Why analysis comes before the decision
Most engineering choices rest on numbers someone measured: how long a part lasted, how much a sensor drifted, how often a build failed. Raw numbers do not make the choice by themselves. Analysis is the step that turns them into something a team can act on, and it is also the step where careless handling can quietly point a team in the wrong direction.
This article follows a practical path: start with the question, check and clean the data, summarise it, say how uncertain the summary is, and only then draw a conclusion. It is general information, not professional engineering advice. Where safety, regulation or contracts are involved, follow the standards and qualified people that apply in your country and industry.
Start with the question
Before opening a spreadsheet, write down what decision the data is meant to inform. “Is the new supplier’s part good enough to replace the old one?” is a usable question. “What does the data say?” is not. A clear question tells you which measurements matter, how precise they need to be and what result would change your mind.
It also helps to decide the acceptance rule in advance. If you pick the threshold after seeing the results, it is easy to choose one that makes the preferred option look good.
Check how the data was collected
Numbers inherit the weaknesses of the process that produced them. Ask a few plain questions:
- Which instrument took the readings, and when was it last checked?
- Were all samples measured the same way, by the same method?
- Were samples chosen at random, or were they the convenient ones?
- Did conditions such as temperature or load change during the test?
A neat table can hide a poor collection process. Note any doubts alongside the data, because they affect how strongly you can state a conclusion later.
Cleaning the data
Cleaning means finding and handling problems without erasing real information. Typical issues include missing values, duplicated rows, mixed units, typing errors and sensor glitches.
A few habits keep this honest:
- Keep an untouched copy of the original data and work on a second copy.
- Look at the values before calculating anything. Sorting a column or plotting it often reveals an impossible number straight away.
- Check units for consistency. A column that mixes millimetres and metres is a common trap, and careful unit conversion at this stage saves trouble later.
- Record every change you make and why.
Be careful with outliers. A reading far from the others might be a recording error, but it might also be the one real sign of an intermittent fault. Removing it because it spoils the average is not a sound reason. Investigate first, and if you do exclude it, say so in your notes.
Basic statistics that do most of the work
For many decisions, a few summaries are enough. The mean and standard deviation describe the centre and spread of a set of readings, and the median is useful when a few extreme values distort the mean.
An illustration with five readings of a length in millimetres: 10.2, 10.4, 10.1, 10.3 and 10.5.
- The sum is 51.5, so the mean is 51.5 / 5 = 10.3 mm. The median is also 10.3 mm.
- The differences from the mean are -0.1, +0.1, -0.2, 0 and +0.2. Squared, they are 0.01, 0.01, 0.04, 0 and 0.04, which add up to 0.10.
- Dividing by n - 1 = 4 gives a sample variance of 0.025, so the sample standard deviation is about 0.16 mm.
The standard deviation describes how much individual readings scatter. The uncertainty of the mean itself is smaller: dividing 0.158 by the square root of 5 (about 2.236) gives roughly 0.07 mm. These numbers are an illustration only; real work may call for other methods depending on the data.
Say how uncertain the result is
A single number such as “10.3 mm” hides how well it is known. Every measurement has some uncertainty from the instrument, the method and natural variation. Reporting a result with its uncertainty, using the ideas behind measurement uncertainty, lets a reader judge whether two values really differ.
Suppose a redesigned part averages 10.3 mm and the old one averages 10.4 mm, with each average uncertain by about 0.07 mm. The gap is only a little larger than the uncertainty, so the data does not clearly show a difference. Reporting figures to a sensible number of significant figures also avoids implying precision you do not have.
Common ways analysis misleads
- Correlation taken as cause. Two quantities rising together does not show that one drives the other. A third factor, such as ambient temperature, may affect both.
- Small samples. Three good results are encouraging but say little about the next thousand units.
- Cherry-picking. Reporting only the runs that worked, or choosing the time window that looks best, produces a misleading picture.
- Hidden changes. If the test setup changed halfway through, comparing early and late results may measure the setup rather than the design.
- False comfort from averages. A mean can look fine while a minority of units fail badly. Where a limit matters, look at the worst cases as well, in the spirit of typical versus worst-case values.
From result to decision
Run simple sanity checks before you trust an output: does the size of the answer make sense, are the units right, does a rough estimate agree? Quick order-of-magnitude estimates are often enough to catch a spreadsheet mistake.
Then write the conclusion in plain terms: what you measured, what you found, how uncertain it is and what you recommend. If the data cannot settle the question, say that and suggest what extra measurement would help. A well-reasoned “not yet clear” is more useful than a confident wrong answer, and it can feed into broader tools such as decision matrices when several factors must be weighed together.