Why Compare Residuals at the Same Concentration?
A direct concentration-effect model assigns the same expected response to the same concentration, after accounting for modeled covariates and other included predictors. That assumption can fail when the response follows changes in exposure with a delay. At a given concentration, the observed response may then differ depending on whether concentration is rising or falling. Comparing residuals in those two directions is a targeted way to look for this pattern.
This comparison is informative only when there is meaningful overlap: observations on both the rising and falling parts of the concentration profile must occur at similar concentrations. A difference between residuals far apart in concentration could simply reflect an incorrect concentration-effect shape. Here, the diagnostic question is specifically whether direction still predicts residuals after concentration is matched.
A positive residual means the observed response is greater than predicted; a negative residual means it is less than predicted. Whether rising-limb residuals should be positive or negative depends on the direction of the modeled concentration effect. The more general signal is a systematic difference between the two limbs, not a universal sign pattern.
Mechanism and Expected Residual Pattern
Consider a direct model whose expected response is \(f(C)\), where \(C\) is measured concentration. Suppose instead that the response follows an unobserved effect-site concentration \(C_e\), which approaches measured concentration gradually:
When concentration is increasing smoothly, \(C_e\) tends to be below \(C\); when concentration is decreasing, \(C_e\) tends to be above \(C\). If the response follows \(f(C_e)\), the direct model predicts \(f(C)\), so the expected residual is approximately \(f(C_e)-f(C)\). For a differentiable response function and a modest lag, a first-order approximation explains the directional pattern.
Thus, for an increasing response function, rising concentrations tend to produce negative residuals and falling concentrations positive residuals. For a decreasing response function, the expected signs reverse. This is a diagnostic justification, not a guarantee: noise, other sources of time dependence, or model misspecification can also create directional patterns.
The approximation assumes a locally smooth concentration trajectory, a positive equilibration rate, and a response function that is locally differentiable. The practical comparison does not require estimating this equation. It asks whether residuals retain a reproducible direction-dependent pattern after accounting for concentration and relevant covariates.
How to Construct the Comparison
Use the same residual definition and prediction scale throughout. If residual variability changes across the response range, inspect appropriately standardized residuals as well as raw residuals.
Use the concentration-time profile to label observations as rising or falling. Avoid labeling a point at a peak or trough as either direction when the slope is effectively zero or uncertain.
Compare observations within a narrow concentration range, preferably also comparable in baseline, dose occasion, and other predictors that affect the response. Do not extrapolate one limb to concentration ranges observed only on the other.
Plot residuals against concentration with rising and falling observations distinguished. Look for consistent separation across the shared concentration range, not just one large pair or an apparent difference driven by one subject.
When observations can be paired meaningfully, define \(D_k=e_{F,k}-e_{R,k}\), where \(e_{F,k}\) and \(e_{R,k}\) are falling- and rising-limb residuals for matched pair \(k\). A positive mean \(D_k\) indicates higher residuals on the falling limb. If pairs are independent and their differences are approximately normally distributed, a paired \(t\) interval is
where \(\bar{D}\) and \(s_D\) are the sample mean and standard deviation of the \(n\) paired differences. If several pairs come from the same subject, they are generally not independent. In that case, do not treat every pair as an independent replicate; summarize within subject or use an analysis that accounts for within-subject dependence.
Worked Example: Paired Residuals on Rising and Falling Limbs
In a hypothetical analysis, six subjects have rising- and falling-limb observations matched at approximately the same concentration. Residuals are measured in biomarker units, with observed minus predicted as the sign convention.
| Subject | Rising residual | Falling residual | Difference \(D=e_F-e_R\) |
|---|---|---|---|
| 1 | -0.5 | 0.7 | 1.2 |
| 2 | -0.3 | 0.5 | 0.8 |
| 3 | -0.4 | 0.6 | 1.0 |
| 4 | -0.2 | 0.4 | 0.6 |
| 5 | -0.6 | 0.5 | 1.1 |
| 6 | -0.1 | 0.8 | 0.9 |
The paired differences sum to \(5.6\), so \(\bar{D}=5.6/6=0.9333\) biomarker units. Their squared deviations from the mean sum to \(0.2333\), giving \(s_D=\sqrt{0.2333/5}=0.2160\) units and \(SE=s_D/\sqrt{6}=0.08819\) units. With \(df=5\), the two-sided 95% critical value is approximately \(2.571\).
The interval is \(0.9333\pm2.571(0.08819)=[0.7066,1.160]\) biomarker units. Under the paired \(t\) assumptions, the data support a positive mean directional contrast for these matched observations. The rising residuals are negative and the falling residuals positive, a pattern consistent with a direct model that predicts the response too high on the rising limb and too low on the falling limb. This supports investigating lag or hysteresis; it does not by itself establish a particular mechanism.
The pairing is appropriate here because each difference compares observations from the same subject at similar concentrations. The interval additionally assumes the six subject-level differences are independent across subjects and approximately normally distributed. With only six pairs, the distributional assumption deserves particular attention; a plot of the individual differences is important, and the numerical interval should not be treated as robust to severe skewness or an influential pair.
Read the Pattern in the Context of the Response
The direction of the contrast must be interpreted relative to the response scale and the sign of the concentration-effect relationship. For a response that increases with concentration, lag commonly yields negative rising-limb residuals and positive falling-limb residuals. For an inhibitory endpoint that decreases with concentration, the expected signs can be opposite.
Worked Example: An Inhibitory Biomarker
Suppose a hypothetical direct model predicts a biomarker value of 40 units at a concentration of 5 mg/L. The concentration-effect relationship is inhibitory: higher concentrations predict lower biomarker values. At approximately 5 mg/L, the observed rising-limb value is 46 units and the observed falling-limb value is 34 units.
Using \(e=Y-\hat{Y}\), the rising-limb residual is \(46-40=+6\) biomarker units. The falling-limb residual is \(34-40=-6\) units. Therefore, \(D=e_F-e_R=-6-(+6)=-12\) biomarker units. The negative contrast is not evidence against a lag pattern: for this decreasing response function, delayed inhibition predicts a higher-than-predicted biomarker on the rising limb and a lower-than-predicted biomarker on the falling limb. The key observation is the systematic directional separation at comparable concentrations.
This interpretation depends on the stated inhibitory relationship and on the observations being genuinely comparable in concentration and relevant covariates. A single pair is illustrative only; it cannot establish that the pattern is reproducible.
Diagnostics, Alternatives, and Failure Modes
The most useful visual diagnostic plots residuals against concentration, distinguishing rising and falling observations by color or symbol. A smooth trend for each limb can show whether the separation is localized to a particular concentration range. A plot of matched differences against their matched concentration can reveal whether the contrast changes with exposure. The earlier material on residuals versus concentration and residuals versus time provides complementary checks: a directional plot should not be interpreted in isolation.
Evidence for a lag-related problem is stronger when the limbs overlap in concentration, the separation is replicated across subjects or occasions, and its sign agrees with the response scale. Evidence is weaker when the apparent difference depends on a few observations, occurs only where there is little overlap, or disappears after accounting for a known covariate or time trend.
Several alternatives answer related but different questions. A residual-versus-time plot can reveal an overall temporal pattern but may not show whether residuals differ at the same concentration. Residuals-versus-concentration plots can reveal a misspecified exposure-response shape but may conceal a direction-dependent difference if rising and falling observations are pooled. A formal comparison of model structures may quantify support for competing models, but it depends on specifying those models and is not a substitute for checking whether the observed residual signature is present. This diagnostic is a focused visual and matched-data check, not a model-selection procedure.
Common failure modes include the following:
- Comparing unmatched concentrations. A larger residual on one limb may reflect an incorrect concentration-effect shape. Compare within a shared concentration range before attributing a difference to direction.
- Using one fixed sign rule. “Rising must be negative” is wrong for a decreasing response function. Derive the expected sign from the response scale and the direction of \(f(C)\).
- Ignoring other time structure. Disease progression, baseline drift, tolerance, rebound, or placebo structure can create time-related residuals. Check whether these explanations are plausible before attributing a pattern to exposure-response lag.
- Overinterpreting a noisy pair. A single large difference can arise by chance. Examine replication, subject-level patterns, and uncertainty rather than relying on one matched comparison.
- Trusting direction labels near turning points. When concentration changes little, the direction can be unstable or dominated by measurement error. Treat such observations cautiously or exclude them from the directional contrast using a prespecified rule.
Worked Example: When the Apparent Contrast Is Not Informative
In a hypothetical data set, rising-limb observations occur only at concentrations from 1 to 3 mg/L, while falling-limb observations occur only from 6 to 8 mg/L. The mean residuals are \(-0.4\) units on the rising limb and \(+0.5\) units on the falling limb, giving a raw difference of \(0.5-(-0.4)=0.9\) units.
That arithmetic is correct, but it is not a matched directional contrast: there are no comparable concentrations on the two limbs. The difference could be due to a concentration-dependent shape error, because the model may underpredict at 6–8 mg/L and overpredict at 1–3 mg/L regardless of direction. The data do not identify a rising-versus-falling difference separately from a concentration effect. The appropriate conclusion is that this plot cannot diagnose hysteresis over the observed ranges; additional sampling that produces overlapping concentrations on both limbs, or another design that separates concentration from direction, is needed.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A response increases with concentration. At a matched concentration, the mean rising-limb residual is \(-0.7\) units and the mean falling-limb residual is \(+0.4\) units. Calculate the falling-minus-rising contrast and interpret its direction.
- For an inhibitory response, why can rising-limb residuals be positive under exposure-response lag?
- Two subjects each contribute ten matched pairs, but an analyst treats all 20 pairs as independent in a paired \(t\) interval. Identify the concern.
- Rising and falling observations have no overlapping concentration range. Can their residual mean difference isolate a directional effect? Explain.
Solutions. (1) The contrast is \(0.4-(-0.7)=1.1\) response units. Falling-limb residuals are higher; for an increasing response, this is consistent with the expected lag signature, provided other predictors and concentration are comparable. (2) With an inhibitory response, rising exposure has not yet produced the full predicted inhibition, so the observed biomarker can be higher than the direct model prediction; observed minus predicted is positive. (3) Pairs from the same subject are dependent, so treating them as 20 independent replicates can understate uncertainty. Analyze subject-level summaries or use a method that accounts for within-subject dependence. (4) No. The limbs are confounded with concentration, so a difference could reflect concentration-effect shape error rather than direction. A comparison requires overlap or other information that separates concentration from direction.