clinicalsignificance (version 2.1.0)

cs_anchor: Anchor-Based Analysis of Clinical Significance

Description

cs_anchor() can be used to determine the clinical significance of intervention studies employing the anchor-based approach. For this, a predefined minimally important difference (MID) for an instrument is known that corresponds to an important symptom improvement for patients. The data can then be analyzed on the individual as well as the group level to estimate, if the change because of an intervention is clinically significant.

Usage

cs_anchor(
  data,
  id,
  time,
  outcome,
  group,
  pre = NULL,
  post = NULL,
  mid_improvement = NULL,
  mid_deterioration = NULL,
  better_is = c("lower", "higher"),
  target = c("individual", "group"),
  effect = c("within", "between"),
  bayesian = TRUE,
  prior_scale = "medium",
  reference_group = NULL,
  ci_level = 0.95
)

Value

An S3 object of class cs_analysis and cs_anchor

Arguments

data

A tidy data frame

id

Participant ID

time

Time variable

outcome

Outcome variable

group

Grouping variable (optional)

pre

Pre measurement (only needed if the time variable contains more than two measurements)

post

Post measurement (only needed if the time variable contains more than two measurements)

mid_improvement

Numeric, change that indicates a clinically significant improvement

mid_deterioration

Numeric, change that indicates a clinically significant deterioration (optional). If mid_deterioration is not provided, it will be assumed to be equal to mid_improvement

better_is

Which direction means a better outcome for the used instrument? Available are

• "lower" (lower outcome scores are desirable, the default) and

• "higher" (higher outcome scores are desirable)

target

String, whether an individual or group analysis should be calculated. Available are

• "individual" (the default) for which every individual participant is evaluated

• "group" for which only the group wise effect is evaluated

effect

String, if target = "group", specify which effect should be calculated. Available are

• "within" (the default), which yields the mean pre-post intervention difference with associated confidence intervals

• "between", which estimates the group wise mean difference and confidence intervals between two or more groups specified with the group argument at the specified measurement supplied with the post- argument The reference group may be supplied with reference_group

bayesian

Logical, only relevant if target = "group". Indicates if a Bayesian estimate (i.e., the median) of group differences with a credible interval should be calculated (if set to TRUE, the default) or a frequentist mean difference with confidence interval (if set to FALSE)

prior_scale

String or numeric, can be adjusted to change the Bayesian prior distribution. See the documentation for rscale in BayesFactor::ttestBF() for details.

reference_group

Specify the reference group to which all subsequent groups are compared against if target = "group" and effect = "within" (optional). Otherwise, the first distinct group is chosen based on alphabetical, numerical or factor ordering.

ci_level

Numeric, define the credible or confidence interval level. The default is 0.95 for a 95%-CI.

Computational details

For the individual-level analyses, the analysis is straight forward. An MID can be specified for an improvement as well as a deterioration (because these must not necessarily be identical) and the function basically counts how many patients fall within the MID range for both, improvement and deterioration, or how many patients exceed the limits of this range in either direction. A patient may than be categorized as:

• Improved, the patient demonstrated a change that is equal or greater then the MID for an improvement

• Unchanged, the patient demonstrated a change that is less than both MIDs

• Deteriorated, the patient demonstrated a change that is equal or greater then the MID for a deterioration

For group-level analyses, the whole sample is either treated as a single group or is split up by grouping presented in the data. For within group analyses, the function calculates the median change from pre to post intervention with the associated credible interval (CI). Based on the median change and the limits of this CI, a group change can be categorized in 5 distinctive categories:

• Statistically not significant, the CI contains 0

• Statistically significant but not clinically relevant, the CI does not contain 0, but the median and both CI limits are beneath the MID threshold

• Not significantly less than the threshold, the MID threshold falls within the CI but the median is still below that threshold

• Probably clinically significant effect, the median crossed the MID threshold but the threshold is still inside the CI

• Large clinically significant effect, the median crossed the MID threshold and the CI does not contain the threshold

If a between group comparison is desired, a reference group can be defined with the reference_group argument to which all subsequent groups are compared. This is usually an inactive comparator such as a placebo or wait-list control group. The difference between the pairwise compared groups is categorized just as the within group difference above, so the same categories apply.

The approach can be changed to a classical frequentist framework for which the point estimate then represents the mean difference and the CI a confidence interval. For an extensive overview over the differences between a Bayesian and frequentist CI, refer to Hespanhol et al. (2019).

Data preparation

The data set must be tidy, which corresponds to a long data frame in general. It must contain a patient identifier which must be unique per patient. Also, a column containing the different measurements and the outcome must be supplied. Each participant-measurement combination must be unique, so for instance, the data must not contain two "After" measurements for the same patient.

Additionally, if the measurement column contains only two values, the first value based on alphabetical, numerical or factor ordering will be used as the pre measurement. For instance, if the column contains the measurements identifiers "pre" and "post" as strings, then "post" will be sorted before "pre" and thus be used as the "pre" measurement. The function will throw a warning but generally you may want to explicitly define the "pre" and "post" measurement with arguments pre and post. In case of more than two measurement identifiers, you have to define pre and post manually since the function does not know what your pre and post intervention measurements are.

If your data is grouped, you can specify the group by referencing the grouping variable (see examples below). The analysis is then run for every group to compare group differences.

References

Hespanhol, L., Vallio, C. S., Costa, L. M., & Saragiotto, B. T. (2019). Understanding and interpreting confidence and credible intervals around effect estimates. Brazilian Journal of Physical Therapy, 23(4), 290–301. https://doi.org/10.1016/j.bjpt.2018.12.006

See Also

Main clinical signficance functions cs_combined(), cs_distribution(), cs_percentage(), cs_statistical()

Examples

cs_results <- antidepressants |>
  cs_anchor(patient, measurement, mom_di, mid_improvement = 8)

cs_results
plot(cs_results)

# Set argument "pre" to avoid a warning
cs_results <- antidepressants |>
  cs_anchor(
    patient,
    measurement,
    mom_di,
    pre = "Before",
    mid_improvement = 8
  )


# Inlcude the MID for deterioration
cs_results_with_deterioration <- antidepressants |>
  cs_anchor(
    patient,
    measurement,
    mom_di,
    pre = "Before",
    mid_improvement = 8,
    mid_deterioration = 5
  )

cs_results_with_deterioration
summary(cs_results_with_deterioration)
plot(cs_results_with_deterioration)


# Group the results by experimental condition
cs_results_grouped <- antidepressants |>
  cs_anchor(
    patient,
    measurement,
    mom_di,
    pre = "Before",
    group = condition,
    mid_improvement = 8,
    mid_deterioration = 5
  )

cs_results_grouped
summary(cs_results_grouped)
plot(cs_results_grouped)

# The plot method always returns a ggplot2 object, so the plot may be further
# modified with ggplot2 code, e.g., facetting to avoid overplotting of groups
plot(cs_results_grouped) +
  ggplot2::facet_wrap(~ group)


# Compute group wise results
cs_results_groupwise <- antidepressants |>
  cs_anchor(
    patient,
    measurement,
    mom_di,
    pre = "Before",
    mid_improvement = 8,
    target = "group"
  )

cs_results_groupwise
summary(cs_results_groupwise)
plot(cs_results_groupwise)


# Group wise analysis, but split by experimentawl condition
cs_results_groupwise_condition <- antidepressants |>
  cs_anchor(
    patient,
    measurement,
    mom_di,
    pre = "Before",
    group = condition,
    mid_improvement = 8,
    target = "group"
  )

cs_results_groupwise_condition
summary(cs_results_groupwise_condition)
plot(cs_results_groupwise_condition)


# Compare all groups to a predefined reference group at a predefined measurement
cs_results_groupwise_between <- antidepressants |>
  cs_anchor(
    patient,
    measurement,
    mom_di,
    post = "After",
    group = condition,
    mid_improvement = 8,
    target = "group",
    effect = "between"
  )

cs_results_groupwise_between
summary(cs_results_groupwise_between)
plot(cs_results_groupwise_between)


# Compare all groups to a predefined reference group with frequentist appraoch
cs_results_groupwise_between_freq <- antidepressants |>
  cs_anchor(
    patient,
    measurement,
    mom_di,
    post = "After",
    group = condition,
    mid_improvement = 8,
    target = "group",
    effect = "between",
    bayesian = FALSE
  )

cs_results_groupwise_between_freq
summary(cs_results_groupwise_between_freq)
plot(cs_results_groupwise_between_freq)