Every number here, and how it was arrived at

Favorable is the top two points of a five-point scale. A respondent scoring 4 or 5 counts as favorable and a respondent scoring 1 or 2 counts as unfavorable. Top-two is the healthcare convention and the only definition for which external norms exist, which is the entire reason to keep it.

What matters more is the base. Every agreement and frequency scale in the library carries an explicit “Does not apply to me” option, and that option scores null. It is not folded into the midpoint and it is not quietly removed from the denominator. It is counted on its own and reported beside the item, so a favorable percentage always says what it is a percentage of. Dropping not-applicable answers from the base silently is the most common way a survey report misleads without lying, and it can move a headline several points.

Skipped and never-shown are also different things and are counted differently. Routing means a respondent is frequently never presented with an item at all, so the item statistic takes the number of respondents the item was actually shown to and reports the skip rate against that. Without it, an item is reported as fully answered when half the sample never saw it.

favorable      score >= 4
unfavorable    score <= 2
n (the base)    respondents with a scored answer
                not-applicable and skipped are counted apart from n
mean and sd     reported beside every favorable percentage

Mean and standard deviation appear beside the favorable percentage every time, because a top-box figure on its own hides a medical staff that has split in two, and a bimodal medical staff is the single most actionable finding a survey can produce.

A composite is the mean of each respondent’s own item mean. It is not the mean of the item means, and the distinction is not pedantic.

The two agree only when every respondent answered every member item, which in a routed instrument never happens. Averaging the item means gives extra weight to the physicians who answered the fewest items, which is exactly backwards: the respondent who worked through the whole instrument should not count for less than the one who answered two questions and closed the tab.

A minimum is enforced per respondent rather than per cut. Someone who answered one of the five engagement items receives no engagement score at all, because a composite built from a single item is an item wearing a composite’s name.

Six composites, 25 member items between them. Compensation fairness is on that list, and it is the one no other physician engagement instrument reports.

Cronbach’s alpha is computed for every composite in your sample rather than in a norm sample, over the respondents who answered every member item, and printed beside the composite.

alpha = ( k / (k - 1) ) * ( 1 - sum(var of each item) / var(total) )

Below 0.70 the composite is flagged rather than printed as though it were as solid as the others. The flag is information rather than an apology. An engagement composite whose items stop hanging together is usually telling you that the medical staff has divided into groups that no longer answer alike, and that is a finding.

The default reporting floor is 5 respondents. Below it, scores are withheld. The floor is printed on every page it applies to, not once in an appendix.

There is exactly one function in the codebase that decides whether a number renders, and every surface calls it. That is deliberate. Suppression implemented separately on each surface is suppression that leaks the first time somebody adds a page in a hurry.

The n itself survives suppression. Knowing that a group is small is not a confidentiality breach, and hiding the count makes a report impossible to audit. A suppressed cell carries a written sentence saying how many responses there were and why the scores are not shown, rather than a blank or a dash.

A floor on its own protects nobody. If a department has twelve respondents, and you can see the department total and every subgroup except one of size three, you can recover the group of three by subtraction. The floor was satisfied and the promise was still broken.

So the check runs across sibling cuts rather than one cut at a time. Where exactly one sibling falls below the floor, the next smallest is suppressed as well. This is the standard complementary suppression rule from statistical disclosure control. It costs a second cell, and it is the difference between a confidentiality promise and a confidentiality claim.

Suppression is correct. Stopping at suppression is not. The respondents exist and their answers are in the file. They belong to a specialty, that specialty belongs to a service line, the service line belongs to a division, and the division belongs to the medical staff. Each of those is a legitimate cut, and each is more anonymous than the one below it.

The taxonomy is therefore a chain rather than a label, and the console walks up it until a cut clears the floor.

request        ortho_surgery
chain          ortho_surgery -> musculoskeletal -> surgery_all -> all
report at      the first level in that chain whose n clears the floor
label          the level it came from, and why it moved

The escalation is never hidden. Any number that came from a level above the one requested is labelled with the level it came from and the reason it moved. A four-physician specialty gets a real number with an honest label on it instead of an empty row. The number is less specific than the chief of staff wanted and it is enormously more useful than nothing, and a rolled-up number presented as a specialty number would be a lie that poisons the whole report it sits in.

This is not a rare case, it is the ordinary case. In the demo campaign, 332 responses from 923 clinicians landed across 38 reporting groups, and 15 of those groups came back with fewer than 5 responses. Without the chain, that is 15 empty rows in the report a chief of staff paid for.

What is usually sold as key driver analysis is one of two things: a bare correlation between each item and the outcome, ranked, or a proprietary index nobody will describe. The second is not analysis, it is a brand. The first is badly wrong whenever the predictors correlate with each other, which on an engagement survey they always do.

Leadership trust and leadership communication correlate at around 0.8. A correlation ranking lists them as the top two drivers, and a system spends a year working on a single underlying thing it has counted twice.

FMF Pulse uses Johnson’s relative weights, which is the standard published answer to exactly this problem. The predictors are projected onto an orthogonal basis maximally related to the original set, and the model R-squared is partitioned across them. The weights sum to R-squared, so “this driver accounts for 14% of explained variance” is a true sentence rather than a figure of speech.

Johnson JW. A heuristic method for estimating the relative weight of predictor variables in multiple regression. Multivariate Behavioral Research 2000;35(1):1-19.

Rxx    =  V L V'                    eigen decomposition
Lambda =  V L^(1/2) V'              orthogonal counterpart
beta   =  Lambda^-1 r_xy            outcome on the orthogonal set
eps_j  =  sum_k Lambda[j][k]^2 * beta[k]^2
R^2    =  sum_j eps_j

R-squared is printed beside the drivers, always. If the model explains 31% of the variance in engagement, the reader is owed the fact that 69% of it is something this survey did not ask about. It is the single most honest figure on the page and it is the one that never appears in a vendor report.

The model refuses to run below 60 complete responses and warns below ten complete responses per predictor, which is the conventional floor for a stable regression solution. A driver ranking computed on forty responses and twenty items is noise with an order, and shipping one anyway is why hospital executives have learned to distrust the whole genre.

Complete cases are taken listwise here, unlike the item statistics, because a regression needs every row to have every column. Where the predictor correlation matrix comes back near singular, usually because two items are measuring the same thing, the model says so and asks that the neighboring items be read as one, rather than returning a confident wrong answer. Where negative eigenvalues have to be clamped to zero, that is reported too, along with the fact that it slightly understates total explained variance.

The action quadrant splits on the median of each axis rather than on a fixed threshold, because a fixed threshold flags everything or nothing depending on how the organization happens to be doing overall.

Every proportion carries a Wilson score interval at 95%: favorable percentages, burnout prevalence, and the response rate itself.

center =  p + z^2 / (2n)
half   =  z * sqrt( p(1-p)/n + z^2/(4n^2) )
lo, hi =  (center -/+ half) / (1 + z^2/n)

Wilson rather than Wald because favorable percentages in small specialty cuts sit near the ends of the range, where a Wald interval runs past 100% and makes a report look unserious. On the demo campaign, 332 responses from 923 invitations gives a response rate of 36% with an interval of 32.9% to 39.1%.

Two means are compared with Welch’s t test rather than Student’s, because survey groups are never equal in size or in variance and pooling pretends otherwise. Degrees of freedom use the Satterthwaite approximation, and p values come from the regularised incomplete beta function, evaluated by a modified Lentz continued fraction and checked against known values in the engine test rather than trusted because it looks right.

A physician survey report compares dozens of segments at once. Without a correction roughly one in twenty of those comparisons is significant by accident, and somebody reorganizes a department because of it. Comparisons are corrected with the Benjamini-Hochberg procedure controlling the false discovery rate at 0.05, and the correction is stated wherever the comparisons appear.

sort p ascending, m tests
largest k with  p(k) <= (k / m) * q     q = 0.05
reject every hypothesis up to and including k

The margin of error is reported at 95% on a fifty-fifty split, with the finite population correction applied.

moe = z * sqrt( p(1-p) / n ) * sqrt( (N - n) / (N - 1) )
      z = 1.959964,  p = 0.5,  N = invited,  n = responded

The correction matters enormously here and is almost always left out. A medical staff is a small closed population, not a sample drawn from an infinite one. One hundred and eighty responses from two hundred physicians has a margin of error near zero; the uncorrected formula reports plus or minus seven points and makes a census look like a poll.

On the demo campaign, 332 of 923 gives 4.3 points.

The target is a governance threshold and not a scientific one, and the product says so wherever it is shown. The defensible claim is about representativeness, which is measured separately: the response rate within each level of a roster attribute is compared against the overall rate, and any group more than ten points below it is named. A 60% response that is entirely employed primary care is worse than a 30% response spread evenly across the medical staff, and the response rate alone cannot tell you which one you are holding.

Burnout is measured with the validated single-item measure, shipped verbatim, and reported as a prevalence with its rule attached.

Rohland BM, Kruse GR, Rohrer JE. Validation of a single-item measure of burnout against the Maslach Burnout Inventory among physicians. Stress and Health 2004;20(2):75-79. Replicated in a large physician sample by Dolan ED et al., Journal of General Internal Medicine 2015;30(5):582-587.

Verbatim is not a stylistic preference. It is the only condition under which the published validation and the national prevalence comparisons apply, and it is the reason a defensible burnout rate can be reported without charging a client for a proprietary inventory license.

A respondent selecting the third anchor or higher is counted as burned out. The fourth and fifth anchors are reported separately as the severe share. That rule is printed next to the number every time the number appears.

Burnout is never reported as a percentile. A percentile would tell a chief medical officer that they sit in the 62nd percentile for physicians wanting to leave medicine, which is a sentence that should never have been written.

Response quality is flagged at submit and reported. Nothing is silently dropped.

straight_line   8 or more scored answers, all identical
low_variance    exactly 2 distinct values across 15 or more answers
speeding        under 2 seconds per item, faster than the stems read

A vendor that quietly removes straight-liners is making an editorial decision about your data without telling you. The console shows the count and lets an analyst choose to exclude them, which is a different and better thing. The same applies to the decline option on sensitive items: it scores null so it never moves a mean, and the decline rate is itself reported, because an inclusion item with fourteen per cent declines is telling you something no favorable percentage will.

The engine refuses rather than guesses, and says why in a sentence a reader can act on. Three of those refusals live in the year-over-year comparison alone.

either side below the reporting floor   no comparison, floor named
item wording changed, no bridge        no trend until the ancestor is
                                       recorded and the change is
                                       reported as a caveat on the line
prior year is an imported aggregate    difference shown, plainly
                                       labelled as not tested

The projected final response rate is labelled a projection everywhere it appears, and its basis is printed with it: no further reminders, and a 40% daily decay in arrivals, which is the observed shape of an unprompted survey tail. A reminder resets the curve and the projection with it. The failure mode of showing a projection is a client who stops sending reminders on the strength of it, so the projection carries its own warning.

And the largest limit of all is the one at the top of every page. Short of the target the console shows you everything it has, states what it would take to reach the threshold the campaign set, and refuses to call the result a mandate.

This instrument can be answered on a screen, over the phone, or read aloud in a browser. Those are not the same measurement, and the differences are one-sided.

The largest is recency. When five options are read aloud, the last one heard is over-chosen; one study of interactive voice response measured the effect at over ten points. This scale is read from strongly disagree to strongly agree, so the pull is toward the favorable end, which is the direction nobody questions. Alongside it: unflattering answers are under-reported, acquiescence and extreme positive responding rise, satisficing rises, and physicians who choose to phone in are not a random sample of a medical staff.

No correction is applied. Reversing the spoken scale to cancel recency would make the spoken instrument a different instrument from the typed one, which trades a measured bias for an unmeasured one. Instead the mode is recorded on every response, the split is printed on the field monitor, and the response export carries a column so the cut can be made in your own tools. Cut by mode before reading a close result.

A driver weight says an item moves together with an outcome inside one year of one organization’s responses. It does not say that moving the item moves the outcome.

The practical consequence is a warning worth stating plainly: acting on an item because it scores high in the driver model, without changing anything a physician actually experiences, moves the item and nothing else. Gaming an item is the failure mode this product is least able to detect and the one a target makes most tempting, which is why the target screen prints its assumptions beside the plan rather than in an appendix.