StatFAS / Blog

PACF plot: What It Shows, When to Use It and How to Interpret the Results

Published November 2026 · Time-series plots for biological, genomic and medical research

The ACF plot shows how strongly a series is tied to the past, but the correlation at lag 5 may simply be the echo of lags 1 to 4. The PACF plot (partial autocorrelation function plot) removes that echo and shows the direct link at each lag.

Skip the setup and make the PACF plot now.
Paste the data, get the tables, the PACF plot and APA 7 wording in seconds. Analysis needs no sign-in; downloads need a free sign up, then log in.
Open the free PACF plot calculator

What is a PACF plot?

A PACF plot shows the partial autocorrelation at each lag: the correlation between xt and xt−k after the linear effect of the observations in between (lags 1 to k − 1) has been removed. The values come from the Durbin–Levinson recursion or the Yule–Walker equations. The same dashed 95% limits as in the ACF plot, ±1.96/√N, are drawn. At lag 1 the partial autocorrelation equals the autocorrelation.

The PACF plot is the main tool for choosing the autoregressive order p of an ARIMA model in the Box–Jenkins approach.

When and why to use a PACF plot

When

Why

Typical biomedical examples

autoregressive structure in daily blood glucose, seizure or event counts across days, daily case counts in an outbreak, heart rate variability, repeated biomarker readings.

How to interpret the results

Sharp cutoff after lag p

When partial autocorrelations are significant up to lag p and then fall inside the limits, while the ACF plot decays gradually, an AR(p) model is suggested.

PACF decays and ACF cuts off

This pattern points to a moving-average model of order q read from the ACF plot.

Both plots tail off

A mixed ARMA structure is likely. Compare candidate models with AIC or BIC instead of relying on the plots alone.

Spikes at seasonal lags

A significant stem at lag 12 (or 24) in monthly data suggests a seasonal autoregressive term.

Difference first

The PACF plot is meant for stationary series. A trending series produces a large lag-1 value and little else, so difference the series and draw the PACF plot again.

Identification is tentative

Treat the order read from the PACF plot as a starting candidate. Fit neighbouring models, compare information criteria and test the residuals with a Ljung–Box test.

Pointwise limits

About 1 stem in 20 crosses the limits by chance. A single stem just past the line at a high lag is rarely meaningful.

Sample size

Short series give unstable partial autocorrelations. Aim for at least 50 observations.

Common mistakes

How to make a PACF plot

The easy way: StatFAS.com

The free PACF plot calculator at StatFAS.com handles the calculations, draws the figure to publication standard and writes the APA 7 results text. The PACF plot calculator page walks through every step.

Make the PACF plot in your browser.
Paste the data, get the tables, the PACF plot and APA 7 wording in seconds. Analysis needs no sign-in; downloads need a free sign up, then log in.
Open the free PACF plot calculator

SPSS

  1. Define the cycle if seasonal differencing is wanted: Data > Define Date and Time.
  2. Choose Analyze > Forecasting > Autocorrelations.
  3. Move the series into Variables.
  4. Under Display tick Partial autocorrelations (SPSS draws the ACF plot and the PACF plot together when both boxes are ticked).
  5. Use Transform for Difference (1) or Seasonally difference, and Options for the maximum number of lags.
  6. Read the cutoff on the partial autocorrelation chart, then fit candidate models under Analyze > Forecasting > Create Traditional Models (ARIMA) and compare the fit statistics.

R

x <- ts(values, frequency = 12)
pacf(x, lag.max = 24, main = "")          # Yule-Walker estimates, white-noise limits
acf(x, lag.max = 24, main = "")           # read both plots together

d <- diff(x)                              # difference first when a trend is present
pacf(d, lag.max = 24)

fit1 <- arima(x, order = c(1, 1, 0))      # candidate AR(1) after one difference
fit2 <- arima(x, order = c(2, 1, 0))      # candidate AR(2)
AIC(fit1, fit2)
Box.test(residuals(fit2), lag = 12, type = "Ljung-Box")

Python

import matplotlib.pyplot as plt
from statsmodels.graphics.tsaplots import plot_acf, plot_pacf
from statsmodels.tsa.arima.model import ARIMA
from statsmodels.stats.diagnostic import acorr_ljungbox

plot_pacf(y, lags=24, zero=False, method="ywm", alpha=0.05)
plot_acf(y, lags=24, zero=False)
plt.show()

d = y.diff().dropna()                     # difference first when a trend is present
plot_pacf(d, lags=24, zero=False, method="ywm")

m1 = ARIMA(y, order=(1, 1, 0)).fit()      # candidate AR(1) after one difference
m2 = ARIMA(y, order=(2, 1, 0)).fit()      # candidate AR(2)
print(m1.aic, m2.aic)
print(acorr_ljungbox(m2.resid, lags=[12]))

How to report the results in APA 7

Example template: “The PACF of the [variable] series (N = [N]) showed significant partial autocorrelation at lag[s] [list] (95% limits ±[limit]) and no significant values beyond lag [p], suggesting an AR([p]) component. Candidate models were compared with AIC.” State any differencing applied before the PACF plot was drawn.

Frequently asked questions

What does a PACF plot show that an ACF plot does not?

The PACF plot shows the correlation at each lag after the effect of the lags in between is removed, so the order of an autoregressive process appears as a clear cutoff.

How is the AR order read from a PACF plot?

Count the lags with significant stems before the first lag that falls inside the limits and stays inside. That count is the candidate order p.

Why does the PACF plot of a trending series look odd?

A trend creates a very large correlation at lag 1 that overwhelms every other lag. Difference the series first.

Is the PACF plot enough to choose a model?

No. Use the PACF plot and the ACF plot to propose candidates, then compare AIC or BIC values and check the residuals.

Key takeaways

Related reading

ACF plot blog · Forecast plot blog · PACF plot calculator