Interview-Based Catch Estimation • tidycreel

Introduction

This vignette extends the effort estimation workflow covered in “Getting Started with tidycreel” to interview-based catch and harvest estimation. The complete workflow involves five steps:

Design: Define your survey calendar and stratification
Counts: Attach instantaneous count observations
Interviews: Attach complete trip interview data
CPUE: Estimate catch per unit effort
Total Catch: Combine effort and CPUE estimates

The example is an access-point survey: anglers are interviewed as they finish their trips. For complete trips the catch rate is the ratio of means, total catch over total effort (Pollock et al. 1994, §15.3.3.5). vignette("estimator-choice") explains how the package picks the estimator from the interview type.

Survey Design and Count Data

We start with the same design and count data workflow from the “Getting Started” vignette:

library(tidycreel)

# Load example calendar and counts
data(example_calendar)
data(example_counts)

# effort_hours is already angler-hours for each sampled day, not an
# instantaneous count, so it goes in through prep_counts_daily_effort()
daily_effort <- prep_counts_daily_effort(example_counts,
  date = date, strata = day_type, daily_effort = effort_hours
)

design <- creel_design(example_calendar, date = date, strata = day_type)
design <- add_counts(design, daily_effort)

# Estimate total effort
effort_est <- estimate_effort(design)
print(effort_est)
#> 
#> ── Creel Survey Estimates ──────────────────────────────────────────────────────
#> Method: Total
#> Variance: Taylor linearization
#> Confidence level: 95%
#> Effort target: sampled_days
#> 
#> # A tibble: 1 × 7
#>   estimate    se se_between se_within ci_lower ci_upper     n
#>      <dbl> <dbl>      <dbl>     <dbl>    <dbl>    <dbl> <int>
#> 1     372.  13.2       13.2         0     344.     401.    14

example_counts records angler-hours for each sampled day, so it goes in through prep_counts_daily_effort(). Passed to add_counts() directly it would be read as instantaneous counts, with a warning that no period length was given. The estimate is in angler-hours, as the data say. The printout has no Unit: line, because the package did not see how those angler-hours were made and so reports the unit as unknown (vignette("unit-rules")). The calendar lists only the 14 sampled days, so this is the total over those days.

Adding Interview Data

Next, we attach interview data containing catch and effort for complete trips:

# Load example interview data
data(example_interviews)
head(example_interviews)
#>         date hours_fished catch_total catch_kept trip_status trip_duration
#> 1 2024-06-01          2.0           5          2    complete           2.0
#> 2 2024-06-01          3.5           8          5    complete           3.5
#> 3 2024-06-02          1.5           2          1    complete           1.5
#> 4 2024-06-02          2.0           3          2  incomplete           1.0
#> 5 2024-06-03          2.5           6          3    complete           2.5
#> 6 2024-06-03          4.0          12          8    complete           4.0
#>   interview_id angler_type angler_method species_sought n_anglers refused
#> 1            1        bank          bait        walleye         2   FALSE
#> 2            2        boat    artificial        walleye         1   FALSE
#> 3            3        bank          bait           bass         3   FALSE
#> 4            4        bank           fly        panfish         2   FALSE
#> 5            5        boat    artificial        walleye         1   FALSE
#> 6            6        boat          bait           bass         4   FALSE

# Attach interviews to the design
design <- add_interviews(design, example_interviews,
  catch = catch_total,
  effort = hours_fished,
  n_anglers = n_anglers,
  harvest = catch_kept,
  trip_status = trip_status,
  trip_duration = trip_duration
)
#> ℹ Added 22 interviews: 17 complete (77%), 5 incomplete (23%)
print(design)
#> 
#> ── Creel Survey Design ─────────────────────────────────────────────────────────
#> Type: "instantaneous"
#> Date column: date
#> Strata: day_type
#> Calendar: 14 days (2024-06-01 to 2024-06-14)
#> day_type: 2 levels
#> Counts: 14 observations
#> PSU column: date
#> Count column: daily_effort
#> Party-size term: not carried
#> Count type: "instantaneous"
#> Survey: <survey.design2> (constructed)
#> Interviews: 22 observations
#> Type: "access"
#> Catch: catch_total
#> Effort: hours_fished
#> Harvest: catch_kept
#> Trip status: 17 complete, 5 incomplete
#> Party size: n_anglers
#> Survey: <survey.design2> (constructed)
#> Sections: "none"

The add_interviews() function maps the interview data columns to the design structure. The design now shows both count and interview data attached. Interview data is treated as a parallel data stream to count data—the two datasets do not need to align on specific dates.

Estimating CPUE

With interview data attached, we can estimate catch per unit effort:

# Estimate CPUE
cpue_est <- estimate_catch_rate(design)
#> ℹ Using complete trips for CPUE estimation
#>   (n=17, 77.3% of 22 interviews) [default]
#> Warning: Small sample size for CPUE estimation.
#> ! Sample size is 17. Ratio estimates are more stable with n >= 30.
#> ℹ Variance estimates may be unstable with n < 30.
#> ℹ Repeats on this design are recorded, not printed: see `get_warnings()`.
print(cpue_est)
#> 
#> ── Creel Survey Estimates ──────────────────────────────────────────────────────
#> Method: Ratio-of-Means CPUE
#> Variance: Taylor linearization
#> Confidence level: 95%
#> Unit: fish/angler-hour
#> 
#> # A tibble: 1 × 5
#>   estimate    se ci_lower ci_upper     n
#>      <dbl> <dbl>    <dbl>    <dbl> <int>
#> 1    0.967 0.110    0.752     1.18    17

The CPUE estimate is the ratio of means: total catch divided by total effort over the 17 complete trips. Its variance is a Taylor linearization of the ratio, which carries the covariance between catch and effort across interviews.

The call warns that 17 trips is a small sample for a ratio estimator. It is: the example dataset is small. The warning concerns precision, and the standard error already reflects it.

For details on the ratio-of-means formula and variance calculation, see ?estimate_catch_rate.

Estimating Total Catch

We can combine the effort and CPUE estimates to compute total catch:

# Estimate total catch
total_catch_est <- estimate_total_catch(design)
#> Warning: `estimate_total_catch()` is pooling over domains the counts do not classify:
#> angler_method and species_sought.
#> ! The rate differs across their levels in these interviews (angler_method:
#>   artificial 1.121, bait 0.8, fly 1.31 and species_sought: bass 1.018, panfish
#>   1.2, walleye 0.933), so the total depends on the interview sample's mix over
#>   those domains.
#> ℹ Without the domain in the counts the total is `E_total * rate_pooled`,
#>   weighted by the interview mix rather than the effort mix. Interview selection
#>   is not proportional to effort by construction (Malvestuto 1996).
#> ℹ This is a risk, not an error: the counts carry no composition to check
#>   against, so it cannot be verified from the data.
#> ℹ Classifying angler_method and species_sought in the count data removes the
#>   assumption -- the total becomes `sum(E_h * rate_h)`.
#> ℹ Repeats on this design are recorded, not printed: see `get_warnings()`.
#> Warning: 1 rate cell in "estimate_total_catch" rests on fewer than 10 trips.
#> • day_type=weekend: n=7
#> ℹ The total multiplies each cell's effort by a rate built from that cell's
#>   trips alone; a ratio from so few trips is unstable, and the rate functions
#>   refuse a group this small.
#> ℹ Combine cells or collect more interviews. Totals warn here rather than refuse
#>   (GH #417).
#> ℹ Repeats on this design are recorded, not printed: see `get_warnings()`.
print(total_catch_est)
#> 
#> ── Creel Survey Estimates ──────────────────────────────────────────────────────
#> Method: Total Catch (Effort × CPUE)
#> Variance: Taylor linearization
#> Confidence level: 95%
#> Effort target: sampled_days
#> 
#> # A tibble: 1 × 5
#>   estimate    se ci_lower ci_upper     n
#>      <dbl> <dbl>    <dbl>    <dbl> <int>
#> 1     364.  47.9     262.     466.    17

The total is built stratum by stratum: each day type’s effort times that day type’s catch rate, then added. The variance of each product uses Goodman’s (1960) unbiased estimator by default, E² Var(R) + R² Var(E) − Var(E) Var(R), with the estimates in place of the true values. The exact variance of a product has + Var(E) Var(R); the estimator subtracts it because each squared estimate already overstates its squared mean by its variance. product_variance = "first_order" drops the last term, which gives a slightly larger, conservative standard error. Both treat the effort and rate estimates as independent. That is an assumption, not a fact of the design: counts and interviews are taken on the same sampled days. vignette("catch-pipeline") discusses it.

This call raises two warnings, both about the example’s small size:

Pooled domains. The interviews record fishing method and target species, and the catch rate differs across them. The counts do not record either, so the total uses one rate per day type, weighted by the interview mix rather than the effort mix. With counts that record method or species, the total would be summed over those domains instead.
A thin rate cell. The weekend rate rests on 7 trips. Each stratum’s effort is multiplied by that stratum’s own rate, and a ratio from so few trips is unstable.

Estimating Harvest

The package distinguishes between total catch (all fish caught) and harvest (fish kept). We can estimate harvest per unit effort (HPUE) and total harvest:

# Estimate HPUE
hpue_est <- estimate_harvest_rate(design)
#> ℹ Filtering to complete trips for HPUE estimation
#>   (n=17, 77.3% of 22 interviews) [default]
print(hpue_est)
#> 
#> ── Creel Survey Estimates ──────────────────────────────────────────────────────
#> Method: Ratio-of-Means HPUE
#> Variance: Taylor linearization
#> Confidence level: 95%
#> Unit: fish/angler-hour
#> 
#> # A tibble: 1 × 5
#>   estimate     se ci_lower ci_upper     n
#>      <dbl>  <dbl>    <dbl>    <dbl> <int>
#> 1    0.595 0.0616    0.475    0.716    17

# Estimate total harvest
total_harvest_est <- estimate_total_harvest(design)
print(total_harvest_est)
#> 
#> ── Creel Survey Estimates ──────────────────────────────────────────────────────
#> Method: Total Harvest (Effort × HPUE)
#> Variance: Taylor linearization
#> Confidence level: 95%
#> Effort target: sampled_days
#> 
#> # A tibble: 1 × 5
#>   estimate    se ci_lower ci_upper     n
#>      <dbl> <dbl>    <dbl>    <dbl> <int>
#> 1     225.  28.1     165.     284.    17

Harvest estimation uses the same ratio-of-means approach as CPUE, but with the harvest column (catch_kept) as the numerator. As expected, total harvest is lower than total catch since not all caught fish are kept.

Grouped Estimation

Like effort estimation, CPUE and total catch functions support grouped estimation using the by parameter:

# Estimate CPUE by day_type (not evaluated - example data too small)
cpue_by_day <- estimate_catch_rate(design, by = day_type)

# Estimate total catch by day_type
total_catch_by_day <- estimate_total_catch(design, by = day_type)

This chunk is not run: the rate functions refuse a group with fewer than 10 trips, and the weekend group here has 7.

Variance Methods

Like effort estimation, CPUE estimation supports multiple variance methods. The default is Taylor linearization, but bootstrap and jackknife are also available:

set.seed(42) # bootstrap replicates are random
cpue_boot <- estimate_catch_rate(design, variance = "bootstrap")
#> ℹ Using complete trips for CPUE estimation
#>   (n=17, 77.3% of 22 interviews) [default]
cpue_jk <- estimate_catch_rate(design, variance = "jackknife")
#> ℹ Using complete trips for CPUE estimation
#>   (n=17, 77.3% of 22 interviews) [default]

cbind(
  method = c("taylor", "bootstrap", "jackknife"),
  rbind(
    tibble::as_tibble(cpue_est)[, c("estimate", "se")],
    tibble::as_tibble(cpue_boot)[, c("estimate", "se")],
    tibble::as_tibble(cpue_jk)[, c("estimate", "se")]
  )
)
#>      method  estimate        se
#> 1    taylor 0.9674419 0.1100611
#> 2 bootstrap 0.9674419 0.1144879
#> 3 jackknife 0.9674419 0.1132936

The estimate is the same under all three; only the standard error depends on the method. Use bootstrap or jackknife to check the linearization, or for statistics where it is awkward.

When to use each method:

Taylor linearization (default): Computationally efficient and appropriate for most smooth statistics. This is the recommended default.
Bootstrap: Use when working with non-smooth statistics or when you want to verify Taylor linearization assumptions. More computationally intensive.
Jackknife: Alternative resampling method that is deterministic (unlike bootstrap). Useful for verification or when bootstrap is too slow.

Age Structure Estimation

Age data collected during interviews can be used to estimate a pressure-weighted age distribution and design-weighted mean age for the catch or harvest. Attach age records with add_ages(), then call est_age_distribution() followed by est_mean_age().

data(example_ages)
data(example_catch)
head(example_ages)
#>   interview_id species age age_type
#> 1            1 walleye   4  harvest
#> 2            1 walleye   3  harvest
#> 3            1 walleye   5  harvest
#> 4            2    bass   2  harvest
#> 5            2    bass   3  harvest
#> 6            6 walleye   6  harvest

# Ages are read from a subsample of the catch, so the age-class totals are
# scaled onto the reported catch. Grouping by species needs this table, which
# is the only record of catch per species.
design <- add_catch(
  design,
  example_catch,
  catch_uid     = interview_id,
  interview_uid = interview_id,
  species       = species,
  count         = count,
  catch_type    = catch_type
)

# Attach age data — one row per aged fish
design <- add_ages(
  design,
  example_ages,
  age_uid      = interview_id,
  interview_uid = interview_id,
  species      = species,
  age          = age,
  age_type     = age_type
)

# Estimate weighted age frequency by species
ad <- est_age_distribution(design, by = species)
#> Warning: ! Age totals were rescaled onto the reported catch.
#> ℹ Measured fish (weighted): 18; reported: 93 -- a factor of 5.17.
#> ℹ estimate, se and the confidence bounds describe the REPORTED catch, estimated
#>   from the measured subsample. Shares (percent) are unaffected.
print(ad)
#>   species age estimate       se   ci_lower  ci_upper percent cumulative_percent
#> 1    bass   2 10.00000 4.233202  1.7030763 18.296924    40.0               40.0
#> 2    bass   3 15.00000 6.524569  2.2120798 27.787920    60.0              100.0
#> 3 panfish   0  3.25000 1.899150 -0.4722657  6.972266    25.0               25.0
#> 4 panfish   1  6.50000 4.023369 -1.3856588 14.385659    50.0               75.0
#> 5 panfish   2  3.25000 3.991201 -4.5726107 11.072611    25.0              100.0
#> 6 walleye   3 12.22222 4.069220  4.2466972 20.197747    22.2               22.2
#> 7 walleye   4 18.33333 6.136191  6.3066208 30.360046    33.3               55.6
#> 8 walleye   5 12.22222 4.069220  4.2466972 20.197747    22.2               77.8
#> 9 walleye   6 12.22222 8.491868 -4.4215332 28.865978    22.2              100.0
#>   n
#> 1 3
#> 2 3
#> 3 2
#> 4 2
#> 5 2
#> 6 3
#> 7 3
#> 8 3
#> 9 3

Ages come from a subsample of the catch, so est_age_distribution() reports a two-phase estimate: the age composition among aged fish, scaled onto the design-estimated reported catch. The call warns when it rescales, naming the factor. Shares (percent) are unaffected by how many fish were aged.

Each row is one integer age class. The estimate column is the survey-design-weighted count of fish at that age, percent is the within-species proportion, and cumulative_percent accumulates from the youngest age class up. The same variance, type, and conf_level arguments available in est_length_distribution() are accepted here.

# Compute weighted mean age from the distribution object
est_mean_age(ad)
#>   species mean_age mean_age_se mean_age_ci_lower mean_age_ci_upper
#> 1    bass 2.600000   0.1456703         2.3144914          2.885509
#> 2 panfish 1.000000   0.3400005         0.3336113          1.666389
#> 3 walleye 4.444444   0.2706522         3.9139759          4.974913

est_mean_age() applies the ratio estimator (Σ a N̂_a / N̂) to the weighted age counts returned by est_age_distribution() and propagates variance via the delta method.

Complete Workflow Example

The full pipeline in one workflow is:

# Load data
data(example_calendar)
data(example_counts)
data(example_interviews)

# Build design and attach data
complete_design <- creel_design(example_calendar, date = date, strata = day_type) |>
  add_counts(prep_counts_daily_effort(example_counts,
    date = date, strata = day_type, daily_effort = effort_hours
  )) |>
  add_interviews(example_interviews,
    catch = catch_total,
    effort = hours_fished,
    n_anglers = n_anglers,
    harvest = catch_kept,
    trip_status = trip_status,
    trip_duration = trip_duration
  )
#> ℹ Added 22 interviews: 17 complete (77%), 5 incomplete (23%)

# Compute all estimates
effort <- estimate_effort(complete_design)
cpue <- estimate_catch_rate(complete_design)
#> ℹ Using complete trips for CPUE estimation
#>   (n=17, 77.3% of 22 interviews) [default]
#> Warning: Small sample size for CPUE estimation.
#> ! Sample size is 17. Ratio estimates are more stable with n >= 30.
#> ℹ Variance estimates may be unstable with n < 30.
#> ℹ Repeats on this design are recorded, not printed: see `get_warnings()`.
hpue <- estimate_harvest_rate(complete_design)
#> ℹ Filtering to complete trips for HPUE estimation
#>   (n=17, 77.3% of 22 interviews) [default]
total_catch <- estimate_total_catch(complete_design)
#> Warning: `estimate_total_catch()` is pooling over domains the counts do not classify:
#> angler_method and species_sought.
#> ! The rate differs across their levels in these interviews (angler_method:
#>   artificial 1.121, bait 0.8, fly 1.31 and species_sought: bass 1.018, panfish
#>   1.2, walleye 0.933), so the total depends on the interview sample's mix over
#>   those domains.
#> ℹ Without the domain in the counts the total is `E_total * rate_pooled`,
#>   weighted by the interview mix rather than the effort mix. Interview selection
#>   is not proportional to effort by construction (Malvestuto 1996).
#> ℹ This is a risk, not an error: the counts carry no composition to check
#>   against, so it cannot be verified from the data.
#> ℹ Classifying angler_method and species_sought in the count data removes the
#>   assumption -- the total becomes `sum(E_h * rate_h)`.
#> ℹ Repeats on this design are recorded, not printed: see `get_warnings()`.
#> Warning: 1 rate cell in "estimate_total_catch" rests on fewer than 10 trips.
#> • day_type=weekend: n=7
#> ℹ The total multiplies each cell's effort by a rate built from that cell's
#>   trips alone; a ratio from so few trips is unstable, and the rate functions
#>   refuse a group this small.
#> ℹ Combine cells or collect more interviews. Totals warn here rather than refuse
#>   (GH #417).
#> ℹ Repeats on this design are recorded, not printed: see `get_warnings()`.
total_harvest <- estimate_total_harvest(complete_design)

# Print key results
print(effort)
#> 
#> ── Creel Survey Estimates ──────────────────────────────────────────────────────
#> Method: Total
#> Variance: Taylor linearization
#> Confidence level: 95%
#> Effort target: sampled_days
#> 
#> # A tibble: 1 × 7
#>   estimate    se se_between se_within ci_lower ci_upper     n
#>      <dbl> <dbl>      <dbl>     <dbl>    <dbl>    <dbl> <int>
#> 1     372.  13.2       13.2         0     344.     401.    14
print(total_catch)
#> 
#> ── Creel Survey Estimates ──────────────────────────────────────────────────────
#> Method: Total Catch (Effort × CPUE)
#> Variance: Taylor linearization
#> Confidence level: 95%
#> Effort target: sampled_days
#> 
#> # A tibble: 1 × 5
#>   estimate    se ci_lower ci_upper     n
#>      <dbl> <dbl>    <dbl>    <dbl> <int>
#> 1     364.  47.9     262.     466.    17
print(total_harvest)
#> 
#> ── Creel Survey Estimates ──────────────────────────────────────────────────────
#> Method: Total Harvest (Effort × HPUE)
#> Variance: Taylor linearization
#> Confidence level: 95%
#> Effort target: sampled_days
#> 
#> # A tibble: 1 × 5
#>   estimate    se ci_lower ci_upper     n
#>      <dbl> <dbl>    <dbl>    <dbl> <int>
#> 1     225.  28.1     165.     284.    17

The same warnings appear here as in the sections above, for the same reasons.

Working with Incomplete Trips

This page uses the 17 complete trips, as an access-point survey does. On an access design the rate functions use complete trips unless told otherwise. Interviews of anglers still fishing (incomplete trips) call for a different estimator, the mean of ratios with short trips dropped; on a roving design the package uses it by default. vignette("incomplete-trips") covers when and how.

Next Steps

For more details on interview-based estimation functions, see:

?add_interviews - Attach interview data to a design
?estimate_catch_rate - Estimate catch per unit effort
?estimate_harvest_rate - Estimate harvest per unit effort
?estimate_total_catch - Estimate total catch
?estimate_total_harvest - Estimate total harvest
?add_ages - Attach age data to a design
?est_age_distribution - Estimate weighted age frequency
?est_mean_age - Compute design-weighted mean age
?example_interviews - Example interview dataset
?prep_counts_daily_effort - Attach effort already worked out for each day
?example_ages - Example age dataset

For the effort estimation workflow, see the “Getting Started with tidycreel” vignette.

References

Goodman, L. A. (1960). On the exact variance of products. Journal of the American Statistical Association 55:708–713.
Pollock, K. H., Jones, C. M., and Brown, T. L. (1994). Angler Survey Methods and Their Applications in Fisheries Management. American Fisheries Society Special Publication 25. Bethesda, Maryland.