Why Model a Fraction?
Many outcomes in applied work are measured as shares, rates, or proportions rather than counts or unrestricted quantities. Examples include the fraction of a hospital’s beds occupied, the share of a firm’s sales made abroad, and the proportion of a patient’s observed days on which a treatment was taken. These outcomes have a natural lower and upper bound. A model for their expected value should respect those bounds, while remaining clear about what it assumes—and what it leaves unspecified.
The central object in this tutorial is the conditional mean of a fractional outcome. Modeling that mean is not the same as choosing a probability distribution for every possible outcome value. This distinction matters when selecting a model and when interpreting its predictions.
What the Fraction Measures
A fractional outcome should correspond to a defined numerator and denominator. For an observation \(i\), one common construction is \(Y_i=A_i/B_i\), where \(A_i\) is the part of interest and \(B_i\) is the relevant whole, with \(0\leq A_i\leq B_i\) and \(B_i>0\). The denominator must be meaningful for the question. For example, a hospital’s occupied-bed fraction uses occupied beds divided by available beds; the ratio is not defined when the number of available beds is zero.
The fraction is recorded on a zero-to-one scale: \(0.25\) is a quarter, or 25 percent. Multiplying by 100 changes the reporting scale, not the underlying quantity. A model of the fraction on the zero-to-one scale has predictions on that scale; multiply a predicted fraction by 100 only when expressing it as a percentage.
A fractional outcome need not be a binary indicator. A binary outcome takes only the values zero and one, whereas a fraction may take many intermediate values. Nor should every ratio of counts automatically be treated as the same kind of outcome: the numerator, denominator, and observational unit determine what the response means.
Worked Example: Constructing an Occupancy Fraction
A hypothetical clinic records the number of occupied rooms and the number of available rooms at closing on three days. Define the fractional outcome for each day and calculate it.
| Day | Occupied rooms | Available rooms | Fraction occupied |
|---|---|---|---|
| Monday | 18 | 30 | 0.600 |
| Tuesday | 21 | 28 | 0.750 |
| Wednesday | 16 | 32 | 0.500 |
For Monday, \(Y=18/30=0.600\); for Tuesday, \(Y=21/28=0.750\); and for Wednesday, \(Y=16/32=0.500\). Each value lies between zero and one and is interpretable as the fraction of available rooms occupied. These are three daily observations, not 90 independent room-level outcomes: the observational unit is the day.
The unweighted average of the three daily fractions is \((0.600+0.750+0.500)/3=1.850/3=0.617\), or about 61.7 percent. The pooled fraction is different: \((18+21+16)/(30+28+32)=55/90=0.611\), or about 61.1 percent. The first gives each day equal weight; the second gives more weight to days with more available rooms. Which summary is appropriate depends on the target being described.
The Conditional Mean and Its Bounds
Let \(Y\) be a fractional response and \(X\) a vector of observed covariates. The function \(m(x)=E[Y\mid X=x]\) describes the average fractional outcome at covariate value \(x\). It is a conditional expectation, not a promise that every observation with those covariates takes the value \(m(x)\). Individual outcomes may vary around that mean.
The bounds on the outcome imply bounds on its conditional mean. This follows from the order-preserving property of conditional expectation: if one random variable is no smaller than another, their conditional expectations preserve that inequality.
Thus any proposed conditional mean function must predict values in the unit interval for the covariate values where the model is intended to apply. This condition is a consequence of the outcome’s support, not an optional preference about prediction. If the conditional mean is exactly zero, a nonnegative outcome must be zero with conditional probability one; if it is exactly one, an outcome no greater than one must be one with conditional probability one. The mean may therefore be at a boundary only when the conditional outcome is concentrated there.
Why an Unrestricted Linear Mean Can Fail
A linear specification \(m(x)=\beta_0+\beta_1x\) is simple to interpret, but without restrictions it can predict values below zero or above one. Such a prediction cannot be the conditional mean of a variable known to lie in \([0,1]\). A linear mean can be appropriate over a limited covariate range if its predictions stay within the bounds there; it is not automatically valid over the full real line.
Worked Example: Checking a Linear Mean’s Range
Suppose a hypothetical analyst proposes \(m(x)=0.12+0.09x\), where \(x\) is a workload index. Check the predictions at \(x=0\), \(x=4\), and \(x=10\).
At \(x=0\), \(m(0)=0.12+0.09(0)=0.12\), a possible mean fraction. At \(x=4\), \(m(4)=0.12+0.09(4)=0.12+0.36=0.48\), also within the unit interval. But at \(x=10\), \(m(10)=0.12+0.09(10)=0.12+0.90=1.02\), which exceeds one.
The proposed line is therefore not a valid conditional mean over a covariate range that includes \(x=10\). This does not by itself show that the observed data are impossible; it shows that this unrestricted linear function is incompatible with the outcome bounds at that covariate value. A different specification or a clearly justified narrower domain would be needed.
A bounded conditional mean can instead be specified using a function whose range is restricted to \([0,1]\), or by imposing parameter restrictions that keep the proposed function in that interval over the relevant covariate support. In either case, the support check is part of model formulation. It should be carried out for the covariate values the model may encounter, not only for the sample average.
Worked Example: A Mean Function Bounded on Its Domain
Suppose a hypothetical process has a covariate \(x\) known to lie between zero and one, and the proposed conditional mean is \(m(x)=0.15+0.60x\). Verify the range and calculate the mean at \(x=0.40\).
Because \(0\leq x\leq 1\), the smallest value is \(m(0)=0.15+0.60(0)=0.15\), and the largest is \(m(1)=0.15+0.60(1)=0.75\). The function is increasing, so every value between those endpoints lies in \([0.15,0.75]\), which is contained in \([0,1]\). At \(x=0.40\), \(m(0.40)=0.15+0.60(0.40)=0.15+0.24=0.39\).
The model’s prediction at \(x=0.40\) is a conditional mean fraction of 0.39, or 39 percent. It does not say that each observation at that covariate value equals 0.39. The boundedness check also depends on the stated domain: if values of \(x\) outside \([0,1]\) were possible, the function would need to be checked there as well.
A Mean Model Is Not a Full Distribution Model
A conditional mean specification describes one feature of the conditional distribution of \(Y\) given \(X\). It does not, by itself, specify the probabilities of individual outcomes, the conditional variance, or the overall shape of that distribution. Distinct distributions can have the same mean.
For example, at a fixed covariate value, one possible outcome is always \(Y=0.40\), giving \(E[Y\mid X=x]=0.40\). Another is \(Y=0\) with probability one-half and \(Y=0.80\) with probability one-half, which also gives \(E[Y\mid X=x]=0.5(0)+0.5(0.80)=0.40\). Both distributions satisfy the fractional bounds and share the same conditional mean, but their variation and probability of observing zero differ.
This distinction is important when a question concerns the expected fraction rather than a complete probability model for individual outcomes. A model for the conditional mean can answer questions about average response across covariate values without claiming to describe every feature of the outcome distribution. Conversely, conclusions about the chance of a particular outcome or about its conditional variance require information beyond the mean alone.
Worked Example: Same Conditional Mean, Different Outcome Variation
Consider two hypothetical groups with the same covariate value. In Group A, each outcome equals 0.40. In Group B, half the outcomes equal 0 and half equal 0.80. Compare their conditional means and the probability of a zero.
For Group A, the conditional mean is \(E[Y\mid X=x]=0.40\), and the probability of a zero is 0. For Group B, the conditional mean is \(0.5(0)+0.5(0.80)=0.40\), while the probability of a zero is 0.5. The means agree, but the distributions—and thus the probability of a zero—do not.
A conditional mean model that predicts 0.40 for both groups captures their shared average response. It cannot distinguish these two distributions without additional assumptions or modeling choices about variation and outcome probabilities.
Common Mistakes and Practical Checks
A frequent mistake is to treat a fitted value of 0.39 as a predicted value for every individual, or to infer the probability of a particular outcome from a conditional mean. The correction is to state exactly what the model gives: an expected fraction conditional on the specified covariates. A mean is not an individual outcome or a complete distribution.
Another mistake is to assume that any regression equation is acceptable because the observed outcomes lie in the unit interval. The response values being bounded does not ensure that a fitted function is bounded. Check the model’s range over the covariate domain, including values used for prediction. Also verify that each ratio has a well-defined denominator and that its observational unit matches the intended question.
These checks separate three decisions: how the fraction is constructed, which conditional mean is being modeled, and whether the model makes claims about more than that mean. Keeping them distinct makes the target of a fractional-response analysis easier to interpret and its assumptions easier to assess.
Check Your Understanding
Answer each question by identifying the relevant outcome, mean, or modeling claim.
- A department records 12 treated cases out of 20 eligible cases. What is the fractional outcome, and what does its value represent?
- Why must \(E[Y\mid X=x]\) lie between zero and one when \(Y\) lies between zero and one?
- A proposed linear mean predicts 1.08 for a covariate value in the intended prediction range. What does this imply about the specification at that value?
- Two groups have the same conditional mean fraction but different probabilities of a zero. Is that possible? Explain.
- When would a pooled ratio and the unweighted average of individual fractions give different answers, and what does each weight equally?