Discretizing Continuous Covariates For Bounds

Why Discretize a Continuous Covariate?

A continuous covariate can help define groups in which treated and untreated units are more comparable. But a finite-support calculation requires a manageable set of groups, so analysts may partition the covariate into bins. The resulting bounds depend not only on the observed outcomes but also on what is assumed about selection bias within each bin. A narrow bin does not automatically make that assumption true, and a wide bin can conceal important differences among units.

This tutorial constructs bounds for the average treatment effect among treated units (ATT). The method is useful when the covariate is continuous but the analyst can defend limits on the difference in untreated potential outcomes between treated and untreated units within each bin. The limits are assumptions: discretizing the covariate does not establish them.

Definition: Let $D\in\{0,1\}$ be treatment, $Y(1)$ and $Y(0)$ the potential outcomes, $Y=D Y(1)+(1-D)Y(0)$ the observed outcome, and $X$ a continuous covariate. Partition the support of $X$ into disjoint bins $B_1,\ldots,B_K$, and let $B=b$ indicate that $X$ lies in bin $b$. The bin-specific untreated-outcome selection bias is $$ \gamma_b=E[Y(0)\mid D=1,B=b]-E[Y(0)\mid D=0,B=b]. $$

Assumptions and the Bin-Specific Bound

Assume treatment and potential outcomes are well-defined, consistency holds, and the outcome has known support $L\leq Y(d)\leq U$ for $d=0,1$. For every bin represented among treated units, require both treated and untreated units in the target population, so that $E[Y\mid D=1,B=b]$ and $E[Y\mid D=0,B=b]$ are defined. This is a conditional overlap requirement for the bins used in the calculation; it does not guarantee adequate sample sizes.

The substantive restriction is a bin-specific bias interval $\ell_b\leq\gamma_b\leq u_b$. It says how far the mean untreated potential outcome for treated units in bin $b$ may differ from the corresponding mean for untreated units in that same bin. The interval can be asymmetric. It must also be compatible with the outcome support: since $E[Y(0)\mid D=1,B=b]$ lies in $[L,U]$, the possible values of $\gamma_b$ must lie in $[L-m_{0b},U-m_{0b}]$, where $m_{0b}=E[Y\mid D=0,B=b]$.

Definition: Write $m_{1b}=E[Y\mid D=1,B=b]$, $m_{0b}=E[Y\mid D=0,B=b]$, and define the support-compatible bias limits $$ c_b^-=\max\{\ell_b,L-m_{0b}\},\qquad c_b^+=\min\{u_b,U-m_{0b}\}. $$ If $c_b^->c_b^+$, the proposed bias restriction and the support or observed mean are incompatible in that bin. Otherwise, the bin-specific ATT is bounded by $$ m_{1b}-m_{0b}-c_b^+ \ \leq\ E[Y(1)-Y(0)\mid D=1,B=b] \ \leq\ m_{1b}-m_{0b}-c_b^-. $$

To see why, consistency gives $E[Y(1)\mid D=1,B=b]=m_{1b}$ and $E[Y(0)\mid D=0,B=b]=m_{0b}$. By the definition of $\gamma_b$, the missing mean for treated units is $E[Y(0)\mid D=1,B=b]=m_{0b}+\gamma_b$. Thus the bin-specific ATT equals $m_{1b}-m_{0b}-\gamma_b$. Subtracting a larger $\gamma_b$ gives a smaller effect, so the upper bias limit produces the lower effect bound. The support restriction clips the bias interval before this calculation.

Let $q_b=P(B=b\mid D=1)$, the share of treated units in bin $b$. Since the overall ATT averages bin-specific effects over the treated population, the aggregate bounds are the treated-share-weighted sums of the bin endpoints:

\sum_{b=1}^K q_b(m_{1b}-m_{0b}-c_b^+) \ \leq\ E[Y(1)-Y(0)\mid D=1]\ \leq\ \sum_{b=1}^K q_b(m_{1b}-m_{0b}-c_b^-), \qquad \sum_{b=1}^K q_b=1.

This aggregation is valid because each bin-specific effect is averaged over the treated units in that bin. It is not generally correct to weight these effects by the overall population share in each bin or by the share of untreated units.

Worked Example: Combining Bounds Across Two Bins

Suppose a hypothetical study has outcomes measured in points and known to lie between $L=0$ and $U=100$. In a lower-$X$ bin, $m_{1,1}=70$, $m_{0,1}=50$, and the assumed bias interval is $[-5,10]$. In a higher-$X$ bin, $m_{1,2}=60$, $m_{0,2}=55$, and the assumed interval is $[-8,4]$. Suppose 60% of treated units are in the first bin and 40% are in the second.

For the first bin, support permits bias values from $0-50=-50$ to $100-50=50$, so the assumed interval remains unchanged: $c_1^-=-5$ and $c_1^+=10$. The observed treated-control difference is $70-50=20$, giving the ATT interval $20-10=10$ to $20-(-5)=25$ points.

For the second bin, support permits bias values from $0-55=-55$ to $100-55=45$, again leaving the assumed interval unchanged. The observed difference is $60-55=5$, so the ATT interval is $5-4=1$ to $5-(-8)=13$ points.

Weighting by treated shares gives the overall lower endpoint $0.60(10)+0.40(1)=6.4$ points and upper endpoint $0.60(25)+0.40(13)=20.2$ points. Under the stated bin-specific assumptions, the ATT is bounded between 6.4 and 20.2 points. These are identification bounds, not a confidence interval.

What Discretization Does to the Assumptions

The bias restriction must match the grouping used in the calculation. A bound on selection bias at each exact value of $X$ is not automatically the same as a bound on $\gamma_b$. Within a bin, treated and untreated units can have different distributions of $X$. Consequently, their average untreated potential outcomes may differ because of that composition difference, even if there is no treated-control difference at any exact value of $X$.

Worked Example: Pointwise Balance Does Not Guarantee Bin Balance

Consider one hypothetical bin containing two covariate values, low and high. Suppose $Y(0)=0$ at the low value and $Y(0)=100$ at the high value for everyone. At each exact covariate value, treated and untreated units have the same untreated potential outcome, so the pointwise difference in mean $Y(0)$ is zero.

Now suppose 90% of treated units in the bin have the high covariate value, while only 10% of untreated units do. Then $E[Y(0)\mid D=1,B=b]=0.10(0)+0.90(100)=90$, while $E[Y(0)\mid D=0,B=b]=0.90(0)+0.10(100)=10$. Therefore $\gamma_b=90-10=80$, not zero. Applying a bin-level restriction such as $|\gamma_b|\leq 5$ would be invalid for these data-generating features, despite zero pointwise differences.

The lesson is not that pointwise assumptions are useless, but that translating them to a bin-level restriction requires an additional argument controlling covariate composition within bins. One defensible option is to state and justify the bias limit directly at the bin level. Another is to use bins narrow enough to make composition differences plausibly small, while recognizing that narrowness alone is not proof.

Bin Width, Sparse Strata, and Informativeness

Narrower bins can make treated and untreated units more similar in $X$, but they also create more strata and fewer observations per stratum. Wider bins usually improve counts while allowing more within-bin heterogeneity. These forces affect different parts of the analysis:

Validity: If the assumed $[\ell_b,u_b]$ fails to contain the actual bin-level bias, the reported bounds need not cover the ATT. A narrower partition can help make a modest bias limit plausible, but cannot ensure it.
Informativeness: A wider bias interval directly widens the effect interval, because its width includes $u_b-\ell_b$ when support does not clip it. Restricting a bias interval without a substantive rationale may make results look precise but does not make them credible.
Data support: Sparse bins make sample means and treated shares unstable. A bin with no controls has no observed $m_{0b}$, so this formula cannot be applied there; the analyst must use outcome-support information, combine bins with a defensible revised assumption, or change the design.

A practical diagnostic is to tabulate, for every proposed bin, treated and untreated counts, observed outcome means, and the implied support-compatible bias range. Also inspect the distribution of $X$ within each treatment group in each bin. Strongly different within-bin distributions warn that a bin-level bias restriction may be difficult to defend. Do not silently merge sparse bins: merging changes the conditioning event and therefore changes the selection-bias assumption as well as the means and weights.

Worked Example: Treated Weights and a Sparse Bin

Consider hypothetical sample summaries for two bins, with outcomes in points. The first bin has two treated outcomes, 60 and 80, and one control outcome, 50. The second has eight treated outcomes averaging 60 and eight control outcomes averaging 55. Suppose the bias limits are $[-5,10]$ in the first bin and $[-8,4]$ in the second, with support $[0,100]$.

In bin 1, $m_{1,1}=(60+80)/2=70$ and $m_{0,1}=50$, so its plug-in effect bounds are $70-50-10=10$ and $70-50-(-5)=25$ points. In bin 2, the bounds are $60-55-4=1$ and $60-55-(-8)=13$ points. The support does not clip either bias interval.

There are 10 treated observations in total, so the treated weights are $q_1=2/10=0.20$ and $q_2=8/10=0.80$. The weighted plug-in endpoints are $0.20(10)+0.80(1)=2.8$ and $0.20(25)+0.80(13)=15.4$ points. The sparse first bin receives only 20% of the treated weight, but its one control observation still makes its estimated control mean especially fragile. These calculations show the plug-in arithmetic; they do not quantify sampling uncertainty or establish that the assumed bias limits are correct.

Alternatives, Limitations, and Common Mistakes

A single broad category for $X$ is simpler, but can combine units with substantially different outcome levels and treatment selection patterns. Finer bins retain more covariate detail, at the cost of sparse strata and a larger number of bias restrictions to justify. Modeling $X$ continuously avoids arbitrary cut points, but any model-based adjustment or restriction then depends on its own functional-form and identification assumptions. These approaches are not interchangeable: a regression fit does not validate a bin-level bias bound, and binning does not by itself remove confounding.

The bounds also rely on the specified outcome support and target population. Incorrect support limits can invalidate the support clipping. The ATT weights describe treated units in the target population; changing the target population can change those weights and the estimand. Finally, identification bounds describe the range compatible with assumptions and population quantities. Sample estimates have uncertainty, which requires separate inference rather than interpreting the endpoints as fixed, known values.

Common mistake: “The bin is narrow, so the treated and untreated units are effectively identical in $X$; therefore a small selection-bias limit is guaranteed.” Narrow bins may reduce some differences, but do not establish balance, eliminate unobserved confounding, or ensure that the bin-level mean difference in $Y(0)$ is small. The correction is to state the bin-level bias restriction explicitly, defend its magnitude, inspect within-bin composition, and report how the bounds change under plausible alternative limits and partitions.

Key takeaway: Discretization is a way to organize a bound calculation, not an identification assumption. The validity of bin-based bounds depends on the bin-level restrictions actually used; their informativeness depends on those restrictions, outcome support, and the treated-share weights.

Check Your Understanding

Answer each question, then compare with the brief solutions.

In a bin, $m_{1b}=40$, $m_{0b}=30$, and the bias interval is $[-2,5]$. Ignoring support clipping, what are the bin-specific ATT bounds?
Why are bin-specific effects aggregated using $P(B=b\mid D=1)$ for the ATT?
Does zero selection bias at every exact value of $X$ necessarily imply zero selection bias within a bin? Explain.
What should be concluded if the proposed bias interval and the support-compatible interval do not overlap?

Solutions. (1) The observed difference is $40-30=10$, so the lower endpoint is $10-5=5$ and the upper endpoint is $10-(-2)=12$. (2) The ATT averages effects among treated units, so each bin contributes in proportion to its share of the treated target population. (3) No. Treated and untreated units can have different within-bin distributions of $X$, creating a difference in their average $Y(0)$ even when the means agree at each exact $X$. (4) The assumptions or supplied quantities are incompatible in that bin; do not report the clipped formula as a valid interval without resolving the conflict.