# Novel object recognition and preference indices: testing against chance and comparing groups

> Test a discrimination index against chance only to show that a group remembers; to show that groups differ, compare the groups directly. Set the exploration exclusion rule in advance and report total exploration. Python and R code.

Canonical: https://www.licklider.ai/guides/preference-index-vs-chance/
Section: Guides › Design, replicates and sample size
Author: Licklider · Published 2026-11-12 · Last reviewed 2026-11-12 (UTC)

> **Quick answer.** A one-sample test of the discrimination index against chance (0, or 50% for a preference index) answers only "**does this group remember?**", and only when the mean is above chance and objects and positions were counterbalanced. To claim that a knockout or treatment **impairs** memory, **compare the groups directly** (for example Welch's t-test on the discrimination index) and report the difference with its 95% confidence interval. "Significantly above chance in wild-type, not in knockout" is not evidence of a difference: in our simulation with identical groups (12 mice each, true discrimination index 0.20), one group was above chance and the other was not in 45.6% of experiments. Set the exclusion rule for low exploration (for example less than 20 s exploring the two objects combined) before testing, and report total exploration for each group.

**Terms used below.** *Novel object recognition (NOR) test*: a memory test in which a mouse that remembers a familiar object spends more time exploring a new one. *Sample (training) session*: the mouse explores two identical objects. *Test session*: one object is replaced by a new one. *Discrimination index (DI)*: (novel − familiar)/(novel + familiar), from −1 to 1, with 0 meaning no preference. *Preference (recognition) index*: novel ÷ total, often as a percentage, with 50% meaning no preference; preference % = 50 × (DI + 1), so DI 0.20 is a preference of 60%. *True DI*: the value set when constructing the example data. *Chance level*: the value expected without memory (DI 0, preference 50%). *One-sample t-test*: tests whether a group's mean differs from a fixed value. *Welch's t-test*: compares two groups' means without assuming equal spreads. *Total exploration*: novel plus familiar exploration time. *ANOVA*: compares means across several groups or factors; an *interaction* tests whether the effect of one factor (genotype) differs between levels of another (sex). *Exclusion criterion*: a rule, set before testing, for leaving out animals that explored too little. *Counterbalancing*: varying which object is novel and where it sits across animals. *Blinding*: scoring without knowing each animal's group. *Paired test*: compares two measurements from the same animal, such as novel and familiar exploration times. *Random seed*: the number that fixes a computer's random draws, so constructed data can be reproduced exactly. *95% confidence interval (CI)*: the range of effects compatible with the data. *Power*: the chance of detecting an effect that is really there. *False positive*: a "significant" result when there is no real effect.

## Which test for your question

| Question | Test | Report |
| --- | --- | --- |
| Does this group show recognition memory? | One-sample t-test of DI against 0 (or preference against 50%); memory only if the mean is above chance, with objects and positions counterbalanced | Mean DI with 95% CI, n, p |
| Does the knockout or treatment change memory? | Welch's t-test between groups on DI (Welch's ANOVA or ANOVA, for more groups or factors) | Difference in DI with 95% CI, p |
| Does the effect differ between sexes or ages? | Two-way ANOVA with the interaction, with group sizes planned for it ([analysing both sexes](/guides/sex-as-a-biological-variable-analysis/)) | Interaction estimate with 95% CI |
| Did groups explore differently? | Welch's t-test on the logarithm of total exploration (which evens out its right skew), or report it descriptively | Mean total exploration per group, before and after exclusions |

Denninger and colleagues state the rule directly: if memory alone is the question, "a one-sample t-test of percent time investigating an object versus the fixed value of 50% can be used", but to compare groups "it is erroneous to conclude that these two groups are significantly different from each other" from separate tests against chance; "those groups must be statistically compared to each other, not only compared to an external standard" ([Denninger et al. 2018](https://doi.org/10.3791/58593)). Nieuwenhuis and colleagues make the general point: "a comparison of two experimental effects requires a statistical test on their difference" ([Nieuwenhuis et al. 2011](https://doi.org/10.1038/nn.2886)). For exclusions, Lueptow's protocol excludes mice that "do not reach a 20 s minimum of exploration for both objects" in the sample or test session, the 20 s being total exploration of the two objects in that session ([Lueptow 2017](https://doi.org/10.3791/55718)), and Leger and colleagues use "a selection criterion based on a minimal time of exploration for both objects during each session" ([Leger et al. 2013](https://doi.org/10.1038/nprot.2013.155)).

## Worked example: wild-type and knockout, 12 mice each

Constructed data: test-session exploration times for 12 wild-type (WT) and 12 knockout (KO) mice. By construction both genotypes have the **same true mean DI, 0.20**; the exclusion rule, set in advance, removes mice with less than 20 s exploring the two objects combined (the example data contain only the test session). The random seed (3) was the first from 1 upward for which the wild-type group was significant against chance, the knockout was not, and at least one mouse was excluded: the situation this page is about. Stored code and data: [nor.csv](/guides/preference-index-vs-chance/nor.csv), [model.py](/guides/preference-index-vs-chance/model.py), [make_data.py](/guides/preference-index-vs-chance/make_data.py), [analysis.py](/guides/preference-index-vs-chance/analysis.py), [analysis.R](/guides/preference-index-vs-chance/analysis.R) and [the analysis output](/guides/preference-index-vs-chance/analysis_output.txt); analysis.py and analysis.R print the same output.

| Analysis | Result |
| --- | --- |
| Excluded (total exploration < 20 s) | WT4 and KO1 (one mouse per group; n = 11 each) |
| WT, one-sample test against 0 | Mean DI 0.293 (95% CI 0.162 to 0.424; preference 64.7%), p = 0.0005 |
| KO, one-sample test against 0 | Mean DI 0.105 (95% CI −0.079 to 0.290; preference 55.3%), p = 0.23 |
| **KO − WT, Welch's t-test** | **−0.188 (95% CI −0.401 to 0.025), p = 0.080** |
| Mean total exploration | All mice: WT 36.0 s, KO 34.4 s; included mice: WT 38.2 s, KO 36.9 s |

Taken as separate tests, the results say the knockout "has no memory". Compared directly, the two groups do not differ clearly: the interval for the difference runs from a large impairment to a slight improvement. In 10,000 simulated experiments of this design ([simulation.py](/guides/preference-index-vs-chance/simulation.py), [the simulation output](/guides/preference-index-vs-chance/simulation_output.txt)):

| Scenario | One group "above chance", the other not | Welch's t-test between groups, p < 0.05 |
| --- | --- | --- |
| Same true DI in both groups (0.20) | **45.6%** | 5.1% (false positives) |
| No memory in either group (true DI 0) | 4.7% | 5.0% (false positives) |
| KO lower (WT 0.25, KO 0.05) | 77.1% | 40.4% (power) |

How often the groups disagree depends on the true DI and the group size: it is high even when both groups have the same moderate memory, as here, and higher still when they truly differ. With 12 mice per group, a difference of DI 0.25 vs 0.05 was detected in only 40.4% of experiments; plan group sizes for the comparison between groups ([how many biological replicates](/guides/how-many-biological-replicates/)).

The code expects one row per mouse with columns mouse, genotype, novel_s and familiar_s; rename your columns and group labels to match.

```python
# Condensed: analysis.py also prints total exploration
import pandas as pd
from scipy import stats

d = pd.read_csv("nor.csv")                           # one row per mouse: mouse, genotype, novel_s, familiar_s
d["total"] = d.novel_s + d.familiar_s
k = d[d.total >= 20].copy()                          # exclusion rule set before testing
k["di"] = (k.novel_s - k.familiar_s) / k.total
for g in ("WT", "KO"):
    t = stats.ttest_1samp(k.di[k.genotype == g], 0)        # recognition memory within each group (DI above 0)
    print(g, t, t.confidence_interval())                 # for preference %, test 100 * novel_s / total against 50
t = stats.ttest_ind(k.di[k.genotype == "KO"], k.di[k.genotype == "WT"], equal_var=False)   # the group comparison, KO - WT
print(t, t.confidence_interval())
```

```r
d <- read.csv("nor.csv"); d$total <- d$novel_s + d$familiar_s
k <- d[d$total >= 20, ]; k$di <- (k$novel_s - k$familiar_s) / k$total
t.test(k$di[k$genotype == "WT"], mu = 0); t.test(k$di[k$genotype == "KO"], mu = 0)   # memory within each group
t.test(di ~ factor(genotype, levels = c("KO", "WT")), data = k)                       # Welch; KO listed first, so KO - WT
```

## Before you test

1. **Write the exclusion rule in advance** (for example at least 20 s exploring the two objects combined, in both the sample and the test session) and report how many mice it excluded in each group ([when can you exclude](/guides/when-can-you-exclude-an-outlier/)).
2. **Counterbalance objects and positions**, and score blind to group; Lueptow's protocol counterbalances "the use of each set of objects" and "the location of the novel object" and asks researchers to "blind the scorer to the experimental conditions" ([Lueptow 2017](https://doi.org/10.3791/55718); blinding is item 5 of the ARRIVE 2.0 animal-research reporting guidelines, [Percie du Sert et al. 2020](https://doi.org/10.1371/journal.pbio.3000410)).
3. **Report total exploration**, before and after exclusions. Ratios adjust for differences in exploration, but Akkerman and colleagues note that "differences in exploration levels may affect the statistical evaluation of group differences" ([Akkerman et al. 2012](https://doi.org/10.1016/j.bbr.2012.03.024)); if totals differ between groups, say so when you interpret the indices. The same paper suggests comparing each group with a constructed group that shows no discrimination, and notes that a one-sided test is acceptable "provided that place- or object biases can be ruled out"; decide on a one-sided test before seeing the data, not after a two-sided p-value misses 0.05.
4. **Choose one index in advance** (DI or preference) and use it throughout; they carry the same information.
5. **Plan the comparison between groups**, not the test against chance, when choosing the group size.
6. **Look at the indices before testing.** DI lies between −1 and +1; if many mice sit at the limits or one value is extreme, add a rank-based test as a check (Wilcoxon signed-rank against 0 within a group, Mann–Whitney between groups). If mice share cages or litters, add cage or litter as a random effect ([nested data](/guides/nested-data-mixed-model/)).

## Check what an AI assistant ran

1. **The assistant concluded that one genotype is impaired** because one group was significant against chance and the other was not. Ask for a direct comparison between the groups.
2. **The assistant excluded mice after seeing the results.** Ask for the rule and when it was set.
3. **The assistant reported only DI**, not total exploration. Ask for both.
4. **The assistant tested raw novel and familiar times with a paired test in each group** and compared p-values. Ask for the DI comparison between groups.

## Report it

> Recognition memory was assessed as the discrimination index (novel − familiar)/(novel + familiar) in the test session. Mice with less than [20] s of total exploration were excluded, as specified in advance ([1] WT, [1] KO). Each genotype's index was compared with 0 using one-sample t-tests, and genotypes were compared with Welch's t-test (R [4.3.3]).

> After exclusions, n = [11] wild-type and [11] knockout mice were analysed. Wild-type mice preferred the novel object (mean DI [0.29], 95% CI [0.16] to [0.42], p = [0.0005]); the knockout mean DI was [0.11] (95% CI [−0.08] to [0.29], p = [0.23]). The difference between genotypes was [−0.19] (95% CI [−0.40] to [0.03]; p = [0.080]), so the data are compatible with anything from a clear impairment to a slight improvement in the knockout. Mean total exploration was [38.2] s (WT) and [36.9] s (KO).

Replace the bracketed values with your own.

## References

- Denninger JK, Smith BM, Kirby ED (2018). Novel object recognition and object location behavioral testing in mice on a budget. *Journal of Visualized Experiments* (141):e58593. [doi:10.3791/58593](https://doi.org/10.3791/58593)
- Nieuwenhuis S, Forstmann BU, Wagenmakers EJ (2011). Erroneous analyses of interactions in neuroscience: a problem of significance. *Nature Neuroscience* 14:1105–1107. [doi:10.1038/nn.2886](https://doi.org/10.1038/nn.2886)
- Lueptow LM (2017). Novel object recognition test for the investigation of learning and memory in mice. *Journal of Visualized Experiments* (126):55718. [doi:10.3791/55718](https://doi.org/10.3791/55718)
- Leger M, Quiedeville A, Bouet V, et al. (2013). Object recognition test in mice. *Nature Protocols* 8:2531–2537. [doi:10.1038/nprot.2013.155](https://doi.org/10.1038/nprot.2013.155)
- Akkerman S, Prickaerts J, Steinbusch HWM, Blokland A (2012). Object recognition testing: statistical considerations. *Behavioural Brain Research* 232:317–322. [doi:10.1016/j.bbr.2012.03.024](https://doi.org/10.1016/j.bbr.2012.03.024)
- Percie du Sert N, Hurst V, Ahluwalia A, et al. (2020). The ARRIVE guidelines 2.0: updated guidelines for reporting animal research. *PLOS Biology* 18:e3000410. [doi:10.1371/journal.pbio.3000410](https://doi.org/10.1371/journal.pbio.3000410)

## Related

- [Analysing both sexes](https://www.licklider.ai/guides/sex-as-a-biological-variable-analysis/)
- [When can you exclude an outlier?](https://www.licklider.ai/guides/when-can-you-exclude-an-outlier/)
- [Randomisation and blinding](https://www.licklider.ai/guides/randomisation-and-blinding-in-animal-studies/)
- [Welch's t-test vs Student's t-test](https://www.licklider.ai/guides/welch-vs-student-t-test/)

## How this page was made

- Written and published by Licklider. Recommendations cite peer-reviewed protocols, methods papers and the ARRIVE 2.0 guidelines listed under References; cited passages were checked against the source text.
- The worked example uses constructed data. Its numbers were produced by stored code (Python 3.13 with SciPy 1.18.1; R 4.3.3, base R), and the two languages print identical output; the simulation was run in Python.
- Separate review passes checked the statistics, the citations and readability before publication.