A New View of Statistics 

Measures of validity are similar to measures of reliability. With reliability, you compare one measurement of a variable on a group of subjects with another measurement of the same variable on the same subjects. With validity, you also compare two measurements on the same subjects. The first measurement is for the variable you are interested in, which is usually some practical variable or measure. The second measurement is for a variable that gives values as close as you can get to the true values of whatever you are trying to measure. We call this variable the criterion variable or measure. The three main measures of reliabilitychange in the mean, withinsubject variation, and retest correlationare adapted to represent validity. I call them the estimation equation, typical error of the estimate, and validity correlation. There is also a measure of limits of agreement. I have a little to say on validity of nominal variables (kappa coefficient, sensitivity, and specificity), and I finish this page with a spreadsheet for calculating validity.
You will find that correlation has a more prominent role in validity than in reliability. Most applications of validity involve differences between subjects, so the betweensubject standard deviation stays in the analysis and can be expressed as part of a correlation. In contrast, most applications of reliability involve changes within subjects; when you compute changes, the betweensubject variation disappears, and with it goes correlation.
Let's explore these concepts with an example similar to the one I used for reliability. Imagine you are a roving applied sport scientist, and you want to measure the weight of athletes quickly and easily with portable bathroom scales. You check out the validity of the bathroom scales by measuring a sample of athletes with the scales and with certified laboratory scales, as shown in the figure. I've shown only 10 points, but in practice you'd use probably 20 or so, depending on how good the scales were.
Note that I have assigned the observed value of the variable to the X axis, and the true value to the Y axis. That's because you want to use the observed value to predict the true value, so you must make the observed value the independent variable and the true value the dependent variable. It's wrong to put them the other way around, even though you might think that the observed value is dependent on the true value.
Notice also that a straight line is a pretty good way to relate the observed value to the true value. You'd be justified in fitting a straight line to these data and using it to predict the true weight (lab scales) from the observed weight (bathroom scales) for any athletes on your travels.
By the way, you won't always get a straight line when you plot true values against observed values. When you get a curve, fit a curve! You can fit polynomials or more general nonlinear models. You know you have the right curve when your points are scattered fairly evenly around it. Use the equation of the curve to predict the true values from the observed values.
You can also use a practical measure that looks nothing like the criterion measure. For example, if you are interested in predicting body fat from skinfold thickness, the practical measure would be skinfold thickness (in mm) measured with calipers, and the criterion measure could be body fat (as percent of body mass) measured with dualemission Xray absorptiometry (DEXA). You then find the best equation to link these two measures for your subjects.
There are sometimes substantial differences in the estimation equation for different groups of subjects. For example, you'd probably find substantially different equations linking skinfold thickness to body fat for subjects differing in such characteristics as sex, age, race, and fitness. Sure, you can derive separate equations for separate subgroups, but it's usually better to account for the effect of subject characteristics by including them as independent variables in the estimation equation. For that you need the technique of multiple linear regression, or you could even go to exotic multiple nonlinear models. A stepwise or similar approach will allow you to select only those characteristics that produce substantial improvements in the estimation of the criterion.
The typical error of the estimate is usually in the output of standard statistical analyses when you fit a straight line to data. If you fit a curve, the output of your stats program might call it the root meansquare error or the residual variation. Some stats programs provide it in the squared form, in which case you will have to take the square root. Your program almost certainly won't give you confidence limits for the typical error, but you should nevertheless calculate them and publish them. See the spreadsheet for confidence limits.
In the sections on reliability, I explained that the withinsubject variation can be calculated as a percent variationthe coefficient of variationby analyzing the log of the variable. The same applies here: take logs of your true and observed values, fit a straight line or curve, then convert the typical error of the estimate to a percent variation using the same formula as for reliability. See reliability calculations for the formula. Analysis of the logs is included in the validity spreadsheet. In the above example the standard deviation is 0.7%. Expressing the standard deviation as a percent is particularly appropriate when the scatter about the line or curve gets bigger for bigger untransformed values of the estimate. Taking logs usually makes the scatter uniform. See log transformation for more.
NewPrediction Error
If a validity study has a small sample size
(<50 subjects), the typical error of the estimate is accurate only for the
subjects in the validity study. When you use the equation to predict
a new subject's criterion value, the error in the new estimatelet's
call it the newprediction erroris larger than the original typical
error of the estimate. Why? Because the calibration equation (intercept and
slope) varies from sample to sample, and the variation is negligible only for
large samples. The variation in the calibration equation for small samples therefore
introduces some extra uncertainty into any prediction for a new subject, so
up goes the error. The uncertainty in the intercept contributes a constant additional
amount of error for any predicted value, but the error in the slope produces
a bigger error as you move away from the middle of the data. Your stats program
automatically includes these extra errors when you request confidence limits
for predicted values. You will find that the confidence limits get further away
from the line as you move out from the middle of the data. The effect is noticeable
only for small samples or only for predicted values that are far beyond the
data.
So, exactly how big is the error in a predicted value based on a validity or calibration study with a small sample size? If you have enough information from the study, you can work out the error accurately for any predicted value. Obviously you need the slope, intercept, and the typical error of the estimate. You also need the mean and standard deviation of the practical variable (the X values). I've factored all these into the formulae for the upper and lower confidence limits of a predicted value in the spreadsheet for analysis of a validity study. I've also included them in the validity part of the spreadsheet for assessing an individual. (In that spreadsheet I've used the mean and standard deviation of the criterion or Y variable, because it's convenient to do so, and the difference is negligible.)
When you don't have access to the means or standard deviations from the validity study, you can work out an average value for the newprediction error, on the assumption that your new subject is drawn randomly from the same population as the subjects in the validity study. One approach to calculating this error is via the PRESS statistic. (PRESS = Predicted REsidual Sums of Squares.) I won't explain the approach, partly because it's complicated, partly because the PRESSderived estimate is biased high, and partly because I have better estimates. For one predictor variable, the exact formula for the newprediction error appears to be the typical error multiplied by root(1+1/n+1/(n3)), where n is the sample size in the validity study. I checked by simulation that this formula works. I haven't yet worked out the exact formula for more than one predictor variable, but my simulations show that the typical error multiplied by root[(n1)/(nm2)] is pretty good, where m is the number of predictor variables.
Researchers in the past got quite confused about the concept of error in the prediction of new values. They used to split their validity sample into two groups, derive the estimation equation for one group, then apply it to the second group to check whether the error of the estimate was inflated substantially. That approach missed the point somehow, because the error was bound to be inflated, although they apparently didn't realize that the inflation was usually negligible. And whether or not they found substantial inflation, they should still have analyzed all the data to get the most precise estimates of validity and the calibration equation. The PRESS approach has a vestige of that datasplitting philosophy. Not that it all matters much, because most validity studies have more than 50 subjects, so the newprediction error from these studies is practically identical to the typical error of the estimate.
A final point about the newprediction error: don't use it to compare the validity of one measure with that of another, even when the sample sizes are small and different. Use the typical error, which is an unbiased and unbeatable measure of validity, no matter what the sample size. (Actually, it's the square of the typical error that is unbiased, but don't worry about that subtlety.)
NonUniform Error of the Estimate
You will recall that calculations for reliability
are based on the assumption that every subject has the same typical error, and
we used the term heteroscedasticity to
describes any nonuniform typical error.
The same assumption and terminology underlies calculations for the validity,
and the approach to checking for and dealing with any nonuniformity is similar.
Start by looking at the scatter of points on the plot of the estimation equation. If every subject has the same typical error of the estimate, the scatter of points, measured in the vertical direction on the graph (parallel to the Y axis), should be the same wherever you are on the line or curve. It's difficult to tell when the points lie close to the line or curve, so you get a much better idea by examining the difference between the observed and the predicted values of the criterion for each subject. These differences are known as the residuals, and it's usual to plot the residuals against predicteds values. I have provided such a plot on the spreadsheet, or click here to see a plot from a later section of this text. If subjects in one part of the plot have a bigger scatter, they have a bigger typical error (because the standard deviation of the residuals is the typical error). The calculated typical error of the estimate then represents some kind of average variation for all the subjects, but it will be too large for some subjects and too small for others.
To get an estimate of the typical error that applies accurately to all subjects, you have to find some way to transform the criterion and practical measures to make the scatter of residuals for the transformed measures uniform. Once again, logarithmic transformation often reduces nonuniformity of the scatter in situations where there is clearly more variation about the line for larger values of the criterion. A uniform scatter of the residuals after log transformation implies that the typical error, when expressed as a percent of the criterion value, is the same for all subjects; the typical error from analysis of the logtransformed measures then gives the best estimate of its magnitude. I have included an analysis of the logtransformed measures in the spreadsheet, although for the data therein it is clear that the scatter of residuals is more uniform for the raw measures than for the logtransformed measures.
If you fit a curve rather than a straight line to your data, the standard deviation of the residuals (the root mean square error) still represents the typical error in the estimate of the criterion value for a given practical value. To estimate the typical error from the spreadsheet might be too difficult, though, because you will have to modify the predicted values according to the type of curve you used. It may be easier to use a stats program. The typical error in the output from the stats program will be labeled either as the SEE, the root meansquare error, or the residual error. Some stats programs provide the typical error as a variance, in which case you will have to take the square root.
When you have subgroups of subjects with different characteristics (e.g., males and females), don't forget to check whether the subgroups have similar typical errors. To do so, you should label the points for each subgroup in the plot of residuals vs predicteds, because what looks like a uniform scatter might conceal a big difference between the subgroups. If there is a big difference, you shouldn't use a composite estimation equation for the two groups; instead, you should derive separate equations and separate typical errors for each subgroup.
Validity Limits of
Agreement
By analogy with reliability
limits of agreement, we can define validity limits of
agreement as the 95% likely range or reference range for the
difference between a subject's values for the criterion and practical
measures. Let's try to understand this concept using the data in the
validity spreadsheet.
The data are from a validity study in which the practical measure was body fat estimated using a Bod Pod, and the criterion measure was body fat measured with a DEXA scan. The units of body fat are percent of body mass (%BM). The limits of agreement (not shown in the spreadsheet) are 2.9 to 7.9 %BM, or 2.5 ± 5.4 %BM. You can interpret these numbers in two ways: there's a 95% chance that a subject's "true" (DEXA) body fat is within 2.5 ± 5.4 %BF of his or her Bod Pod value; or, if you measured a large number of subjects in the Bod Pod, 95% of them would have a DEXA body fat within 2.5 ± 5.4 %BF of their Bod Pod values. The value 2.5 in this example is the mean of the criterionpractical difference (or the difference between the means of the criterion and practical measures); it is sometimes known as the bias in the practical measure, but don't confuse this concept with the smallsample bias I described in connection with measures of reliability. The value ±5.4 on its own is usually referred to as the limits of agreement; it is ±2.0x the standard deviation of the criterionpractical difference (= 2.7). The standard deviation of the criterionpractical difference is itself known as the pure error or total error.
Limits of agreement are related to the typical error of the estimate. When the slope of the estimation equation is exactly 1, the pure error is the same as the typical error, so in this special case the limits of agreement are twice the typical error. If the slope differs from 1, the limits of agreement are greater than twice the typical error. If the calibration equation is a curve rather than a straight line, the limits of agreement will also be greater than twice the typical error.
Advocates of limits of agreement encourage authors to plot the criterionpractical differences against the mean of these measures (or against the criterion). The resulting plot is similar to a plot of the residuals against the predicteds from the analysis of the estimation equation: if the estimation equation is a straight line of slope close to 1, the criterionpractical differences are the same as the residuals, and the mean of the criterion and practical is near enough to the predicted value. The plot will therefore allow you to check for heteroscedasticity. If the calibration equation is a straight line with slope different from 1, or if it is a curve, the scatter of points in the plot of the criterionpractical differences will show a trend towards a straight line or a curve, so it will be harder to tell if heteroscedasticity is present.
Validity limits of agreement suffer from problems similar to those of reliability limits of agreement: they are harder to understand than the typical error, and they are too large as a reference range for making a decision about a subject's true (criterion) measurement. The fact that the nature of the estimation equation affects the magnitude of the limits is also a serious problem. Unfortunately some authors have published limits of agreement without an estimation equation or the typical error, so readers cannot properly assess the practical measure and the published data cannot be used to recalibrate the practical measure.
When it comes to calculating the validity correlation, you don't have much choice: if you fit a straight line to the data, the correlation is a Pearson correlation coefficientthere is no equivalent intraclass correlation coefficient. If you fit a curve, the stats program should provide you with a goodnessoffit statistic called the variance explained or the Rsquared. Just take the square root of this statistic and you have the equivalent of the Pearson correlation coefficient.
An estimate of validity correlation can also be obtained by taking the square root of the concurrent reliability correlation. By concurrent reliability I mean the immediate retest reliability, rather than the retest reliability over the time frame of any experiment you may be planning. This relationship between validity and reliability comes about because reliability is the correlation of something with itself (and there is error in both measurements), whereas validity is something correlated with the real thing (so there is error in only one measurement). The relationship can be derived from the definition of correlation (covariance divided by product of standard deviations) applied to the validity and reliability correlations.
The square root of concurrent reliability represents the maximum possible value for validity. The actual validity correlation could be less, because a measure can have high reliability and low validity. To put it another way, a measure can produce nonsense consistently!
Validity can be difficult to measure, because the true value of something can be difficult to assay. Measures other than the true value are called surrogates. These measures usually result in underestimates of validity when they are correlated with observed values, for obvious (I hope) reasons. Here's an example. Body density obtained by underwater weighing is often referred to as the gold standard for estimating percent body fat, but it is only a surrogate for true percent body fat. So if you are validating a skinfold estimate of body fat against the value obtained by underwater weighing, the validity correlation will be lower than if you validated the skinfold estimate against a more accurate method than underwater weighing, for example, a DEXA scan. Similarly the typical error of the estimate will be smaller when you validate skinfolds against DEXA rather than underwater weighing.
The spreadsheet estimates the calibration equation and the following measures of validity: typical error of the estimate, newprediction error, correlation coefficient, and limits of agreement (but don't use them!). Analysis of logtransformed data is included for estimation of errors as percents of the mean.

