Standardization In Regression: What Actually Makes Sense
Unraveling The Mystery Of Starks And Menchinger A Comprehensive Guide
Most people who run regression models standardize every continuous variable before they even check their results. They do it because a blog post told them to, or because they saw someone else do it on Stack Overflow. The result is usually a mess of coefficients that are harder to interpret than the raw ones, with no actual gain in model fit or predictive power. Menchinger and Starks pointed this out pretty directly in their work on standardization procedures. The core confusion comes from mixing up three separate goals: making coefficients comparable across variables, dealing with multicollinearity, and reporting effect sizes. They are not the same thing. You pick one and standardize accordingly. Everything else is usually noise.
What Starks and Menchinger Actually Argue
Their main point, stripped of academic padding, is that standardization is a presentation decision, not a model-building decision. Running your OLS on standardized data gives you the exact same fitted values, the exact same R-squared, the exact same p-values. It changes nothing about the model itself. It only changes how the coefficients look when you write them down. That matters less than most people think. They break the problem into two types of standardization that people routinely blur together. Full standardization means subtracting the mean and dividing by the standard deviation for every continuous predictor. Semi-standardization means leaving the independent variables in their original units and only standardizing the dependent variable. Most real-world applications that claim to use standardized coefficients are actually doing semi-standardization without admitting it.
Why People Standardize Everything Anyway
The most common reason is comparison. If your dataset has income measured in thousands and age measured in years, the raw coefficients sit on completely different scales. A beginner looks at that and assumes the smaller coefficient means a weaker effect. That assumption is wrong. The scale of the predictor determines the scale of the coefficient by design. Dividing income by a thousand and age by its standard deviation makes them numerically adjacent, which looks useful until you need to explain the model to anyone who doesn't already know the math. There is also the matter of numerical stability. Some people standardize to keep matrix calculations from blowing up. In practice, modern regression routines handle wide-ranging scales without breaking. I have run models with predictors spanning six orders of magnitude and never seen a convergence issue that wasn't already handled by the software's default scaling. Unless you are doing something unusual like writing your own solver from scratch, this concern is mostly theoretical.
Get the Full Details
When Standardization Actually Helps
The one clear case where standardization improves interpretability is when you have interaction terms or polynomial terms in your model. If you include X squared or X times Y, centering the variables first dramatically reduces non-essential multicollinearity between the main effects and the higher-order terms. This is different from the correlation between X and Y themselves. It is the artificial inflation of correlation caused by not centering. The fix is simple: subtract the mean before creating the interaction or squared term. Standardizing beyond that point is optional and usually unnecessary. A second legitimate use case is meta-analysis or cross-study comparison. If you are aggregating results from five different papers that measured the same construct with different instruments, the only way to put those coefficients on a common footing is to report them in standard deviation units. This is where semi-standardized coefficients come in handy. You leave the predictors alone and divide the dependent variable by its standard deviation. The resulting coefficient tells you how many standard deviations the outcome changes per unit change in the predictor. That is a single number that survives across studies with different measurement scales.
The Multicollinearity Myth
Here is where most guides get it wrong. Standardization does not fix multicollinearity. I ran into this explicitly last year when a client came to me with a model that had an variance inflation factor above twelve for two of their predictors. They had already standardized everything because a consultant told them to. The VIF was exactly the same before and after standardization. It is still the same now. The correlation structure of the predictors does not change when you rescale them. Centering helps with polynomial terms, but full standardization leaves pairwise correlations untouched. If you actually have a multicollinearity problem, the solutions are ridge regression, principal component regression, dropping or combining variables, or collecting more data. Standardization is not among them. Telling someone to standardize their variables to reduce multicollinearity is like telling someone to wash their hands to fix a leaky pipe. It addresses a different problem entirely.
What The Coefficients Actually Mean After Standardization
A fully standardized coefficient tells you the expected change in the dependent variable, measured in standard deviations of the dependent variable, for a one-standard-deviation increase in the predictor. This sounds intuitive until you remember that the standard deviation of your dependent variable is specific to your sample. Change the sample and the coefficient changes, even if the underlying relationship is identical. That is why standardized coefficients are fragile across populations and why researchers who compare them across datasets are often comparing apples to oranges. A semi-standardized coefficient tells you the change in the dependent variable in its own standard deviation units per one-unit change in the predictor. This is more stable because the predictor stays in its natural units. You can still compare it across studies if the predictors are measured the same way, which is why it is the preferred format for meta-analytic work. The tradeoff is that it only works well when the predictor has a meaningful zero point and a consistent scale across contexts.
A Practical Workflow
Start with the raw model. Fit it, check the diagnostics, look at the residuals. Only after you have a model you are reasonably happy with do you consider whether standardization serves a purpose. If you are reporting to an audience that expects effect sizes in standard deviation units, run the semi-standardized version. If you are building a model with interactions, center your variables first. If you are doing neither of those things, there is no compelling reason to standardize at all. I keep a simple checklist in my notes for this now. First, is there a valid reason to standardize for this specific analysis? Second, if yes, which type: full or semi? Third, does the standardization change any substantive conclusions, or does it only change the numbers on the page? The answer to the third question is almost always that it changes only the numbers on the page. That is useful information in itself.
Common Mistakes To Avoid
Standardizing before splitting your data into training and test sets is a mistake that costs people credibility. If you compute the mean and standard deviation on the full dataset and then split, your test set has leaked information from the training set. The fix is to fit the scaler on the training data only and apply it to both. This is basic machine learning hygiene but it gets ignored constantly in academic papers. Another mistake is standardizing a binary variable. There is no point. A dummy coded 0 and 1 already has a natural interpretation. Standardizing it turns a clean interpretable coefficient into something that requires a conversion table to explain. Leave binaries as they are. A third one is treating standardized coefficients as a substitute for checking model assumptions. A standardized model with heteroscedastic errors is still a standardized model with heteroscedastic errors. The diagnostic work is identical. The only difference is the labels on the axis.
When Standardization Fails Completely
Standardized coefficients become nearly meaningless when your dependent variable has a very small standard deviation relative to its range, or when it is heavily skewed and the standard deviation is inflated by outliers. In those cases, the coefficient suggests a tiny effect simply because you divided by a large number, not because the relationship is weak. Robust standardization using the median absolute deviation can help here, but most packages do not offer it by default and most practitioners do not know to ask for it. They also break down when you compare results across models with different dependent variables. Each dependent variable has its own standard deviation, so a coefficient from model one is not comparable to a coefficient from model two even if the predictors are identical. This is a routine problem in longitudinal studies where the outcome measure changes over time, and it is another area where the literature is surprisingly sloppy.
The Bottom Line
Starks and Menchinger's contribution is not a new technique. It is a reminder that standardization is a reporting choice with real consequences, not a universal best practice. Use it when it clarifies your results. Skip it when it does not. The model itself does not care either way.
