All-in-one tables for regressions, with each modelled effect beside its observed one

Fits one regression model per column and returns a tabxplor table of the per-family effect measure — a linear mean difference (gaussian), odds ratios (binomial), incidence-rate ratios (poisson), one odds-ratio column per outcome category (nominal 3+ level), a cumulative odds ratio (ordinal) — one row per predictor level, grouped by predictor, with the observed (crude) effect beside each adjusted one. Each cell stores its estimate, interval and p-value, so the table prints with stars, greys what is not significant, and exports like any tabxplor crosstab.

Usage

tab_reg(
  data,
  outcome,
  predictors = NULL,
  tab_vars = NULL,
  wt = NULL,
  family = "auto",
  link = "auto",
  measure = "auto",
  effect = "auto",
  outcome_level = NULL,
  trials = NULL,
  empirical = TRUE,
  n = NULL,
  color = "measure",
  color_signif = NULL,
  stars = TRUE,
  ref = NULL,
  multiplier = "2sd",
  shape = NULL,
  stats = "auto",
  conf_level = NULL,
  na = c("drop_by_outcome", "drop_by_model", "drop_all", "keep_for_predictors"),
  display = NULL,
  digits = 0,
  cleannames = NULL,
  subtext = "",
  caption = NULL,
  ...
)

Arguments

data

A data frame, or a prebuilt survey design (survey::svydesign()). A design's weights, clustering, stratification and calibration drive the estimation, and wt is ignored.

outcome

<tidy-select> The outcome variable(s) — bare names, quoted names, or any selection helper, exactly as in tab() — or a model formula (the escape hatch; leave predictors unset). Several names give one effect column per outcome; with a predictors list, a single name is required. reg_formulas() shows what was fitted.

predictors

<tidy-select> The predictors of one model — or a named list, one model per element, its name labelling the column, each element selected on its own (list(m1 = c(race, age), m2 = starts_with("inc"))), which is how models are compared. Leave NULL when outcome is a formula. A bare name is a column of data first, then an object, so a variable holding names works without all_of().

a*b is an interaction, R's own spelling, bare or quoted — "a's effect, allowed to vary with b". Two categorical variables give one row per cell of the pair; a continuous a gives its slope within each level of b. An interaction supplies both its variables, so do not list them beside it, which is what makes "with and without" an ordinary model comparison. a:b, which drops the main effects, is refused. See the regression vignette.

tab_vars

<tidy-select> Optional. One grouping variable — the same argument as tab()'s: one sub-table per group, the same model(s) fitted within each level. Two readings of "does this effect hold in every subgroup?" come with it: color = "between_groups" colours and tests each effect against the first group's, row by row, and stats = c(..., "group_interaction") adds the aggregated test, once per predictor. For an interaction between two PREDICTORS of one model, write it in predictors as a*b instead.

wt

<tidy-select> Optional. One weight column, switching to design-based survey estimation (survey::svyglm()). For clustering, stratification, a finite-population correction or calibration, build the design with survey::svydesign() and pass it as data. See vignette("tabxplor-weights").

family

The model family, resolved per outcome so several outcomes with different families can share one table. "auto" (default) detects each one and says so: a binary outcome gives "binomial", an ordered 3+ level "ordinal", a nominal 3+ level "multinomial", any other numeric "gaussian" — a genuine count is yours to name. Or set it: "gaussian" (linear), "binomial" (logistic), "poisson" / "quasipoisson" (counts), "multinomial", "ordinal". A scalar applies to every outcome; a vector aligned to outcome, or a named vector keyed by outcome (e.g. c(income = "poisson", satisfied = "binomial")), sets one family per outcome. Mixed families need a character predictors.

It answers one question — what kind of number the outcome is — and never picks a link behind your back: on a binary outcome family = "poisson" is refused, naming the two things it could have meant, link = "ratio" and measure = "ratio".

link

Which measure the model estimates — the only argument that changes the model. A link is a measure (the one a model gives you directly), so it takes measure's own words:

• "auto" (default) — the family's own: logistic for a binary outcome, linear for a quantity, Poisson for a count.

• "odds_ratio" — the logit fit (binomial, multinomial, ordinal).

• "ratio" — the log link: the modified Poisson on a binary outcome (a conditional risk ratio), Poisson on a count, Poisson pseudo-likelihood on a continuous one.

• "difference" — the identity link; on a binary outcome the risk difference. It can fail to converge, and the linear probability model then takes over, with a message.

Reach for it when you want the model's coefficient to be that measure; to report a measure without changing the model, set measure instead. "log" is the one word the two arguments do not share: here the log link, on measure a spelling of "raw_coefficient".

measure

Which measure of deviation is reported — a deviation being how far a group sits from the reference, the measure which of the ways of expressing it you read. The one argument most readers ever set, and the one that never changes the model. "auto" (default) is the model's own. The full word is canonical, the discipline's acronym a synonym:

• "odds_ratio" ("OR") — the odds of the outcome, times what.

• "ratio" ("RR", "IRR", "RoM") — how many times as likely, as frequent, as large. Reach for it when the outcome is common, where an odds ratio is far from the risk ratio people hear in it, and because a risk ratio stays comparable across nested models.

• "difference" ("RD", "diff") — how much more, in the outcome's own units.

• "raw_coefficient" ("coef", "log", ...) — the model's own coefficient, un-transformed.

On an ordered outcome the first three read the whole predicted distribution rather than one category, so they stay in one column: Somers' D and the win ratio. Where the measure IS the model's own it is read off the coefficients, otherwise from its predictions — so it is available whichever model you fit. "auto" never lands on a predicted odds ratio, a specialist quantity to be asked for by name. Call reg_measures() to see what an outcome offers.

effect

Where the reported number comes from, once the model and the measure are fixed. "auto" (default) needs no thought: the model's own coefficients when the reported measure is the model's, its predictions otherwise. The other values name a reading:

• "conditional" — read off the coefficients ("holding the other predictors constant"). Only where measure is the model's own; otherwise the abort names the two cures.

• "marginal" — the average marginal effect: the model's prediction for every observed person, averaged. Comparable across models (Mood 2010), and always available.

• "at_reference" — the same at one profile, every other predictor at its reference.

The contrast is a marker on the measure in the header, so the acronym stays the one thing to look up: Model_OR, Model_mRR, Model_refRD. The observed companion carries the measure alone (Obs_RR), a univariable effect having no adjustment to be marginal over.

outcome_level

Which level of the outcome to single out, as a named vector keyed by outcome name — outcome_level = c(married = "Married"). It is the twin of ref: ref names the level you compare AGAINST, outcome_level the one you MODEL.

• binomial: the level whose probability is estimated; it becomes the column header. Defaults to the outcome's first level. A 0/1 numeric outcome may be named either way.

• multinomial: the baseline category the other categories' columns are compared to.

• ordinal and numeric outcomes: refused, with the reason.

trials

Grouped-binomial (summed-score) outcomes only. The number of items behind the score, fitting cbind(score, trials - score) as a binomial. NULL (default) fits an ordinary binary logit; an integer (or a vector named by outcome) sets the item count; TRUE uses each outcome's observed maximum. Requires family = "binomial".

empirical

Show the observed, unadjusted (crude) effect beside each modelled one — the same quantity fitted with a single predictor, on exactly the same people. The distance between the two is what adjustment changed, read left to right; it is the feature the package exists for. TRUE (the default) or FALSE; three spellings say where it goes, and in every one but "no" it is stored in the obs field and read by $obs, color = "adjustment", forest_plot() and the hover tooltip:

• TRUE — a crude column beside the model one, except where that would double a table already wide (tab_vars groups, a 3+ level outcome), which take "tooltip".

• "column" — always the column, per outcome category if that is what it takes.

• "tooltip" — computed, printed nowhere. The narrowest table.

• "cell" — inside the model cell, (1/1.69) 1/1.63***. display overrides it.

The two columns are the same column twice: same estimand, same colour ladder, one legend block. Each cell prints the effect with the level it sits on — the observed percentage or mean on the crude side, the adjusted prediction on the model side. A continuous predictor has no levels, so its crude cell is the univariable slope, which assumes linearity: check that with shape first.

n

How many people the table is about. NULL (default) reads options(tabxplor.n) — "range", which adds an n column holding the unadjusted count behind each predictor level, on the model's own complete cases. Where several models rest on different people it prints the whole range (5 139-9 862), so an unequal base cannot pass unnoticed; "min" shows the smallest count only, "no" none. Continuous predictors are left blank: their count is the model N, the first footer row.

color, color_signif

Colouring of the effect cells. color = "measure" (default, TRUE equivalently) grades each cell on its own measure, so the ladder follows what the column estimates; color = FALSE turns colouring off. color_signif is the significance policy — NULL (default) is "grey_non_signif" here, where tab() defaults to "ignore".

What is left to choose is what each effect is compared to. color is positional, c(text, background), so color = c("measure", "adjustment") answers "how strong is this effect?" and "how much did the model change it?" in one glance:

• "adjustment" — how far each modelled effect sits from its observed (crude) counterpart: what adjusting for the other predictors did to it. It turns empirical = TRUE on. Set its thresholds with set_color_breaks() (adj_ratio, adj_diff, adj_diff_std).

• "between_groups" — with tab_vars, how far each group's effect sits from the first group's: a per-predictor reading of effect modification.

The two are mutually exclusive, and each always tests its own gap: a gap whose interval covers zero is greyed whatever color_signif says — so a cell can be filled while neither estimate carries a star, which is correct rather than odd. A conditional odds ratio is not collapsible, so there the colours stay descriptive and are not tested. Read a coloured cell as "adjustment moved this effect", not as "this variable is a confounder". See the regression vignette.

stars

Logical (default TRUE for regression tables, where significance stars are standard). When FALSE, the per-cell p-value is dropped and no stars are shown (colours still read the interval).

ref

The reference every effect is measured from — one argument, one meaning per kind of predictor. For a factor it is the level the others are compared against (a level name, or "first" (default) / "last"); for a continuous predictor the value it is anchored at (a number, or "mean" (default), "median", "min", "max").

Anchoring a continuous predictor does not change its own effect — a slope is the same wherever you start reading it from — but it does move the Constant row; its own row says where the anchor sits, per SD/13.5 (at mean/42.4). The default is the mean because zero is usually outside the data: nobody is 0 years old.

ref, multiplier and shape share one grammar: a value on its own is the default for every predictor it can apply to, a named one overrides that variable — ref = c("median", "last", race = "Black").

For the level of the outcome, see outcome_level: ref names the level you compare AGAINST, outcome_level the one you MODEL.

multiplier

How a continuous predictor's effect is scaled — the unit its row reports. One unit is rarely a readable amount (a one-year change in age barely moves the odds), so the default is two standard deviations: roughly the span a binary predictor's own contrast covers, which is what makes a continuous row and a factor row comparable at a glance (Gelman 2008). Values: "2sd" (default), "sd", or a number of units (10 = per decade). Same grammar as ref: multiplier = c("sd", age = 10).

The estimate, its interval and the observed companion all scale together; the p-value does not move. Because the default is not 1, a continuous predictor's Model_* cell does not equal exp(coef(glm(...))) unless you pass multiplier = 1.

shape

How a continuous predictor enters the model, when one straight line is not enough. The Linearity footer row and the little curve drawn in the predictor's n cell tell you whether a line is enough; this argument is how you fix it without leaving the framework. Same grammar as multiplier and ref — shape = "quintiles" cuts every continuous predictor, shape = c(age = "quadratic") only that one, and anything unnamed stays linear:

"linear"

one slope (the default).

"quintiles" / "quartiles" / an integer k

cut into k quantile groups, so the predictor becomes an ordinary factor: one estimate per group, its own observed companion, counts and colours per group — the non-linearity becomes visible in the printed numbers. Start here; it is the most readable answer.

"sd_bands"

cut at the mean and one standard deviation either side — the classic low / average / high reading, whose cut points mean the same thing across sub-samples, where quantiles move with each one. The bands are not balanced: prefer quantiles when the group sizes matter.

"quadratic"

adds a curvature term, so the predictor takes two rows — the slope at the mean, and whether it flattens or accelerates away from it.

"log" / "sqrt"

fit log(x) / sqrt(x) instead of x — diminishing returns.

Example: shape = c(age = "quadratic", income = "log"). The observed companion is fitted with the same shape, so the comparison stays like with like. It is the vocabulary tab() takes (shape_numeric_var()) plus "quadratic", which is a model term.

stats

The statistics shown in the model-summary footer (one block per model). "auto" (default) uses the per-family set — R square, adjusted R square, the overall F-test and the residual SD for a linear model, the likelihood-ratio test against the null model, McFadden's pseudo-R square, AIC and BIC otherwise — plus the default model checks (see below). A weighted model shows the survey-appropriate set. Pass a character vector to pick them: "n", "lr_null", "mcfadden_r2", "aic", "bic", "phi", "r2", "r2_adj", "f_model", "sigma", "global", "interaction", "group_interaction", "linearity", "proportionality", "dispersion", "influence", "collinearity"; "all" for everything, or NULL / FALSE / "no" to hide the footer.

Model comparison happens by default wherever it means anything — when predictors is a list of several models. Where each nests in the next, every model is tested against the previous one; otherwise each against the first. "compare_sequential" and "compare_baseline" (optionally naming the model) override that, and naming any footer statistic drops it.

Three tests are asked for by name: "global", one overall test per predictor — "is this variable associated with the outcome at all?", which a block of stars against a reference category cannot answer; "interaction", whether each crossed pair in predictors is real or the additive model is enough; and "group_interaction", one aggregated effect-modification test per predictor across tab_vars groups.

conf_level

Confidence level for the intervals. NULL (default) reads options(tabxplor.conf_level) — 0.95. It drives every interval, the significance stars, the greying under color_signif and the model-versus-observed gap interval, and each column records the level it was built at.

na

Which rows each model is fitted on — the grain at which missing values are dropped. "drop_by_outcome" (default) gives every model of one outcome the same complete-case population, which is what makes the comparisons honest: the observed columns are computed on exactly the model's rows, and nested models get equal N. "drop_by_model" lets each model use its own complete cases — more rows, at the price of comparability. "drop_all" shares one population across the whole call.

"keep_for_predictors" drops nothing but a missing outcome: every predictor keeps its missing values as an ordinary NA level, with its own row, count and effect — often the fastest way to find out whether non-response is itself patterned. A number has no level to put them in, so a numeric predictor that has any is cut into bands.

display

What each effect cell shows — tab()'s display grammar, same names, same meaning, on every family and on the crude column as well as the model one. NULL (default) shows the plain estimate, or, with empirical, the estimate with the level it sits on beside it. The whole vocabulary — the named layouts, the {} templates and the per-token precision "{est:3} ({base:1})" — is in tabxplor-display; the ones this table uses most are "est_ci" (1/2.22*** [1/2.47;1/1.99]), "est_base" (1/2.22 (32.8%)), "est_coef" and "base_est_mdiff" / "base_est_mratio", which read the same comparison the other way. The Constant row holds the quantity the column's effects operate on, so it is read in one step: a baseline odds on an odds-ratio column, the level itself on an additive one.

display is post-hoc: every quantity it can name is already stored, so choosing a layout never triggers a computation and never changes a number — set_display() on a built table gives the same result. It never changes the estimand, which is measure's job alone.

digits

The number of decimals. A single integer sets every cell (0, the default, means "each measure's own"), and a measure finer than the level it sits on keeps its own precision. Name a display field to set just that one, an aside included — digits = c(ratio = 3), digits = c(1, or = 3); a template may carry its own, display = "{est:3} ({base:1})". Left alone, a column reading in the outcome's own units (a mean, a mean difference) follows that outcome's magnitude, so a six-figure salary prints no decimals and a rate prints two; digits = always raises it back.

cleannames

Logical. If TRUE, strips numeric prefixes from factor levels for display. Uses getOption("tabxplor.cleannames") when NULL.

subtext

Optional character. A note shown below the table.

caption

A title for the table, stored on it and carried into every export. Without one a regression table titles itself from the model it shows.

...

One rarely-typed argument, plus internal plumbing.

ci_method — how the interval and p-value are computed: the same argument, and the same named-vector grammar, as in tab(), whose fifth slot is this producer's. "wald" (default) matches standard software output and is the only option for weighted models; "profile" uses the profile-likelihood interval and the likelihood-ratio test — more accurate near separation, unweighted binomial / poisson only.

Every argument removed or renamed while tab_reg() was in development is still accepted here, and gives an error naming its replacement rather than R's bare "unused argument". The dot-prefixed names are the jamovi live-cache plumbing, not user arguments.

Value

A tabxplor_grouped_tab (grouped by predictor), one effect column per model / outcome.

Details

To learn what such a table says, read Reading a regression (vignette("tabxplor-reading-a-regression")); to look something up, the regression vignette (vignette("tabxplor-reg")).

New to regressions with tabxplor? A first model needs three arguments: data, outcome and predictors. The model follows the outcome's type — a two-level factor gives logistic odds ratios, a numeric a linear mean difference, a count Poisson rate ratios, a 3+ level factor multinomial or ordinal odds ratios — so you rarely set family by hand.

The estimand is a cascade: family -> link -> measure -> effect, where "auto" means follow from the left. family says what kind of number the outcome is; link which measure the model estimates (a link is a measure, so it takes the same words); measure which one is reported; effect where that number comes from. Set any one and everything to its right re-derives, so most tables set none of them — and the one most people ever set is measure.

A coefficient exists only where the reported measure IS the model's own. Ask for another and it is worked out from the model's predictions instead, averaged over the sample. So measure = "ratio" on a binary outcome gives a marginal risk ratio from the logistic fit, while link = "ratio" fits the modified Poisson and gives its conditional one — two different quantities, and now two different arguments.

Model checks

Beside the fit statistics the footer carries five model checks, each naming an assumption and the instrument that measured it: Linearity and Proportionality (p-values) say whether the estimate means what it claims, Dispersion whether the intervals are wide enough, Influence whether one respondent carries the result, Collinearity why the intervals are wide. Four are shown by default; stats = "linearity" adds the fifth, and shape is the cure for what it flags. reg_check_plots() draws them all. What each one asks, and how to read it: the regression vignette.

Out of scope

tab_reg() covers linear, logistic, Poisson, multinomial and ordinal models, with survey designs. Three families of models are deliberately not supported, and are unlikely to be: survival / Cox models, mixed / multilevel models, and pooling over multiply-imputed datasets. Fit those with their own packages.

The header acronyms

A column header names the measure; the contrast is a marker on it — no marker for a conditional effect, m for a marginal one, ref at the reference profile — and measure = "log" wraps it (Model_mRR, Model_refRD, Model_log(OR)). The observed companion carries the measure alone (Obs_RR).

• OR — odds ratio (binomial, multinomial)

• cumOR — cumulative odds ratio (ordinal)

• RR — risk ratio (binomial, multinomial)

• RD — risk difference (binomial, multinomial)

• IRR — incidence-rate ratio (poisson)

• RoM — ratio of means (gaussian)

• diff — mean difference (gaussian, poisson)

• D — Somers' D (ordinal)

• WR — win ratio (ordinal)

References

Clogg, C. C., Petkova, E. & Haritou, A. (1995). Statistical Methods for Comparing Regression Coefficients between Models. American Journal of Sociology, 100(5), 1261-1293 — the comparison color = "adjustment" implements.

Zou, G. (2004). A Modified Poisson Regression Approach to Prospective Studies with Binary Data. American Journal of Epidemiology, 159(7), 702-706 — link = "ratio".

Altman, D. G. & Bland, J. M. (2003). Interaction revisited: the difference between two estimates. BMJ, 326, 219 — the color = "between_groups" test.

See also

reg_formulas() shows the formula each column was fitted with, and reg_measures() what an outcome can be modelled as. forest_plot() draws the finished table — every effect with its interval, its stars and its colour, and (with empirical = TRUE) the observed effect beside it with the margin of error of the gap. reg_check_plots() draws the model checks. tabxplor-display says what a cell can show, tab() builds cross-tables.

Examples

# The shape table a continuous predictor draws under the footer is noise in an example:
.opt <- options(tabxplor.shape_table = "no")

# Logistic: the odds of being released, adjusted, beside the observed (crude) odds ratio.
tab_reg(car_arrests, "released", c("colour", "checks"))
#> ℹ "released": binary outcome detected -> `family = "binomial"` (logit).
#> This message is displayed once per session.
#> | predictors     | Model fit                    | released |
#> |:---------------|:-----------------------------|---------:|
#> | colour, checks | N                            |    5 226 |
#> |                | Dispersion (robust/model SE) |     1.02 |
#> |                | Collinearity (max VIF)       |     1.01 |
#> |                | Influence (max dfbetas)      |     0.06 |
#> |                | LR vs null                   |   <0.01% |
#> |                | McFadden R2                  |    0.076 |
#> |                | AIC                          |    4 420 |
#> |                | BIC                          |    4 440 |
#> 
#> # A tabxplor tab: 4 × 5
#> # Outcome:        released
#> # Groups:         var [3]
#>   var      levels                             n          Obs_OR        Model_OR
#>                                             <n>     <(obs%) OR>     <OR (adj%)>
#> 1 Constant Reference profile                                      6.51*** (87%)
#> 
#> 2 colour   White                          3 938 (86%)      1         1    (85%)
#> 3 colour   Black                          1 288 (74%) 1/2.11*** 1/1.72*** (77%)
#> 
#> 4 checks   per 3.08 (2SD), at 1.64 (mean)             1/3.71*** 1/3.47***      
#> # Model: logistic regression; OR: odds ratio (vs the reference category); obs%: observed proportion; adj%: adjusted/predicted proportion.
#> # Obs_OR, Model_OR: OR (ref.): 1/4 1/2 1/1.5 1/1.2 1.2 1.5 2 4 [grey: non-significant or under ×1.2]
#> # ***: significantly different from the reference category (in bold) at the 99% confidence level (from 1 for the Constant); **: at the 95% level; *: at the 90% level; no star: not significant.

# \donttest{
# Linear: a mean difference in dollars.
tab_reg(car_salaries, "salary", c("sex", "discipline", "rank"))
#> | predictors       | Model fit                    |    salary |
#> |:-----------------|:-----------------------------|----------:|
#> | sex, discipline, | N                            |       397 |
#> | rank             | Dispersion (robust/model SE) |      1.00 |
#> |                  | Collinearity (max VIF)       |      1.02 |
#> |                  | Influence (max dfbetas)      |      0.22 |
#> |                  | F                            |    <0.01% |
#> |                  | R2                           |     0.447 |
#> |                  | Adjusted R2                  |     0.441 |
#> |                  | Residual SD                  | 22 640.99 |
#> 
#> # A tabxplor tab: 8 × 5
#> # Outcome:        salary
#> # Groups:         var [4]
#>   var        levels              n             Obs_diff           Model_diff
#>                                <n>    <(obs mean) diff>    <diff (adj mean)>
#> 1 Constant   Reference profile   6                                 68 224   
#> 
#> 2 sex        Female             39 (101 002)       0          0    (109 656)
#> 3 sex        Male              358 (115 090) +14 088***  +4 492    (114 148)
#> 
#> 4 discipline A                 181 (108 548)       0          0    (106 248)
#> 5 discipline B                 216 (118 029)  +9 480*** +13 709*** (119 957)
#> 
#> 6 rank       AsstProf           67 ( 80 776)       0          0    ( 79 733)
#> 7 rank       AssocProf          64 ( 93 876) +13 100*** +13 723*** ( 93 456)
#> 8 rank       Prof              266 (126 772) +45 996*** +47 403*** (127 136)
#> # Model: linear regression; diff: mean difference (vs the reference category); obs mean: observed mean; adj mean: adjusted/predicted mean.
#> # Obs_diff, Model_diff: diff in SD (ref.): -0.8 -0.4 -0.2 -0.1 +0.1 +0.2 +0.4 +0.8 [grey: non-significant or under ±0.1 SD]
#> # ***: significantly different from the reference category (in bold) at the 99% confidence level; **: at the 95% level; *: at the 90% level; no star: not significant.

# A count outcome: incidence-rate ratios.
tab_reg(car_arrests, "checks", c("colour", "employed"), family = "poisson")
#> | predictors       | Model fit                    | checks |
#> |:-----------------|:-----------------------------|-------:|
#> | colour, employed | N                            |  5 226 |
#> |                  | Dispersion (robust/model SE) |   1.23 |
#> |                  | Pearson dispersion (phi)     |   1.40 |
#> |                  | Collinearity (max VIF)       |   1.02 |
#> |                  | Influence (max dfbetas)      |   0.07 |
#> |                  | LR vs null                   | <0.01% |
#> |                  | McFadden R2                  |  0.030 |
#> |                  | AIC                          | 18 058 |
#> |                  | BIC                          | 18 078 |
#> 
#> # A tabxplor tab: 5 × 5
#> # Outcome:        checks
#> # Groups:         var [3]
#>   var      levels                n            Obs_IRR          Model_IRR
#>                                <n> <(obs mean) ratio> <ratio (adj mean)>
#> 1 Constant Reference profile 3 200                                1.3   
#> 
#> 2 colour   White             3 938     (1.5)     1            1    (1.5)
#> 3 colour   Black             1 288     (2.1) ×1.41***     ×1.34*** (2.0)
#> 
#> 4 employed Yes               4 111     (1.4)     1            1    (1.5)
#> 5 employed No                1 115     (2.4) ×1.63***     ×1.57*** (2.3)
#> # Model: Poisson regression; IRR: incidence-rate ratio (vs the reference category); obs mean: observed mean; adj mean: adjusted/predicted mean.
#> # Obs_IRR, Model_IRR: IRR (ref.): ÷4 ÷2 ÷1.15 ×1.15 ×2 ×4 [grey: non-significant or under ×1.15]
#> # ***: significantly different from the reference category (in bold) at the 99% confidence level; **: at the 95% level; *: at the 90% level; no star: not significant.

# `measure` reports another measure WITHOUT changing the model: a MARGINAL risk ratio,
# averaged over the sample, still from the logistic fit.
tab_reg(car_arrests, "released", c("colour", "checks"), measure = "ratio")
#> | predictors     | Model fit                    | released |
#> |:---------------|:-----------------------------|---------:|
#> | colour, checks | N                            |    5 226 |
#> |                | Dispersion (robust/model SE) |     1.02 |
#> |                | Collinearity (max VIF)       |     1.01 |
#> |                | Influence (max dfbetas)      |     0.06 |
#> |                | LR vs null                   |   <0.01% |
#> |                | McFadden R2                  |    0.076 |
#> |                | AIC                          |    4 420 |
#> |                | BIC                          |    4 440 |
#> 
#> # A tabxplor tab: 4 × 5
#> # Outcome:        released
#> # Groups:         var [3]
#>   var      levels                             n         Obs_RR      Model_mRR
#>                                             <n> <(obs%) ratio> <ratio (adj%)>
#> 1 Constant Population average             5 226                        83%   
#> 
#> 2 colour   White                          3 938 (86%)     1        1    (85%)
#> 3 colour   Black                          1 288 (74%) ÷1.16*** ÷1.10*** (77%)
#> 
#> 4 checks   per 3.08 (2SD), at 1.64 (mean)             ÷1.40*** ÷1.36***      
#> # Model: logistic regression; mRR: marginal risk ratio (the ratio of adjusted proportions, sample-averaged); obs%: observed proportion; adj%: adjusted/predicted proportion.
#> # Obs_RR, Model_mRR: RR (ref.): ×2 [grey: non-significant or under ×2]
#> # ***: significantly different from the reference category (in bold) at the 99% confidence level; **: at the 95% level; *: at the 90% level; no star: not significant.

# `link` changes the model: the CONDITIONAL risk ratio of a modified-Poisson fit.
tab_reg(car_arrests, "released", c("colour", "checks"), link = "ratio")
#> | predictors     | Model fit                    | released |
#> |:---------------|:-----------------------------|---------:|
#> | colour, checks | N                            |    5 226 |
#> |                | Dispersion (robust/model SE) |     1.00 |
#> |                | Collinearity (max VIF)       |     1.02 |
#> |                | Influence (max dfbetas)      |     0.06 |
#> |                | Wald vs null                 |   <0.01% |
#> 
#> # A tabxplor tab: 4 × 5
#> # Outcome:        released
#> # Groups:         var [3]
#>   var      levels                             n         Obs_RR       Model_RR
#>                                             <n> <(obs%) ratio> <ratio (adj%)>
#> 1 Constant Reference profile                                           84%   
#> 
#> 2 colour   White                          3 938 (86%)     1        1    (85%)
#> 3 colour   Black                          1 288 (74%) ÷1.16*** ÷1.11*** (77%)
#> 
#> 4 checks   per 3.08 (2SD), at 1.64 (mean)             ÷1.26*** ÷1.24***      
#> # Model: modified Poisson regression; RR: risk ratio (vs the reference category); obs%: observed proportion; adj%: adjusted/predicted proportion.
#> # Obs_RR, Model_RR: RR (ref.): ×2 [grey: non-significant or under ×2]
#> # ***: significantly different from the reference category (in bold) at the 99% confidence level; **: at the 95% level; *: at the 90% level; no star: not significant.

# A named list of predictor sets: one column per model, compared in the footer.
tab_reg(car_salaries, "salary",
        list("sex alone" = "sex",
             "+ field"   = c("sex", "discipline"),
             "+ rank"    = c("sex", "discipline", "rank")),
        empirical = FALSE)
#> | Model fit                    | sex alone |   |   + field |   |    + rank |
#> |:-----------------------------|----------:|:-:|----------:|:-:|----------:|
#> | N                            |       397 |   |       397 |   |       397 |
#> | Dispersion (robust/model SE) |      0.87 |   |      1.00 |   |      1.00 |
#> | Collinearity (max VIF)       |           |   |      1.00 |   |      1.02 |
#> | Influence (max dfbetas)      |      0.32 |   |      0.29 |   |      0.22 |
#> | F                            |    0.567% |   |   0.0159% |   |    <0.01% |
#> | R2                           |     0.019 |   |     0.043 |   |     0.447 |
#> | Adjusted R2                  |     0.017 |   |     0.039 |   |     0.441 |
#> | Residual SD                  | 30 034.61 |   | 29 699.35 |   | 22 640.99 |
#> | F vs previous                |           |   |    0.172% |   |    <0.01% |
#> 
#> # A tabxplor tab: 8 × 6
#> # Outcome:        salary
#> # Groups:         var [4]
#>   var        levels                    n `sex alone`  `+ field`   `+ rank`
#>                                <n_range>      <diff>     <diff>     <diff>
#> 1 Constant   Reference profile      6-39  101 002     95 914     68 224   
#> 
#> 2 sex        Female                   39        0          0          0   
#> 3 sex        Male                    358  +14 088*** +14 029***  +4 492   
#> 
#> 4 discipline A                       181                   0          0   
#> 5 discipline B                       216              +9 449*** +13 709***
#> 
#> 6 rank       AsstProf                 67                              0   
#> 7 rank       AssocProf                64                        +13 723***
#> 8 rank       Prof                    266                        +47 403***
#> # Model: linear regression of salary; diff: mean difference (vs the reference category).
#> # diff in SD (ref.): -0.8 -0.4 -0.2 -0.1 +0.1 +0.2 +0.4 +0.8 [grey: non-significant or under ±0.1 SD]
#> # ***: significantly different from the reference category (in bold) at the 99% confidence level; **: at the 95% level; *: at the 90% level; no star: not significant.

# A continuous predictor cut into groups, on French survey data:
tab_reg(questionr_hdv, "cinema", c("qualif", "age"), shape = c(age = "quartiles"))
#> ℹ "cinema": binary outcome detected -> `family = "binomial"` (logit).
#> This message is displayed once per session.
#> | predictors  | Model fit                    | cinema |
#> |:------------|:-----------------------------|-------:|
#> | qualif, age | N                            |  1 653 |
#> |             | Dispersion (robust/model SE) |   1.07 |
#> |             | Collinearity (max VIF)       |   1.03 |
#> |             | Influence (max dfbetas)      |   0.22 |
#> |             | LR vs null                   | <0.01% |
#> |             | McFadden R2                  |  0.160 |
#> |             | AIC                          |  1 904 |
#> |             | BIC                          |  1 958 |
#> 
#> # A tabxplor tab: 12 × 5
#> # Outcome:        cinema
#> # Groups:         var [3]
#>    var      levels                     n          Obs_OR        Model_OR
#>                                      <n>     <(obs%) OR>     <OR (adj%)>
#>  1 Constant Reference profile         49                   6.26*** (86%)
#> 
#>  2 qualif   Cadre                    260 (65%)      1         1    (67%)
#>  3 qualif   Ouvrier specialise       203 (22%) 1/6.63*** 1/9.39*** (23%)
#>  4 qualif   Ouvrier qualifie         292 (25%) 1/5.67*** 1/7.72*** (26%)
#>  5 qualif   Technicien                86 (43%) 1/2.50*** 1/3.73*** (40%)
#>  6 qualif   Profession intermediaire 160 (46%) 1/2.25*** 1/2.76*** (47%)
#>  7 qualif   Employe                  594 (44%) 1/2.39*** 1/3.40*** (42%)
#>  8 qualif   Autre                     58 (47%) 1/2.17*** 1/3.28*** (43%)
#> 
#>  9 age      18 to 34                 368 (62%)      1         1    (62%)
#> 10 age      35 to 47                 457 (55%) 1/1.33**  1/1.43**  (54%)
#> 11 age      48 to 59                 432 (32%) 1/3.39*** 1/3.90*** (32%)
#> 12 age      60 to 97                 396 (18%) 1/7.37*** 1/9.45*** (18%)
#> # Model: logistic regression; OR: odds ratio (vs the reference category); obs%: observed proportion; adj%: adjusted/predicted proportion.
#> # Obs_OR, Model_OR: OR (ref.): 1/4 1/2 1/1.5 1/1.2 1.2 1.5 2 4 [grey: non-significant or under ×1.2]
#> # ***: significantly different from the reference category (in bold) at the 99% confidence level (from 1 for the Constant); **: at the 95% level; *: at the 90% level; no star: not significant.
# }
options(.opt)