Handling Missing Values in C++

June 2026 · 10 min read

Real datasets have holes: a sensor dropped out, a form field was left blank. A model that expects a full numeric matrix cannot consume a gap, so the gaps must be filled first. The simplest reliable strategy is imputation: replace each missing entry with a summary of the column it lives in, computed from the training data.

Marking the Gaps

First we need a value that means "missing." A floating-point NaN is ideal: it survives in a double column and is easy to test for with isnan, so a missing entry never gets mistaken for a real $0$.

#include <bits/stdc++.h>
using namespace std;
static const double NA = nan("");          // the "missing" marker

Mean and Median, Skipping NaNs

Both statistics are computed over the present values only. The mean is the obvious fill, but it is pulled toward extreme values; the median is robust to outliers, which often makes it the safer default.

double columnMean(const vector<double>& col) {
    double sum = 0; int n = 0;
    for (double v : col) if (!isnan(v)) { sum += v; n++; }
    return sum / n;
}

double columnMedian(vector<double> col) {
    vector<double> present;
    for (double v : col) if (!isnan(v)) present.push_back(v);
    sort(present.begin(), present.end());
    int n = present.size();
    return n % 2 ? present[n/2] : (present[n/2 - 1] + present[n/2]) / 2;
}

vector<double> impute(const vector<double>& col, double fill) {
    vector<double> out = col;
    for (double& v : out) if (isnan(v)) v = fill;
    return out;
}

Running It

Take a column with two gaps and one outlier ($100$), and fill it both ways:

int main() {
    vector<double> col = {1, 2, NA, 4, 100, NA};
    double mean   = columnMean(col);
    double median = columnMedian(col);
    printf("train mean (skip NA) = %.2f   median = %.2f\n", mean, median);
    // print impute(col, mean) and impute(col, median)
}

The lone $100$ drags the mean up to $26.75$, so mean-imputation fills the gaps with a value larger than every non-outlier in the column, a poor guess. The median, $3$, ignores the outlier and fills with a typical value. That robustness is why median imputation is a sensible default whenever a column may contain extreme values.