Your residual plot is funnel-shaped. What does that mean?

Residuals fanning out as fitted values increase is heteroscedasticity: the variance of the errors is not constant.

Why it matters

Your coefficients are still unbiased. The problem is the standard errors, which are wrong, usually too small. That makes your confidence intervals too narrow and your p-values too small, so you find significance that is not there.

What causes it

Multiplicative relationships. Very common in biology and economics, where error scales with magnitude. A measurement with 10% error produces small absolute errors at low values and large ones at high values, which is exactly a funnel.

A missing predictor, particularly one that interacts with something already in the model.

Count or proportion outcomes fitted with ordinary linear regression. Their variance is tied to their mean by definition.

The fixes

Transform the outcome. A log transform is the standard answer for a funnel opening to the right, and it turns a multiplicative relationship into an additive one. Square root works for count-like data. Remember to interpret coefficients on the transformed scale.

Use robust standard errors. Keeps the model and the interpretation, and corrects the inference. Often the least disruptive option.

Fit the right model family. Poisson or negative binomial for counts, logistic for binary outcomes. If your outcome is a count, a linear model was the wrong tool from the start, and transforming it is a patch on the wrong choice.

Weighted least squares, if you can model how the variance changes.

Check the other patterns too

Curvature in the residuals means a missing nonlinear term, not heteroscedasticity. The two are often present together and have different fixes.

In Plotara

Residual diagnostics produce residual and Q-Q plots alongside the model, with Poisson and negative binomial available when a count outcome is the real answer.