Multi show exhibitor retention and whether the second show protects the first
Multi show exhibitor retention is genuinely higher, at 91 per cent against 72 per cent in a typical portfolio, though the raw gap mostly reflects who buys a second show. Match multi show accounts to single show accounts on space, tenure and category using propensity scores, and a nineteen point gap falls to about four.
Slide eleven of the portfolio review says exhibitors holding two or more shows renew at 91 per cent, against 72 per cent for exhibitors holding one. Nineteen points. The cross-sell budget gets approved on the strength of it, and the number goes into the strategy deck and stays there for three years.
The number is real. Somebody pulled it correctly from the warehouse and it will reproduce, assuming the team works from a written retention rate definition, which is G5's subject. The claim attached to it, that selling a second show protects the renewal on the first, does not follow from it at all, and the gap between those two statements is where a lot of portfolio strategy quietly lives.
Where did the nineteen points come from?
Start by writing down every story consistent with the observation, because at least three are.
The second show causes better retention. An exhibitor with stands on two of your shows has more people internally invested, a bigger annual commitment to defend, and a rep relationship that spans two teams. Plausible.
Better retention causes the second show. An exhibitor who is delighted with show one, growing, and confident in the category buys a second stand for the same reasons they were always going to renew. The second show is a symptom.
Something else causes both. Size is the obvious candidate. A company taking 60 square metres has a trade show programme, a dedicated events manager, and the budget to run several. A company taking 9 square metres has one show and a marketing manager who also does the website. Size drives the second purchase and it drives renewal, and neither causes the other.
All three can be operating at once. What the nineteen points cannot tell you is the mix, and the mix is the entire decision, because only the first story justifies spending cross-sell money to protect retention.
What you are actually trying to estimate
The question has a precise form. For an exhibitor holding one show, what would their renewal probability have been if they had also held a second one.
You never observe that. Every account is either in the second-show group or not, and the counterfactual is unavailable. This is the standard setup in causal inference, and the clean solution is randomisation, which is unavailable here for an obvious reason: you cannot randomly assign exhibitors to buy a second stand.
So the estimate has to come from the data you have, which means finding, for each multi-show account, one or more single-show accounts that resemble it closely enough on everything you can observe that the second show is the main remaining difference.
Doing that on one variable is easy. Compare multi-show and single-show accounts within the 30 to 40 square metre band, and the size story is dealt with. Doing it on eight variables at once is where it falls apart, because the cells empty out. Space band, times tenure band, times category, times country, times sponsorship attach gives you hundreds of cells and most of them contain nobody.
The propensity score, and the word doing the work
Rosenbaum and Rubin solved the dimensionality problem in Biometrika in 1983. Their propensity score is the conditional probability of receiving the treatment given a vector of observed covariates, and their result is that adjusting for that single number is sufficient to remove the bias due to all of the observed covariates.
Eight variables collapse into one. Fit a model predicting the treatment, use the predicted probability as a matching key, and two accounts with the same score are comparable on the whole covariate set even where they differ on individual variables.
The word observed is doing all of the work in that sentence, and it is worth sitting with before you build anything. The result is a statement about the covariates you put in the model. It says nothing about the ones you did not measure, and in an exhibition business the unmeasured ones are not exotic. Whether the account has an internal champion. Whether they have a personal relationship with a rep who has been there eleven years. Whether their 2024 hall position was good. None of that is in your warehouse and all of it plausibly drives both buying a second show and renewing the first.
Matching on your own file
The mechanics are less work than they sound. Here is the shape on a portfolio of 3,400 exhibiting accounts, 2,760 holding one show and 640 holding two or more. Splitting those two groups correctly needs the resolved parent view that G25 builds, because an unresolved file puts the same company in both.
Raw first: renewal on the original show is 91.2 per cent for the multi-show group and 72.4 per cent for the single-show group. That is the 18.8 point gap from slide eleven.
Now look at what the two groups are. Mean space on the original show is 41 square metres for multi-show accounts and 19 for single-show. Median tenure is 6 editions against 3. Sponsorship attach is 44 per cent against 12 per cent. These are not similar populations and no honest comparison survives that.
Fit a logistic regression predicting whether an account holds a second show, using space on the original show, tenure in editions, product category, country, sponsorship attach, service ticket count, and renewal outcome in the prior edition. Freeze every covariate at the point immediately before the second show was purchased, which matters more than it sounds and is the subject of the timing paragraph below. The predicted probability is the propensity score.
Then match. For each multi-show account, find the closest single-show account by propensity score, inside a caliper so that a bad match is refused rather than forced. Of the 640, 71 find nothing inside the caliper. Those 71 are almost all at the top of the size distribution: accounts at 200 square metres and above for which no single-show comparison exists anywhere in the portfolio, because every exhibitor that large already holds three of your shows. That leaves 569 matched pairs.
Reporting those 71 is not a footnote. It is the finding that the estimate does not apply to your largest accounts at all, and if someone extends the conclusion to them anyway, the extension has no evidence behind it.
Matched result: 89.6 per cent renewal among the 569 matched multi-show accounts, against 85.4 per cent among their matched single-show controls. A gap of 4.2 points.
Nineteen points raw. Four matched. Roughly three quarters of the headline number was the population, and one quarter is a candidate for being an effect.
How do you check the match before believing the answer?
A matched estimate is only worth the balance it achieved, and balance is checkable in about twenty minutes.
The standardised difference is the tool. Austin's 2011 introduction to propensity score methods in Multivariate Behavioral Research sets out the four ways of using a propensity score: matching, stratification, inverse probability of treatment weighting, and covariate adjustment. It uses the standardised difference to compare the mean of continuous and binary variables between treatment groups, notes there is no universally agreed criterion for what counts as an important imbalance, and reports that under 0.1 has been taken to indicate a negligible difference.
Compute it on each covariate before and after matching. On this file, space starts at 0.94 and falls to 0.06. Tenure starts at 0.71 and falls to 0.04. Sponsorship attach starts at 0.33 and falls to 0.09. Every covariate lands under 0.1, so the matched sets are comparable on what you measured.
If one covariate refuses to balance, the model is misspecified for it, usually because the relationship is not linear in the logit. Add a squared term or band the variable and refit. Do not proceed with an unbalanced covariate and a note in the appendix, because the whole method rests on the balance and nothing else.
Publish the before-and-after table with the result. A matched estimate without a balance table is an unaudited number, and the person who has to defend the cross-sell budget next year will need the table more than they need the estimate.
What four points buys, and what it cost to find out
Four points is not nothing, and it is worth pricing properly before anyone gets deflated.
On the 569 matched accounts, 4.2 per cent is 24 additional renewals on the original show. At an average first-show value of 26,000, that is 624,000 of retained revenue in one cycle, and it recurs while those accounts stay. Set that against the cost of running the cross-sell programme, which is rep time plus whatever incentive sits on second-show acquisition, and the comparison is a real one that a finance lead can work with.
What changes is the argument. Nineteen points justified a programme aimed at everyone. Four points justifies a programme aimed at accounts where the second show is actually sellable, and it moves the primary justification back to where it belonged, which is the revenue from the second stand itself. The retention effect is a genuine second-order benefit worth about 624,000, and the difference between selling that internally and selling nineteen points is the difference between a programme that survives audit and one that does not.
Where this stops
Unobserved confounding is the limit and it does not go away with more covariates.
Ask the question directly: how strong would an unmeasured variable have to be, in its association with both buying a second show and renewing the first, to move 4.2 points to zero. On a gap that small the answer is usually not very strong, which is an uncomfortable and honest thing to put in the deck. A long-tenured rep relationship could plausibly account for several points on its own.
Timing is the second trap and it is the one I have seen break these analyses most often. If you measure covariates at the point of the renewal decision rather than immediately before the second show was bought, you have let post-treatment information into the matching model, and the estimate is contaminated in a direction you cannot sign. Freeze the covariates before the treatment. Always.
The third is that this measures a portfolio average over a specific period. If your cross-sell in that window went almost entirely to one category, the estimate is about that category wearing a portfolio label. And multi year contracts sit underneath both renewal figures, changing what a renewal even is for the accounts that hold them, which is G38's argument.
Pull the two groups this week and compute nothing more than the means. Mean space, median tenure, and sponsorship attach rate for multi-show accounts against single-show accounts. If those differ by more than a standard deviation, and they will, you already know the headline gap on slide eleven is mostly population and you can say so before anyone builds a model.
Questions people ask about multi show exhibitor retention
- Do exhibitors on two shows renew better than exhibitors on one?
- They do, and most of the difference is who they are. In a portfolio of 3,400 accounts, renewal runs 91.2 per cent for multi show and 72.4 per cent for single show. Mean space is 41 square metres against 19, median tenure 6 editions against 3. Those are not comparable populations.
- How do you test whether a second show protects renewal on the first?
- Match. Fit a logistic regression predicting whether an account holds a second show, using space, tenure, category, country, sponsorship attach and prior renewal, all frozen before the second show was bought. Use the predicted probability as a matching key inside a caliper. Rosenbaum and Rubin showed in Biometrika in 1983 that adjusting for that one number removes the bias from all observed covariates.
- How do you check a propensity score match worked?
- Compute the standardised difference on every covariate before and after matching. Space might start at 0.94 and fall to 0.06, tenure at 0.71 falling to 0.04. Austin's 2011 paper reports that a standardised difference under 0.1 has been taken to indicate a negligible difference. Publish the before and after table alongside the result.
Related reading
- Writing an exhibitor retention rate definition two analysts can both reproduce
- Cross sell across an event portfolio starts with one account view