Package 'cdCAT'

Description

Filters the candidate item pool so that the next item is drawn from the content domain most under-represented relative to the target blueprint (Kingsbury & Zara, 1991).

Usage

apply_content_balancing(candidate_items, administered, content, content_prop)

Arguments

Details

The gap for each domain is defined as:

$\text{gap}_d = \text{prop\_target}_d - \text{prop\_observed}_d$

The domain with the largest gap is selected, and only items from that domain are returned. If no candidate items belong to the target domain, the full candidate pool is returned as a safe fallback (no items are left un-selectable).

Value

Integer vector of candidate item indices, restricted to the most under-represented domain (or the full candidate_items if no balancing is needed or no candidates belong to that domain).

References

Kingsbury, G. G., & Zara, A. R. (1991). A comparison of procedures for content-sensitive item selection in computerized adaptive testing. Applied Measurement in Education, 4(3), 241–261. doi:10.1207/s15324818ame0403_4

See Also

select_next_item(), CdcatSession

Examples

content      <- c("A", "A", "A", "B", "B", "B")
content_prop <- c(A = 0.5, B = 0.5)

# No items administered yet -- returns all candidates unchanged
apply_content_balancing(1:6, integer(0), content, content_prop)

# After two "A" items: gap favours "B"
apply_content_balancing(3:6, c(1L, 2L), content, content_prop)

Description

Selects the next item from the candidate pool according to the exposure control regime encoded in exposure.

Usage

apply_exposure_control(scores, available, exposure, n_administered)

Arguments

Details

Two regimes are supported and detected automatically from the values of exposure:

Sympson-Hetter (Sympson & Hetter, 1985) – triggered when all values in exposure are in $[0, 1]$. Each item has an acceptance probability equal to its exposure value. Items are visited in descending score order; the first one that passes the random draw is selected. Values near 1 are almost always selected; values near 0 are rarely selected.

Randomesque – triggered when all values in exposure are $\geq 1$. At position $k$ in the test (i.e. when selecting the $k$-th item), the top exposure[k] candidates by score are collected and one is drawn at random. exposure[k] = 1 is equivalent to greedy selection.

When exposure is NULL, greedy selection (which.max(scores)) is used, equivalent to no exposure control.

Value

Integer scalar – global index of the selected item.

References

Sympson, J. B., & Hetter, R. D. (1985). Controlling item-exposure rates in computerized adaptive testing. Proceedings of the 27th annual meeting of the Military Testing Association (pp. 973–977).

See Also

apply_sympson_hetter(), apply_randomesque(), select_next_item(), CdcatSession

Examples

scores    <- c(0.8, 0.5, 0.3, 0.9)
available <- c(2L, 4L, 6L, 8L)

# Greedy (no exposure control)
apply_exposure_control(scores, available, exposure = NULL, n_administered = 0L)

# Sympson-Hetter: item 8 (score 0.9) has probability 0.2 of being accepted
set.seed(1)
exposure_sh <- rep(0.8, 10)
exposure_sh[8] <- 0.2
apply_exposure_control(scores, available, exposure_sh, n_administered = 2L)

# Randomesque: at position 3, draw from top-2 candidates
exposure_rq <- rep(1, 10)
exposure_rq[3] <- 2
set.seed(1)
apply_exposure_control(scores, available, exposure_rq, n_administered = 2L)

Description

Selects an item by drawing uniformly at random from the top-n candidates by score. When n = 1 the result is identical to greedy (deterministic) selection.

Usage

apply_randomesque(scores, available, n)

Arguments

Value

Integer scalar – global index of the selected item.

See Also

apply_exposure_control()

Description

Selects an item by iterating over candidates in descending score order and accepting the first one that passes a Bernoulli draw with probability p[item] (Sympson & Hetter, 1985).

Usage

apply_sympson_hetter(scores, available, p)

Arguments

Value

Integer scalar – global index of the selected item.

References

Sympson, J. B., & Hetter, R. D. (1985). Controlling item-exposure rates in computerized adaptive testing. Proceedings of the 27th annual meeting of the Military Testing Association (pp. 973–977).

See Also

apply_exposure_control()

Description

Generates all 2^K binary skill profiles for K attributes.

Usage

build_skill_patterns(K)

Arguments

Value

A matrix of 2^K rows and K columns.

Description

Constructs and validates the item bank object used in cognitive diagnosis computerized adaptive testing (CD-CAT) sessions.

Usage

cdcat_items(q_matrix, model, slip = NULL, guess = NULL, gdina_params = NULL)

Arguments

Details

The item bank is defined by a Q-matrix and model-specific item parameters.

The Q-matrix is a binary matrix where:

$q_{jk} = \begin{cases} 1 & \text{if item } j \text{ requires attribute } k, \\ 0 & \text{otherwise.} \end{cases}$

Three cognitive diagnosis models are supported:

DINA model (Junker & Sijtsma, 2001):

$P(X_j=1|\alpha) = (1-s_j)^{\eta_j(\alpha)} g_j^{1-\eta_j(\alpha)},$

where $s_j$ is the slip parameter and $g_j$ is the guess parameter.

DINO model shares the same functional form but uses a compensatory OR-type mastery indicator.

GDINA model (de la Torre, 2011) is a saturated model allowing arbitrary interaction effects among required attributes.

All models assume:

• Binary item responses;

• Conditional independence of item responses given the latent profile;

• Fixed item parameters during the adaptive session.

Value

A list of class cdcat_items containing the Q-matrix, item parameters, and model specification.

References

Junker, B. W., & Sijtsma, K. (2001). Cognitive assessment models with few assumptions, and connections with nonparametric item response theory. Applied Psychological Measurement, 25(3), 258–272. doi:10.1177/01466210122032064

de la Torre, J. (2011). The generalized DINA model framework. Psychometrika, 76(2), 179–199. doi:10.1007/s11336-011-9207-7

Description

Constructs and validates a prior probability vector over all 2^K latent skill profiles, using pattern names as keys. The pattern names follow the same ordering as build_skill_patterns(): expand.grid with the first attribute varying fastest.

Usage

cdcat_prior(..., K = NULL, normalize = TRUE)

Arguments

Value

A named numeric vector of length 2^K whose entries sum to 1, ordered consistently with build_skill_patterns().

See Also

estimate_alpha(), CdcatSession

Examples

# K = 2: four profiles -- "00", "10", "01", "11"
prior <- cdcat_prior("00" = 0.4, "10" = 0.2, "01" = 0.2, "11" = 0.2)

# Use in a session
Q <- matrix(c(1,0, 0,1, 1,1), nrow = 3, ncol = 2, byrow = TRUE)
items <- cdcat_items(Q, "DINA", slip = c(0.1,0.1,0.1), guess = c(0.2,0.2,0.2))
session <- CdcatSession$new(items, prior = prior)

Description

A named list of 16 simulated datasets for Cognitive Diagnostic Computerized Adaptive Testing (CD-CAT) research. The 16 elements cover all combinations of:

• K (number of attributes): 3 or 5

• Model: DINA or GDINA

• Quality: HD-LV, HD-HV, LD-LV, or LD-HV (see Item quality section below)

Usage

cdcat_sim

Format

A named list with 16 elements. Element names follow the pattern ⁠k{K}_{MODEL}_{QUALITY}⁠, e.g. "k3_DINA_HDLV" or "k5_GDINA_LDHV". Each element is itself a list with four components:

Q

Integer matrix ($J \times K$). Q-matrix mapping items to attributes. Constructed by replicating each of the $2^K - 1$ non-zero attribute profiles 10 times, yielding $J = 10 \times (2^K - 1)$ items (K = 3: $J = 70$; K = 5: $J = 310$). Rows are named item1, ..., itemJ; columns A1, ..., AK.

alpha

Integer matrix ($N \times K$). True latent attribute profiles of the examinees. Constructed by replicating each of the $2^K$ possible binary profiles 20 times, yielding $N = 20 \times 2^K$ examinees (K = 3: $N = 160$; K = 5: $N = 640$). Rows are named examinee1, ..., examineeeN; columns A1, ..., AK.

parameters

List of length $J$. Item parameters whose format depends on the model:

DINA

Each element is a list with two scalars: slip (probability that a master answers incorrectly) and guess (probability that a non-master answers correctly).

GDINA

Each element is a named list of $2^{K_j}$ scalars, where $K_j$ is the number of attributes required by item $j$. Names are binary strings representing the latent response pattern (e.g. "0", "1" for $K_j = 1$; "00", "10", "01", "11" for $K_j = 2$). Each value is the probability of a correct response given that pattern of required-attribute mastery.

responses

Integer matrix ($N \times J$). Dichotomous item responses (0 / 1) simulated from alpha and parameters using GDINA::simGDINA().

Item quality

Parameters were sampled from uniform distributions on ranges that reflect four levels of item quality, defined by the probability of a correct response for non-masters ($P_0$) and masters ($P_1$):

For DINA/DINO: slip = 1 - P1 and guess = P0, independently sampled per item and clipped to $[0.001, 0.999]$.

For GDINA: $P_{min} \sim P_0$ and $P_{max} \sim P_1$ are independently sampled per item; the probability of a correct response increases linearly with the count of mastered required attributes: $P_{min} + (m / K_j) \times (P_{max} - P_{min})$, where $m$ is the number of required attributes mastered.

Reproducibility

All datasets were generated with set.seed(2025). The generation script is available in data-raw/generate_cdcat_sim.R.

Dataset dimensions

Source

Generated with GDINA::simGDINA() (Ma & de la Torre, 2020). See data-raw/generate_cdcat_sim.R for the full script.

References

Ma, W., & de la Torre, J. (2020). GDINA: An R package for cognitive diagnosis modeling. Journal of Statistical Software, 93(14), 1–26. doi:10.18637/jss.v093.i14

Examples

# List all available datasets
names(cdcat_sim)

# Access a specific dataset
d <- cdcat_sim[["k3_DINA_HDLV"]]
dim(d$Q)          # 70 x 3
dim(d$alpha)      # 160 x 3
dim(d$responses)  # 160 x 70

# Inspect DINA parameters for the first item
d$parameters[[1]]

# Inspect GDINA parameters for the first item
cdcat_sim[["k3_GDINA_HDLV"]]$parameters[[1]]

# Use in a CD-CAT session
d <- cdcat_sim[["k3_DINA_HDLV"]]
items <- cdcat_items(
  d$Q, "DINA",
  slip  = sapply(d$parameters, `[[`, "slip"),
  guess = sapply(d$parameters, `[[`, "guess")
)
session <- CdcatSession$new(items, criterion = "PWKL", max_items = 10L)

Description

CD-CAT Session Object

CD-CAT Session Object

Details

An R6 class implementing the adaptive testing cycle for a single cognitive diagnosis computerized adaptive testing (CD-CAT) session.

This class orchestrates the sequential CD-CAT procedure:

1. Initialize the posterior distribution over latent profiles;

2. Select the next item according to a specified selection criterion;

3. Observe the response and update the posterior;

4. Evaluate stopping criteria;

5. Repeat until termination.

The adaptive algorithm assumes:

• Binary item responses;

• Conditional independence of item responses given the latent profile;

• A calibrated item bank defined by a Q-matrix and fixed item parameters;

• A specified estimation method (MLE, MAP, or EAP);

• A specified item selection criterion (e.g., KL, PWKL, MPWKL, SHE).

The session object maintains the full state of the adaptive test, including responses, administered items, posterior distribution, and stopping status.

Implementation details

Each call to next_item() evaluates stopping criteria, computes item selection (with full diagnostic details), and caches the result in private$pending. Each call to update(item, response) records the response, re-estimates the attribute profile, and appends a complete step record to self$history.

Per-step history

The history field stores one named list per administered item. Use session$history_df() to convert to a flat data.frame. Each record contains:

step

Integer. 1-based position in the test.

item

Integer. Item index administered.

response

Integer. 0 or 1.

candidate_items

Integer vector. Items available after content balancing filter.

criterion_scores

Named numeric vector. Criterion score per candidate item (NULL for SEQ/RANDOM).

item_pre_exposure

Integer. Item that would have been chosen without exposure control (NA in shadow/non-adaptive mode).

exposure_redirected

Logical. Whether exposure control redirected the selection.

target_domain

Character. Content domain targeted by balancing at this step (NULL if inactive).

domain_gap

Named numeric. Gap (target - observed) per domain at this step (NULL if inactive).

stop_check

List with stop (logical) and reason (character) from the stopping evaluation before this item.

posterior

Numeric vector of length 2^K. Posterior after this response.

alpha_hat

Integer vector of length K. Estimated profile.

alpha_hat_index

Integer. Row index of alpha_hat in the skill patterns matrix.

mastery_marginal

Named numeric of length K. P(alpha_k = 1) for each attribute after this response.

timestamp

POSIXct. Time at which update() was called.

response_time

Numeric. Seconds elapsed since the previous update() call (or session start for the first item).

Public fields

Methods

Public methods

• CdcatSession$new()

• CdcatSession$next_item()

• CdcatSession$update()

• CdcatSession$history_df()

• CdcatSession$result()

• CdcatSession$print()

• CdcatSession$clone()

Method new()

Create a new CD-CAT session.

Usage

CdcatSession$new(
  items,
  method = "MAP",
  criterion = "PWKL",
  min_items = 1L,
  max_items = 20L,
  threshold = 0.8,
  prior = NULL,
  content = NULL,
  content_prop = NULL,
  exposure = NULL,
  constr_fun = NULL,
  initial_profile = NULL,
  start_item = NULL
)

Arguments

Method next_item()

Select the next item to administer.

Evaluates stopping criteria, computes item selection, and caches selection details internally for update() to incorporate into the history record.

Usage

CdcatSession$next_item()

Returns

Integer index of the next item, or 0L if stopped.

Method update()

Register a response and update session state.

Records the response, re-estimates the attribute profile, and appends a complete step record to self$history.

Usage

CdcatSession$update(item, response)

Arguments

Method history_df()

Convert the history list to a flat data.frame.

Vector-valued fields (posterior, mastery_marginal, alpha_hat, criterion_scores, candidate_items, domain_gap) are kept as list-columns.

Usage

CdcatSession$history_df()

Returns

A data.frame with one row per administered item, or an empty data.frame if no items have been administered.

Method result()

Extract final session results.

Usage

CdcatSession$result()

Returns

Named list with estimation results and session metadata.

Method print()

Print a summary of the session state.

Usage

CdcatSession$print(...)

Arguments

Method clone()

The objects of this class are cloneable with this method.

Usage

CdcatSession$clone(deep = FALSE)

Arguments

Description

Evaluates whether the computerized adaptive test should terminate based on the current posterior distribution over latent profiles and operational constraints.

Usage

check_stopping(
  est,
  n_administered,
  min_items = 1L,
  max_items = 20L,
  threshold = 0.8
)

Arguments

Details

This function implements three stopping mechanisms within a cognitive diagnosis computerized adaptive testing (CD-CAT) framework:

1. Fixed-length rule

The test stops when the number of administered items reaches max_items. This corresponds to the traditional fixed-length stopping rule described in the CD-CAT literature.

2. Attribute-level posterior threshold rule (Tatsuoka, 2002)

Let $\pi_t(\alpha_c)$ denote the posterior distribution over latent profiles after $t$ administered items.

The marginal posterior probability of mastery for attribute $k$ is:

$P(\alpha_k = 1 \mid x^{(t)}) = \sum_{c: \alpha_{ck}=1} \pi_t(\alpha_c).$

The test stops when all attributes are classified with sufficient confidence:

$P(\alpha_k = 1 \mid x^{(t)}) \ge \tau \quad \text{or} \quad P(\alpha_k = 1 \mid x^{(t)}) \le 1 - \tau \quad \text{for all } k.$

This rule performs classification at the attribute level rather than at the full latent class level.

3. Dual posterior threshold rule (Hsu, Wang, & Chen, 2013)

Let $\pi_{(1)}$ and $\pi_{(2)}$ denote the largest and second-largest posterior probabilities among all latent profiles.

The test stops when:

$\pi_{(1)} \ge \tau_1 \quad \text{and} \quad \pi_{(2)} \le \tau_2,$

where $\tau_1 > \tau_2$. This conservative rule avoids stopping when two latent profiles remain similarly probable.

In addition to these theoretical criteria, the function enforces:

• A minimum number of administered items (min_items);

• A maximum number of administered items (max_items).

All posterior computations assume conditional independence of item responses given the latent profile.

Value

A list with:

• stop: Logical. Whether the test should stop.

• reason: Character string describing the stopping condition.

References

Tatsuoka, C. (2002). Data analytic methods for latent partially ordered classification models. Journal of the Royal Statistical Society Series C: Applied Statistics, 51(3), 337–350. doi:10.1111/1467-9876.00272

Hsu, C. L., Wang, W. C., & Chen, S. Y. (2013). Variable-length computerized adaptive testing based on cognitive diagnosis models. Applied Psychological Measurement, 37(7), 563–582. doi:10.1177/0146621613488642

Description

Estimates the latent attribute profile given observed item responses under a specified cognitive diagnosis model.

Usage

estimate_alpha(responses, items, method = "MAP", prior = NULL)

Arguments

Details

The likelihood for latent profile $\alpha_c$ is:

$L(\alpha_c) = \prod_{j \in \mathcal{J}_t} P(X_j = x_j | \alpha_c)$

MLE selects:

$\hat{\alpha}_{MLE} = \arg\max_c L(\alpha_c)$

MAP selects:

$\hat{\alpha}_{MAP} = \arg\max_c L(\alpha_c)\pi(\alpha_c)$

EAP estimates attribute mastery probabilities:

$\hat{\alpha}_{EAP,k} = \sum_c \pi_t(\alpha_c)\alpha_{ck}$

Value

A list of class cdcat_est.

Description

Computes the matrix of item response probabilities $P(X_j = 1 | \alpha)$ for all items and all latent attribute profiles under the specified cognitive diagnosis model.

Usage

get_prob_matrix(q_matrix, parameters, skill_patterns, model)

Arguments

Details

This function computes item response probabilities under three cognitive diagnosis models:

DINA model:

$P(X_j=1|\alpha) = (1-s_j)^{\eta_j(\alpha)} g_j^{1-\eta_j(\alpha)}$

DINO model: Same functional form as DINA, but with $\eta_j(\alpha) = 1$ if at least one required attribute is mastered.

GDINA model: A saturated model allowing arbitrary interaction effects among required attributes.

Value

A matrix of dimension $J \times 2^K$.

Description

Computes the Kullback-Leibler (KL) information index for a candidate item under a cognitive diagnosis computerized adaptive testing (CD-CAT) framework.

Usage

KL_criteria(item_index, alpha_hat_index, prob_matrix)

Arguments

Details

This criterion selects the item that maximizes the Kullback-Leibler divergence between the item response distribution induced by the current estimated latent profile and the response distributions induced by all alternative latent profiles.

Let $\hat{\alpha}^{(t)}$ denote the current latent profile estimate after $t$ administered items, and let $\alpha_c$, $c = 1,\dots,2^K$, denote all possible latent attribute profiles.

For a candidate item $h$, the KL index is defined as:

$KL_h(\hat{\alpha}^{(t)}) = \sum_{c=1}^{2^K} \sum_{x=0}^{1} P(X_h = x \mid \hat{\alpha}^{(t)}) \log \frac{P(X_h = x \mid \hat{\alpha}^{(t)})} {P(X_h = x \mid \alpha_c)}$

Thus, the divergence is computed between conditional response distributions for the candidate item under different latent profiles, not between the attribute vectors themselves.

The logarithm is computed using the natural logarithm.

Assumptions:

• Conditional independence of item responses.

• A fixed cognitive diagnosis model defining $P(X=1|\alpha)$.

• Equal weighting of all latent profiles.

Computational complexity per candidate item is $O(2^K)$.

Value

Numeric scalar representing the KL information value.

Description

Computes the symmetric posterior-weighted KL divergence across all pairs of item response distributions induced by latent profiles.

Usage

MPWKL_criteria(item_index, prob_matrix, posterior)

Arguments

Details

The MPWKL index for item $h$ is defined as:

$MPWKL_h = \sum_{d=1}^{2^K} \sum_{c=1}^{2^K} KL_h(\alpha_d \Vert \alpha_c) \pi_t(\alpha_d)\pi_t(\alpha_c)$

where $KL_h(\alpha_d \Vert \alpha_c)$ denotes the KL divergence between the conditional response distributions of item $h$ under latent profiles $\alpha_d$ and $\alpha_c$.

This formulation evaluates the expected global discrimination of the item under the full posterior distribution.

Computational complexity per candidate item is $O((2^K)^2)$.

Value

Numeric scalar representing the MPWKL information value.

References

Cheng, Y. (2009). When cognitive diagnosis meets computerized adaptive testing: CD-CAT. Psychometrika, 74(4), 619–632. doi:10.1007/s11336-009-9123-2

Description

Computes the posterior-weighted KL information index for a candidate item in CD-CAT.

Usage

PWKL_criteria(item_index, alpha_hat_index, prob_matrix, posterior)

Arguments

Details

Unlike the standard KL method, this criterion weights each latent profile by its current posterior probability rather than assuming equally likely latent states.

Let $\pi_t(\alpha_c)$ denote the posterior probability of latent profile $\alpha_c$ after $t$ administered items.

The PWKL index for item $h$ is defined as:

$PWKL_h(\hat{\alpha}^{(t)}) = \sum_{c=1}^{2^K} KL_h(\hat{\alpha}^{(t)} \Vert \alpha_c) \pi_t(\alpha_c)$

where $KL_h(\cdot)$ represents the KL divergence between the conditional response distributions induced by two latent profiles.

This weighting reduces early-stage instability and avoids the implicit assumption of uniformly distributed latent states.

Assumptions:

• Correct posterior updating.

• Conditional independence of responses.

Computational complexity per candidate item is $O(2^K)$.

Value

Numeric scalar representing the PWKL information value.

References

Cheng, Y. (2009). When cognitive diagnosis meets computerized adaptive testing: CD-CAT. Psychometrika, 74(4), 619–632. doi:10.1007/s11336-009-9123-2

Description

Computes the item selection criterion score for all available items and returns the index of the best candidate.

Usage

select_next_item(
  items,
  responses,
  administered,
  criterion,
  est = NULL,
  method = "MAP",
  prior = NULL,
  content = NULL,
  content_prop = NULL,
  exposure = NULL,
  constr_fun = NULL,
  initial_profile = NULL,
  return_details = FALSE
)

Arguments

Details

Greedy mode (constr_fun = NULL, default):

1. Filter candidates by content balancing;

2. Compute adaptive criterion scores for filtered candidates;

3. Select via exposure control.

Shadow mode (constr_fun provided):

1. Compute criterion scores for all J items (full bank);

2. Delegate selection to constr_fun(scores, items, administered).

Value

When return_details = FALSE (default): integer scalar – item index.

When return_details = TRUE: named list with fields:

item

Integer. Selected item index.

candidate_items

Integer vector after content filter.

criterion_scores

Named numeric vector of scores (names are item indices), or NULL for non-adaptive criteria.

item_pre_exposure

Integer. Item that would have been selected without exposure control, or NA if not applicable.

exposure_redirected

Logical. Whether exposure control changed the selected item.

References

Kingsbury, G. G., & Zara, A. R. (1991). A comparison of procedures for content-sensitive item selection in computerized adaptive testing. Applied Measurement in Education, 4(3), 241–261. doi:10.1207/s15324818ame0403_4

van der Linden, W. J., & Veldkamp, B. P. (2004). Constraining item exposure in computerized adaptive testing with shadow tests. Journal of Educational and Behavioral Statistics, 29(3), 273–291. doi:10.3102/10769986029003273

van der Linden, W. J. (2022). Review of the shadow-test approach to adaptive testing. Behaviormetrika, 49(2), 169–190. doi:10.1007/s41237-021-00150-y

See Also

apply_content_balancing(), apply_exposure_control(), CdcatSession

Description

Computes the expected posterior Shannon entropy after administering a candidate item.

Usage

SHE_criteria(item_index, prob_matrix, posterior)

Arguments

Details

The selected item minimizes the expected posterior entropy after observing the potential response to the candidate item.

Let $\pi_t(\alpha_c)$ denote the posterior distribution over latent profiles after $t$ administered items.

Shannon entropy is defined as:

$SH(\pi_t) = -\sum_{c=1}^{2^K} \pi_t(\alpha_c)\log \pi_t(\alpha_c)$

For candidate item $h$, the expected posterior entropy is:

$E[SH(\pi_{t+1})] = \sum_{x=0}^{1} SH(\pi_{t+1} \mid X_h = x) P(X_h = x \mid x^{(t)})$

Thus, the criterion evaluates the expected reduction in uncertainty about the latent profile induced by administering item $h$.

Computational complexity per candidate item is $O(2^K)$.

Value

Numeric scalar representing the expected posterior entropy.

References

Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27(3), 379–423. doi:10.1002/j.1538-7305.1948.tb01338.x