratio exceeds a critical value for your sample size and chosen confidence level, the test suggests the point is discordant.7PubMed. Estimation of type I error probability from experimental Dixon’s “Q” parameter on testing for outliers within small size data sets The Q-test is most commonly applied to small data sets, typically between 3 and 12 observations.
For larger data sets or situations requiring more statistical power, the Grubbs test tends to perform better. A comparative study of four common outlier tests found that the Grubbs test and a kurtosis-based test outperformed the Dixon test across a wide range of sample sizes and contamination scenarios.8PubMed Central. Comparative performance of four single extreme outlier discordancy tests from Monte Carlo simulations
Regardless of which test you use, the most important practice is to document everything. If you remove a data point, state which test you applied, what the result was, and why you believe the exclusion is justified. If you can identify a concrete reason the measurement went wrong (a bubble in the solution, a power fluctuation during the reading), note that too. Never silently delete data points to make your results look better.
Human Factors and Observer Bias
Not all errors come from instruments or statistics. The person doing the measuring introduces biases that are surprisingly hard to eliminate. One well-documented example is digit preference: when recording values, people tend to round to certain numbers, especially zeros and fives. A study of birthweight recordings found that preference for the terminal digit 0 increased progressively with increasing birthweight, and correcting for this bias led to a 1.8% increase in the number of babies classified as low birthweight.9PubMed. Observer error and birthweight: digit preference in recording
This kind of rounding bias might seem trivial, but it can shift distributions and change conclusions in ways that accumulate across large datasets. In your own lab work, digit preference typically shows up when you are reading analog instruments: graduated cylinders, analog thermometers, spring scales. You unconsciously favor certain values over others. Being aware of this tendency is the first step toward reducing it. Recording to one extra decimal place than you think is warranted and then rounding later can help, as can having a second observer independently record readings.
Confirmation bias is another human factor. If you expect a result near a certain value, your eye will tend to interpolate readings in that direction. Blinding yourself to the expected outcome when possible, or at least being explicitly aware of your expectations, reduces this effect. These are not exotic laboratory concerns; they are everyday realities of measurement that no statistical technique can fully correct after the fact.
How to Actually Report Your Error
This is where most students and many professionals stumble. Calculating uncertainty is only useful if you communicate it clearly. A result reported as “5.27 g” tells the reader almost nothing about its quality. A result reported as “5.27 ± 0.03 g” tells them the measurement is precise to roughly the hundredths place. A result reported as “5.27 ± 0.03 g (95% confidence)” tells them that and also specifies the confidence level, which is the gold standard.
The ± symbol should always be defined. Does it represent one standard deviation? Two standard deviations? The standard error of the mean? A 95% confidence interval? These are all different things with different numerical values, and failing to specify which one you are using makes your reported uncertainty ambiguous. In clinical and experimental research, both standard deviation and standard error of the mean are widely used to present data characteristics and analysis results.10PubMed Central. Standard deviation and standard error of the mean But if you do not say which one you are reporting, the reader cannot interpret your number correctly.
A few practical conventions to follow:
- Match significant figures: Your uncertainty should have one or two significant figures, and your result should be rounded to the same decimal place as the uncertainty. Reporting 5.2734 ± 0.03 implies false precision in the result.
- State the confidence level: If you report an expanded uncertainty (like a 95% interval), say so. If you report one standard deviation, say that.
- Use the right metric for the question: If you are describing how variable your measurements are, report standard deviation. If you are claiming a best estimate of the mean, report the standard error or a confidence interval around the mean.
- Include units: The uncertainty has the same units as the measurement. This sounds obvious, but it gets lost surprisingly often in tables with many columns.
The Standard Error Confusion
Misuse of the standard error of the mean is one of the most persistent problems in published science, and it is worth understanding why. Because the standard error is always smaller than the standard deviation (for any sample larger than one), reporting SEM instead of SD makes your data look less variable than it really is. In some fields, this has been a chronic problem.
An evaluation of four anaesthesia journals found that between roughly 12% and 28% of articles published in 2001 used the standard error of the mean incorrectly, typically by reporting SEM where the context called for standard deviation.11PubMed. Misuse of standard error of the mean (SEM) when reporting variability of a sample. A critical evaluation of four anaesthesia journals A more recent study of manual medicine journals found that about 82% of articles used SD and SEM correctly, but roughly 1.4% showed clear inappropriate use of SEM, and about 2.5% failed to define the ± symbol at all.12PubMed Central. Reporting the standard error of the mean: a critical analysis of three journals in manual medicine The improvement over two decades is real but incomplete.
The rule of thumb is simple. If you want to describe how spread out your individual data points are, use standard deviation. If you want to express how precisely you have estimated the mean, use the standard error or build a confidence interval. If you are plotting error bars on a graph, state in the figure caption which metric they represent. Ambiguous error bars are worse than no error bars, because they give the reader false confidence in an interpretation they may be making incorrectly.
Putting It All Together for a Lab Report
If you are writing up a student lab report or a professional measurement result, a clean error analysis follows a natural sequence. Start by identifying your sources of uncertainty: which are random (you can see scatter in repeated measurements) and which are systematic (calibration offsets, known biases, instrument specs). Quantify the random component from your data. Estimate the systematic components from calibration records or manufacturer specifications. Propagate all uncertainties through your calculation to get a combined uncertainty on your final result. Then report the result with that combined uncertainty, clearly stating what the ± value represents.
If your experiment has a known accepted value, calculate your percent error and compare it to your uncertainty. If the accepted value falls within your uncertainty range, your result is consistent with the known value. If it falls outside, either your uncertainty estimate is too small or there is an unaccounted systematic error. Both conclusions are informative, and stating them honestly is far more valuable than fudging numbers to make the percent error look small.
The broader point is that experimental error is not a penalty or a failure. It is information. A well-characterized uncertainty tells the next researcher exactly how much to trust your result and where the weak points in the measurement are. A result with no reported uncertainty is, in a meaningful sense, incomplete. It is a number floating without context, and no one can build on it reliably.
Least Squares Fitting and Residual Analysis
When your experiment generates a set of data points that should follow a trend, like a straight line on a graph, fitting a curve and analyzing the residuals is another way to assess error. The method of least squares finds the line (or curve) that minimizes the sum of squared differences between your measured points and the fitted values. Those differences, the residuals, tell you how well the model describes your data.13ScienceDirect. A tutorial history of least squares with applications to astronomy and geodesy
If your residuals scatter randomly above and below zero with roughly equal magnitude, your model is a reasonable fit and the remaining scatter reflects random measurement error. If the residuals show a pattern, like curving systematically or growing larger at one end of your data range, something is off. Either your model is wrong (the relationship is not really linear, for instance) or there is a systematic error that depends on the measurement conditions.
Residual plots are one of the most underused diagnostic tools in student labs. They can reveal problems that are invisible when you just look at the fitted line overlaid on your data. A line can look like a good fit by eye while the residuals shout that it is not. Getting in the habit of plotting residuals alongside your main graphs catches these issues early and strengthens your error analysis substantially.