Coca-Cola, a rumor and a different constituency

In the four weeks from February 3, 2025, Coca-Cola's share of soda units fell 4.25 percentage points in ZIP codes more than half Hispanic and 0.63 points in ZIPs under 10% Hispanic. Both figures are measured against the same four weeks of 2024 and against each ZIP's own gap to 2024 in the three weeks before. The data are convenience-store sales in 2,469 ZIPs. A regression of the same change on Republican vote share alone gives a slope of +0.24 points per 10 points of Republican share (standard error 0.04), which reads as Democratic-leaning ZIPs cutting back more. With Hispanic share in the regression the slope is +0.07 (0.04), and comparing ZIPs only within the same county it is −0.06 (0.07). This case builds the comparison one step at a time and shows how an omitted variable produced the first number.

A boycott list, a day of action and a false rumor

In late January 2025 a "Latino Freeze" spending boycott circulated on TikTok. Its do-not-shop list named Coca-Cola alongside Target, Walmart and Amazon. Monday, February 3 was a national "Day Without Immigrants", a call to skip work, school and shopping. Between February 5 and 8, posts in English and Spanish claimed that Coca-Cola had called immigration agents on Latino workers at a Texas bottling plant. Fact-checkers at Snopes, PolitiFact and Lead Stories found no evidence, and the plant named in the posts does not exist.

Case 3 found Bud Light's loss concentrated in Republican-leaning ZIPs after the brand's own marketing. Here Coca-Cola had changed nothing about its products, prices or marketing, and the people called on to act were defined by ethnicity rather than party. This case asks whether the method from case 3 still works.

The data: 2,469 ZIP codes and 79 weeks of soda sales

The unit is a ZIP code and a week. The sales are point-of-sale records from convenience stores, with at least two stores pooled in every ZIP in every week.Data. Weekly share of carbonated soft drink units sold, January 1, 2024 to June 30, 2025, from convenience-store point-of-sale data (Dewey). ZIP demographics are Census and election-return estimates supplied with the case files. No survey weights: each ZIP counts once.

Coca-Cola's level differs by neighborhood before anything happens. Over the 57 weeks before February 3, 2025 it averaged 24.0% of soda units in ZIPs under 10% Hispanic and 39.1% in ZIPs more than half Hispanic. A comparison of levels would mix the boycott with that permanent gap, so every comparison below is a change.

Step 1: compare each week with the same week of 2024

Soda sales may have calendar patterns, and the Super Bowl falls in the first days of February. The cleanest correction subtracts, for each ZIP, its share in the same week 52 weeks earlier:

Anything that recurs in the same week every year cancels. The effect for ZIP i is the average of Δ over weeks 0 to 3 minus its average over the three baseline weeks, −5 to −3 (December 30 to January 19):

Weeks −2 and −1 are in neither period. The inauguration fell in week −2, and a dip had begun in week −1.

In these data the calendar correction is small. In every week from January 1 to July 1, 2024 the share in ZIPs 50% or more Hispanic stayed between 38.4% and 39.8%, and subtracting the same calendar weeks of 2024 moves a group's estimate by at most 0.2 points.

The year-over-year version costs some precision: the standard error of the slope rises from 0.04 to 0.05 because the 2024 weeks add their own noise. In return it gives a placebo. Run on 2024, when nothing happened, the same arithmetic gives a slope of +0.03 (0.03) per 10 points of Hispanic share, so the 2025 gradient is not a calendar pattern that recurs every February.³

Step 2: the loss rises with Hispanic share and the first dip comes before the rumor

Figure 2 plots the weekly change for four groups of ZIPs, each relative to its own baseline weeks. Over weeks 0 to 3 the mean change is −0.63 points in ZIPs under 10% Hispanic, −2.44 in ZIPs 10 to 30% Hispanic, −3.45 in ZIPs 30 to 50% and −4.25 in ZIPs 50% or more. The 50%-or-more group is 4.5 points below its baseline in week 0 and 5.0 in week 1, its lowest point.

The groups begin to separate before the rumor. In the week of January 27, which contains the weekend of January 31 to February 2, the 30-to-50% group was 1.5 points below its baseline (95% interval 0.8 to 2.1) and the 10-to-30% group 0.7 below (0.3 to 1.0). The weekly series is noisy for the smaller groups. The 30-to-50% group, with 137 ZIPs, was already 1.0 below in week −2, and the 50%-or-more group, with 102, was 0.5 below in week −1 with an interval of −1.2 to +0.3. The posts of February 5 to 8 fall inside week 0, together with the Day Without Immigrants, so the weekly data cannot give the rumor's own contribution to the 4.5-point drop in that week.

Step 3: one dot per ZIP

As in case 3, each ZIP is reduced to one number, its effect over weeks 0 to 3, and plotted against its Hispanic share. One number per ZIP keeps a ZIP's seven weeks from counting as seven independent observations, which would make the standard errors too small.² The fitted line is

or −0.71 points for each 10 points of Hispanic share (standard error 0.05). A ZIP that is 50% Hispanic lost about 3.6 points more than a ZIP with almost no Hispanic residents. Hispanic share accounts for 10.7% of the variation in ZIP changes, which have a standard deviation of 3.3 points, so the line is a tendency with wide scatter around it.

The menu in Figure 3 refits the line within bands of Republican vote share. The slope is between −0.58 and −0.82 per 10 points in each of the four bands, which is the first sign that the Hispanic gradient does not depend on partisanship.

Step 4: the partisan slope is Hispanic share in disguise

An analyst fresh from case 3 regresses the same effect on Republican vote share instead:

That is +0.24 per 10 points, with a standard error of 0.04. ZIPs that voted more Republican in 2020 lost less, which seems to support the headline "Democratic-leaning ZIPs boycotted Coca-Cola". Put Hispanic share into the regression as well and the Republican slope falls to +0.07 (0.04), 72% smaller and within two standard errors of zero. Comparing each ZIP only with ZIPs in the same county gives −0.06 (0.07) for Republican share and −0.43 (0.11) for Hispanic share.

Columns 1 to 3: ZIP-level OLS with robust standard errors. Column 4: ZIP-by-week panel (weeks −5 to −3 and 0 to 3) with ZIP and county-by-week fixed effects and standard errors clustered by ZIP; only counties with at least two ZIPs enter.

The workbench below runs these regressions in your browser on the 2,469 ZIP changes. Start with Republican share alone, then add Hispanic share, then everything. With all seven regressors the Republican coefficient is −0.12 (0.07) and the Hispanic coefficient −0.63 (0.08). Foreign-born share alone gives −1.16 (0.10) per 10 points because it moves with Hispanic share across ZIPs (correlation 0.73); in the full model it falls to −0.48 (0.17), and the split between the two is imprecise.

Why the short slope is too large

The omitted-variable formula says by how much.¹ Leave Hispanic share out and the slope on Republican share absorbs part of Hispanic share's relation to the outcome:

where δ is the slope from regressing Hispanic share on Republican share. Here δ = −0.25: ZIPs 10 points more Republican have, on average, 2.5 points less Hispanic population. In units of 10 points the formula reads

The product, +0.173, is 72% of the short slope. The bias needs two conditions: Hispanic share must move with the outcome (b_H is not zero) and with the included variable (δ is not zero). Both are negative here, so their product is positive and the short slope is too high. In per-point units the same check is +0.0067 + (−0.0687) × (−0.2519) = +0.0240. The decomposition panel in Figure 4 recomputes it for any set of regressors.

Step 5: Pepsi's gain was about three-quarters of Coca-Cola's loss

Shares add to 100, so every point Coca-Cola lost went to another brand. Per 10 points of Hispanic share, Pepsi's share rose 0.54 points (standard error 0.03) over weeks 0 to 3 while Coca-Cola's fell 0.71, so Pepsi's gain was 76% of the loss. With Republican share also in the regression the figures are +0.49 and −0.69, or 72%. All other sodas together gained 0.17 per 10 points (0.03). The Mexican-import sodas, the brands most closely tied to Hispanic buyers, gained 0.02 (0.01), and none of Sprite, Fanta, Mountain Dew or Dr Pepper moved by more than 0.03.

Step 6: the Hispanic gradient shrank in every block of weeks and was mostly gone by June

Averaged over the 22 weeks from February 3 to June 30, the Hispanic gradient is −0.35 per 10 points (standard error 0.04), half the −0.71 of the first four weeks. Figure 6 splits the period into four-week blocks. The gradient is −0.47 in weeks 4 to 7, −0.31 in weeks 8 to 11 and −0.12 in weeks 20 to 21 (95% interval −0.23 to −0.01). By week 20 the 50%-or-more group, with a mean change of +0.2 points (standard error 0.5), was back to its 2024 pattern, while the 30-to-50% group was still 0.8 points down (0.4). Unlike Bud Light, which case 3 finds had not recovered through 2024, the Coca-Cola gap closed in about five months.

Two boycotts compared

The Bud Light loss followed the Republican vote and had not recovered through 2024. The Coca-Cola loss followed Hispanic share, and its gradient was −0.12 per 10 points by weeks 20 to 21.

What these data cannot show

ZIP characteristics describe places, not people. A ZIP that is 50% Hispanic has residents who are not Hispanic, and the data do not say who bought less Coca-Cola there. A slope across ZIPs compares places with different Hispanic shares; reading it as the behavior of Hispanic shoppers infers individual behavior from group data and can be wrong in either direction.

The stores are convenience stores, and the ZIPs lie mostly in the South (1,309) and Midwest (753), with 227 in the West and 180 in the Northeast. Grocery chains, restaurants and fountain sales are not covered.

Sales in these stores cannot separate a boycott from anything else that changed in the same weeks in the same places. Tariff news, a promotion, a distribution change or a shift in which stores report would appear as a 2025 change in Coca-Cola's share if it moved with Hispanic share. The 2024 placebo rules out a calendar pattern that repeats each year, not an event specific to 2025. The year-over-year comparison also assumes that, without the boycott, each ZIP's gap to 2024 would have stayed near its baseline level, a form of the parallel-trends assumption that the baseline weeks can check only in part.³

The first week of the loss partly pre-dates the rumor, and week 0 contains both the Day Without Immigrants and the posts of February 5 to 8, so these data give no separate estimate of the rumor. The groups of 102 and 137 ZIPs give single-week estimates with standard errors of 0.3 to 0.6 points. Standard errors in the workbench treat ZIPs as independent, but ZIPs in the same county are correlated: clustering by county instead of by ZIP raises the within-county Hispanic standard error from 0.11 to 0.12 per 10 points. Finally, a fall in Coca-Cola's share and a rise in Pepsi's describe the mix of units sold, and say nothing here about whether total soda sales changed.

Questions for discussion

1. The rumor was false and the plant does not exist, yet sales fell within days. What should Coca-Cola have done in the first 72 hours, who should have heard it from the company, and what would a public denial risk?
2. Using only weekly store sales by ZIP, design a monitoring report that would have flagged the break in the week of February 3 or earlier. Which neighborhoods would it group, what threshold would trigger a call, and how many false alarms would you accept?
3. A colleague sees column 1 of the regression table and proposes a campaign in Democratic-leaning ZIPs. What do you ask before approving it, and which column do you show in reply?
4. Suppose Hispanic share were positively correlated with Republican share (δ = +0.25) and the other numbers stayed the same. Compute the short slope on Republican share from the formula and say whether the short regression would then understate or overstate the long one.
5. The weekly data show a dip in the week of January 27, before the rumor. List two explanations that do not involve the rumor and say what additional data, such as daily sales, would distinguish them.
6. Bud Light's loss had not recovered through 2024 and Coca-Cola's gradient was −0.12 per 10 points by weeks 20 to 21. Which features of the two episodes might explain the difference, and which of them can these data speak to? Does Pepsi's gain in the same ZIPs change how you would measure the recovery?

@misc{singh2026coke,
  author = {Singh, Vishal},
  title  = {Coca-Cola, a rumor and a different constituency},
  year   = {2026},
  note   = {Teaching case, NYU Stern School of Business},
  url    = {https://vishalsingh.org}
}