Signal Detection Theory and Z-scores — Study Notes

Z-scores, standard deviation, and why they matter

• Z-score intuition: a z-score tells you how far an observation is from the mean in units of standard deviation. It’s the ratio of the deviation to the data’s variability.
  • Concept: for any data point x, with distribution mean μ and standard deviation σ, the z-score is $z = \frac{x - \mu}{\sigma}$.
  • When we convert measurement differences into z-scores, units cancel out, giving a unitless, comparable measure.
• Why standard deviation is fundamental:
  • It quantifies how much your data vary trial-to-trial.
  • A big difference between conditions is more meaningful if variability (σ) is small; a big σ can make the same difference look trivial.
  • Statistics often compare an effect size (difference) to variability (noise) to judge meaningfulness.
• Practical note from the lecture:
  • You don’t necessarily need to calculate σ on an exam, but you should understand what σ (the spread) represents and why a standard deviation matters conceptually.
  • Z-scores let you compare effects across different measurement units (milliseconds, volts, etc.).

Introduction to Signal Detection Theory (SDT)

• Core idea:
  • There are real physical events (the stimulus) and perceptual decisions our noisy nervous system makes about them.
  • Perception is noisy; we cannot measure the stimulus directly in the brain, only via behavior.
• Reality vs. perception:
  • Reality: the target is either present or absent (binary).
  • Perception: we may perceive it or not, with errors due to internal noise and external conditions.
• The normal curve premise:
  • When information from a stimulus is processed by the brain, the resulting perceptual strength is assumed to be normally distributed across trials.
  • Noise alone gives a distribution; the presence of a signal shifts that distribution upward (signal+noise).
• Two distributions:
  • Noise distribution: represents perceptual strength with no target.
  • Signal+Noise distribution: represents perceptual strength with the target present.
• The four possible outcomes (yes/no task):
  • Hit: target present and reported present.
  • Miss: target present but reported absent.
  • False alarm: target absent but reported present.
  • Correct rejection: target absent and reported absent.
• Decision criterion (the “threshold”):
  • A single criterion (threshold) is used to decide between “target present” and “target absent” on each trial.
  • If perceptual strength exceeds the criterion, respond “present”; otherwise respond “absent.”
• Why this matters:
  • Different people can have the same ability to distinguish signals (same info extraction) but different response biases (tendency to say yes or no).
  • SDT provides metrics that separate perceptual sensitivity from response bias.

Yes/No task and SDT geometry

• Target present vs absent trials:
  • Noise-only trials (absent trials) generate False Alarms and Correct Rejections.
  • Target-present trials generate Hits and Misses.
• Visual intuition (two overlapping distributions):
  • Noise distribution sits below, signal+noise sits shifted to the right (toward higher perceptual strength).
  • The decision criterion sits somewhere along the information axis; moving it changes the balance of Hits and False Alarms without changing the underlying distributions.
• What changes with bias:
  • A liberal bias (lower criterion) increases Hits but also increases False Alarms.
  • A conservative bias (higher criterion) reduces False Alarms but also reduces Hits.
• Practical takeaway:
  • Accuracy alone mixes sensitivity and bias; SDT aims to separate these components.

Two key SDT metrics: d′ and criterion c

• Hit rate and False Alarm rate:
  • Hit rate: H = \frac{\text{Hits}}{\text{N_present}}
  • False alarm rate: FA = \frac{\text{False Alarms}}{\text{N_absent}}
• Z-transform of rates (assuming normal distributions):
  • $z(H) = \text{the z-score corresponding to the cumulative probability } H$
  • $z(FA) = \text{the z-score corresponding to the cumulative probability } FA$
• d′ (d-prime): perceptual sensitivity (distance between the two distributions in SD units)
  • Definition: $d' = z(H) - z(FA)$
  • Interpretation: larger d′ means greater separation between noise and signal+noise distributions; better perceptual discrimination.
• Criterion c (response bias):
  • Definition (common convention): $c = -\tfrac{1}{2} \big( z(H) + z(FA) \big)$
  • Interpretation: negative c = liberal bias; positive c = conservative bias; zero = unbiased (balanced) criterion.
• Relationship between d′ and c:
  • d′ measures perceptual sensitivity independent of bias.
  • c captures the decision bias (where the criterion lies relative to the two distributions).
• How to interpret a fixed d′ when bias changes:
  • If you shift the criterion (bias) but keep the same underlying sensitivity, d′ remains the same.
  • Accuracy can go up or down with bias even if d′ stays constant.

Worked examples (yes/no task)

• Example 1 (200 trials: 100 present, 100 absent):
  • Hits = 80; False Alarms = 35
  • Hit rate: $H = 0.80$; FA rate: $FA = 0.35$
  • Compute z-scores: $z(H) = \text{Φ}^{-1}(0.80) \approx 0.842$, $z(FA) = \text{Φ}^{-1}(0.35) \approx -0.385$
  • d′: $d' = z(H) - z(FA) \approx 0.842 - (-0.385) \approx 1.23$
  • c: $c = -\tfrac{1}{2} [ z(H) + z(FA) ] \approx -\tfrac{1}{2} (0.842 - 0.385) \approx -0.23$
  • Interpretation: d′ ≈ 1.23 (moderate-to-good sensitivity); c ≈ -0.23 (liberal bias: more willing to say “present”).
• Example 2 (200 trials: 100 present, 100 absent):
  • Hits = 75; False Alarms = 22
  • Hit rate: $H = 0.75$; FA rate: $FA = 0.22$
  • Compute z-scores: $z(H) = \text{Φ}^{-1}(0.75) \approx 0.674$, $z(FA) = \text{Φ}^{-1}(0.22) \approx -0.772$
  • d′: $d' = z(H) - z(FA) \approx 0.674 - (-0.772) \approx 1.45$
  • c: $c = -\tfrac{1}{2} [ z(H) + z(FA) ] \approx -\tfrac{1}{2} (0.674 - 0.772) \approx 0.049$
  • Interpretation: d′ ≈ 1.45 (greater sensitivity than Example 1); c ≈ 0.05 (slightly conservative bias).
• Key takeaway from the examples:
  • d′ values reflect perceptual information strength; higher d′ means better discrimination.
  • c reflects bias toward saying “present” or “absent.”
  • It is possible for a person to have higher accuracy with a stronger bias that makes more hits but also more false alarms; d′ is not affected by this bias.

Edge cases and practical adjustments

• When hit rate or false alarm rate is 0% or 100%:
  • Directly converting 0% or 100% to z-scores is problematic (they map to ±∞).
  • Common correction: add a small imaginary/trial correction by adding 0.5 to each cell in the 2x2 table (present/absent × target/response), effectively making Npresent and Nabsent incremented by 1 each and adjusting both H and FA slightly.
  • Rationale: prevents infinite z-scores and yields a finite d′; should be applied to all participants consistently and planned before data collection.
• Negative d′ (theoretical only):
  • d′ < 0 would imply you are systematically responding in the wrong direction (confusing noise for signal more often than signal for noise) in literal terms.
  • In normal perception tasks, d′ < 0 is rarely meaningful unless describing an illusion or reversed task; typically, we interpret d′ as >= 0.
• Alternative measures when variances differ: d′ assumes equal-variance normal distributions.
  • When this assumption is violated, an approach called d′-of-a (d′a) or ROC-based methods (e.g., d′ with unequal variances) can be used.
  • d′a comes from fitting an ROC curve with rating data and can handle unequal variances between noise and signal+noise distributions.

Rating scales and ROC methods

• Beyond binary yes/no responses, you can collect confidence or multiple response categories (e.g., definitely present, probably present, guess present, guess absent, probably absent, definitely absent).
  • Each category yields a different hit and false alarm rate, enabling multiple z-scores and an ROC curve with more points.
  • You can compute d′ using these points and fit a line (ROC) to summarize sensitivity.
• Benefits of rating-based SDT:
  • Provides richer data (more than a single hits/FA pair).
  • Allows modeling of unequal variances (via d′a) and a more nuanced view of decision processes.
• Practical example from research:
  • A memory strength study used rating-scale SDT to examine how memory traces differ under conditions; d′a was used to account for unequal variances between memory strengths.

Applications and extensions of SDT concepts

• Bias-free memory and perception:
  • d′ is a measure of perceptual or memory sensitivity independent of bias (criterion c).
  • Higher d′ implies stronger discriminability or memory strength, regardless of where the respondent tends to say “present.”
• Real-world and research examples mentioned in the lecture:
  • Social psychology and stereotypes: a yes/no task using names (e.g., NBA player names) to study bias and perception; a bias shift can manifest as changes in d′ or in criterion depending on context.
  • Stereotype bias can lead to a shifted criterion (pseudo-d′), illustrating that bias can affect response tendencies even when underlying sensitivity is constant.
• How SDT connects to broader science and practice:
  • Provides a principled framework for separating perceptual/memory strength from decision criteria.
  • Helps interpret performance changes due to environment (noise) vs. instruction/goal (criterion shifts).
  • Useful across domains: perception, memory, psychometrics, clinical decision-making, and even weather/event detection.

Quick recap and takeaways

• Z-scores convert raw performance into unitless measures that reflect how far observed signals are from noise, in SD units.
• Signal Detection Theory decomposes performance into two components:
  • Sensitivity (d′): how well you can distinguish signal from noise.
  • Bias/criterion (c): your default tendency to say “present” vs. “absent.”
• Key formulas:
  • Hit rate: H = \frac{\text{Hits}}{\text{N_present}}
  • False alarm rate: FA = \frac{\text{False Alarms}}{\text{N_absent}}
  • $d' = z(H) - z(FA)$
  • $c = -\tfrac{1}{2} \big( z(H) + z(FA) \big)$
• Examples illustrate how d′ and c can diverge: higher d′ means better discrimination, while c reflects liberal vs. conservative response style.
• Practical data issues: zero/one rate corrections, potential unequal variances (d′a), and the use of rating scales to build ROC curves.
• SDT concepts extend beyond simple perception tasks to memory strength, stereotypes, and many decision-making contexts.